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
|
|---|---|---|---|---|---|---|---|---|---|
1609.02202 | 1 | 1609 | 2016-09-07T21:50:38 | Deconstructing Odorant Identity via Primacy in Dual Networks | [
"q-bio.NC"
] | In the olfactory system, odor percepts retain their identity despite substantial variations in concentration, timing, and background. We propose a novel strategy for encoding intensity-invariant stimuli identity that is based on representing relative rather than absolute values of the stimulus features. Because, in this scheme, stimulus identity depends on relative amplitudes of stimulus features, identity becomes invariant with respect to variations in intensity and monotonous non-linearities of neuronal responses. In the olfactory system, stimulus identity can be represented by the identities of the p strongest responding odorant receptor types out of a species dependent complement. We show that this information is sufficient to recover sparse stimuli (odorants) via elastic net loss minimization. Such a minimization has to be performed under constraints imposed by the relationships between stimulus features. We map this problem onto the dual problem of minimizing a functional of Lagrange multipliers. The dual problem, in turn, can be solved by a neural network whose Lyapunov function represents the dual Lagrangian. We thus propose that networks in the piriform cortex compute odorant identity and implement dual computations with the sparse activities of individual neurons representing the Lagrange multipliers. | q-bio.NC | q-bio | Deconstructing Odorant Identity via Primacy in Dual
Networks
Daniel Kepple1, Hamza Giaffar1, Dmitry Rinberg2, and Alexei
Koulakov1
1 Cold Spring Harbor Laboratory, Cold Spring Harbor, NY
2 NYU Neuroscience Institute, NYU Langone Medical center, New York, NY
Abstract
In the olfactory system, odor percepts retain their identity despite substantial
variations in concentration, timing, and background. We propose a novel
strategy for encoding intensity-invariant stimuli identity that is based on
representing relative rather than absolute values of the stimulus features.
Because, in this scheme, stimulus identity depends on relative amplitudes of
stimulus features, identity becomes invariant with respect to variations in
intensity and monotonous non-linearities of neuronal responses. In the olfactory
system, stimulus identity can be represented by the identities of the p strongest
responding odorant receptor types out of a species dependent complement. We
show that this information is sufficient to recover sparse stimuli (odorants) via
elastic net loss minimization. Such a minimization has to be performed under
constraints imposed by the relationships between stimulus features. We map this
problem onto the dual problem of minimizing a functional of Lagrange
multipliers. The dual problem, in turn, can be solved by a neural network whose
Lyapunov function represents the dual Lagrangian. We thus propose that
networks in the piriform cortex compute odorant identity and implement dual
computations with the sparse activities of individual neurons representing the
Lagrange multipliers.
1 Introduction
Sensory systems face the problem of computing stimulus identity invariant to several features. The
olfactory system, for example, has to compute stimulus identity despite substantial variation in the
absolute concentrations of molecules present in the stimulus. This computation is necessary to
enable navigation in chemical gradients or within variable odorant plumes. How can the olfactory
system robustly represent odorant identity despite variable stimulus intensity? The first step of
olfactory processing involves odorants binding to and activating a set of molecular sensors known
as olfactory receptors. Olfactory receptors are proteins expressed by olfactory sensory neurons
(OSNs) located in the olfactory epithelium. Most mammalian olfactory systems contain ~1000
types olfactory receptors, while humans rely on the responses of only 350 [1; 2; 3]. Importantly,
every OSN expresses only a single type of olfactory receptor chosen randomly out of the large
ensemble. Odorant identity is therefore represented in the patterns of activation of olfactory
receptor proteins and, by extension, OSNs. How the olfactory system represents odorant identity is
not clearly understood. Here we examine the hypothesis that stimulus identity is inferred on the
basis of the relative amplitudes of responses of OSNs. In particular, we propose that odorant
identity can be defined by specifying the p strongest responding. We call this type of
representation the primacy model. According to this primacy coding scheme, in which the
olfactory system uses relative rather than absolute receptor responses, odorant representations
become independent of absolute odorant concentration, leading to a concentration-invariant
representation of odor identity. We formulate the identity decoding scheme using a dual Lagrange-
1
Karush-Kuhn-Tucker problem. We show that this dual problem can be solved by a neural network
that we call a dual network. Dual networks that solve the primacy problem share many features
with real olfactory networks. We therefore derive gross structure of olfactory circuits from first
principles, based on the primacy model.
2 Dual Networks
2.1 Representing odorants by sparse vectors.
6
M
~ 10
. Each component of this vector
Ethologically important odorant stimuli are mixtures of monomolecular components. Such stimuli
can be represented by a vector of concentrations x
jx is equal to
the concentration of an individual monomolecular component numbered by the index j . The
number of potential monomolecular components, which is equal to the dimensionality of vector x
, will be denoted here by M . This number has been estimated to be on the order of several
million,
, based on the count of potentially volatile molecules lighter than 300 Daltons
in the popular database PubChem [4]. However, odorant mixtures cannot contain all of these
molecules at the same time and are necessarily represented by sparse vectors x
. For the purposes
of the olfactory system, the sparseness of the concentration vector is further increased by the
inability of the system to detect or recognize individual components. Indeed, psychophysical
studies suggest that human observers can detect roughly 12 monomolecular components of the
vector x
[5]. Therefore, we suggest that realistic odorants can be defined by concentration vectors
x
of high dimensionality (
Odorant mixtures are further represented by the responses of olfactory receptor neurons r
responses can be approximated by linear nonlinear functions of vector x
model of receptor with a single binding site and no cooperativity, the law of mass action yields
(1)
. These
. Indeed, in the simplest
) and very few non-zero elements (
~ 10
~ 10
).
M
K
6
r F y
(
i
i
)
y
i
,
A x
ij
j
j
i
1...
N
Here index
enumerates the olfactory receptor types. The total number of olfactory
receptors N varies between 350 in humans and 1100 in rodents. Matrix ijA contains affinities of
molecules of type j to the receptors of typei .
is the nonlinear function
describing the activation of a receptor. The problem solved by the olfactory system can be
jx given the set of responses of olfactory sensory
formulated as follows: find the odorant stimulus
neurons ir . Because the responses
iy via a simple monotonic
function F , we can assume that networks analyzing responses of receptor neurons have the linear
component of response iy available to them. Overall, the problem of olfactory decoding is this:
find the sparse vector
iy containing olfactory receptor responses.
ir are related to their inputs
jx given vector
F y
(
/ (1
y
y
)
)
given y
2.2 Sparse olfactory stimulus recovery.
The problem formulated above (find x
) can be reduced to solving the system of linear
equations (1). This problem is not entirely trivial, because the number of unknowns (components
of x
) is substantially larger than the number of equations (components of y
M
,
N
3
). Some relief arrives from the fact that the vector of unknowns is sparse. A sparse
vector can be recovered from a system of linear equations using arguments from compressed
sensing [6; 7; 8]. The vector x
can be found exactly, despite containing more components than
,
~ 10
~ 10
6
2
j
x
j
y Ax
x
0
j
argmin
, one can use the method of sparse signal recovery via
exactly, the following certain condition has to be met relating
equations. To determine x
given y
minimization of the 1l norm [6; 7; 8]:
x
To be able to reconstruct x
parameters M , N , and K :
This condition is the necessary condition for sparse signal recovery using 1l norm minimization
obtained by Donoho and Tanner [6; 7; 8]. For
we obtain the following
condition for the number of olfactory receptor types necessary for the recovery of the stimulus:
N
). Thus,
equation (1) can in principle be solved with the existing parameters in the olfactory system. This
means that the olfactory system can reconstruct
monomolecular components given the
responses of
, which is satisfied for both humans (
olfactory receptors.
) and mice (
N K
1000
~ 10
~ 10
N
N
~ 10
200
350
N
1000
and
log
M
(2)
(3)
K
M
6
K
results in the doubling of the vector of receptor responses y
The solution (2) is limited in that it is not concentration invariant. Doubling the concentration
vector x
. Solving equation
(2) under the constraint of doubled vector y
in turn reconstructs the doubled vector of
concentrations x
. Our goal, however, is to build a representation that is concentration invariant.
This means that we would like to obtain a framework in which doubling of the receptor responses
y
does not affect the reconstructed stimulus, i.e. a concentration-invariant odorant identity. We
therefore propose a primacy-based decoding model that makes this possible.
Ax
, and the resulting receptor response vector y
1000
2.3 Concentration invariant decoding algorithm via primacy. To make odorant representation
invariant to concentration, we propose to reformulate the sparse recovery problem. In its simplest
form, for each input vector x
Ax
, we propose to
isolate the set of p strongest responding receptor neurons (
) . We call this set the
primacy set P . The complementary set P includes weaker responding receptor types. For these
two sets we can state that
y
Here y means the set of components of vector y
these constraints, the sparse recovery problem (2) can be reformulated as follows:
P
that belong to the primacy group P . With
argmin
p
(5)
(4)
x
x
y
P
When seeking a minimum in this equation, we have to use the relationship y
Ax
, however, we
do not use knowledge about exact values of the vector y
, only the relationships between its
components. This approach can be used to recover the concentration vector x
, however, the
jy . This
amplitude of this vector does not depend on the absolute values of the neural responses
is due to the non-parametric formulation of relative rather than absolute values of receptor
responses in the sparse signal recovery problem (5). For example, doubling each component of
vector y
, and, thus, the reconstructed stimulus
x
is not affected. This decoding scheme therefore yields a reconstruction of the stimulus x
that
is concentration invariant: many x
stimuli with the same direction and different length will yield
a reconstructed concentration vector x
that is exactly the same. We therefore suggest that the
does not change the set of inequalities y
y
j
j
y
x
y
P
0
j
P
3
y
, thus leading to the same minimum 1l norm x
primacy constraints result in recovery of concentration-invariant stimuli, leading to the decoding
of the concentration-invariant odorant identity.
Similarly, computing a monotonically increasing non-linear function of each component of y
,
such as that given by equation (1), does not affect the reconstructed vector x
. Such an operation
does not affect the constraint y
. Within
our approach the odorant identity is therefore invariant with respect to the non-linear monotonic
transformations of receptor responses. One can therefore use linear and non-linear models of
receptor responses i.e. y
interchangingly, since this choice does not affect the solution.
Overall, using relative rather than absolute values of sensor responses to reconstruct stimuli makes
reconstruction invariant to a set of transformations, such as stimulus intensity etc. Although we
demonstrated this idea for the particular example of primacy coding, we propose that this neural
relativity rule could be used more generally by neural systems to transmit and recover signals
invariant to various transformations. Instead of minimizing the 1l norm alone to find a sparse
solution, one can minimize the elastic net functional:
and r
arg min
x
j
x
2
j
2
(6)
x
P
j
y
x
y
P
0
j
~
y
that differ in their primacy code
0 . For
The equation (6) is the same as the 1l norm minimization (5) when the parameter
sufficiently small , equation (6) yields sparse solutions similar to equation (5). Problem (6),
however allows a more straightforward formulation in the dual space. We will therefore use
equation (6), i.e. elastic net minimization, to recover olfactory stimuli for the rest of the paper,
with a sufficiently small .
2.4 Number of receptors needed to implement primacy coding. To estimate the number of
primary receptors p needed to recover the stimulus we estimate a possible number of
combinations in y
. The estimate for the
number of combinations in x
number of combinations of y
following condition:
For a typical mammal, such as a mouse, for which the number of olfactory receptors is
1000
approximately
. For humans, the number of functional olfactory
N
and, therefore the primacy number (the number of primacy receptor types)
receptors is
required is somewhat higher,
. In both of these estimates we assumed that the potential
6
. If one assumes
number of types of molecules available in the environment is close to
~ 12
that the number of discernable molecular components in each mixture is
, as follows
which is substantially less
from human psychophysics [5], we obtain the primacy number
than the total number of receptors
is.
has to exceed the number of combinations in x
. Assuming that for an exact reconstruction of x
the
, we arrive to the
N
350
, we obtain
x M
~ 10
K
p K
K
/ log
NC
p
~p
log
2.4
1000
N
(7)
30
M
M
N
K
N
~
p
p
.
K
2
p
2.5 Duality transformation. Since problem (6) includes various inequality constraints, one can
use the Lagrange multiplier method to solve it. To present the problem in Lagrange form, one has
to reformulate the conditions (4) in a more convenient form. Indeed, minimization of 1l norm or
elastic net functional of x
connected to y
Ax
)
x
0
. To obtain a non-zero solution, one has to introduce a finite scale
results in the solution
into the conditions (4). A set of equivalent conditions with constrained magnitude is
via a set of positive coefficients A ( y
(8)
,
y
y
P
P
4
j
j
(
i
1
y
i
(9)
A x
ij
0
iu for i P and
0 is the scale parameter. By introducing a sign variable
1
where
iu for i P , equation (8) can be rewritten as a single set of conditions
Here
u y
)
i
with receptor i to odorant j affinities given by a set of non-negative
ijA . Problem (6) combined with constraints (9) represents the primal optimization
numbers
problem.
To transform the primal minimization problem (6) to the dual problem, we introduce the
i, with
Lagrangian with two sets of Lagrange multipliers,
i
N
jx
0
correspond to the constraints (9), while the latter,
1..
for
. The full Lagrangian for the elastic net minimization problem (6) is as follows
L x
)
( ,
,
(10)
i
j are supposed to have non-negative values,
In this equation, the Lagrange multipliers
j [9]. To transform the primal problem into the dual problem, we minimize the
i.e.
Lagrangian (10) with respect to x
if no constraints were present to compute the optimal value of
x
denoted here as x
. We then compute the value of the Lagrangian in the minimum
,
)
,
)
,
(
:
L x
( ,
,
j . The former multipliers,
j , enforce the constraint
x
i and
i and
M
L x
(
1..
j
u
i
i
A x
ij
(11)
/ 2)
1) /
x
0
)
j
) /
x
(
x
,
i
x
j
0
(
j
2
j
j
j
j
j
j
j
(
i
i
j
u A
i
ij
i
u A
ij
i
1
2
j
2
2
j
(
j
)
(12)
,
)
(
G
ik
j
Here
k
u G
i
ik
i
ik
u
k
1
ij
(
u
1
A
ij
)
1
2
i
A A
ij
kj
i
i
in the Gramm matrix for rows of matrix A . According to optimization
j
,
)
and
. These values can be used to find the solution of the primal problem using equation (11). The
is to be maximized to find the optimal values of
theory [9] the dual Lagrangian
dual problem can therefore be formulated as follows
(13)
)
(
)
(
,
,
arg max (
j
0,
0
i
,
The reason why the dual problem (13) is preferable to the primal problem (6) lies in the simplicity
j are easier to implement than inequalities (9). The
of the constraints. The constraints
main observation that we make in this study is that, for the nervous system, it is especially easy to
impose non-negativity constraints, because neural responses are described by firing rates that
cannot fall below zero. Motivated by this observation, we will argue below that Lagrange
multipliers
could be represented by the responses of different types of olfactory
neurons that solve the dual rather than primal representation problem.
and
0
i
Another motivation for relating
to neural activity can be derived from the Karush-
Kuhn-Tucker theorem (KKTT) [9]. According to KKTT, at the allowed maximum of the dual
Lagrangian (12), the Lagrangian contributions in equation (10) vanish, i.e.
and
5
u
i
i
(
)
0
,
jx
0
j
j
(14)
The former equality means that if
A x
ij
j
0
(inactive constraint), the corresponding
0
i . KKTT therefore has significant implications from the standpoint of
Lagrange multiplier
i are identically equal zero (for inactive constraints)
neural representations. That some values of
makes the vector of responses sparse, which is one of the broadly observed features of neural
responses [10; 11; 12] including in the olfactory system [13; 14; 15; 16]. The mapping between
the dual problem and neural responses, if established, could explain the sparsity of neural
responses as a corollary of KKTT.
0
i and
H
,
)
2.6 Dual network. We will now describe the neural networks that solve the dual problem (12) and
(13). We associate vectors
and
with firing rates of two groups of neurons (cell types). The
0
j are then satisfied automatically as firing rates cannot be negative.
conditions
We then assume that the neurons are connected into a network and that the Lyapunov function of
H
)
the network
[cf. equation (12)]. To generate network equations from the Lyapunov function, for each neuron
in the network, we define internal variables that can be viewed as total synaptic input current, for
jb respectively. Consider
example. For
the following equations for
a
i
j units, these currents will be denoted
is proportional to the negative dual Lagrangian:
u
i
ia and
i and
(15)
A
ij
a
i
u
i
(
(
)
(
,
,
j
j
ia and
jb
W
ik
k
b
b
k
j
j
u A
i
ij
j
i
i
u A
ij
i
1
(16)
j
k
(17)
(18)
external excitatory drive equal to
i
Here and throughout this paper we use
with firing rates
These cells are also connected to
W u G u
ik
i
ik
ik
b
a
[
] ,
]
[
j
i
dx dt
x
/
. The first equation describes inputs into cells
ikW
. Connectivity between cells is symmetric.
i connected by weights
. cells also receive
j cells that are
j cells with synaptic weights
. Equation (16) describes
connected symmetrically to cells. For both cell types, their firing rates (
jb ) by rectifying threshold-linear relationships (18) ([ ]x
connected to their inputs (
0
0
for
rewritten as forms of gradient descend
,
j ) are
x
x ). It is straightforward to show that these equations can be
i
is indeed the Lyapunov function of equations (15)-(18), i.e. a function
To show that
that is not increasing when our network evolves according to these equations, we observe that
H
i
H
j
x and [ ]
x for
i and
ia and
iju A
,
(19)
A
ij
) /
u
i
) /
u
i
b
0
,
(
(
j
i
j
) /
dt
i
H
) /
(
,
i
j
H
) /
(
,
j
(
a
H
,
)
(
,
dH
j
A x
ij
j
i
a
i
i
j
i
b
j
j
i
f a a
)
'(
i
2
i
6
j
j
f b b
'(
)
j
2
j
0
(20)
0
f x
( )
x
[ ]
,
f
0
x . Because
'( )
Here
, our network will minimize the
Lyapunov function, and, by extension, maximize the dual Lagrangian (12). Because of the
constraints imposed on the firing rates (18), these variables will stay non-negative in the course of
this optimization, thus automatically satisfying the constraints imposed on the variables of the dual
Lagrangian. We conclude therefore that our network, which includes only two cell types, can
solve the dual constraint optimization problem, leading to the accurate computation of the
molecular composition of a mixture in dual space.
iu is to identify the set of primary variables
The purpose of variables
components of vector y
that are larger than all others. To compute these variables one could use
the network with winner-takes-all architecture, except one has to ensure that there are p winners.
This network can be implemented by connecting N neurons with an unstructured global
inhibition, in the manner described previously by us [17]:
iy , i.e. the set of p
,
c
v
i
v
i
u
k
y
i
k
,
(
) /
dt
dH
sgn(
u
i
ensure that a given number of u - cells is active.
v
i
)
. Here, c is the strength of global inhibition. By adjusting parameter c , one can
Figure 1. Simulations with the network described. (A) The structure of the network. We propose
that cells implementing the dual representation of the concentration vector reside in the
piriform cortex (PC). This is because PC is known to have extensive recurrent connectivity.
cells, which implement the non-negativity constraints on the concentration vector, could be related
to the granule cells of the olfactory bulb [14; 18]. (B) Firing rate simulation of the dual model
(blue) successfully identifies the non-zero components of the stimulus (top, red circles). Small
false positive components of x result from our use of the elastic net measure and disappear as
0 . They can be removed by thresholding the reconstruction. In the simulation, we used
p
60
.
3 Discussion
Herein we propose a novel model for the intensity-invariant encoding of olfactory stimuli.
According to the primacy model, an odor can be identified on the basis of the identities of the p
strongest responding olfactory receptors. In the case of the human olfactory system, we have
estimated p to be ~30. Because the primacy model relies on the relative rather than absolute
strengths of receptor responses, the recovered stimulus is independent wrt the absolute stimulus
7
concentration and depends only on the relative concentrations of individual molecules in the
mixture. Although we demonstrated this idea for the particular example of primacy coding, we
suggest that this rule of neural relativity could be used to produce intensity invariant signal
recovery for more complex conditions and in other modalities.
We then attempted to formulate the solution of the decoding problem from the first principles. To
implement primacy decoding, we cast this problem as a dual Lagrange problem, which required
introducing two sets of Lagrange multipliers. The first set () enforced the primacy conditions,
while the second set () ensured the non-negativity of the individual molecular concentrations.
By assuming that the dual Lagrangian corresponds to a Lyapunov function of a network, we were
able to derive the network structure. Interestingly, we found the cells were found to be
recurrently connected, while cells are connected to cells via inhibitory circuitry. These
features are reminiscent of the cells in real olfactory networks. Pyramidal cells in the piriform
cortex (PC) form extensive recurrent connections [19], while granule cells of the olfactory bulb
receive inhibitory connections from PC [18]. The connection from the granule cells to the PC is
indirectly inhibitory, because granule cells inhibit the mitral cells within the olfactory bulb
through dendrodendritic synapses. Mitral cells are excitatory cells projecting to PC. These
considerations allow us to place cells and related u cells into the piriform cortex and associate
cells with the granule cells of the olfactory bulb. Interestingly, the number of cells matches
the number of monomolecular components that the system can resolve, because these cells
implement the conditions of non-negativity of the molecular concentrations. The number of
granule cells found in the olfactory bulb is a few million, which matches with our estimate of the
number of volatile molecules [18]. This coincidence provides a further argument in favor of the
association of cells with granule cells.
Lengyel et al. [20] have proposed that the maximum a posteriori (MAP) solution to the inference
problem of recovering a sparse N-dimensional odor vector x can be achieved in a low-dimensional
measurement space, reflecting the known biology of olfactory processing. In their study, the
compressed sensing problem was formulated as an 1l norm minimization of the odor vector x
subject to the linear equality encoding constraint y = Ax and solved by considering the problem in
dual space. The resulting generalized energy function for the network reflects the equality
constraints and, as such, their proposed network implementation differs significantly from our
solution, which relies on Karush-Kuhn-Tucker-type inequality conditions implementing primacy.
Two features of dual networks are worth highlighting. First, according to Karush-Kuhn-Tucker
theory, dual Lagrangians are optimized under constraints of non-negativity of the Lagrange
coefficients. Neuronal responses can naturally enforce these constraints, because they are
described by firing rates that cannot fall below zero. Secondly, due to KKTT [9], a large number
of the Lagrange coefficients are bound to be zero, drawing an interesting parallel with SVMs. This
observation is consistent with the observed sparsity of neuronal responses, both in olfaction [13;
14; 15; 16] and beyond [10; 11; 12; 21], further strengthening the possible association between
neuronal responses and Lagrange coefficients. We therefore provide arguments in favor of relating
neuronal responses and Lagrange multipliers implementing several constraints present in the
stimulus.
8
4 References
[1] S. Firestein, How the olfactory system makes sense of scents. Nature 413 (2001) 211-
8.
[2] A. Koulakov, A. Gelperin, and D. Rinberg, Olfactory coding with all-or-nothing
glomeruli. J Neurophysiol 98 (2007) 3134-42.
[3] X. Zhang, and S. Firestein, The olfactory receptor gene superfamily of the mouse. Nat
Neurosci 5 (2002) 124-33.
[4] S. Kim, P.A. Thiessen, E.E. Bolton, J. Chen, G. Fu, A. Gindulyte, L. Han, J. He, S. He, B.A.
Shoemaker, J. Wang, B. Yu, J. Zhang, and S.H. Bryant, PubChem Substance and
Compound databases. Nucleic Acids Res 44 (2016) D1202-13.
[5] A. Jinks, and D.G. Laing, A limit in the processing of components in odour mixtures.
Perception 28 (1999) 395-404.
[6] R.G. Baraniuk, Compressive sensing. Ieee Signal Proc Mag 24 (2007) 118-+.
[7] D.L. Donoho, and J. Tanner, Sparse nonnegative solution of underdetermined linear
equations by linear programming. P Natl Acad Sci USA 102 (2005) 9446-9451.
[8] D.L. Donoho, and J. Tanner, Thresholds for the recovery of sparse solutions via L1
minimization. 2006 40th Annual Conference on Information Sciences and
Systems, Vols 1-4 (2006) 202-206.
[9] S.P. Boyd, and L. Vandenberghe, Convex optimization, Cambridge University Press,
Cambridge, UK ; New York, 2004.
[10] M.R. DeWeese, and A.M. Zador, Non-Gaussian membrane potential dynamics imply
sparse, synchronous activity in auditory cortex. J Neurosci 26 (2006) 12206-18.
[11] T. Hromadka, M.R. Deweese, and A.M. Zador, Sparse representation of sounds in
the unanesthetized auditory cortex. PLoS biology 6 (2008) e16.
[12] W.E. Vinje, and J.L. Gallant, Sparse coding and decorrelation in primary visual cortex
during natural vision. Science 287 (2000) 1273-6.
[13] L.M. Kay, and G. Laurent, Odor- and context-dependent modulation of mitral cell
activity in behaving rats. Nat Neurosci 2 (1999) 1003-9.
[14] A.A. Koulakov, and D. Rinberg, Sparse incomplete representations: a potential role
of olfactory granule cells. Neuron 72 (2011) 124-36.
[15] D. Rinberg, A. Koulakov, and A. Gelperin, Sparse odor coding in awake behaving
mice. J Neurosci 26 (2006) 8857-65.
[16] D.D. Stettler, and R. Axel, Representations of odor in the piriform cortex. Neuron 63
(2009) 854-64.
[17] H. Sanders, B.E. Kolterman, R. Shusterman, D. Rinberg, A. Koulakov, and J. Lisman, A
network that performs brute-force conversion of a temporal sequence to a
spatial pattern: relevance to odor recognition. Front Comput Neurosci 8 (2014)
108.
[18] G.M. Shepherd, The synaptic organization of the brain, Oxford University Press,
Oxford ; New York, 2004.
[19] K.M. Franks, M.J. Russo, D.L. Sosulski, A.A. Mulligan, S.A. Siegelbaum, and R. Axel,
Recurrent circuitry dynamically shapes the activation of piriform cortex. Neuron
72 (2011) 49-56.
[20] S.a.L. Tootoonian, M. , A Dual Algorithm for Olfactory Computation in the Locust
Brain. Advances in Neural Information Processing Systems 27 (NIPS 2014)
(2014).
9
[21] S.R. Lehky, T.J. Sejnowski, and R. Desimone, Selectivity and sparseness in the
responses of striate complex cells. Vision Res 45 (2005) 57-73.
10
|
1611.10252 | 1 | 1611 | 2016-11-29T18:11:00 | SeDMiD for Confusion Detection: Uncovering Mind State from Time Series Brain Wave Data | [
"q-bio.NC",
"cs.AI",
"cs.LG"
] | Understanding how brain functions has been an intriguing topic for years. With the recent progress on collecting massive data and developing advanced technology, people have become interested in addressing the challenge of decoding brain wave data into meaningful mind states, with many machine learning models and algorithms being revisited and developed, especially the ones that handle time series data because of the nature of brain waves. However, many of these time series models, like HMM with hidden state in discrete space or State Space Model with hidden state in continuous space, only work with one source of data and cannot handle different sources of information simultaneously. In this paper, we propose an extension of State Space Model to work with different sources of information together with its learning and inference algorithms. We apply this model to decode the mind state of students during lectures based on their brain waves and reach a significant better results compared to traditional methods. | q-bio.NC | q-bio |
JMLR: Workshop and Conference Proceedings 1 (2016) 1–11
NIPS 2016, Time Series Workshop
SeDMiD for Confusion Detection: Uncovering Mind State
from Time Series Brain Wave Data
Jingkang Yang
International School
Beijing University of Posts and Telecommunications
Beijing, 100876, China
Haohan Wang
Language Technologies Institute
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA, 15206, the USA
[email protected]
[email protected]
Jun Zhu
State Key Laboratory of Intelligent Technology and Systems
Tsinghua National Laboratory for Information Science and Technology
Tsinghua University
Beijing, 100084, China
[email protected]
Eric P. Xing
Machine Learning Department
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA, 15206, the USA
Editor: List of editors' names
[email protected]
Abstract
Understanding how brain functions has been an intriguing topic for years. With the recent
progress on collecting massive data and developing advanced technology, people have be-
come interested in addressing the challenge of decoding brain wave data into meaningful
mind states, with many machine learning models and algorithms being revisited and devel-
oped, especially the ones that handle time series data because of the nature of brain waves.
However, many of these time series models, like HMM with hidden state in discrete space
or State Space Model with hidden state in continuous space, only work with one source of
data and cannot handle different sources of information simultaneously. In this paper, we
propose an extension of State Space Model to work with different sources of information
together with its learning and inference algorithms. We apply this model to decode the
mind state of students during lectures based on their brain waves and reach a significant
better results compared to traditional methods.
Keywords: Sequence Data based Mind-Detecting (SeDMiD) Model, Time Series, Brain
Wave, Mind State Reading
c(cid:13) 2016 J. Yang, H. Wang, J. Zhu & E.P. Xing.
Yang Wang Zhu Xing
1. Introduction
Understanding how human brain functions has been an attractive research question in
recent years Mitchell et al. (2008). One important progress is on collecting a large amount
of brain wave data with different technologies, like fMRI Wehbe et al. (2014), MEG Sudre
et al. (2012) and EEG Wang et al. (2013). The nature of these data collecting technologies
have introduced a variety of substantial challenges in understanding these data with machine
learning techniques. For example, MEG and EEG technologies can describe the brain with
considerable temporal granularity, but with a relatively low spatial resolution. Therefore,
machine learning techniques that can handle temporal dependencies are highly appreciated.
Fortunately, in recent years, there is an increasing trend towards the use models to work
with time series problem. For example, Khaleghi and Ryabko (2013) find the points in time
where the probability distribution generating the data has changed given a heterogeneous
time-series sample. Anava et al. (2013) use regret minimization techniques to develop ef-
fective online learning algorithms for predicting a time series using autoregressive moving
average model. Alon et al. (2003) fit a finite mixture of HMMs in motion data, using
the expectation maximization (EM) framework, aiming at discovering groupings of similar
object motions that were observed in a video collection. Also, Hidden Markov Models or
other graph-based methods are often used for the purpose like speech recognition, pattern
recognition and neural networks, similar to Nahar et al. (2016), Schwenk (1999), Bar-Joseph
(2004) and LeCun and Bengio (1995). In a more rigorous setting, a lot more theoretical
questions are being asked and solved for time series machine learning these days. For exam-
ple, Khaleghi and Ryabko (2013, 2014) derive theories for estimation of highly dependent
time series data. Kuznetsov and Mohri (2014, 2015) push the theoretical work further for
non-stationary time series problems. Rakhlin and Sridharan (2013); ? naturally combines
the problem of analyzing time series data with online learning, which opens the door to a
whole area of new problems.
With the guidance of previous work on time series data.
In this paper, we present
Sequence Data based Mind-Detecting (SeDMiD) Model , a novel time series method to un-
cover the state of brain using more than a typical source of EEG/MEG recordxings. Others
sources like video recordings or audio data can be supplemented for better performance on
brain state estimation.
We also develop the learning algorithm for SeDMiD based on sparsity regularized linear
system, and the inference algorithm as an extension of Viterbi algorithm under Gaussian
assumption for continuous space. The results show that, the performance of SeDMiD can
uncover the mind state based on brain waves with a significant better results than traditional
methods.
Our contribution of this paper is three-fold:
• We propose a SeDMiD that can analyze time series brain wave data with extra sources
of information.
• We improve the existing vertibi algorithm to enable the inference of SeDMiD model.
• We show the possibility of deciphering students' mind state of understanding lectures
with brain wave data.
2
SeDMiD for Confusion Detection
The rest of this paper is organized as following. Section 2 describes some work that
others did to solve the prediction problem using MEG data. Section 3 raises the novel
model and describe its learning and inference method, while section 4 shows the result of
our experiment. Finally, Section 5 concludes and suggests future work.
2. Related Work
There are many works implementing time series technology to deal with the problem about
MEG/EEG data. It is worth noting that both electroencephalography (EEG) and mag-
netoencephalography (MEG) provide a more direct measure of the electrical activity in
the brain, as is described professionally in the work of Michel (2009) and Proudfoot et al.
(2014). He et al. (2008) measure the difference in electric potentials on the scalp and cap-
tures high frequency oscillations on the millisecond timescale that is most relevant for the
characterisation of cognitive processes. There are many existing work to do experiments
and discover new knowledge in several fields. Moghadamfalahi et al. (2015) use abstract-
noninvasive EEG-based braincomputer interfaces (BCI) for intent detection, specifically for
EEG-based BCI typing systems. Phillips et al. (1997) have developed a Bayesian framework
for image estimation from combined MEG/EEG data.
EEG signal is a kind of voltage signal that can be measured on the surface of the scalp,
arising from large areas of coordinated neural activity manifested as synchronization (groups
of neurons firing at the same rate), described by Niedermeyer and da Silva (2005). This
neural activity varies as a function of development, mental state, and cognitive activity,
and the EEG signal can measurably detect such variation Marosi et al. (2002), Lutsyuk
et al. (2006), Berka et al. (2007), Wang et al. (2013) which in turn are important for and
predictive of learning Baker et al. (2010).
On the other hand, hidden Markov model is also widely implemented to make best use
of MEG/EEG recordings. Rukat et al. (2016) analyses the temporal and spatial dynamics
of physiological substrate of cognitive processes as measured by EEG, with a hidden Markov
model. Liu et al. (2010) combine kernel principal component analysis (KPCA) and HMM
to differentiate mental fatigue states with the help of EEG data. Ko and Sim (2011)
describe a procedure of classification of motor imagery EEG signals using HMM, which can
tell the person is performing left, right hand or foots motor imagery based on the current
EEG recordings. What is more, another bio-signal named electrocardiogram (ECG) is also
suitable for Markov model. Coast et al. (1990) and Andreao et al. (2006) describe a new
approach to ECG arrhythmia analysis based on HMM. Inspired by these work, we propose
a novel model to analyze time series brain wave data with extra sources of information.
3. SeDMiD Model and Its Learning and Inference Algorithm
We propose a novel state-space model for inferring people's state using several sources of
information simultaneously including MEG/EEG recordings, named Sequence Data based
Mind-Detecting (SeDMiD) Model. In this section, we will first introduce the model, and
then show its learning and inference algorithm respectively.
3
Yang Wang Zhu Xing
Figure 1: The proposed model. For the time-stamp at i, Mi corresponds to EEG/MEG
recording and Ci corresponds to mind states (e.g. happy/sad, confused/non-
confused). Si is extra source information (e.g. video/audio data). A, B, E,
F and G are stationary transition matrices that describe the dynamics among
states in the graph.
3.1. SeDMiD
The aim of our work is to raise a model for brain-state estimation using MEG/EEG record-
ings while considering other essential time-series sources at the same time. A pictorial
representation of our model can be found in Figure 1. Notations of this paper is illustrated
in the caption of Figure 1.
In the SeDMiD model, we firstly assume that the extra time-series sources can match
exactly with brain wave signals in aspect of time stamps, which means Si, Ci and Mi
are sampled at the same time. And we simplify the mind state inferring task as a binary
decision of states Ci, indicating that people is either happy or sad, or students are either
confused or not. We also assume that MEG/EEG recordings Mi have linear dependence
on current external source and current mind state, since brain signal is easily affected
by people's internal mind situation and outside influence on them Baker et al. (2010).
Additionally, current mind state Mi depends linearly on previous status and current video
contents, because brain state often has a close connection with its state few seconds ago,
and other sources like videos also affect the mind state. Finally, the complementary source
is a continuous time-series process, thus current recording is dependent on the former one,
and to simplify the model, we assume linear dependency again.
Further, we assume that the supplemental source data {Si}n
i=1 can be described by
Gaussian distribution.
We also assume that mind states {Ci}n
i=1 is in form of Gaussian distribution.
S1 ∼ N (µs, σs)
Si ∼ N (A · Si−1, Λi)
C1 ∼ N (µc, σc)
4
SeDMiD for Confusion Detection
Ci ∼ Nl(E · Si + B · Ci−1), Ωi
Considering the assumption that only on-and-off state exist in space C, we use the function
l() to map the score of positive state (E · Si + B · Ci−1) into two dimensions above.
(cid:20) 1
(cid:21)
1+ex
ex
1+ex
l(x) =
Thus brain wave recordings {Mi}n
linear combination of {Si}n
co-variance is Σi.
i=1 and {Ci}n
i=1 also follows Gaussian distribution because it is the
i=1, as the mean of Mi is G · Si + F · Ci and
Mi ∼ N (G · Si + F · Ci, Σi)
The practical meaning of notations in the model is described by Table 1.
Table 1: Practical Meaning of Notations in SeDMid model
Notation Description
Si
Ci
Mi
A
B
E
F
G
Features of supplemental source information at time i
Mind States at time i
Brain wave signal recordings (EEG/MEG) at time i
Linear Relationship between adjacent supplemental source information
Linear Relationship between adjacent mind states
Linear Relationship between supplemental source and mind state
Linear Relationship between mind state and brain wave signal
Linear Relationship between supplemental source and brain wave signal
3.2. Learning
Here we introduce the parameter learning algorithm of A, B, E, F and G for SeDMiD.
Because of the nature of the task, a supervised training training procedure, with S, C, M
known, is sufficient. At training, at time-stamp i, we can observe MEG/EEG recording
Mi, supplemental source data Si and current mind state Ci. Thus, parameter learning
is a maximum likelihood estimation (MLE) problem, in which case linear regression is the
solution. However, supplemental source data generally comes with a higher dimension than
response variable space. Therefore, sparsity regularize is required for transition matrix S
by Equation (3) in a sparse form.
Mi =(cid:2)G F(cid:3)(cid:20)Si
Ci =(cid:2)B E(cid:3)(cid:20) Ci
Ci
(cid:21)
(cid:21)
Si+1
Si+1 = A · Si
5
(1)
(2)
(3)
Yang Wang Zhu Xing
3.3. Inference
In inference, observation only contains brain wave recordings, so we want to estimate mind
state Ci and even further inference complementary data Si. With the observation Mi and
five transition matrices that learned in the learning phase, we formulate an inference method
to find the best sequence of mind states C1, C2, . . . , CT and features of complementary data
S1, S2, . . . , ST . A natural choice for inference on the model is Viterbi algorithm Forney
(1973). However, the problem is that the state space S and M are infinite, and we assume
that Si and Mi follow Gaussian distribution, we formulate the inference mathematically
in Gaussian form. Consider the Viterbi function at the time-stamp t is calculated by
Equation (4), given by the factorization of graphic model shown in Figure 3.1.
Vt(St, Ct) = max
St−1,Ct−1
[P (MtSt, Ct)P (StSt−1)P (CtSt, Ct−1)Vt−1(St−1, Ct−1)]
(4)
To begin with, we consider to calculate the Viterbi function V1 at the start point of the
whole process when t = 1.
V1(S1, C1) = P (M1S1, C1)P (S1)P (C1S1)
= N (G · S1 + F · C1, Σ) · N (µs, σs) · N (l(E · µs), σc)
(5)
According to the property of multiplication in Gaussian distribution, we obtain that V1
is also in form of Gaussian with its mean and variance shown in (6) and (7) respectively.
(cid:20)σ−1
(cid:20)
s
0
(cid:21)
(cid:21)
(cid:20)GT
(cid:21)
F T
Σ−1(cid:2)G F(cid:3)−1
(cid:21)
(cid:20)GT
Σ−1M1
F T
+
c
+
0
σ−1
σ−1
s µs
c l(E · µs)
σ−1
ΣV1 =
µV1 = ΣV1
(6)
(7)
For now, the mean and covariance of Viterbi function for the beginning point have been
calculated, and in order to know all the Viterbi function at every time-stamp, we find when
t (cid:62) 2:
Vt(St, Ct) = max
St−1,Ct−1
P (CtSt, Ct−1)P (StSt−1)Vt−1(St−1, Ct−1)P (MtSt, Ct)
(8)
Also, we can calculate the mean and covariance since for time t, St and Ct are in the form
of Gaussian distribution.
ΣV2 =
µV2 = ΣV2
F T
Φ−1 +
(cid:20)GT
(cid:21)
Φ−1 · µV1 +
(cid:20)(A−1)T
Σ−1(cid:2)G F(cid:3)−1
(cid:20)GT
(cid:21)
(cid:20)A−1
(cid:21)
Σ−1M2
F T
Σ−1
(B−1)T
V1
0
0
0
0 B−1
(9)
(10)
(11)
(cid:21)−1
where
(cid:20)Λ−1
0
0 Ω−1
(cid:21)
Φ =
+
6
SeDMiD for Confusion Detection
Once the means and co-variances are calculated for all supplemental source data and
mind state sequence, the best sequence of S and C can be inferred by calculating backwards.
So we start the inference at the end of the time sequence.
(cid:21)
(cid:20)ST
CT
= arg max
ST ,CT
N (µVT , ΣVT )
(12)
Note that the mean of the final state is calculated exactly, we proceed to infer the former
state that lead to the current one, so the formula to calculate the former state based on
current state is shown in (13).(cid:20)ST−1
(cid:21)
(cid:20)(E · A)T
(cid:21)
(cid:20)(E · A)T
Σ−1
Σ−1
µVT−1 +
CT−1
where
γ = τ
VT−1
VT−1
τ =
BT
+
BT
= arg max
N (γ, τ )
ST −1,CT −1
Ω−1(cid:2)E · A B(cid:3) +
(cid:21)
(cid:21)
(cid:20)ST
Ω−1
+
(cid:21)−1
(cid:20)AT Λ−1A 0
(cid:21)−1
(cid:21)(cid:20)ST
(cid:20)AT Λ−1A 0
0
0
0
0
CT
CT
(13)
(14)
(15)
Finally, by calculate all the Si and Ci from end to the first one, we can infer mind states
for each time-stamp t ∈ {1, . . . , T}.
4. Experiment
To solve the problem that we raised in Section 4.1, we need to extract features of lecture
videos. Using SeDMid model that we raise in Section 3,
4.1. Experiment Setting
In our experiment, we set up a task to estimate students' mind states in a given lecture
session, finding out they are confused or not with the data from Wang et al. (2013). Every
participants is asked to watch 10 lecture videos, the length of which is around 2 minutes.The
two-minute period is called an 'experiment period'. In all there are 10 students participate
in the experiments so there are 100 data points in total. During every experiment period,
their EEG signals are recorded in the frequency of 1 Hz, and EEG signals here have 11
features, including Proprietary measure of mental focus, 1-3 Hz of power spectrum and so
on. Students are asked to annotate whether they are confused or not based on every whole
experiment period, and every second in the period by the annotation is noted. The beginning
and the end of every experiment is cut off with consideration of reducing noise, only left the
middle 112 seconds for analysis. Finally, lecture videos are served as supplemental source
data.
4.2. Video Feature Extraction
We use the tool kit OpenCV developed by Bradski and Kaehler (2008) to extract video
features like optical features and object movement information, and openSMILE developed
7
Yang Wang Zhu Xing
by Eyben et al. (2010) to extract features in audio data, such as lecturer's speech speed and
intonation. As a result, we obtain video image features with 1440 dimensions while audio
features have 6669 dimensions, thus we get 8109 dimensions for video features in total.
Since both MEG recordings and students' status vector is collected in the frequency of 1
Hz, we sample the feature vector every 1 second in order to alignment the data for SeDMiD
model.
4.3. Performance Comparison
We use the simple state space model (SSM) as baseline, which only make use of EEG data,
and current mind state only depends on the former state, while current EEG recording
depends on current mind state. SSM does not make use of video features. We also compare
our model with logistic regression, which employ brain wave signals without other sources.
As is shown in the ROC curve in Figure 2, we find that our SedMid model outperform the
simple HMM model and logistic regression, with the accuracy of ours reaches 87.76% while
simple HMM only gets 53% and logistic regression 60%. In the experiment, We find that
errors that SeDMiD makes always exist at the beginning of experiment period, and it will
lead to correct result in few time. It is intuitive that SeDMiD model often makes mistakes
at beginning of every experiment period because it can perform better with more data gets
concluded.
Figure 2: The ROC curve for comparing the performance of SeDMiD model, Simple State
Model (SSM) and logistic regression. SeDMiD greatly outperforms the others.
8
SeDMiD for Confusion Detection
5. Conclusion
In this work, we propose an novel state space model called Sequence Data based Mind-
Detecting (SeDMiD) Model, which analyzes time series brain wave data with extra sources
of information, improving the existing vertibi algorithm to enable the inference. We evaluate
the effectiveness of SeDMiD model by comparing with simple Markov model and logistic
regression. The performance of our model has a 30% higher in accuracy than the normal
one.
Apart from proposing the SeDMiD model, our contribution includes showing the possi-
bility of deciphering students' mind state of understanding lectures with brain wave data.
This work is also useful in real world implementation, since teachers can modify their teach-
ing strategy based on audiences' status if it can be inferred.
Acknowledgments
This work was supported by International School, Beijing University of Posts and Telecom-
munications. We thank Hao Ding for advising the method of extracting videos' features
and dormitory administrator in Student Building 5, BUPT, for providing electronic power
for running experiment at night.
References
Jonathan Alon, Stan Sclaroff, George Kollios, and Vladimir Pavlovic. Discovering clusters in
motion time-series data. In Computer Vision and Pattern Recognition, 2003. Proceedings.
2003 IEEE Computer Society Conference on, volume 1, pages I–375. IEEE, 2003.
Oren Anava, Elad Hazan, Shie Mannor, and Ohad Shamir. Online learning for time series
prediction. In COLT, pages 172–184, 2013.
Rodrigo Varejao Andreao, Bernadette Dorizzi, and J´erome Boudy. Ecg signal analysis
through hidden markov models. IEEE Transactions on Biomedical engineering, 53(8):
1541–1549, 2006.
Ryan SJd Baker, Sidney K D'Mello, Ma Mercedes T Rodrigo, and Arthur C Graesser.
Better to be frustrated than bored: The incidence, persistence, and impact of learners
cognitive–affective states during interactions with three different computer-based learning
environments. International Journal of Human-Computer Studies, 68(4):223–241, 2010.
Ziv Bar-Joseph. Analyzing time series gene expression data. Bioinformatics, 20(16):2493–
2503, 2004.
Chris Berka, Daniel J Levendowski, Michelle N Lumicao, Alan Yau, Gene Davis, Vladimir T
Zivkovic, Richard E Olmstead, Patrice D Tremoulet, and Patrick L Craven. Eeg corre-
lates of task engagement and mental workload in vigilance, learning, and memory tasks.
Aviation, space, and environmental medicine, 78(Supplement 1):B231–B244, 2007.
Gary Bradski and Adrian Kaehler. Learning OpenCV: Computer vision with the OpenCV
library. " O'Reilly Media, Inc.", 2008.
9
Yang Wang Zhu Xing
Douglas A Coast, Richard M Stern, Gerald G Cano, and Stanley A Briller. An approach
IEEE Transactions on
to cardiac arrhythmia analysis using hidden markov models.
biomedical Engineering, 37(9):826–836, 1990.
Florian Eyben, Martin Wollmer, and Bjorn Schuller. Opensmile: the munich versatile and
fast open-source audio feature extractor. In Proceedings of the 18th ACM international
conference on Multimedia, pages 1459–1462. ACM, 2010.
G David Forney. The viterbi algorithm. Proceedings of the IEEE, 61(3):268–278, 1973.
Biyu J He, Abraham Z Snyder, John M Zempel, Matthew D Smyth, and Marcus E Raichle.
Electrophysiological correlates of the brain's intrinsic large-scale functional architecture.
Proceedings of the National Academy of Sciences, 105(41):16039–16044, 2008.
Azadeh Khaleghi and Daniil Ryabko. Nonparametric multiple change point estimation
in highly dependent time series. In International Conference on Algorithmic Learning
Theory, pages 382–396. Springer, 2013.
Azadeh Khaleghi and Daniil Ryabko. Asymptotically consistent estimation of the number
of change points in highly dependent time series. In ICML, pages 539–547, 2014.
Kwang-Eun Ko and Kwee-Bo Sim. Hsa-based hmm optimization method for analyzing eeg
pattern of motor imagery. Journal of Institute of Control, Robotics and Systems, 17(8):
747–752, 2011.
Vitaly Kuznetsov and Mehryar Mohri. Generalization bounds for time series prediction with
non-stationary processes. In International Conference on Algorithmic Learning Theory,
pages 260–274. Springer, 2014.
Vitaly Kuznetsov and Mehryar Mohri. Learning theory and algorithms for forecasting non-
In Advances in Neural Information Processing Systems, pages
stationary time series.
541–549, 2015.
Yann LeCun and Yoshua Bengio. Convolutional networks for images, speech, and time
series. The handbook of brain theory and neural networks, 3361(10):1995, 1995.
Jianping Liu, Chong Zhang, and Chongxun Zheng. Eeg-based estimation of mental fa-
tigue by using kpca–hmm and complexity parameters. Biomedical Signal Processing and
Control, 5(2):124–130, 2010.
NV Lutsyuk, EV ´Eismont, and VB Pavlenko. Correlation of the characteristics of eeg
potentials with the indices of attention in 12-to 13-year-old children. Neurophysiology, 38
(3):209–216, 2006.
Erzs´ebet Marosi, Oscar Baz´an, Guillermina Yanez, Jorge Bernal, Thal´ıa Fern´andez, Mario
Rodr´ıguez, Juan Silva, and Alfonso Reyes. Narrow-band spectral measurements of eeg
during emotional tasks. International Journal of Neuroscience, 112(7):871–891, 2002.
Christoph M Michel. Electrical neuroimaging. Cambridge University Press, 2009.
10
SeDMiD for Confusion Detection
Tom M Mitchell, Svetlana V Shinkareva, Andrew Carlson, Kai-Min Chang, Vicente L
Malave, Robert A Mason, and Marcel Adam Just. Predicting human brain activity
associated with the meanings of nouns. science, 320(5880):1191–1195, 2008.
Mohammad Moghadamfalahi, Umut Orhan, Murat Akcakaya, Hooman Nezamfar, Melanie
Fried-Oken, and Deniz Erdogmus. Language-model assisted brain computer interface for
typing: a comparison of matrix and rapid serial visual presentation. IEEE Transactions
on Neural Systems and Rehabilitation Engineering, 23(5):910–920, 2015.
Khalid MO Nahar, Mohammed Abu Shquier, Wasfi G Al-Khatib, Husni Al-Muhtaseb,
and Moustafa Elshafei. Arabic phonemes recognition using hybrid lvq/hmm model for
continuous speech recognition. International Journal of Speech Technology, pages 1–14,
2016.
Ernst Niedermeyer and FH Lopes da Silva. Electroencephalography: basic principles, clinical
applications, and related fields. Lippincott Williams & Wilkins, 2005.
James W Phillips, Richard M Leahy, John C Mosher, and Bijan Timsari. Imaging neural
activity using meg and eeg. IEEE Engineering in Medicine and Biology Magazine, 16(3):
34–42, 1997.
Malcolm Proudfoot, Mark W Woolrich, Anna C Nobre, and Martin R Turner. Magnetoen-
cephalography. Practical neurology, pages practneurol–2013, 2014.
Alexander Rakhlin and Karthik Sridharan. Online learning with predictable sequences. In
COLT, pages 993–1019, 2013.
Tammo Rukat, Adam Baker, Andrew Quinn, and Mark Woolrich. Resting state brain
arXiv preprint
networks from eeg: Hidden markov states vs. classical microstates.
arXiv:1606.02344, 2016.
Holger Schwenk. Using boosting to improve a hybrid hmm/neural network speech rec-
In Acoustics, Speech, and Signal Processing, 1999. Proceedings., 1999 IEEE
ognizer.
International Conference on, volume 2, pages 1009–1012. IEEE, 1999.
Gustavo Sudre, Dean Pomerleau, Mark Palatucci, Leila Wehbe, Alona Fyshe, Riitta
Salmelin, and Tom Mitchell. Tracking neural coding of perceptual and semantic fea-
tures of concrete nouns. NeuroImage, 62(1):451–463, 2012.
Haohan Wang, Yiwei Li, Xiaobo Hu, Yucong Yang, Zhu Meng, and Kai-min Chang. Using
eeg to improve massive open online courses feedback interaction. In AIED Workshops,
2013.
Leila Wehbe, Brian Murphy, Partha Talukdar, Alona Fyshe, Aaditya Ramdas, and Tom
Mitchell. Simultaneously uncovering the patterns of brain regions involved in different
story reading subprocesses. PloS one, 9(11):e112575, 2014.
11
|
1612.02243 | 1 | 1612 | 2016-12-07T13:47:58 | How structure sculpts function: unveiling the contribution of anatomical connectivity to the brain's spontaneous correlation structure | [
"q-bio.NC"
] | Intrinsic brain activity is characterized by highly structured co-activations between different regions, whose origin is still under debate. In this paper, we address the question whether it is possible to unveil how the underlying anatomical connectivity shape the brain's spontaneous correlation structure. We start from the assumption that in order for two nodes to exhibit large covariation, they must be exposed to similar input patterns from the entire network. We then acknowledge that information rarely spreads only along an unique route, but rather travels along all possible paths. In real networks the strength of local perturbations tends to decay as they propagate away from the sources, leading to a progressive attenuation of the original information content and, thus, of their influence. We use these notions to derive a novel analytical measure, $\mathcal{T}$ , which quantifies the similarity of the whole-network input patterns arriving at any two nodes only due to the underlying topology, in what is a generalization of the matching index. We show that this measure of topological similarity can indeed be used to predict the contribution of network topology to the expected correlation structure, thus unveiling the mechanism behind the tight but elusive relationship between structure and function in complex networks. Finally, we use this measure to investigate brain connectivity, showing that information about the topology defined by the complex fabric of brain axonal pathways specifies to a large extent the time-average functional connectivity observed at rest. | q-bio.NC | q-bio |
How structure sculpts function: unveiling the contribution of anatomical
connectivity to the brain's spontaneous correlation structure
R. G. Bettinardi,1, 2, a) G. Deco,1, 2, 3 V. M. Karlaftis,4 T. J. Van Hartevelt,5, 6 H. M. Fernandes,5, 6 Z. Kourtzi,4
M. L. Kringelbach,5, 6 and G. Zamora-L´opez1, 2, b)
1)Center for Brain and Cognition, Universitat Pompeu Fabra, Barcelona, Spain.
2)Department of Information and Communication Technologies, Universitat Pompeu Fabra, Barcelona,
Spain.
3)Instituci´o Catalana de la Recerca i Estudis Avan¸cats, Universitat Pompeu Fabra, Barcelona,
Spain.
4)Department of Psychology, University of Cambridge, United Kingdom.
5)Department of Psychiatry, University of Oxford, United Kingdom
6)Center for Music in the Brain, Aarhus University, Aarhus, 8000 Aarhus C,
Denmark
Intrinsic brain activity is characterized by highly structured co-activations between different regions, whose
origin is still under debate.
In this paper, we address the question whether it is possible to unveil how
the underlying anatomical connectivity shape the brain's spontaneous correlation structure. We start from
the assumption that in order for two nodes to exhibit large covariation, they must be exposed to similar
input patterns from the entire network. We then acknowledge that information rarely spreads only along an
unique route, but rather travels along all possible paths. In real networks the strength of local perturbations
tends to decay as they propagate away from the sources, leading to a progressive attenuation of the original
information content and, thus, of their influence. We use these notions to derive a novel analytical measure,
T , which quantifies the similarity of the whole-network input patterns arriving at any two nodes only
due to the underlying topology, in what is a generalization of the matching index. We show that this
measure of topological similarity can indeed be used to predict the contribution of network topology to
the expected correlation structure, thus unveiling the mechanism behind the tight but elusive relationship
between structure and function in complex networks. Finally, we use this measure to investigate brain
connectivity, showing that information about the topology defined by the complex fabric of brain axonal
pathways specifies to a large extent the time-average functional connectivity observed at rest.
I.
INTRODUCTION
In the last two decades a large body of research
has demonstrated that spontaneous brain activity forms
structured patterns of consistent co-activations across
different subsets of brain regions 1 -- 10. Contrary to what
was somehow implicitly assumed, intrinsic brain activity
cannot be considered as a simple sum of rather unpre-
dictable and noisy fluctuations. Spontaneous collective
dynamics of different brain regions, measured either with
EEG, fMRI or MEG, can be clustered into highly orga-
nized and reproducible spatial patterns, referred to as
resting-state networks (RSNs), that strikingly resemble
those activations observed during the performance of dif-
ferent tasks7,11,12. Furthermore, growing evidences sug-
gest that this spontaneous large-scale structure is char-
acterized by a marked temporal organization, mirrored
by recurrent alternations of subnetworks that concur in
generating a rich dynamical repertoire at different time
scales13,14. Although the origin and the purpose of spon-
taneous brain dynamics are still under debate, there is
wide agreement that mental states and brain malfunc-
tion alter the patterns of spontaneous activity, e.g. the
a)Electronic mail: [email protected]
b)Electronic mail: [email protected]
dynamical repertoire of the brain at rest tends to decrease
during sleep15 and under anaesthesia16 -- 18.
From its discovery, great efforts have been invested in
trying to reproduce resting-state brain activity through
the use of computational models, in order to obtain a
mechanistic explanation of this intriguing but elusive
phenomenon. Early models based on the structural con-
nectomes of cats and macaques extensively explored the
emerging patterns of correlations in those networks at dif-
ferent spatial and temporal scales19 -- 21. With the arrival
of structural human connectomes obtained through trac-
tography, computational models could finally attempt
to fit the empirical correlation structure observed via
resting-state fMRI22 -- 26. Despite these attempts, we still
lack of a precise understanding of the relationship be-
tween the shape of the brain's connectome and the emer-
gent patterns of correlations observed during rest. One
of the major reasons is that the collective dynamics of
a network do not only depend on the shape of the un-
derlying connectivity, but also on the model chosen to
simulate the local dynamics of the cortical regions26,27.
In this paper we aim at unveiling what is the precise
contribution of the anatomical connectivity on the cor-
relation structure observed during rest. To this aim, we
first need to derive a theoretical estimate of the expected
correlation between nodes due to the network's topology
alone; therefore we will try to avoid, as much as possible,
the contribution of other factors. To do so, in Section II
we introduce a novel graph theoretical quantity measur-
ing the topological similarity of the entire input profiles
that two nodes receive from the whole network. This
measure, that we named topological similarity, is a gener-
alization of the concept of matching index that explicitly
accounts for the fact that in networks information trav-
els along all possible paths, not only along the shortest
ones, and that information content tends to decay as it
moves away from its source28 -- 31. This measure is based
on the concept of network's communicability introduced
by Estrada and Hatano 32, a function that quantifies the
strength of the influence that one node exerts over an-
other through all paths of any length assuming an ex-
ponential decay of influence with path length. We then
demonstrate that the topological similarity function, al-
tough based on pure topological information about the
underlying path structure of the network, can indeed be
used to approximate the expected correlation structure
of networks of time-varying coupled units.
In Section III we systematically investigate the con-
tribution of three topological primitives: the weight of
the links, the length of the path and the presence of
redundant alternative paths through which information
can travel. Topological primitives are fundamental fea-
tures underlying the network architecture which deter-
mine how influence spreads, thus sculpting the similarity
of the input profiles of nodes. To do so, we study simple
graphs in which these primitive features can be manip-
ulated. We show how these three features are in fact of
primary importance in modulating the strength of the
influence and the relative topologies into which different
nodes are embedded.
Finally, in Section IV we investigate the spontaneous
correlation structure of the human brain, demonstrat-
ing that taking into account the similarity of the whole-
network influences shaped by the underlying anatomi-
cal topology it is possible to understand and predict the
large-scale functional organization observed during rest.
II. HOW TOPOLOGY SCULPTS THE CORRELATION
STRUCTURE OF NETWORKS
In order to properly address the contribution of net-
work's topology on the emerging spontaneous correlation
structure, we first need to acknowledge that the collective
behaviour of a set of coupled dynamical units depends on
three principal ingredients: (i) the structure of the net-
work, (ii) the local dynamics of the nodes and (iii) the
coupling function determining how information is passed
from one node to another. In fact, for a fixed network,
changing the local dynamical model of the nodes and
the coupling function usually leads to different collective
dynamics26,27,33. Therefore, in order to estimate the con-
tribution of structure alone we need to set apart the role
of the other two factors.
Typically, the activity of two nodes exhibits statistical
2
dependence either if they are connected by means of a
direct link, or if the aggregate inputs they receive from
the entire network are similar, independently of whether
there is a link between the two or not. Because infor-
mation in a network rarely travels exclusively along the
shortest paths34,35 but instead diffuses along the whole
network, we realise that the total influence of one node
over another mainly depends on three topological fea-
tures: (i) the strength of the coupling between them, usu-
ally represented by the weights of the links, (ii) the graph
distance between the two nodes and (iii) the propagation
of the influence along multiple alternative paths36,37. We
refer to these three topological features as the topologi-
cal primitives because, in combination, they encode the
shape of the network's path structure. As such the anal-
ysis of the building blocks underlying larger complex net-
works is a necessary step to understand the contribution
of network structure to the emergent activity.
In general, the influence of a direct link is greater than
the influence exerted over longer paths, as the latter is
mediated through third nodes. In fact, in real systems
the "power" of the signals or their information content
naturally decays along the path28 -- 31, unless there exists
an active mechanism which amplifies the incoming signal
at the cost of energy. Following the same rationale, it
is unlikely that influence or information from one node
to another propagates only along a single, selected path,
unless there are specific gating mechanisms controlling
the detailed routing of information over all existing paths.
The total number of paths (of all lengths) between any
two nodes in a network is in fact infinite. It is well known
that the total number of paths ml
ij of length l between
nodes i and j in a graph grows with l. This number is
given exactly by the lth power of the adjacency matrix
A, mij = (Al)ij
38,39. Thus, the total number of paths
leaving from node i and arriving at node j is given by
the sum:
∞(cid:88)
l=0
(Al)ij = 1 + Aij + (A2)ij + (A3)ij + . . .
(1)
which the series(cid:80)∞
This number typically diverges and thus, for the dy-
namics within a network to remain bounded, the amount
of influence needs to decay faster with the length than
the growth in the number of paths. Mathematically, the
problem consists in finding a set of coefficients {kl} for
l=0 kl Al converges for any adjacency
matrix, A. While the solution to this problem is not
unique, Estrada and Hatano 32 proposed an exponential
decay of the influence with path length and introduced
the communicability measure C. The communicability
function thus corresponds to the matrix exponential of
A, which can be expanded into a series of powers with
coefficients kl = 1/ l!:
C ≡ eA =
∞(cid:88)
l=0
Al
l!
= 1 + A +
A2
2!
+
A3
3!
+ . . .
(2)
From a physical perspective, the communicability is
analogous to the Green's function of the network32,40 and
expresses how local perturbations propagate along the
system. Perturbation of a given node's dynamics in fact
first propagates to its direct neighbours, affecting their
activity; the activity of the neighbours will in turn prop-
agate to their neighbours (as well as back to the node
that was perturbed in the first place), and interact with
their intrinsic local dynamics. This simple propagation
mechanism implies that the effect of a local perturbation
could be perceived also by distant nodes but attenuated
and modulated by the dynamics of each node along the
route. Due to its correspondence to the Green's function,
the communicability can be tuned using a constant global
coupling parameter that uniformly scales the weights of
all links in A40,41, allowing to search through the emerg-
ing collective dynamics over multiple scales. The gener-
alised, tunable, communicability is then:
∞(cid:88)
l=0
C = egA =
glAl
l!
= 1 + gA +
g2A2
2!
+
g3A3
3!
+ . . . (3)
When g is weak, perturbations quickly decay, produc-
ing local correlations only around the node's neighbour-
hood. As g grows perturbations propagate deeper into
the network, giving raise to stronger correlations over
more distant nodes. Zamora-L´opez et al. 41 have shown
that considering the communicability as the propagator
kernel for the diffusion of Gaussian noise along the net-
work, it reveals an equivalent correlation structure as net-
works of generic and widely used models, e.g. Kuramoto
oscillators and neural masses (see Supplementary Infor-
mation in Ref. 41). Under these assumptions (Gaussian
white noise sources and exponential decay), in Ref 41 it
was shown that the covariance matrix of the system, Σ,
can be analytically estimated as Σ = C · CT where CT
is the transposed communicability. The cross-correlation
matrix R of the system is then calculated, as usual, nor-
malising the rows of the covariances Σij by the auto-
covariances Σii as, Rij = Σij/Σii. Despite the merit
of being analytical, this estimate of the network's cross-
correlation matrix still relied on assuming very simple
dynamical model, the diffusion of Gaussian white noise,
and did not fully disentangle the unique contribution of
the topology. In the present work we want to address the
question whether it is possible to estimate the most likely
correlation structure that a network gives rise, based only
on its topological properties. As mentioned before, it is
legitimate to assume that the activities of two nodes will
exhibit statistical dependence either if they are connected
by means of a direct link, or if the aggregate input they
receive from the entire network is similar. With this in
3
mind, we realised that the column vectors cj of the com-
municability matrix C indeed represent the input profile
of the influences node j receives from all nodes in the
network along all possible paths, including the influence
a node has on itself due to recurrent paths. Therefore,
it should be possible to predict analytically the magni-
tude of the expected correlation between any two nodes
in a network by comparing the similarity of their input
profiles. The resemblance between two input profiles can
in principle be calculated using any measure of similarity
between multidimensional vectors. Here we choose the
cosine similarity because it returns results bounded be-
tween −1 and 1, equivalent to the cross-correlation mea-
sure we aim at comparing. Consequently, we define the
topological similarity, Tij, between two nodes as the co-
sine of the angle between the corresponding columns ci
and cj in the communicability matrix:
Tij =
(cid:104)ci, cj(cid:105)
(cid:107)ci(cid:107)(cid:107)cj(cid:107) ,
(4)
being (cid:104) , (cid:105) the inner product and (cid:107)(cid:107) the vector norm.
The definition of T depends uniquely on the topologi-
cal constraints of the network encoded in the adjacency
matrix A, plus the realistic assumption (embedded in C)
that the influence or the information content decays with
the length of the path. Despite being a purely topolog-
ical measure, we recognized that, when the variance of
the nodes is the same for all nodes, T corresponds ex-
actly to the correlation matrix R of the Gaussian diffu-
sion dynamics studied in Ref 41. Therefore, T formally
closes the cycle for the search of a direct relation be-
tween the structure of a network and the expected corre-
lation pattern that this structure will tend to generate.
Furthermore, this analytical measure unveils the funda-
mental mechanism behind the contribution of network
structure to the emerging function, which can be under-
stood in terms of the similarity of influences that two
nodes receive from the whole network via its complete
path structure.
Finally, we shall notice that T can be estimated for
both directed and undirected graphs, as well as for
weighted networks. In the case of undirected graphs C is
symmetric but if the links are directed, then the columns
of C determine the input profiles and the rows represent
the profile of output influences of the nodes.
It must
be noted that despite the measure can be computed for
any weighted adjacency matrix, it does not always make
sense to do so. Because communicability is a measure
of influence along the paths, it only has a direct physi-
cal meaning when the weights of the links represent the
coupling strength between the nodes, the flow capacity
of the link or a compatible physical sense.
Summarising, here we have introduced a topological
estimator of a network's correlation structure. Although
it ignores any specific dynamics of the nodes, it accounts
for the fact that information or influence within a net-
work propagates along all possible alternative paths and
that naturally decays for longer paths. The measure ac-
tually quantifies the similarity between the input profiles
of nodes. In the following section we systematically in-
vestigate how three fundamental features of a network
(the weights of the links, the path length and the redun-
dancy of paths) influence both the communicability and
the topological similarity between nodes that, as shown
above, can in fact be used as a proxy of their expected
correlation.
III. EFFECT OF TOPOLOGICAL PRIMITIVES
In this section, we will investigate how three topolog-
ical primitives we defined before, namely links' weights,
graph distance and the presence of multiple alternative
paths, sculpts the influence that one node exerts over an-
other, setting aside the role of local nodes' dynamics. To
this aim, we will focus on three simple classes of graphs:
(i) chains, (ii) cycles and (iii) path-redundant networks.
We will evaluate how manipulating critical parameters
of these graphs leads to changes in the influence between
given pairs of nodes a and b (measured by their commu-
nicability, Cab) and in their topological similarity Tab.
The contribution of link weight in the simplest case can
be understood when analyzing how changes in the weight
modulate the mutual influence of two nodes directly con-
nected by a single link (see Supplemental Figure S1). As
expected, incresing link weight is associated with increase
in both the influence that one node exert over the other,
as well as in their topological similarity.
In general,
A simple manner to obtain a better intuition of the
behavior of both communicability and topological simi-
larity is by studying how increasing the graph distance
between two reference nodes affects both Cab and Tab in
simple chain topologies. See top-left panels of Figure 1.
it is possible to see how increasing the
length of the path separating the two nodes correspond
to a decrease in both their communicability and topolog-
ical similarity, behavior directly determined by the de-
cay formalized in the definition of the communicability.
From the example, it is indeed evident how, for increasing
lenghts of the chain, the input profiles of the two nodes
at the ends of the chain (highlighted with green rectan-
gles in the communicability matrices, top-left panels of
Figure 1) become more and more antithetic, reflecting
opposed whole-network influences. This behavior is cap-
tured by the corresponding decrease in the nodes topo-
logical similarity. The effect of uniformly varying the
weights of all links in a chain is mainly quantitative: in
fact, augmenting the weights in chains of fixed length in-
creases the strength of the influence of each node over
all other nodes in a way that is inversely proportional
to the distance separating them, top-right panel of Fig-
ure 1. See also Supplemental Figure S2 to appreciate the
interaction of chain length and links' weight on C and T
between the nodes at the two ends of a chain).
4
Due to the definition of communicability, the length of
the path separating any two nodes as well as the weights
of the links are the most important parameters defining
the strength of their mutual influence and therefore their
topological similarity as well. As such, chain topologies
can be thought as a baseline to compare how more com-
plex motifs such as cycles and path-redundant topolo-
gies, which indeed incorporate chain topologies, modu-
late both the resulting communicability and the topolog-
ical similarity between given nodes.
With this in mind, we compared Cab and Tab of the
three model graphs (chains, cycles and path-redundant
motifs) having corresponding graph diameter, i.e. having
same longest path L. Bottom-left panel of Figure 1 pro-
vides a schematic representation of different motifs hav-
ing equivalent longest paths. For the case of chains and
path-redundant topologies, we computed Cab and Tab for
those two nodes that were more distant, in other words,
those at the two extremes of (each) path, whereas for
cyclic topologies, we always selected two adjacent nodes:
this choice allowed us to clearly disentangle the contribu-
tion that the direct link has on both the commmunicabil-
ity and on the topological similarity, above and beyond
the modulation produced by chains of different length.
in chains both Cab and Tab decay as
the distance between the extremal nodes increases, see
bottom-right panels of Figure 1, blue lines. On the other
hand the effect of the direct link is well illustrated for
the case of cyclic architectures (bottom-right panels in
Figure 1, red lines). In fact, the presence of a direct link
importantly enhances and poses a lower bound for both
Cab and Tab , that however exhibit a chain-like decay as
the indirect path between them increases.
As expected,
This approach makes the difference between chains and
cyclic primitives straightforward; however, in order to
being able to properly understand how increasing path
redundancy affects the communicability and hence topo-
logical similarity, it can be of great help to directly quan-
tify the difference between Cab and Tab calculated from
chains and from redundant topologies for comparable
path length. The two lowest-right panels in Figure 1
illustrate three examples of these differences, namely the
cases with two (dark green lines), three (light green lines)
and four (orange lines) redundant paths. From this anal-
ysis, we appreciate that increasing the number of alterna-
tive paths does increase both the total influence and the
topological similarity between the two reference nodes at
both ends of the paths, but that the magnitude of this
increase decays with the length of the path, vanishing for
paths longer than 5 links.
Studying these simple networks (single link, chain, cy-
cles and redundant paths) gives the opportunity to un-
derstand the central relevance of the three topological
primitives in shaping how the influences of nodes unfould
through the graph. This information in turn determines
how strong will be the expected correlation between any
pair of nodes, approximated by their topological similar-
ity. These simplified network models can thus be used
5
FIG. 1. Behavior of Communicability and Topological Similarity in simple network motifs. (Top-left)
Communicability and Topological Similarity matrices of chains of different lengths. In the upper matrices, the red
dots indicate the matrix entry corresponding to the communicability between the nodes at the two ends of the
chains, whereas the green rectangles marks their whole-network input profiles (column vectors), that are used to
calculate the topological similarity of the corresponding nodes (marked with the green dots in the lower matrices.
(Top-Right) Communicability and Topological Similarity matrices of chains of constant length (L =21) for different
links' weights, w. See top-left caption for the legend of red dots, green rectangles and green dots. (Bottom-left)
Schematic representation of different graphs (chains, cycles and path-redundant architectures) having comparable
longest path. The reference nodes for which both the communicability (Cab) and the topological similarity (Tab)
where calculated are highlighted in yellow. (Bottom-Right) Upper panels: comparison of Cab and Tab of the three
different graphs having comparable longest path. Line colors correspond to those in the schematic representation in
the bottom-left region of the figure (Light blue lines: chains; Red lines: cycles; Dark green lines: two redundant
paths; Light green lines: three redundant paths; Orange lines: four redundant paths). Lower panels: diffference
between Cab and Tab of chains and redundant topologies. All results were obtained for constant links' weights and
global coupling w = g = 1.
to summarize some of the simplest forms of interactions
sustained by these topological features.
IV. UNDERSTANDING THE BRAIN'S SPONTANEOUS
CORRELATION STRUCTURE
In the previous section we have analyzed how simple
network architectures could affect both communicability
and topological similarity between a given pair of nodes.
However, real networks are made of interwined assem-
blies of those topological motifs that form intricate ar-
chitectures and, together with the particular dynamical
properties characterizing the system at hand, determine
the emergence of complex patterns of interactions.
In
this section, we will try to find out) how much of the
complex pattern of spontaneous correlations empirically
observed in the resting brain can be explained just by the
topology of the underlying anatomical structure.
The main assumption we make is the following: if two
brain regions receive similar influences from the entire
network, then the probability that they will exhibit con-
sistent co-activations is high; on the other hand, if the
inputs they receive from the whole network are very dif-
ferent, it is more likely that their covariance is weak. In
Sections II and III, we showed how the the structure of
a network can be used to estimate the strength of the
influence that one node exerts over another trough the
communicability C, and that C can be further used to
quantify the topological similarity, T , of the input pro-
files between any pair of nodes. We then demonstrated
that T can in fact be interpreted as the expected correla-
tion between the nodes in the network. As such, we will
use T to estimate the unique contribution of topology
to the empirical time-average correlation structure of the
brain's spontaneous activity measured using resting-state
fMRI.
To this aim, we will compute the topological similar-
ity matrix from the group-average structural connectiv-
ity matrix (SC), and use it to estimate the correlation
structure mirrored by the group-average empirical func-
tional connectivity matrix (FC) obtained from resting-
state fMRI. See Section VI for details about the em-
pirical SC and FC matrices. The SC matrix stores the
information about axonal pathways reconstructed using
whole-brain diffusion tensor imaging and tractography,
thus defines the whole-brain anatomical wiring diagram
of the brain. It should be noted that, despite their re-
producibility, current methods used to reconstruct fiber
bundles from diffusion imaging are characterized by in-
trinsic limitations constraining their accuracy: in fact, it
is well-known that these reconstruction algorithms tend
to favor the shortest, straightest and simplest path be-
tween any two reference voxels42, which in turn impair
their ability to accurately detect crossing fibers and long
inter-hemispheric axons43,44. With this in mind, we fo-
cused only on intra-hemipsheric structural and functional
connectivity, in order to avoid the confounding effects of
6
FIG. 2. Effect of the global coupling. The figure
illustrate the effect of the global coupling, g, onto the
resulting topological similarity matrix, T . The three
matrices have been obtained from the same empirical
structural connectivity matrix, SC.
accumulating errors due to the above mentioned limita-
tions in sampling inter-hemispheric pathways.
As mentioned in Section II, the communicability can
be scaled by a factor g controlling the emerging collective
dynamics, thus leading to the possibility of obtaining as
many topological similarity matrices as the values of g
used to scale C (see Figure 2 for three instances of T
obtained for increasing values of g).
For each hemisphere, we thus searched for the topo-
logical similarity matrix T that best explained the ob-
served intra-hemispheric correlation structure, by opti-
mizing the communicability matrix according to a global
scaling factor, g (Panel G and H in Figure 3).
As a mesure of model fitting, we used the mean ab-
solute error (MAE), an outlier-robust alternative of the
mean squared error (MSE), a classic statistic to quantify
the goodness of an estimator. From the scatter plots
(Panels I,L of Figure 3) and the best-fitting T matrices
(Panels M,N of Figure 3), it is possible to appreciate
how properly accounting for the overall input pattern
sustained by a given topology can indeed reduce the
error in estimating the emergent correlations from the
raw structural connectivity alone; in fact, in the example
we pass from a E(SC,FC)≈ 0.42 to E(T ,FC)≈ 0.15).
These results demonstrate that knowledge of the topol-
ogy of whole-network input patterns of different brain re-
gions, sustained by direct and indirect routes of multiple
interweaved axonal bundles, can very much predict the
time-average correlation structure observed from sponta-
neous BOLD fluctuations, above and beyond the infor-
mation about direct anatomical connections stored in the
SC matrices.
In the following, we will investigate how explicitly in-
troducing local nodes' dynamics could affect the predic-
tion of the average correlations. To this aim, we will
make use of the Hopf normal model, that have success-
fully used to predict mesoscopic brain activity45,46, and
evaluate their ability to explain brain functional connec-
tivity. By this comparison, we will be able to gain some
insight about the contribution of adding local dynam-
ics in determining the observed correlation structure, as
well as to have a better understanding of the role of the
7
FIG. 3. Using topology to approximate the brain's spontaneous correlation structure. The figure shows
results obtained separately for the left and right hemisphere. (A,B) Intra-hemispheric structural connectivity
matrices (SC). (C,D) Scatterplot depicting the relationship between SC and the intra-hemispheric empirical
correlation structure, summarized also in panels E and F. (E,F) Empirical functional connectivity matrices (FC) for
the left and the right hemisphere. (G,H) Mean absolute error (MAE) between the empirical FC and the topological
similarity T computed for different value of the global coupling parameter, g. For both hemispheres, minimal MAE
(≈0.15) corresponded to g ≈1. The dotted lines correspond to the mean absolute error between the SC matrix and
the empirical FC, which ≈0.45. (I,L) Scatterplots of the empirical FC and the best-fitting topological similarity for
the two hemispheres. (M,N) Best-fitting Topological similarity matrices of the two hemispheres.
underlying topology, shared either by the two models.
A.
Introducing local dynamics
We will now explicitly introduce local dynamics to sim-
ulate large-scale brain activity using the connectional ar-
chitecture defined by the empirical SC matrix. By do-
ing this, we will be able to estimate, through numeri-
cal simulations, the correlation between different brain
regions, and then compared these results with the aver-
age empirical correlation matrix (Empirical FC) obtained
from resting-state fMRI. Numerical simulations will be
achieved through the use of the Hopf normal model, a dy-
namical model able to reconcile noise-based approaches
with models based on oscillators. This formalism is based
on the normal form of a Hopf bifurcation45,47, a type
of bifurcation that occurs when a system characterized
by a stable fixed point loses its stability by exhibiting
oscillations. As such, this model allows transitions be-
tween asynchronous noise activity and oscillations, thus
making it a good candidate to reproduce empirical data
as observed either with EEG, MEG or fMRI45 -- 47. This
model rely on the choice of two parameters, namely g,
the global scaling of the strength of all the links in the
network, and α, the parameter controlling the dynami-
cal regime of each node. The bifurcation parameter will
be set to 0, meaning that all nodes lie at the bifurca-
8
V. SUMMARY AND DISCUSSION
The brain is a complex system and as such its overall
dynamics cannot be fully understood without taking into
account the rich patterns of interactions into which its
components are inherently embedded into. The collective
dynamics in a network results from the complex interplay
between its underlying structure, the nature of the local
dynamics of individual nodes, their specific working point
and the manner in which they are coupled26,27,33. Here,
we aimed at understand to what extent does the structure
of a network drive the emerging patterns of interactions.
Is it possible, knowing the complete wiring diagram of a
network, to estimate its most likely correlation structure?
Setting aside the large influence from the local dy-
namics, we propose that the topological features which
play major roles in shaping the interaction between
two nodes are:
(i) the strength of the links, (ii) the
length of the path between the nodes, and (iii) the
routing of information along multiple and redundant
paths. A strong direct connection between two nodes
is usually a reliable indicator of the strength of their
functional interaction.
In fact, in general direct links
lead to more effective communications since the content
of information (e.g., the amplitude of the perturbations)
tends to decay over longer paths28. However, the flow of
information from one node to another does not follow a
unique path, but rather spreads along several. Therefore,
the total influence that one node exerts over another
is accumulated over all possible paths, of all lengths29 -- 31.
The goal of this paper is that of answering the question
of how the structure of a network contributes to sculpt
the expected pattern of correlations that emerge when
the network hosts a dynamical process. A direct appli-
cation is that of understanding the extent to which the
structural connectivity of the brain determines the large-
scale correlation structure observed during rest. Accord-
ingly, we have developed a graph measure, the topological
similarity T , which estimates the expected correlation
between two nodes based on the similarity of the "in-
fluences" both receive from the whole network.
If two
nodes receive the same sets of inputs, then they should
be strongly correlated. On the contrary, if they share
no common inputs, their correlation tends to be weaker.
Hence, this measure can be considered as a generaliza-
tion of the matching index. In the analysis of graphs the
matching index is typically used to quantify the number
of common neighbours shared by two nodes. However,
the matching index only accounts for the direct neigh-
bours and ignores the rest of the complex fabric of inter-
actions sustained by the entire network. To perform a
complete comparison of the inputs accumulated over all
paths, while accounting for the natural decay of signal
power or information content in any real system, we con-
sidered the columns of the communicabiilty measure32
as they directly correspond to the whole-network input
profiles of the nodes. Although the actual decay rate of
FIG. 4. Numerical simulations. The figure
illustrate results for the two hemispheres obtained from
numerical simulations using the Hopf normal model.
(A,B) The model returned best results (MAE≈0.1) for
values of the global coupling, g ≈ 0.12 for both
hemispheres. As for data shown in Section IV, model
fitting was performed using the mean absolute error
(MAE). MAE between empirical structural (SC) and
functional connectivity (FC) matrices is indicated by
the dotted lines (≈ 0.42 in both hemispheres). (C,D)
scatter plots of the empirical functional connectivity
(FC) versus the simulated one obtained at the
best-fitting global coupling. (E,F) In both panels, the
upper triangles store the empirical FC values, whereas
the lower triangles the corresponding ones obtained
from simulations at the best fitting value of g.
tion working point, region that has been demonstrated
to give good approximate of the empirical resting-state
FC46, and we will thus optimize for g. Results from nu-
merical simulations are depicted in Figure 4. The in-
troduction of complex local dynamics does indeed in-
crease the capacity of predicting the time-average corre-
lation structure (MAE(T )≈ 0.15, whereas MAE(Hopf)≈
0.11)), even though the magnitude of this increase is rel-
atively small, suggesting that most of the average struc-
ture observed in correlated spontaneous BOLD fluctua-
tions are in fact largely determined by the topology of
the underlying network's architecture.
the signals may differ across real systems, our choice of
the communicability guarantees (due to its exponential
decay) that the accumulation of perturbations over all
possible paths, Eq. (1), converges for all adjacency ma-
trices A. However, any other formalism implementing a
decay of the influence between nodes as a function of their
graph distance can be used to compute the topological
simmilarity T , with the constrain that such formalism
must be based on a set of coefficients assuring series con-
vergence for any adjacency matrix.
Recently, Zamora-L´opez et al. 41 found that consider-
ing the communicability as the propagator kernel for the
diffusion of Gaussian noise along the network allows to
analytically estimate the time-average cross-correlation
matrix R of the system. We realise that both approaches,
the one starting from a dynamical system describing
the propagation of perturbations41, and the other based
uniquely on topological constrains and the assumption
of exponential decay of influences, are indeed equivalent.
In fact, we find that T ≡ R when the variance ξi of the
Gaussian noise is the same for all nodes in the Gaus-
sian diffusion system. Therefore we can conclude, with
a high confidence, that T represents the most likely ex-
pected correlation structure of a network only due to its
underlying topology.
As argued before, the other crucial ingredients for the
collective dynamics on a network are the local dynamics
of the nodes and the coupling functions, both of which
can be either linear or nonlinear. The topological simi-
larity T thus captures the tendency of the nodes of the
network to correlate or to synchronise with each other.
The contribution of the local dynamics and of the cou-
pling function is to modulate this original background
tendency set by the connection topology and either en-
hance or disrupt the underlying patterns of correlation.
In order to gain understanding on how each of the three
topological features, namely, link weight, path length and
path redundancy, precisely affect how influence propa-
gates through the network, we have first investigated the
behaviour of C and T between selected nodes in simple
networks: chains and cycles of varying length, and path-
redundant graphs. Not surprisingly, we have found that
the strength of a direct link between two nodes is a major
contributor to the intensity of their expected correlation.
However, the presence of common inputs or redundant
paths between them enhances the intensity of their in-
teraction beyond the baseline determined by the strength
of the link. If there is no direct link between two nodes,
common inputs and redundant paths can trigger strong
correlations between them. However, this influence tends
to decay with the length of the paths. Although the pre-
cise decay rate of the influence depends on the charac-
teristics of the real system, it is worth noting that real
networks with diameter larger than five are rather un-
common. Specially in brain and neural networks which
are dense. From a more general perspective, we shall no-
tice that large efforts have been devoted in the literature
to investigate which of the common network properties,
9
e.g., degree distribution, clustering coefficient, communi-
ties or motifs, determine more prominently the collective
dynamics on a network, particularly its capacity to syn-
chronise48 -- 53. After our observations we can conclude
that what truly matters for the collective dynamics is
the path structure of the network which determines how
far do perturbations reach and how their effect accumu-
lates over redundant and recurrent paths. Thus, we fore-
see that, to precisely understand how the typical network
properties determine the collective dynamics, future work
needs to identify how those properties alter the underly-
ing network's path structure.
Limitations and outlook
The results we have here presented come with some
limitations which shall be underlined. (A) The topologi-
cal similarity T we have introduced represents an estima-
tion of the time-averaged cross-correlation. Thus, it only
estimates the spatial correlations between brain regions
and does not capture temporal correlations. (B) Commu-
nicability measure assumes an exponential decay of the
influence with pathlength. This is not the only choice
possible and other sets of coefficients kl exists which will
lead the weighted version of the sum in Eq. (1) converge.
The precise decay rate of influence, perturbations or in-
formation in real systems will vary according to the sys-
tem's nature. Thus, in some real applications it might be
possible to identify the decay rate with pathlength and
identify the "right" set of coefficients kl. As such, T is
not restricted by the set of coefficients used in the com-
municability measure proposed by32 .(C) We found here
more convenient to quantify topological similarity as the
cosine similarity of the input profiles, Eq. (4), as we found
it more accurate in estimating the similarity between in-
put profiles of small size, as in the case of short chains.
However, other measures are possible, e.g., correlation or
euclidean distance, without conceptually modifying the
meaning of topological similarity. (D) for the analysis on
the brain's connectivity and spontaneous correlations, we
considered both hemispheres as if they were independent.
The reason was to avoid biases due to the unreliability
of tractography to identify inter-hemispheric fibers. Still,
our comparisons are biased to some extent because the
SC-based T and simulations consider both hemispheres
as independent while the empirical resting-state measure-
ments reflect the activity of the brain regions, which are,
certainly, embedded on the whole network. Finally, we
want to highlight the usefulness of both topological sim-
ilarity T and the correlation matrix R of the Gaussian
diffusion system41 to explore the functional connectiv-
ity of synthetic and empirical networks. Because they
estimate the expected correlation analytically, they are
very fast to compute. For example, they allow to com-
pare the effects of network perturbations, e.g., node or
link lessions, without the need to run computationally
expensive simulations. The only difference between T
and R is that the formalism of R allows also to explore
the effects of simulated inputs by increasing of decreasing
the variance of the Gaussian noise at selected nodes.
VI. MATERIALS AND METHODS
To obtain the population average structural and func-
tional connectomes twenty-one healthy volunteers (mean
age 21.56 years; standard deviation 1.84 years; all males;
all right handed) participated in five (5) resting-state and
two (2) DTI scanning sessions. All participants signed in-
formed consent to participate. The study was conducted
at the School of Psychology, Birmingham and was ap-
proved by the University of Birmingham Ethics Com-
mittee.
Data acquisition
The scanning sessions were conducted at the Birm-
ingham University Imaging Centre using a 3T Philips
Achieva MRI scanner with a 32-channel SENSE head
coil. T1-weighted anatomical data (175 slices; 1 × 1 × 1
mm3 resolution) were collected during the first ses-
sion only. DTI data were collected in two sessions
(23.3 2.5 days apart). The DTI acquisition consisted
of 60 isotropically-distributed diffusion weighted direc-
tions (b=1500 smm-2; TR=9.5s; TE=78ms; 75 slices;
2 × 2 × 2 mm3 resolution) plus a single volume without
diffusion weighting (b = 0 smm-2, denoted as b0). The
DTI sequence was repeated twice during each session,
once following the Anterior-to-Posterior phase-encoding
direction and once the Posterior-to-Anterior direction, to
correct for susceptibility-induced geometric distortions54.
Resting-state data were collected in five sessions (the first
and the last collected in the same scanning session as the
DTI data) using whole brain echo-planar imaging (EPI)
(TR = 2s; TE = 35ms; 32 slices; 2.5 × 2.5 × 4 mm3 res-
olution). Participants were instructed to have their eyes
open and maintain fixation to a white dot presented at
the centre of the screen.
Whole-brain DTI tractography
We
5.0.8
(FMRIB
in
the
DTI
processed
data
Software
FSL
version
Library,
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) on a Red Hat
Linux operating system. We corrected the data for
susceptibility distortions, eddy currents and motion
artifacts55. We subsequently rotated the gradient direc-
tions (bvecs) to correct them for motion rotation42,56,57.
We then used the corrected gradients in the Bayesian
Estimation of Diffusion Parameters Obtained using
Sampling Techniques (BedpostX) tool to generate a
distribution model in each voxel58. We used the default
parameters in BedpostX for the diffusion modelling: 2
10
fibers per voxel, weight of 1 for the secondary fibers,
discard of the first 1000 iterations before sampling.
We parcellated the brain into 116 areas using the Auto-
mated Anatomical Labeling (AAL) atlas59. We followed
a 4-step registration procedure to align the AAL atlas
from MNI to native space: (a) align the non-weighted
diffusion volume (b0) of each session to their midspace
and create a midspace-template, (b) align the midspace-
template to the anatomical (T1) scan, (c) align the T1 to
the MNI template of FSL, and (d) invert and combine all
the transformation matrices of the previous steps to ob-
tain the MNI-to-native registration. The results of each
step were visually inspected to ensure that the alignment
was successful. Step (a) controls for potential bias to-
wards the first session, when the T1 was acquired (similar
methodology to Ref 60. The first two registrations are 6-
dof linear transformations (rigid-body) since we aligned
images of the same subject, while the third is 12-dof non-
linear to warp the participants brain around the MNI
template. The final matrix of step (d) was applied to
the AAL atlas using nearest-neighbour interpolation to
preserve the labels of the areas.
We calculated the number of probabilistic streamlines
starting from each AAL area and reaching any other AAL
area by feeding the BedpostX model to the Probabilis-
tic Tracking algorithm (ProbtrackX)61. The parameters
we used in ProbtrackX are: 5000 samples per voxel, 2000
steps per sample until conversion, 0.5mm step length, 0.2
curvature threshold, 0.01 volume fraction threshold and
loopcheck enabled to prevent streamlines from forming
loops. We normalised the number of streamlines by the
size of the seed area and thresholded streamlines lower
than 1% of maximum (i.e. setting them to zero). We sub-
sequently computed the undirected structural connectiv-
ity matrix by averaging the normalised streamlines from
area i to area j and from area j to area i. The results for
each subject (in each DTI session) were organised into
42 weighted adjacency matrices A of size 116 × 116.
Population-average structural connectome
To estimate the population average structural connec-
tivity (SC) we pooled the 42 SC matrices together (2
per subject) and considered only the reduced parcella-
tion into 90 brain areas (45 per hemisphere) by exclud-
ing the 26 regions of the cerebellum and the vermis.
The 42 SC matrices contained a variable number L of
undirected links ranging from L = 895 for the spars-
est case (density ρ = 0.22) to L = 1279 for the dens-
est (ρ = 0.32). We noticed that the simple average of
the matrices into a single SC matrix by averaging the
42 values each link takes along the pool leads to an av-
erage connectome with strongly biased network proper-
ties. For example, this plain average SC matrix contained
L = 1967 links, which is almost twice the number of
links as in the individual matrices. In order to avoid this
problem we have devised a method which automatically
removes outlier links before performing the average. For
ij}
each link (i, j) we have initially a set of 42 weigths {ws
where s = 1, 2, . . . , 42. The method searches for outlier
weights (data-points falling out of 1.5 times the inter-
quartile range) and removes them from the data pool.
The search is iteratively repeated until no further outliers
are detected and then the population-average SC weight
for the link (i, j) is calculated as the average weight of
the surviving values. In practice, the method converges
very rapidly and it rarely performs more than 2 iterations
per link. This method allows to clean the data without
having to set an arbitrary hard threshold62 for the min-
imally accepted prevalence of the link. Full details of
the method are currently in preparation and will be pre-
sented somewhere else. The resulting population-average
SC matrix out of our iterative pruning method contains
L = 1189 links (ρ = 0.30), which lies within the range
of connectivity for the individual 42 matrices. For the
simulations we treated the left and the right hemispheres
independently, as two matrices of N = 45 ROIs because
tractography is known to largely miss interhemispheric
connections.
Resting-state time-courses and functional connectome
We
the EPI
resting-state
pre-processed
data
in FSL version 5.0.8 (FMRIB Software Library,
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) on a Red Hat
Linux operating system using MELODIC (Multivariate
Exploratory Linear Optimized Decomposition into
Independent Components). We corrected the data for
motion and slice scan timing, removed the non-brain
tissue, applied 5mm FWHM spatial smoothing and
removed spike motion artifacts using WaveletDespike63.
We subsequently applied high-pass temporal filtering
and then extracted the average time-course from each
AAL area. To estimate the population-average func-
tional connectivity (FC) matrix we considered again
only N = 90 brain areas excluding the cerebellum and
the vermis. We concatenated the 105 sequences of
resting-state signals (21 subjects, 5 sessions per subject)
into a single long multivariate time-series and computed
the Pearson correlation (z-Fisher corrected) for every
pair of signals. The opposite procedure, to compute an
FC matrix per session and averaging over the 105 FC
matrices leads to almost identical results.
Hopf normal model
Within this model, the temporal evolution of the ac-
tivity z of node j is given in the complex domain as:
= [αj + iωj − z2] + σηj(t)
dzj
dt
zj = ρjeiθj = xj + iyj
(5)
(6)
11
Where ω is the node's intrinsic frequency of oscilla-
tion, α is the local bifurcation parameter (local because
the model allows the possibility to assign a different value
of α for each node in the network) and η is additive Gaus-
sian noise with standard deviation σ. This system above
shows a supercritical bifurcation at α = 0. Specifically, if
αj is smaller than 0, then the local dynamic has a stable
fixed point at zj = 0, while for αj values larger than 0
there exists a stable limit-cycle oscillation of frequency
f = ω/2π . Whole-brain dynamics are described by the
following coupled equations:
= [αj − x2
j − y2
j ]xj − ωjyj +
Cij(xi − xj) + σηxj(t)
dxj
dt
dyj
dt
N(cid:88)
i=1
N(cid:88)
g
g
= [αj − x2
j − y2
j ]yj + ωjxj +
Cij(yi − yj) + σηyj(t)
(7)
(8)
i=1
Where Cij
is the anatomical connectivity between
nodes i and j, g is the global coupling factor and the
standard deviation of gaussian noise is σ = 0.02. In this
model the simulated activity corresponds to the BOLD
signal of each node. The intrinsic frequency of each
node was estimated as the peak frequency in the associ-
ated narrowband (i.e., 0.04 - 0.07 Hz64) of the empirical
BOLD signals of each brain region. We simulated, for ech
of the two hemispherese (45 ROIs each), 330000 points
using Euler's method for integration (dt = 0.001). The
connectivity between all the regions of interest was de-
fined using the empirical structural connectivity matrix
(SC), and obtained timeseries were then used to compute
the simulated correlation matrix (Simulated FC).
ACKNOWLEDGMENTS
We would like to thank Rui Wang and Caroline
di Bernardi Luft for their help in collecting the data
used in this study.
This work was supported by
(RGB) the FI-DGR scholarship of the Catalan Gov-
ernment through the Ag`encia de Gesti´o d'Ajuts Uni-
versitari
i de Recerca, under agreement 2013FI-B1-
00099, (GZL) the European Union's Horizon 2020 re-
search and innovation programme under grant agreement
No. 720270 (HBP SGA1), (GD) the European Research
Council Advanced Grant: DYSTRUCTURE (295129)
and the Spanish Research Project PSI2013-42091-P,
(ZK) European Community's Seventh Framework Pro-
gramme [FP7/2007-2013] under agreement PITN-GA-
2011-290011, (VMK) European Community's Seventh
Framework Programme [FP7/2007-2013] under agree-
ment PITN-GA-2012-316746 and (MLK) by the Euro-
pean Research Council Consolidator Grant: CAREGIV-
ING (615539).
1B. Biswal, F. Zerrin Yetkin, V. M. Haughton, and J. S. Hyde,
"Functional connectivity in the motor cortex of resting human
brain using echo-planar MRI," Magn. Reson. Med. 34, 537 -- 541
(1995).
2A. Arieli, A. Sterkin, A. Grinvald, and A. Aertsen, "Dynamics
of ongoing activity: explanation of the large variability in evoked
cortical responses," Science 273, 1868 -- 1871 (1996).
3D. A. Gusnard and M. E. Raichle, "Searching for a baseline:
Functional imaging and the resting human brain," Nat. Rev.
Neurosci. 2, 685 -- 694 (2001).
4B. Mazoyer, L. Zago, E. Mellet, S. Bricogne, O. Etard, O. Houd´e,
F. Crivello, M. Joliot, L. Petit, and N. Tzourio-Mazoyer, "Corti-
cal networks for working memory and executive functions sustain
the conscious resting state in man," Brain Res. Bull. 54, 287 -- 298
(2001).
5R. G. Shulman, D. L. Rothman, K. L. Behar, and F. Hyder,
"Energetic basis of brain activity:
implications for neuroimag-
ing," Trends Neurosci. 27, 489 -- 495 (2004).
6M. D. Fox, A. Z. Snyder, J. L. Vincent, M. Corbetta, D. C. V.
Essen, and M. E. Raichle, "The human brain is intrinsically or-
ganized into dynamic, anticorrelated functional networks," Proc.
Nat. Acad. Sci. USA 102, 9673 -- 9678 (2005).
7C. F. Beckmann, M. DeLuca, J. T. Devlin, and S. M. Smith,
"Investigations into resting-state connectivity using independent
component analysis," Phil. Trans. R. Soc. B 360, 1001 -- 1013
(2005).
8P. Fransson, "How default is the default mode of brain function?:
Further evidence from intrinsic BOLD signal fluctuations," Neu-
ropsychologia 44, 2836 -- 2845 (2006).
9S. M. Smith, P. T. Fox, K. L. Miller, D. C. Glahn, P. M. Fox,
C. E. Mackay, N. Filippini, K. E. Watkins, R. Toro, A. R. Laird,
and C. F. Beckmann, "Correspondence of the brain's functional
architecture during activation and rest," Proc. Nat. Acad. Sci.
USA 106, 13040 -- 13045 (2009).
10B. B. Biswal, M. Mennes, X.-N. Zuo, S. Gohel, C. Kelly, S. M.
Smith, C. F. Beckmann, J. S. Adelstein, R. L. Buckner, S. Col-
combe, A.-M. Dogonowski, M. Ernst, D. Fair, M. Hampson, M. J.
Hoptman, J. S. Hyde, V. J. Kiviniemi, R. Kotter, S.-J. Li, C.-
P. Lin, M. J. Lowe, C. Mackay, D. J. Madden, K. H. Madsen,
D. S. Margulies, H. S. Mayberg, K. McMahon, C. S. Monk,
S. H. Mostofsky, B. J. Nagel, J. J. Pekar, S. J. Peltier, S. E.
Petersen, V. Riedl, S. A. R. B. Rombouts, B. Rypma, B. L.
Schlaggar, S. Schmidt, R. D. Seidler, G. J. Siegle, C. Sorg, G.-
J. Teng, J. Veijola, A. Villringer, M. Walter, L. Wang, X.-C.
Weng, S. Whitfield-Gabrieli, P. Williamson, C. Windischberger,
Y.-F. Zang, H.-Y. Zhang, F. X. Castellanos, and M. P. Milham,
"Toward discovery science of human brain function," Proc. Nat.
Acad. Sci. USA 107, 4734 -- 4739 (2010).
11H. Laufs, K. Krakow, P. Sterzer, E. Eger, A. Beyerle, A. Salek-
Haddadi, and A. Kleinschmidt, "Electroencephalographic signa-
tures of attentional and cognitive default modes in spontaneous
brain activity fluctuations at rest," Proc. Nat. Acad. Sci. USA
100, 11053 -- 11058 (2003).
12M. J. Brookes, M. Woolrich, H. Luckhoo, D. Price, J. R. Hale,
M. C. Stephenson, G. R. Barnes, S. M. Smith, and P. G. Mor-
ris, "Investigating the electrophysiological basis of resting state
networks using magnetoencephalography," Proc. Nat. Acad. Sci.
USA 108, 16783 -- 16788 (2011).
13A. Ponce-Alvarez, G. Deco, P. Hagmann, G. L. Romani, D. Man-
tini, and M. Corbetta, "Resting-State Temporal Synchronization
Networks Emerge from Connectivity Topology and Heterogene-
ity," PLOS Comput Biol 11, e1004100 (2015).
14S. D. Keilholz, J. C. W. Billings, K. Wang, A. Abbas, C. Hafe-
neger, W. J. Pan, S. Shakil, and M. Nezafati, "Multiscale net-
work activity in resting state fMRI," in 2016 38th Annual Inter-
national Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC) (2016) pp. 61 -- 64.
12
15G. Deco, P. Hagmann, A. Hudetz, and G. Tononi, "Modeling
resting-state functional networks when the cortex falls asleep:
Local and global changes," Cereb. Cortex 24, 3180 -- 3194 (2013).
and J. Binder, "Spin-glass model
predicts metastable brain states that diminish in anesthesia,"
Front. Syst. Neurosci. 8, 234 (2014).
16A. Hudetz, C. Humphries,
17A. Hudetz, X. Liu, and S. Pillay, "Dynamic repertoire of intrin-
sic brain states is reduced in propofol-induced unconsciousness,"
Brain Connectivity 0, 1 -- 13 (2014).
18R. G. Bettinardi, N. Tort-Colet, M. Ruiz-Mejias, M. V. Sanchez-
Vives, and G. Deco, "Gradual emergence of spontaneous corre-
lated brain activity during fading of general anesthesia in rats:
Evidences from fMRI and local field potentials," NeuroImage
114, 185 -- 198 (2015).
19C. S. Zhou, L. Zemanov´a, G. Zamora-L´opez, C.-C. Hilgetag,
and J. Kurths, "Hierarchical organization unveiled by functional
connectivity in complex brain networks," Phys. Rev. Lett. 97,
238103 (2006).
20C. J. Honey, R. Kotter, M. Breakspear, and O. Sporns, "Net-
work structure of cerebral cortex shapes functional connectivity
on multiple time scales," Proc. Nat. Acad. Sci. USA 104, 10240 --
10245 (2007).
21J. G´omez-Gardenes, G. Zamora-L´opez, Y. Moreno, and A. Are-
nas, "From modular to centralized organization of synchroniza-
tion in functional areas of the cat cerebral cortex," PLoS ONE
5, e12313 (2010).
22G. Deco, V. Jirsa, A. R. McIntosh, O. Sporns, and R. Kotter,
"Key role of coupling, delay, and noise in resting brain fluctua-
tions," Proc. Nat. Acad. Sci. USA 106, 10302 -- 10307 (2009).
23G. Deco, V. K. Jirsa, and A. R. McIntosh, "Emerging concepts
for the dynamical organization of resting-state activity in the
brain," Nat. Rev. Neurosci. 12, 43 -- 56 (2011).
24G. Deco, V. K. Jirsa,
and A. R. McIntosh, "Resting brains
never rest: computational insights into potential cognitive ar-
chitectures," Trends Neurosci. 36, 268 -- 274 (2013).
25G. Deco and M. L. Kringelbach, "Great expectations: Using
whole-brain computational connectomics for understanding neu-
ropsychiatric disorders," Neuron 84, 892 -- 905 (2014).
26A. Mess´e, D. Rudrauf, H. Benali, and G. Marrelec, "Relating
structure and function in the human brain: relative contributions
of anatomy, stationary dynamics, and non-stationarities," PLoS
Comput. Biol. 10, e1003530 (2014).
27G. Schmidt, G. Zamora-L´opez, C. Zhou, and J. Kurths, "Sim-
ulation of large scale cortical networks by individual neuron dy-
namics," Int. J. Bif. & Chaos 20, 859 -- 867 (2010).
28V. Latora and M. Marchiori, "Efficient behavior of small-world
networks," Phys. Rev. Lett. 87, 198701 (2001).
29B. A. Huberman and L. A. Adamic, "Information dynamics in
the networked world," in Complex Networks, Lecture Notes in
Physics No. 650, edited by E. Ben-Naim, H. Frauenfelder, and
Z. Toroczkai (Springer Berlin Heidelberg, 2004) pp. 371 -- 398.
30D. J. Ashton, T. C. Jarrett, and N. F. Johnson, "Effect of conges-
tion costs on shortest paths through complex networks," Phys.
Rev. Lett. 94, 058701 (2005).
31A. Trusina, M. Rosvall,
and K. Sneppen, "Communication
boundaries in networks," Phys. Rev. Lett. 94, 238701 (2005).
32E. Estrada and N. Hatano, "Communicability in complex net-
works," Phys. Rev. E 77, 036111 (2008).
33L. Huang, Q. Chen, Y.-C. Lai, and L. Pecora, "Generic be-
haviour of master-stability functions in coupled nonlinear dy-
namical sstems," Phys. Rev. E 80, 036204 (2009).
34S. P. Borgatti, "Centrality and network flow," Soc. Networks 27,
55 -- 71 (2005).
35V. Colizza, A. Flammini, M. A. Serrano, and A. Vespignani,
"Detecting rich-club ordering in complex networks," Nat. Phys.
2, 110 -- 115 (2006).
36J. Goni, M. P. v. d. Heuvel, A. Avena-Koenigsberger, N. V. d.
Mendizabal, R. F. Betzel, A. Griffa, P. Hagmann, B. Corominas-
Murtra, J.-P. Thiran, and O. Sporns, "Resting-brain functional
connectivity predicted by analytic measures of network commu-
13
tainty in diffusion weighted mr imaging." Magn. Reson. Med.
72, 227 -- 236 (2003).
59N. Tzourio-Mazoyera, B. Landeaub, D. Papathanassioua,
F. Crivelloa, O. Etarda, N. Delcroixa, B. Mazoyerc, and M. Jo-
liot, "Automated anatomical labeling of activations in SPM using
a macroscopic anatomical parcellation of the MNI MRI single-
subject brain," Neuroimage 1 (200215).
60S. M. Smith, N. De Stefano, M. Jenkinson,
and P. M.
Matthews, "Normalized accurate measurement of longitudinal
brain change." J. Comput. Assist. Tomogr. 25, 466 -- 475 (2001).
61T. Behrens, H. Berg, S. Jbabdi, M. Rushworth, and M. Woolrich,
"Probabilistic diffusion tractography with multiple fibre orienta-
tions: What can we gain?" Neuroimage 34 (2007).
62M. A. de Reus and M. van den Heuvel, "Estimating false positives
and negatives in brain networks," NeuroImage 70, 402 -- 409
(2013).
63A. X. Patel, P. Kundu, M. Rubinov, P. S. Jones, P. E. V´erts,
K. D. Ersche, J. Suckling,
and E. T. Bullmore, "A wavelet
method for modeling and despiking motion artifacts from resting-
state fmri time series." NeuroImage 95, 287 -- 304 (2014).
64E. Glerean, J. Salmi, J. M. Lahnakoski, I. P. Jaaskelainen, and
M. Sams, "Functional magnetic resonance imaging phase syn-
chronization as a measure of dynamic functional connectivity,"
Brain connectivity 2, 91 -- 101 (2012).
nication," Proc. Nat. Acad. Sci. USA 111, 833 -- 838 (2014).
37A. Avena-Koenigsberger, B. Misi´c, R. Hawkins, A. Griffa, P. H.
andJoaqu´ın Goni, and O. Sporns, "Path ensembles and a tradeoff
between communication efficiency and resilience in the human
connectome," Brain Struct. Funct. (2016).
38F. Harary and A. Schwenk, "The spectral approach to determin-
ing the number of walks in a graph," Pacific Journal of Mathe-
matics 80, 443 -- 449 (1979).
39J. Bang-Jensen and G. Z. Gutin, Digraphs: Theory, Algorithms
and Applications (Springer, 2008).
40E. Estrada, N. Hatano, and M. Benzi, "The physics of communi-
cability in complex networks," Phys. Reps. 514, 89 -- 119 (2012).
41G. Zamora-L´opez, Y. Chen, G. Deco, M. L. Kringelbach, and
C. Zhou, "Functional complexity emerging from anatomical con-
straints in the brain: the significance of network modularity and
rich-clubs," Scientific Reports 6, 38424 (2016).
42D. K. Jones and M. Cercignani, "Twenty-five pitfalls in the anal-
ysis of diffusion mri data." NMR Biomed. 23, 803 -- 820 (2010).
43C. Thomas, Q. Y. Frank, M. O. Irfanoglu, P. Modi, K. S. Saleem,
D. A. Leopold, and C. Pierpaoli, "Anatomical accuracy of brain
connections derived from diffusion mri tractography is inherently
limited," Proceedings of the National Academy of Sciences 111,
16574 -- 16579 (2014).
44S. Jbabdi, S. N. Sotiropoulos, S. N. Haber, D. C. Van Essen,
and T. E. Behrens, "Measuring macroscopic brain connections
in vivo," Nature neuroscience 18, 1546 -- 1555 (2015).
45F. Freyer, J. A. Roberts, R. Becker, P. A. Robinson, P. Ritter,
and M. Breakspear, "Biophysical mechanisms of multistability in
resting-state cortical rhythms," The Journal of Neuroscience 31,
6353 -- 6361 (2011).
46G. Deco and M. L. Kringelbach, "Metastability and coherence:
extending the communication through coherence hypothesis us-
ing a whole-brain computational perspective," Trends in neuro-
sciences 39, 125 -- 135 (2016).
47F. Freyer, J. A. Roberts, P. Ritter,
and M. Breakspear, "A
canonical model of multistability and scale-invariance in biologi-
cal systems," PLoS Comput Biol 8, e1002634 (2012).
48A. E. Motter, C. S. Zhou, and J. Kurths, "Enhancing complex-
network synchronization," Europhys. Lett. 69, 334 -- 340 (2005).
49M. di Bernardo, F. Garofalo, and F. Sorrentino, "Effects of de-
gree correlation on the synchronizability of networks of nonlinear
oscillators," (Proc. 44th IEEE Conference on Decision and Con-
trol, and the European Control Conference WeA14.1, 2005).
50F. Atay, T. Biyikoglu, and J. Jost, "Network synchronization:
Spectral versus statistical properties," Physica D 224, 35 -- 41
(2006).
51C. Zhou and J. Kurths, "Hierarchical synchronization in networks
of oscillators with heterogeneous degrees," Chaos 16, 015104
(2006).
52X. Wu, B. Wang, T. Zhou, W. Wang, M. Zhao, and H. Yang,
"Synchronizability of highly clustered scale-free networks," Chi-
nese Phys. Lett. 23 (4), 1046 -- 1049 (2006).
53A. Arenas, A. D´ıaz-Guilera, J. Kurths, Y. Moreno, and C. S.
Zhou, "Synchronization in complex networks," Phys. Reps. 469,
93 -- 153 (2008).
54J. L. Anderson, S. Skare, and J. Ashburner, "How to correct
susceptibility distortions in spin echo-panar images: application
to diffusion tensor imaging." NeuroImage 20, 870 -- 888 (2003).
55J. Anderson and S. Sotiropoulos, "An integrated approach to
correction for off-resonance effects and subject movement in dif-
fusion mr imaging." NeuroImage 125, 1063 -- 1078 (2016).
56A. Leemans and D. Jones, "The b-matrix must be rotated when
correcting for subject motion in dti data." Magn. Reson Med 61
(2009).
57A. Ersiz, V. E. Arpinar, S. Dreyer, and L. T. Muftuler, "Quanti-
tative analysis of the efficacy of gradient table correction on im-
proving the accuracy of fiber tractography." Magn. Reson. Med.
72, 227 -- 236 (2014).
58T. E. Behrens, H. J. Berg, S. Jbabdi, M. F. S. Rushworth, and
M. W. Woolrich, "Charaterization and propagation of uncer-
"How structure sculpts function: unveiling the contribution of anatomical
connectivity to the brain's spontaneous correlation structure." by
Supplementary material for:
Ruggero G. Bettinardi, Gustavo Deco, Vasilis M. Karlaftis, Timothy J. Van Hartevelt, Henrique M. Fernandes, Zoe
Kourtzi, Morten L. Kringelbach and Gorka Zamora-L´opez
1
2
FIG. S1. Effect of Single Link Weight. (A) graph and matrix representation of two nodes connected by a
direct link. (B) Communicabity between the two nodes, Cab, for diffferent values of link's weight. (C) Topological
Similarity between the two nodes, Tab, for diffferent values of link's weight. On top of the panel are exemplified the
communicability matrices obtained for different values of link's weight. Remember that Tab is computed as the
cosine similarity between the two columns of the communicability matrix.
3
FIG. S2. Effect of Link Weight in chain and cyclic topologies. (A) Communicability Cab between the two
nodes a and b at the ends of chains of different length for varying links' weights. (B) Topological similarity Tab
between the two nodes a and b at the ends of chains of different length for varying links' weights. The blue line in
Panels A and B correspond to the links' weight value used in chain topologies in the main text of the paper (w =1.)
(C) Communicability Cab between two adjacent nodes a and b for cycles of increasing perimeter (N , number of
nodes in the cycle) for varying links' weights. (D) Topological similarity Tab between two adjacent nodes a and b for
cycles of increasing perimeter (N , number of nodes in the cycle) for varying links' weights. The red line in Panels C
and D correspond to the links' weight value used in cyclic topologies in the main text of the paper (w =1.)
|
0910.2741 | 2 | 0910 | 2010-04-07T22:23:55 | Fractals in the Nervous System: conceptual Implications for Theoretical Neuroscience | [
"q-bio.NC"
] | This essay is presented with two principal objectives in mind: first, to document the prevalence of fractals at all levels of the nervous system, giving credence to the notion of their functional relevance; and second, to draw attention to the as yet still unresolved issues of the detailed relationships among power law scaling, self-similarity, and self-organized criticality. As regards criticality, I will document that it has become a pivotal reference point in Neurodynamics. Furthermore, I will emphasize the not yet fully appreciated significance of allometric control processes. For dynamic fractals, I will assemble reasons for attributing to them the capacity to adapt task execution to contextual changes across a range of scales. The final Section consists of general reflections on the implications of the reviewed data, and identifies what appear to be issues of fundamental importance for future research in the rapidly evolving topic of this review. | q-bio.NC | q-bio |
Fractals in the Nervous System: conceptual implications for Theoretical Neuroscience.
Gerhard Werner
[email protected]
Department of Biomedical Engineering
University of Texas at Austin, TX.
Abstract
This essay is presented with two principal objectives in mind: first, to document the
prevalence of fractals at all levels of the nervous system, giving credence to the notion of their
functional relevance; and second, to draw attention to the as yet still unresolved issues of the
detailed relationships among power law scaling, self-similarity, and self-organized criticality. As
regards criticality, I will document that it has become a pivotal reference point in
Neurodynamics. Furthermore, I will emphasize the not yet fully appreciated significance of
allometric control processes. For dynamic fractals, I will assemble reasons for attributing to
them the capacity to adapt task execution to contextual changes across a range of scales . The
final Section consists of general reflections on the implications of the reviewed data, and
identifies what appear to be issues of fundamental importance for future research in the
rapidly evolving topic of this review.
Contents
1. Introduction
2. Power-law scaling in neuronal structures and processes
2.1 Neuronal Morphology
2.2 Peripheral nervous system
2.2.1 Ion channels
2.2.2 Point process analysis in peripheral nerves and neurons
2.3 The mesoscopic level
2.4 The macroscopic level
2.4.1 Fractals in Brain Networks
2.4.2 Fractals and Criticality of Brain States
2.4.3 Significance of Brain Criticality
3. Psychological and Behavioral Processes
3.1 Symbol processing and Fractals
3.2 Motor Behavior and Allometric control processes
4. Processes that generate power-law distributions
5. Fractals in Action
5.1 Complexity Matching Effect
5.2 Linking across many scales of space and time
6. Summary and final thoughts
7. References
1. Introduction
Fractals, introduced by Mandelbrot in 1977, are in the spatial domain considered to be
self-similar geometric objects with features on an infinite number of scales. In the analysis of
time series, fractal time describes highly intermittent self-similar temporal behavior that does
not possess a characteristic time scale. Their statistical analysis can provide access to
understanding the dynamics of complex systems. Not possessing a single characteristic scale,
static and dynamical fractals, measured on different scales of space and time, respectively, can
be characterized by power functions whose (usually non-integer) exponents are their fractal
Dimensions. In this essay, principal emphasis is on random fractals which include a stochastic
element in their generation.
Fractals are signals which display scale-invariant, self-similar behavior; they can be
analyzed by decomposing a signal into a hierarchy of temporal and spatial scales that may cover
the wide range between coarse-scale long-term and high-frequency fine-scale fluctuations. If
the relationship of the property under consideration is simple with respect to change of scale,
the process is considered monofractal , the process can then be characterized by a single
scaling exponent which is related to the fractal dimension or the spectral exponent of the
process. It can be expressed as the Hurst exponent of the process (Mandelbrot, 1968;
Koutsoyiannis, 2002). However, in some instances the scaling behavior may not be adequately
characterized by a single, stationary scaling exponent. In such cases, several scaling exponents
may be required, each exponent locally pertaining to some portion of the data stream. Such
multifractal signals are represented by the histogram of the Hoelder exponents, also known as
the singularity spectrum (Muzy et al., 1993).
Power law scaling and other manifestations of fractal and self-similar patterns in space
and/or time can be identified at all levels of neural organization. With few exceptions, these
observations remained largely islands in the otherwise rapidly advancing theoretical
Neuroscience with different priorities. However, recent advances in methodology of
measurement of fractal connectivity at higher levels of brain organization have led to a
proliferation of new data. This now calls for integrating fractality with other insights into brain
organization and complexity, notably in the light of the substantial evidence for the brain being
a complex system in a regime of criticality, as understood in statistical physics (Chialvo,2004,
2008; Kitzbichler et al, 2009; Fraiman et al, 2009; Werner, 2007b, 2009a,b). Like in other
physiological systems manifesting fractal patterns (see for instance: Bassigthwaighte et al,
1994; West and Deering, 1995; Iannacone and Khoka, 1995; West, 2006) the question of
ubiquity of power law scaling needs to be addressed in relation to other features of brain
organization. Similarly, is there a relation between fractal organization and the propensity for
phase transitions of critical systems? Is there a bridge between coarse graining (including
renormalization group transformation) and fractality ? And, most importantly, can fractal
properties be viewed as playing a role for the functional integration among different levels of
neuronal organization as Andersen (2000) suggests in an article entitled “From Molecules to
Mindfulness”. Generalizing from a comprehensive theory of organization and interactions at
the molecular level, Agnati et al (2004, 2009) view the Central Nervous System as a nested
network at all levels of organization, in the image of the Russian Matryoshka dolls: self-similar
structures being embedded within one another. This theory is elaborated in great detail, with a
“Fractal (self-similarity) Logic” operating on a set of identical rules which would govern the
relation between successive levels of the nested system.
Giesinger’s ( 2001 ) comprehensive overview of scale invariance in Biology provides the
background of this review, as do the insights gained in Physics through the work of Wilson
(1979 ) and Kadanoff (1990), amongst many others. In addition, Turcotte (1999) discussed at
great length the relation between aspects of self-organized criticality and fractal scaling from
the points of view of “avalanche” behavior and systematic properties of correlation length,
specifically directing attention to inverse cascade and site-percolation models of the well-
known forest fire paradigm. While none of the issues discussed in the following will receive a
definitive answer, I will aim at an explicit formulation of the network of interrelated factors
that constitute the territory in which new perspectives and potential solutions may lie.
With the agenda set forth in the foregoing, the organization of the presentation is as
follows: I will first briefly review the neuroscience literature on fractals, organized by level of
neuronal organization, from ion channels to cortical networks , and to psychological and
behavioral functions. Criticality of brain states, Allometric control and the origin of fractals by
phase transitions of complex systems will be addressed in more detail. This will be followed by
a brief overview of the essentials features of the theory of fractal generators, including random
walk theory and fractional differential operators. Having laid out the background in this
manner, I will consider relations between renormalization group transformation and fractals as
having some potential bearing on the apparent ubiquity and universality of power law scaling in
neural structures and processes, and its relation to criticality. Finally, I will direct attention to
the amazing consequence of self-similarity which assures the telescoping of different levels of
structural and functional organization to constitute a fractal object or time series. This will lead
me to posing the ultimate question: is there a process for unpacking interactions between
different levels of the fractal object, responsive to circumstances and conditions, which eludes
us entirely? If it existed, fractals would surely be a most extraordinary design principle for
operational economy in complex systems.
2. Power-law scaling in neuronal structures and processes.
This section is intended to summarize essential aspects of fractal properties at each of the
conventionally designated organizational levels, as the basis for conceptual consideration of
relations across these levels. However, a word of caution is in order: the sketches of
observational data in this section encompass a vast variety of biological substrates, conditions
of observation, and methods of measurement. This heterogeneity imposes limits on
generalizations, as do the differences of criteria for identifying fractal or self -similar features in
the data. Potential pitfalls were discussed and illustrated in LaBarbera’s (1989) useful (largely
pedagogic) publication. More recently, Eke et al (2002), Deligniers et al (2006) and Clauset et al
(2009) set forth stringent criteria for design, collection and interpretation of data for
identifying and categorizing fractal properties. The latter authors are very specific in
formulating a principled statistical framework, combining maximum likelihood fitting methods
with goodness of fit tests; they demonstrate examples of data that had been conjectured to
represent power law fits, but did not withstand the rigor of their tests. Touboul and Destexhe
(2009) also suggest that apparent power law scaling may in some instances not be supported
by more stringent statistical criteria. Of particular relevance to the topic of Section 2.3 is their
claim that experimentally observed power law scaling must not considered proof of self-
organized criticality, lest there be other supporting evidence available. The analysis, synthesis
and estimation of fractal-rate stochastic point processes is reviewed and illustrated with
examples by Thurner et al. (1997). Note also that the distinction between mono- and
multifractal scaling is sometimes difficult to draw (Kadanoff et al, 1989).
Conceivably, some of the variations among the reports reviewed in subsequent sections
may be attributable to procedural differences among studies; other reports may not meet the
rigorous statistical criteria of Clauset et al. Nevertheless, I submit that the majority of
experimental data on fractals in neural structures give collectively adequate reason for
ascribing to them wide-spread functional significance. At least the results based on wavelet
analysis appear immune to methodological criticism (see Section 2.4).
2.1 Neuronal morphology
In the foundational work “The fractal geometry of Nature”, Mandelbrot (1977) wrote “it
would be nice if neurons - he mentioned specifically Purkinje cells in the cerebellum- turned out
to be fractal”: Nature obliged abundantly as the following sample of findings with dendrites,
neuron cell bodies and glia cells indicates. Studying the branching pattern of dendritic trees of
retina neurons, Caserta et al (1990) identify by box counting fractal shapes with a fractal
dimension of approximately 1.7, which can be explained by a diffusion limited aggregation
model (Witten and Sander, 1981); but fractal dimension measured by different methods (for
instance comparing box counting with cumulative mass method) gives appreciably different
values (Caserta et al, 1955). A fractal structure was observed by Kniffki et al (1994 ) for the
branching dendrite patterns of thalamic neurons in Golgi impregnated specimens. In a
separate series, a scaling relation for bifurcations within the dendrite trees was ascertained
(Kniffki et al, 1993). Significant species differences in fractal dimensions of dendrite
arborizations in dorsal horn spinal cord neurons (Milosevic et al, 2007) may be attributable to
species differences in peripheral somesthetic sensibility (the dorsal horn neurons being the first
receiving station of this type of afferent input). Fractal analysis also reveals a distinct
differentiation of neuron types in the different laminae of the dorsal horn (Milosevic et al,
2005). Differences in regional connectivity and functional capacity amongst different regions in
visual cortex pyramidal neurons are also associated with marked variation in the fractal
dendrite branching structure (Zietsch and Elston, 2005). Fractal analyses provide a measure of
space filling of dendrite arbors which, in a study by Jelinek and Elston (2001) , differentiates in
the macaque visual cortex the two known processing streams between primary and secondary
visual area by differences in fractal properties. These investigators had undertaken a meticulous
examination of criteria for ‘quality control’ in studies of this nature, from the stage of pre-
processing of tissue specimens to comparative evaluation of methods for determining fractal
dimension (Jelinek et al, 1995).Examining the connectivity repertoire of basal dendrite arbors of
pyramidal neurons, Wen et al (2009) determined a universal power law scaling for dendrite
length and radius, suggesting that the dendrite arbors are constructed by statistically similar
processes; moreover, fragments of an arbor are statistically similar to the entire arbor, thus
displaying self-similarity. These design features are thought to maximize functionality for a
fixed dendrite cost.
Additional evidence comes from digital image analysis which enabled Smith et al (1989)
to determine the fractal dimension of neuron contours. Results obtained with conventional
methods of scaling analysis are corroborated by Wavelet Packet fractal analysis (Jones and
Jelinek, 1989). Multifractals were identified for cortical pyramidal cells while, in comparison,
neurons of synRas transgenic mice display less complex arborization patterns (Schierwagen,
2008.). Shape complexity of neurons and elements of microglia in human brain can be
classified over a range for fractal dimensions which is different for normal and pathological
brains (Karperien et al, 2008 ). The sequence of developmental stages of oligodendrocytes,
tracked the basis of their immunoreactivity, parallels changes in fractal dimension (Bernard et
al, 2001). Fractal analysis of cell ramification and space filling patterns differentiate microglia
cells into two categories, depending on whether their fractal dimension did, or did not increase
after brain injury (Soltys et al, 2001).
The scaling law for the cortical magnification factor in primate visual cortex is an
illustrative example with functional significance: as is well known, the part of the visual scene
corresponding to the eye center is represented densely at the cortex, becoming progressively
sparser towards the periphery. It turns out that the scaling law for the sampling density away
from area centralis is a power function which assures locating a peripheral target in the
shortest time (Koulakov, 2010). As brain size increases, the cortex thickens only slightly, but the
degree of sulcal convolution increases dramatically, indicating that human cortices are not
simply scaled versions of one another (Im et al., 2008). Changizi (2003) infers several
organizing principles of Neocortex from scaling relations among its components: e.g. diameters
of neural structures predispose for efficient transport through neuron arborizations;
economical wiring reflects well-connectedness within given volumes of neural tissue. These
relations are viewed to represent a universal law for scaling that applies to hierarchical
complexity and combinatorial systems, generally (Changizi, 2001 b).
As one among several instances of scaling in the cerebral cortex, Changizi (2001 a) also
shows that axon cross sectional area increases in Phylogeny with brain size, presumably
compensating increase of conduction distances with conduction velocity (see Section 3.1). A
synthesis of comparative neuroanatomy with biophysics leads Harrison et al (2002) to conclude
that scaling trends in morphological specializations at the cellular level may constitute
functional adaptations. One of the examples in support of their thesis is the role of component
scaling for managing the conflicting developmental trends of increasing brain size and surface
folding on the one hand, and the requirement for optimizing energy requirements and
processing speed, on the other. The role of scaling relations for brain growth becomes evident
when its scaling relation is disrupted by preterm birth in a dose-dependent, sexually dimorphic
fashion ; it directly parallels the incidence of neuro-developmental impairments in preterm
infants (Kapellou et al., 2006).
Taken together, the observations surveyed in the foregoing two paragraphs suggest that
fractal dimension of neuronal and glia elements bear some relations to developmental,
functional and pathological conditions of neural tissue. This warrants a few conceptual
considerations: Bieberich (2002) attaches neural-computational significance to the self-
similarity of dendritic branching as a platform for economical information compression and
recursive algorithms. On the same self-similarity principle, Pellionisz (1989) envisages a fractal
growth model of dendritic arbors by iterated code repetition as process for global construction
of fractals (see for instance: Barnsley & Dempko, 1985): the essential underlying theme is to
both reduce complexity of generating, and at the same time conserving the full richness of the
dendrite arbor. I will expand on this principle in later section of this essay. Among the not yet
explored implications of dendrite fractal arborizations are the effect they may induce on the
dynamics of processes and critical phenomena in dendrite spines for which they are a
supporting platform: In Statistical Physics, such effects obtain when the neighborhood
relations among interacting elements (for instance: Ising spins or coupled maps (Cosenza and
Kapral, 1992) are themselves provided by a self-similar fractal lattices, such as the Sierpinsky
Gasket (Gefen et al 1980), rather than an Euclidean geometric base.
In an extension of fractal analysis to features of complex neural structures, Zhang (2006)
determined the Magnetic Resonance image-based fractal dimension of white matter of human
brain. This method was shown to accurately quantify white matter structural complexity in
three dimensions, and detect age-related degenerative changes. Tractography based on
Diffusion Tensor Imaging enabled Katsaloulis and Vergenelakis (2009) to determine fractal
dimension, self-similarity and lacunarity of neuron tracts in human brain. The lacunarity analysis
is understood to indicate the distribution of fractal neuron tracts of different length scales, as
evidence of connections between different neuron ensembles. Another extraordinary technical
advance made it possible to determine the fractal properties of receptor density and
distribution in human brain, using Positron Emission Tomography (PET) and Single -photon
Emission Tomography (SPET) (Kuikka and Tiihonen, 1998).
2.2 The peripheral nervous system: ion channels, point process analysis of activity in peripheral
nerves and individual neurons
2.2.1: Ion Channels.
Turning to primarily functional aspects of fractality in neural systems, attention focuses
in this section on temporal aspects of ion channel gating and its relation to time series of
neuronal discharge patterns. The following collage of data obtained with different
experimental conditions as well as modeling studies consistently supports the dominant
presence of fractal features in the functional manifestations at the levels under consideration.
The kinetics of Ion transport across neuronal membranes occurs, in part, via ion channels .
Application of the Patch clamp technique made it possible to follow the time course of channel
opening and closing precisely. Typically, the rate of channel opening and closing opening
fluctuates, changing at times suddenly from periods of great to periods of slow activity. This
pattern served as clue to surmise an underlying fractal process with infinite variance. On this
basis Liebowitch et al (1987; 2001) asked how the switching probabilities at one time scale of
observation are related to those at another time scale. It turned out that these probabilities
(defined as effective kinetic rate) at a given time scale are characterized by fractal scaling, and
that effective kinetic rates for different time scales of observation display self -similarity: there
are bursts within bursts of openings and closings. The suggestion is that energy barriers in
stochastically switching protein conformational states are the underlying mechanism (for a
detailed account, see Ch. 8 in Bassingthwhaite et al, 1994). A different version that also
accounts for the power law relationship of ion channel gating kinetics assumes that ion channel
proteins have a very large number of states, all of similar energy, making the gating process
more akin to a diffusion (Millhauser 1988). Recent theoretical modeling defined more precisely
the conditions that give rise to the power law distributions in relation to the activation barriers,
compatible with the known Physics of proteins (Goychuck and Hanggi, 2002). Roncaglia et al
(1993) developed on theoretical grounds a stringent criterion for ascertaining the validity of the
fractal theory by evaluating the experimental distribution of channel closing times in terms of
the Hurst phenomenon. A few years thereafter, Varanda at al. (2000) delivered the evidence
for Ca-activated K channels in the form of long term correlations of open and closed dwell
times, expressed as Hurst coefficients of the order of 0.6, which alternative Markovian models
failed to satisfy.
Before proceeding to discuss the implication of channel kinetics for the patterning of
trains of neuron spikes, a brief remark on the fractal activity at the site of neural impulse
transmission at the neuromuscular junction. As is well known from the work of DelCastillo and
Katz (1954), the neural transmitter substance acetylcholine is released from the nerve terminal
in small packages: the miniature end potentials (MEPP) are considered manifesta tions of the
exocytosis of humoral transmitters. In departure from initial textbook accounts of the MEPP
release reflecting a set of homogeneous stationary Bernoulli trials, Perkel and Feldman (1979)
categorically reject a purely binomial model of (quanta l) transmitter release. For the frog
neuromuscular junction, Rothshenker & Rahaminoff (1970) could show that excocytosis can
exhibit correlations (memory) extending over periods of seconds, suggesting self-similar
characteristics. When sampled over prolonged periods, Lowen et al (1997) collected conclusive
data at the neuromuscular junction and synapses in hippocampal tissue culture that frequency
and amplitudes of MEPP’s display fractal scaling .Takeda et al (1999) also reported comparable
findings for the vertebral neuromuscular junction. The detailed analysis of quantitative features
of the recorded data led Lowen at al (1997) to conclude that traditional renewal models of
vesicular exocytosis as a memoryless stochastic process are entirely inadequate for
representing many of its salient features. Instead, their recommendation is that a new class of
models should be considered that relies on fractal-rate stochastic point processes: fractal rate
activity represents a kind of memory in that occurrence of an event at a given point in time
increases the likelihood of another event to occur at a later point in time, with that likelihood
persisting for some time.
2.2.2 Point process analysis in peripheral nerves and neurons
In this section, neuron discharge trains are viewed as mathematical objects, belonging
to the class of point processes (Thurner et al, 1997; Lowen and Teich, 2005): events occurring at
a point in time or space. Werner and Mountcastle (1963, 1964) determined scaling of neural
responses in primary cutaneous afferent nerve fibers with the magnitude of mechanical stimuli
applied to receptors. The implications of their findings in Psychophysics will be taken up in
Section 3. Adaptation in neural structures serves to extend their dynamic range: the
significance of this function is discussed in Section 5.2.
The statistics of action potential trains recorded from single neurons in the cochlear
nucleus of anaesthetized cats formed the basis of a mathematical analysis by Gerstein and
Mandelbrot (1964 ). The principal result was that a random walk model towards an absorbing
and a reflecting barrier can account for a wide range of fractal neuronal activity patterns,
assuming no more than the known physiological mechanisms of a threshold for membrane
depolarization, and the summation of excitatory and inhibitory post synaptic potentials.
Except for a thesis by Johannesma in 1969, it took almost 20 years of hegemony of Poisson and
Gaussian distributions until fractal approaches to spike train statistics were resumed: this time
by Wise (1981) in a study of spike interval distributions of data that had been recorded
primarily by Bloom (1969) in the cerebral cortex, and in respiratory neurons recorded by
Smolders and Folgering (1977. Wise found that plots of the spike interval histograms on log-log
scales showed negative powers on time with long tails, which he attributed to the neuron
membrane potential undergoing a random walk while the firing threshold fluctuates. Re-
working some of Wise’s data, West and Deering (1994) identified fractal (hyperbolic) spike
interval distributions. Taking an entirely different approach to conceptualizing irregular
behavior in neuron spike trains led Shahverdian and Apkarian (1998) to discuss self -affinity,
powerlaw dependence and computational complexity of spike trains in terms of a
multidimensional Cantor space with zero Lebesgue measure as attractor.
The turning point in the history of identifying fractal neuronal firing is associated with
the work of Teich and Lowen, beginning in the early 1980s (Ch 22, in McKenna, 1992) with
invalidating the then prevalent notion of Poisson point processes. More recently, the
shortcoming of Poisson spike interval statistics was also pointed out by Kass and Ventura (2001)
and by van Vreeswick (2001) who critizised experimental (Richmond et al, 1990) and
theoretical (Ohlshausen and Field, 1998) reports for unwarrantedly assuming either Poisson
neurons or rate based neurons with rate independent Gaussian noise; instead he considered a
renewal model as biologically more plausible.
Teich and Lowen’s essential realization was that determining long-time correlations in
spike trains requires sample sizes to be appreciably larger than conventionally used. On this
basis, Teich et al (1990) identified the following essential features of the time series of neural
spikes recorded from cat auditory nerve fibers and the lateral superior olivary nucleus:
discharge rates determined with different averaging times can exhibit self -similarity; the
variance-to-mean ratio of spike number increases with sufficiently large counting time in a
fractional power law fashion, with the exponent in the power law varying with the stimulus
level. With these data in hand, Lowen and Teich (1993) suggested that the fractal action
potential patterning in auditory nerve may be related to fractal activity in the ion channels of
the sensory organs feeding into the auditory nerve: that is, the hair cells in the cochlea. This
idea required to show that ion channel gating and neuronal spiking patterns are indeed causally
related. Lowen et al (1999) succeeded with demonstrating this causal dependence in
computational models, thus adding for the special case of the cochlear hair cells some credence
to their proposal that gating patterns in sensory organ ion channels can affect discharge
patterns in the sensory nerve tracts they feed. In an elegant experimental design, Teich (1977)
not only ascertained a power function for the activity in retina ganglion cells and neurons in the
lateral geniculate body when studied independently, but also succeeded with recording from
synaptically connected pairs of retina ganglion cells and geniculate neurons. In this situation,
fractal exponents for retina and target neurons in the lateral geniculate body were nearly
identical. This was interpreted to mean that fractal behavior is either transmitted across
synapses, or has a common origin for the synaptically connected pre- and postsynaptic
structure. On the other hand, fractal activity of medullary sympathetic premotor and the
synaptically connected pregangionic synmpathetic neurons is apparently generated
independently (Orer et al, 2003).
More support for the notion that ion channel properties play an important role for
determining neuron performance comes from demonstrating a kind of memory mechanism for
traces of prior activity in voltage-gated Na channels (Toib, 1998): time constants of channel
recovery stand in a power function relation to duration of prior activation. The question of
primary interest is of course how the dynamics of ion channels relates to the functional
characteristics of a whole neuron. Gilboa et al (2005) addressed this question in a
computational model of an ensemble of ion channels. In analogy to a ‘real’ neuron, this model
neuron exhibits various dynamics at different time scales: a power law function recovery time
scale after stimulation, temporal modulation of discharge pattern during maintained
stimulation, and the dependence of adaptation to a stimulus step on the duration of the
priming stimulus. The suggestive implication is that the ensemble of ion channels can exhibit
properties on many scales, comparable to ‘real’ neurons, thus supporting the notion that the
‘macroscopic behavior’ of the ‘real’ neuron is, in fact, the result of cooperative fractal channel
kinetics.
In addition to the studies cited in foregoing paragraphs, there are numerous reports
documenting fractal-rate behavior in single neuronal point processes. However, these data
were generally obtained for examining spike trains for encoding stimulus properties, and they
are quite heterogeneous as regards species, neural structure examined, use of anesthetics and
experimental conditions. Although this imposes serious limitations on drawing inferences on
general principles, I select here a few studies which applied several of the commonly agreed
upon and typical indicators of fractal properties, such as self similarity of firing rate with
different averaging time, increase of spike number variance-to-mean ratio with counting time,
and power law scaling relating the variable of interest to the resolution of measurement. In a
series of publications, Grueneis et al (1993) reported fractal properties in spike trains recorded
under various conditions including REM sleep of cats. In visual cortical areas of cats and
macaques, Baddeley et al (1997) observed consistently non-Poisson spike train statistics, with
some displaying self-similarity. Other neural structures examined included medullary
sympathetic neurons (Lewis et al, 1993) and dorsal horn of the spinal cord (Salvador and Biella,
1994). A common feature of these and other like reports not cited here, was the lack of
agreement on a consistent mathematical model that would satisfactorily describe the fractal
process underlying the experimental data. In a study of retina ganglion cells, Teich & Saleh
(1981) suggest a shot-noise driven self exciting point process; in a later study of the same
experimental object, Teich found a modulated gamma-r-renewal process satisfactory while
Grueneis et al(1993 ) favor a clustering Poisson process. Mandelbrot and van Ness (1968)
considered Fractal Brownian motion as candidate. Clearly, the goal of determining whether a
common principle governing spike train variability could be identified, and if not then for what
reason, eluded this group of investigators.
Without examining specifically for manifestations of fractality, a number of
investigators attempted statistical characterization of neural point processes, primarily
motivated to reconcile irregularity of spike trains with their presumptive function as “code” of
neural signals. But recall that Harris (2005) attributes irregularity of spike train discharges to
cell assembly organization. In various modifications, the general approach chosen by Sakai et
al (1999), Cateau and Reyes ( 2006), and Feng and Zhang (2001) consisted in designing
model neurons to generate spike trains whose statistics would match that of “real” neurons
recorded in animal experiments. Shinomoto et al (2003) recorded spike sequences from
different cortical areas in awake macaques which they classified phenomenologically into
different groups. Salinas & Sejnowski (2002) and Stevens and Zador (1998) assigned the
principal source of discharge variability to correlations in the input feeding the examined
neuron. None of these results warranted the allocation of observed or simulated spike train
data to one of the probability distributions in the conventional repertoire of statistics, but
Maimon and Assad (2009) at least excluded Poisson – like randomness from being a universal
feature of spike time distributions in primate parietal cortex. In an exquisitely elegant
experiment, Evarts (1967) followed the changes of interspike interval (ISI) histograms in
premotor cortex pyramidal neurons in wakefulness, sleep and the phase of sleep associated
with low-voltage fast EEG. Regrettably, his characterization of the ISI histograms is limited to
rejecting Poisson distributions. However, inspecting the histograms displayed in Fig. 12 of his
publication arouses one’s suspicion of a long-tail distribution for sleep activity.
In a notable and very extended comparison of cortical neuron discharges in alert
macaques, Shadlen & Newsome (1998) attributed to single neurons the ability to perform
simple algebraic operations resembling averaging by combining inputs from several sources but
they cautiously concluded that irregularity of the interspike interval distribution precludes
them from reflecting information about the actual temporal structure of the synaptic input.
They rejected random walk models of the kind applied by Gerstein and Mandelbrot as
inadequate for capturing the statistical features of spike interval distributions, and found
Poisson and various renewal processes likewise failing to yield satisfactory and consistent
correspondence with recorded data.
If there is one conclusion that can be drawn from the extant data on the statistics of
spike interval distributions, then it is that demonstrating fractal properties in spike trains
requires carefully selected conditions. Multiple convergences from incoming pathways appears
to obscure characteristic statistical properties of discharges in the recipient neurons. Thus, a
neuron’s intrisic connection pattern carries the burden of discharge variability. This view is
reminiscent of Harris’s (2005) view that spike discharge variability may be a signature of cell
assembly organization. This is perhaps also the source of futility of assigning any information
bearing capacity to discharge patterns of individual neurons (see for instance: Werner, 2007a).
On the other hand, the more direct a neuron’s connection pattern to peripheral sensor s is, the
more distinctly are fractal discharge properties demonstrable. But the opposite also seems to
be the case when neurons are embedded in a network, as the observations of El Boustani et al
(2009) in Section 2.3 show. The place to look for consistent fractal properties is apparently the
macroscopic, global level of brain organization (see Section 2.3). Whether and how the
mesoscopic level of the next Section bridges bridges the gap is the subject of the next section
2.3 The mesoscopic level of organization
Despite their relative simplicity, in vitro cultured neuronal networks are here viewed
as mesocopic in the sense of representing neuron ensembles which exhibit rich spontaneous
dynamical activity in the form of periodic bursting (Robinson et al, 1993; Nakanishi and Kukita,
1998; van Pelt et al, 2004; earlier references are cited in: Beggs and Plenz, 2003, 2004 ). At
superficial inspection, the brief burst of activity lasting tens of milliseconds are separated by
quiescence lasting up to several seconds (Corner et al, 2002; Tateno et al, 2002). The
spontaneous emergence of patterns in the discharge trains was also noted by Giugliano et al
(2004) and replicated in computational models . In extension of earlier work that led to
identification of scale invariant Levy distributions and long range correlations in cultured
neuronal networks (Segev et al, 2002; see Section 4 ), Segev et al (2004) attributed the activity
bursts to be associated with the formation of statistically distinguishable subgroups of neurons
, each with its own distinct pattern of interneuronal spatiotemporal correlations. Wagenaar et
al (2006 a) emphasized the extremely rich repertoire of bursting patterns during the
development of cortical cultures. The cultured cortical networks spontaneously generated a
hierarchical structure of periodic activity with a stereotyped population -wide spatiotemporal
structure. These recurring patterns ( called by the authors ‘superburst’ ) converged periodically
to a dynamic attractor orbit, and were taken to imply large-scale self-organization of neurons
in vitro , refuting the commonly held view of having random connectivity (Wagenaars, 2006 b).
The recorded data of these authors are available for distribution to interested investigators.
Departing from the practice of focusing on spontaneous activity, Breskin et al (2006) explored
the propagation of stimulus evoked activity in neuron cultures. A graph theoretic analysis of
their data attributed the dynamic evolution of the network connectivity to a percolation
transition with power law characteristics, while the degree distribution of the grown network
failed to meet power law criteria.
Working with mature organotypical cultures and acute slices of rat cortex, Beggs and
Plenz (2003, 2004) concluded that the cascades of propagating neuron discharges they
observed were indicative of the neural culture being in a state of self-organized criticality. The
importance of this claim, and a recent disputes of its validity (see below), warrant close
attention to methodological details: Beggs and Plenz based their analysis on recording
spontaneously appearing negative local field potentials (NLFP), apparently occurring
preferentially in cortical layers 2/3 (Gireesh and Plenz, 2008). The peaks of NLFP were
considered indicative of synchronized population discharges occurring near the recording
electrode tip (Plenz and Aertsen, 1993), on the rational that bursts of multiunit activity are
more likely to generate large NLFP’s than are single neuron discharges (Plenz and Thiagarajan,
2007). Accordingly, brief bursts of synchronized action potentials were the units of analysis in
their experiments.
The alternation between brief burst of NLFP activity and quiescent periods remained in
their experiments stable with a high degree of temporal precision, extending over periods of
many hours. Beggs and Plenz set out to examine the idea that these cascades of neural
activity may constitute a special mode of network activity, possibly of the character of
“avalanches”, indicative of SOC (Bak et al., 1987). They had specifically in mind the kind of
self-organizing branching process discussed by Zapperi et al, ( 1995) and de Carvalho and Prado
(2000). To examine this idea required determining the statistical properties of the observed
activity patterns. For carrying out this analysis, they first defined the spatial pattern of signal-
carrying electrodes during one time bin as frame; a sequence of consecutively active frames,
preceded and ended by a blank period was called an avalanche. The NLFP did not propagate in
the network in a spatially contiguous manner, thus excluding wave-like propagation. With these
definitions and precautions in place, distributions of avalanche size and lifetimes were found to
scale in cultures and acute cortex slices with a power law exponent -3/2, this value being
resilient to various choices of scales (Plenz and Thiagarajan, 2007). The branching parameter
was determined as sigma =1.04 . Being statistically indistinguishable from the ideal value of 1 ,
it signifies a critical state in the sense that activity starting at one electrode would initiate
activity in one other electrode, on the average, keeping the network at the edge of criticality
(Harris, 1989). This complements the evidence for fractal properties and supports the validity
of the working hypothesis Beggs and Plenz started out with : to view avalanches as
manifestations of the collective critical dynamics of SOC. More recently , fractal scaling of
patterned neural activity was reported to also occur in cultivated neurons of leech ganglia and
rat hippocampus (Mazzoni et al, 2007) ; and Pasquale et al, (2008) describe avalanches in
dissociated neuronal cultures of cortex from embryonic rats.
Avalanches were subsequently also studied in superficial layers of rat prefrontal cortex
(Stewart and Plenz, 2006) and during development of cortical layer 2/3 where they display
nested theta- and beta/gamma oscillations ( Gireesh and Plenz, 2008). The theory of critical
states predicts (see Section 2.4.3) and experiments confirm that neuronal avalanches display a
maximized dynamic range of responses to stimuli (Shew et al, 2009). Functional architecture of
avalanches conformed to Small World Topology (Pajevic and Plenz, 2009). Comparing NLFP
activity in vitro cortex preparations with in vivo activity of awake macaque monkey cortex,
Petermann et al (2009) established that high fidelity propagation of local synchronized scale-
invariant activity patterns is a robust and universal feature of cortex in awake monkeys.
Furthermore, large amplitude negative field potentials (like those constituting the avalanches)
spread in a cascade-like fashion through the cortical network without distortion , much like
action potentials: Thiagarajan et al (2010) described these stereotypical waveforms as
‘coherence potentials’. They occur often in rapid succession as a stream of dynamical
associations, suggesting the switching of the cortical network from one dynam ical state to
another.
The reason for viewing neuronal avalanches as manifestation of self-organized criticality
was based on their fractal scaling properties for size and duration. Here just a brief reminder of
the amply documented ‘family resemblance’ of fractality and SOC, to which the publications
of Grinstein (1995), Chen et al (1995), Paczuki et al. ( 1996), Tebbens and Burroughs (2003) and
Cessac (2004) speak, as do the model computations of Papa and da Silva (1997) , da Silva et al
(1998), De Arcangelis et al (2006) and Levina et al (2007 . However, despite the ‘avalanche’ of
research on mechanisms leading to scale invariance, there exist questions about the necessary
conditions for establishing the self-organized critical state (Kinouchi and Prado, 1999). A non-
conserving critical branching model was proposed by Juanico et al,(2007) to demonstrate that
SOC can occur in mean-field sand piles, provided the branching process is coupled to a
background activity of spontaneous switching between refractoriness and quiescence among
system components; in the stationary state, the system can undergo a transition from a
subcritical to a critical state. In an elaborate recent study, Bonachela and Munoz ( 2009) claim
that non-conserving (dissipative) dynamics does not lead to bona fide criticality. Non-
conserving systems are in their view not truly critical models. Instead, non-conserving systems
(such as the brain) would just hover around a critical point (presumably after some form of fine-
tuning) with broadly distributed fluctuations which do not disappear at the thermodynamic
limit. Such systems could be fluctuating in the vicinity of the critical point, but not at it. The
authors call this condition ’self-organized quasi-criticality’. Whether this is the last word in the
long standing debate of conditions for criticality in SOC remains yet to be seen.
Conditions for universality of 1/f scaling in dissipative self-organized criticality models
were established by De Los Rios and Zhang (1999). Models predicting avalanches of neural
activity include the work of Herz and Hopfield (1995) and were noted by Eurich et al (2002) in a
network of globally coupled nonlinear threshold elements. Models of neural networks of non -
leaky integrate-and-fire neurons exhibit over a wide range of connectivity patterns power law
avalanches with an exponent closely approximating that reported by Beggs and Plenz for tissue
cultures (Levina et al, 2007). In a comment to this paper, Beggs (2007) gives a lucid account of
how neural networks could self-organize to operate at criticality. Critical avalanche networks
can be computationally constructed by simple network growth models (Abbott and
Rohrkemper, 2007). An exponent of the experimentally determined value -1.5 of avalanche size
and lifetime scaling is predicted by the neural field theory of Buice and Cowan (2007).
A field theoretic approach was also investigated by Freeman (2005): analysis of high
resolution electroencephalograms of rabbits revealed neural fields in the form of spatial
patterns in amplitude and phase modulation of gamma and beta carrier waves which
distinguished positive and negative conditioning stimuli. The goal of applying field theory was in
these experiments to model states and state transitions as large-scale spatial patterns of
neural activity for quick changes in adjustment to different behavioral situations. The cortical
states were viewed as “wave packets”, resembling frames in motion picture, stabilized in a
scale free state of self-organized criticality. Recall, however, the reservations raised by
Touboul and Destexhe (2009) that substantiating interdependence of fractality and self -
organization requires additional evidence. In an entirely different context (namely fossil
extinction), Newman (1996) shows that fractality need not be a unique indicator of SOC and
criticality since an alternative simple model can account for the empirical power law relation
(for an extended discussion of this, see Sections 4 and 5)
Seeking to strengthen the conjecture of self-organized criticality of avalanches, Plenz
and Chialvo (2010) acquired experimental evidence that the neural avalanches in superficial
layers of cortex exhibit five additional quantitative aspects of their dynamics which are
consistent with critical dynamics. These were: separation of time scales between triggering and
the avalanching event itself; stationarity of size distribution; temporal clustering prior to and
following large avalanches , resembling Omori’s law; power law decay for avalanche size in the
wake of preceding large avalanches; and a fractal dimension for scaling spatial spread. The
importance of these results lies first, in affirming evidence for avalanches displaying robust
critical behavior; and second, in suggesting that their scale-invariant (fractal) properties do in
fact reflect cortical networks being in a state of criticality. This is also supported by the
observation that the exponents of simulated branching processes at near-critical branching are
similar to scaling exponents characterizing oscillations in the MEG recorded alpha frequency
band of humans at rest (Poli et al, 2008).
Criticality in cortex will be take up in Section 2.4.3, but let me merely stress at this
point that it implies the brain being poised to undergo sudden second order phase transitions
to new configurations with long range correlations among its disparate constituting elements.
The dynamics is universal in its independence of details at the microscopic level (Sornette,
2000; Stanley, 1987). Viewing neural activity in this framework is a fundamental departure
from a wave-type oscillatory or stochastic dynamics as the more commonly considered
theoretical approaches in Neuroscience. To underscore this distinction, early-stage activity in
the developing retinal network can be cited as an illustrative example (Hennig et al, 2009) of
the essential feature of critical dynamics: rhythmic bursts of action potential in retina ganglion
cells , propagating as wave-like events across the retina surface, arise at a very specific network
state which meets the criteria of the classical percolation model of statistical mechanics (Essam,
1980): the phase transition consists in separating states of purely local from global functional
connectedness, the latter displays in addition conspicuous fractal properties (Stauffer and
Aharony ,1991/1994).
It may be revealing to contrast the failure of consistently finding fractal activity patterns
in individually sampled neurons (other than those receiving relatively direct input , see Section
2.2.2) with the abundance of fractal patterns of (mesoscopic) neuron ensembles. It raises the
question whether the origin of the latter may be a matter of assembly organization: note that
in the records of neuron cultures, it is the concurrent activity of interconnected neurons
sampled by different electrodes that forms the fractal pattern; this is of course quite different
from the sampling of neurons in point process analysis, guided by chance encounters of a
microelectrode with one active neuron at the time. The puzzle posed at the end of section
2.2.2 thus finds perhaps its resolution in network architecture: Teramae and Fukai (2007)
describe a model that shows how the fractal property of a few individual neurons can turn into
an organized communal property of an ensemble. This lesson can also be learned from models
of dynamic pattern formation in neuron populations, forming fractal power spectra and power
law pulse distribution (Usher and Stemmler, 1995). Similarly, network amplification of local
fluctuations causes fractal firing patterns and oscillatory field potentials in two -dimensional
models of leaky Integrate-and-Fire neurons ; feedback connectivity of local excitation and
surround inhibition being the essential prerequisites ( Usher et al., 1994).
Bedard et al. (2006 b) accept existing claims for 1/f scaling of global variables of neural
activity (e.g. EEG: Freeman et al, (2003); EMG: Novikov et al., (1997); see Section 2.3) , and
acknowledge them as evidence for self-organized critical states with power law distributions ,
much as the models of De Arcangelis et al (2006) and Levina et al (2007) suggest. They also
accept the validity of claims for fractality discussed in Section 2.2.2 for various point process
analyses. But they contest the legitimacy of generalizing from these disparate sources of data.
The connection between 1/f frequency scaling of global variables and critical states of neural
activity is, in their mind, far from firmly established. Moreover, they emphasize that the
association of 1/f spectra with criticality may not be obligatory (for a review: see Giesinger,
2001; Newman, 1999).
Having raised this warning flags, Bedard et al (2006) not only confirmed in their own
investigations the presence of 1/f frequency scaling in EEG of cat parietal cortex (in absence of
anesthesia), but showed in addition 1/f frequency scaling in bipolar records of Local Field
Potentials (LFP). That bipolar LFP recordings sample relatively localized populations of neurons
was shown by Destexhe et al. (1999). They turned then their attention to investigating whether
this 1/f scaling reflects self-organized critical states with the result that avalanche size
distributions did in their experiments not follow power law scaling, nor did interspike interval
distributions of concurrently recorded single neuron activity show 1/f scaling ; rather, the
distributions were consistent with a Poisson process. Accordingly, Bedard et al. reject the
evidence for critical state dynamics. Instead, they proposed a model that would account for
1/f frequency scaling without being associated with critical states. Their model shows that
the observed 1/f scaling can indeed be produced by a band pass filtering effect of
extracellular media. This means that extracellular field potentials (such as LFP) can show power
law scaling while the underlying neuronal activity per se need not be critical. El Boustani et al
(2007) found in experimental data and models of cortical “activated” states also evidence for
Poisson spike distributions, and absence of avalanche dynamics. The irregular states of cortical
networks are thought to stem from a very high dimensional deterministic chaos. However, as
alternative, it is worth recalling that Harris (2005) views spike train irregularity as one of the
signatures of cell assembly organization.
Investigating further the discrepancy between the Beggs-Plenz and the Bedard et al
observations, Touboul and Destexhe (2009) recorded avalanches in cortex of (awake) cats,
paying careful attention to the conditions of data collection: avalanche statistics of negative
peaks LFP (linked to neuronal firing), positive peaks (unrelated to neuron firing) and surrogate
data (obtained by random shuffling experimental data) were analyzed; avalanche criteria were
those of the Beggs-Plenz studies. Essentially, time and amplitude distributions of NLFP showed
power law distributions, preferably at high detection thresholds. But, positive NLFP and
surrogate data (randomly shuffled peak times -- essentially equivalent to a threshold
stochastic process-- can also show power-law distributed amplitude distributions. The
conclusion of this study, then, was that apparent power law scaling cannot be considered as
proof of self-organized criticality .
At this point, a comment seems in order: the publication of Gireesh and Plenz (2007)
seems to suggest that cortical layers 2/3 are the preferential site of avalanche occurrence. The
data analyzed in the Touboul and Destexhe study were obtained in earlier work of Destexhe et
al (1999) in parietal cortex of awake cats; there is no indicat ion that a possible role of cortical
layer was considered. Whether layer specific patterns of neuronal arborization (Callaway,
2002) could be a source of the discrepancy can at this point not be determined
The discrepancy between presence of 1/f scaling in ensemble neural activity (EEG and
LFP) and, yet, 1/f scaling not being an intrinsic property of the individual neuron itself , that
Bedard and El Boustani et al claimed to have identified, motivated El Boustani et al (2009) to
adopt yet another experimental approach: to this end, they measured the scaling properties of
the power spectrum of the intracellularly recorded membrane potential of individual neurons.
The experiments were conducted in cat primary visual cortex in vivo , the animals being
anesthetized and paralyzed. Full-field visual stimuli of varying characteristics were presented to
the dominant eye to drive the cortical region under study to states with different firing
characteristics. The remarkable result was that the frequency scaling of individual cells was
largely determined by the visual stimulus statistics. There was no consistent relation between
individual neurons’ scaling exponent and the visual stimuli, neither was there any correlation
between membrane potential and the spiking scaling exponents. Various control tests and a
computer model corroborate the authors’ conclusion that statistical correlations in a neuron’s
input (i.e. its presynaptic activity) can modify the power law exponent of its spiking activity.
Hence, it appears that modulation in a neuron’s power law exponent may reflect changes in the
correlation state of the network activity. According to these findings, intrinsic cellular
properties do not seem to play a major role for its scaling which reflects in the authors’ s view
primarily the network context.
Regarding self-organization, El Boustani and Destexhe (2009) follow the lead taken in
Destexhe’s Doctoral Thesis of 1992 and observations of Korn and Faure (2003), and present
new evidence in support of chaotic dynamics in EEG: sensitivity to initial conditions is of course
prominent; it is also associated with broad-band power spectra and a fractal attractor
dimension. The authors confront at length the puzzle that coherence and low dimensiona lity
at the macroscopic level of EEG is associated with stochastic neuronal dynamics at the
microscopic level. Is this comparable to conditions obtaining in thermodynamics ?
Whence criticality ? In peripheral neurons, it seems to be favored by closeness to input
from peripheral receptors (Section 2.2.2). At the mesoscopic level, Plenz and Chialvo’s ( 2009)
analysis of avalanches in primate cortex seem to assure legitimate criticality at the mesoscopic
level; yet, the work of Bedard et al (2006) and El Boustani et al (2009) raises the possibility that
scaling properties of neuron activity may not be of intrinsic neuronal origin, but a consequence
of network activity. The next Section will continue to ask: if and where in the nervous system,
and under what conditions, does fractality and criticality in the brain originate ?
2.4 The macroscopic level of neural organization:
2.4.1: Fractals in brain networks
Fractality at the macroscopic brain level should be viewed in the context of, and in
reference to, the two major conceptual and observational frameworks that have come to guide
neuroscience research: the network structure of cortical connectivity, and the brain’s state of
criticality resulting from the complexity of nonlinear dynamic interactions among its
constituents. Advances in network theory (Albert and Barabasi , 2002; Dorogovtsev, 2002;
Park and Newmann, 2004 ) influenced the application of computational and graph-theoretical
methods for characterizing structural brain connectivity in accord with statistical and
topological criteria (Hilgetag et al 2002). Examining the columnar organization of neocortical
cortex in detail, Roerig and Chen (2002) found that the number of connections to a central
neuron has the shape of a long-tailed histogram, fitting a power law. On the basis of this “bio-
power-law connection probability function”, Stoop and Wagner (2007) tested a range of
network types for spread of synchronization among cortical columns: the superiority of the
power-law connection was evident. In general, interaction among neurons and neuron
ensembles by synchronization is constrained by network topology (Arenas et al, 2008), hence
the relevance of network architecture for Neurodynamics. The potential role of neural
synchrony for perceptual organization and conscious experience is a subject of a recent review
by Uhlhaas et al (2009). There is considerable evidence that anatomical and functional
connections between different cortical areas possess an intricate organization in the form of
“small world networks” (Watts & Strogatz, 1998) , forming clusters of nearby cortical areas with
short links, which in turn have long range connections to other clusters (Hilgetag and Kaiser,
2004; Sporns and Zwi, 2004; Sporns et al, 2004; Stam, 2004; Stam and Reijneveld, 2007).
Neuroanatomical data sets permit identifying a repertoire of characteristic structural building
blocks (motifs) (Sporns and Koetter, 2004).
A hierarchical cluster architecture is thought to provide the structural basis for stability
and diversity of functional patterns in cortical networks (Kaiser et al., 2007; Kaiser, 2008).
Moreover, hierarchical modular topologies assure sustained activation in neural networks,
intermediate between rapid fading and generalized activity spread. This can be considered a
prerequisite for the occurrence of criticality (Kaiser and Hilgetag, 2010). Hierarchical graphs can
switch between different dynamic activity patterns, depending on the level of ongoing
(spontaneous) background activity (Muller-Linow et al, 2008; Hutt and Lesne, 2009). In terms of
hierarchy theory, these investigations do not specifically address the implications of nested
hierarchies which, however, are suggested by the finding of inter- and intra-cluster network
hubs (Sporns et al, 2007). Moreover, fMRI data obtained from subjects in resting state identify
strong functional connections between regions for which no direct structural connections are
known (Honey et al., 2009). This finding may be an indicator of nested clustering (see Sections
2.4.3 and 6).
In the absence of deliberate external stimulation, neuronal cortical dynamics displays
complex spatial and temporal patterns of activity. In simulations of networks that mimick the
large-scale inter-areal connection patterns of cortex, activity takes place spontaneously at
multiple time scales, punctuated by episodes of inter-regional phase locking of oscillations
(Honey et al, 2007). Significantly, the connections link neural populations of multiple levels of
scale, from whole brain regions to local cell columns: this suggests that cortical connections
may be arranged in fractal, possibly self-similar patterns. Statistical measures of a
computational model of a fractal connection pattern did in fact resemble those of a real
neuroanatomical data set (Sporns 2006). The computational models also show that varying
fractal patterns induce strongly correlated changes in several structural and functional
measures of network properties, as evidence of their interdependence.
In general, scale free complex networks display self-similarity under length-scale
transformations (Song et al, 2005) but not necessarily with regard to degree distribution (Kim
et al, 2007), but models of scale-free networks need not necessarily be fractal. How, then, can
the fractality of many naturally occurring networks come into being ? Song et al. (2006) account
for the simultaneous emergence of fractality, modularity and small-world effect, as well as the
scale-free property of real world networks by a multiplicative growth process: the network
growth dynamics is conceived as the inverse of a renormalization procedure, whereby the
network hubs accrete connections by linking with less connected nodes, which leads to a robust
fractal topology.
Within the small-world network clusters, functional Magnetic Imaging (fMRI) identifies
a scale-free connection pattern inasmuch as the number of links per network node (the node
degree) satisfies a power law relationship (Eguiluz et al, 2005). Likewise, van den Heuvel et al
(2008) find In an imaging study of the resting brain, that inter-voxel connections follow power
law scaling as evidence for scale free network topology, possibly associated with a small -world
organization. This form of organization is associated with conserved wiring length and
conducive to synchronization of activity across the network (Zhou et al,2007; see also Changizi,
2003; Section 2.1).
Although citing merely a small fraction of the numerous publications concerned with
relations between network topology and dynamics, this section underscores two points of
relevance for the objective of this review: first, the presence of, and effect on network
dynamics of hierarchic network organization (itself being of several types); and, second, effects
of network fractality on network dynamics; but the functional implications of the latter,
notably for criticality, require further investigation, as does the possibility of self–similar
modularity in brain networks. In the case of metabolic networks, the latter is shown to affect
path connections for diffusion and resistance of flows (Gallos et al., 2007).
2.4.2: Fractals and Criticality of Brain States
Criticality, listed in the foregoing as the second notable feature in current thinking about
global brain function designates the view that brain is under normal circumstances at the verge
of undergoing a second order phase transition. This is attributed to its complex organization of
a large number of components interacting via nonlinear dynamic functions.
Measuring the fractal dimension of EEG records, Babloyantz (1986) related different
values with differences in sleep states. With subjects acting as their own controls, inhalation
anesthesia causes a noticeable increase in EEG dimensionality (Mayer-Kress and Payne, 1987).
Multichannel MEG records, obtained with a SQUID show scaling with varying degrees of scale
similarity , decreasing with the distance between recording channel locations (Novikov et al,
1997). Studying dynamical synchronization in the brain, Gong et al (2003) find scale invariant
fluctuations of dynamical synchronization in human EEG. Linkenkaer-Hansen et al (2001) report
long-range temporal correlations and scaling with 10-20 Hz brain oscillations. Pursuing this
observation in more detail, Linkenkaer-Hansen et al (2003, 2004) suggest that the long-term
spatial-temporal structure of the complex ongoing EEG activity may reflect a memory of the
system’s dynamics extending beyond just a few seconds, possibly by a continuous modification
of functional brain networks in the sense of SOC . In these tests, somatosensory stimuli
attenuate temporal correlations and power law scaling behavior, suggesting that stimuli
degrade the network memory of its past. The relationship to SOC was also the subject of the
work of Freeman et al (2003) in measurements of temporal and spatial power spectral densities
that identify EEG phenomena as fractal. Moreover, Freeman (2005) proposed a field-theoretic
approach to account for scale-free neocortical dynamics. In five frequency ranges (extending
from 0.5 to 48 Hz), detrended fluctuation analysis of EEG show global synchronization time
series with scale free features (Stam and de Bruin, 2004); the scaling exponent differs for
conditions of eye open and eye closed. Stam (2005) also reviewed the nonlinear dynamical
analysis of EEG and EMG at great length. Positive and negative feedback affect the scaling
exponent of EEG differentially; this was determined in a detrended fluctuation analysis (Biuatti
et al., 2007). Performance in Stimulus detection of weak stimuli is best accounted for by
modulation of the power law component in the power spectrum of MEG record : Shimono et
al, (2007) attribute this phenomenon to the brain operating in a state of self -organized
criticality which modulates the power spectral exponent to optimize responsiveness to external
stimuli.
Transients in EEG records can be detected as differences in fractal dimension of EEG
(Arle and Simon, 1990), as can be neuropathological conditions (Paramanathan and
Uthayakumar, 2008 ) , and differences in age and gender (Nikulin and Brismar, 2005). Nonlinear
spectral analysis enabled Kulish et al (2006) to determine in EEG a set of generalized fractal
dimensions and fractal spectra which reveal differences in subjects when replying to questions
with either YES or NO. In a study of human development from infancy to 16 years of age,
Thatcher et al (2009) measured phase shift duration and phase locking intervals of the EEG for
computing instantaneous phase differences between pairs of electrodes; the log-log spectral
plots showed 1/f distributions. The data revealed increased phase stability in local systems,
paralleled by lengthened periods of unstable phase relation between distant connections.
These results were taken to reflect progression towards self-organized criticality, accompanying
the growth spurts from infancy to adolescence. When Listening to music Bhattacharya and
Petsche (2001) find homogeneous scaling in the gamma band EEG over distributed brain areas,
whereas the homogeneity is reduce at rest, or when reading text or during spatial imagination.
As is well known, music has been under scrutiny for fractal properties for quite some time, see
for instance : Voss, 1975; Hsu and Hsu, 1991; Boon and Decroly, 1995 ; see also below:
Bianco,2007).
Long-range temporal correlations in spontaneous discharge patterns of hippocampal -
amygdala complex neurons show a power-law relation in epileptic patients (Bhattacharya et al,
2005): activity of individual neurons was in this study recorded by means of micro-wire
electrodes that had been implanted for localization of epileptic foci , and records were taken in
inter-ictal periods, with the subjects being awake . Neuronal activity in substantia nigra exhibits
fractal activity in anaesthetized rats, but was strikingly absent in the dopaminergic nigrostriatal
neurons with relatively constant discharge rate (Rodiguez et al, 2003). The authors consider the
possibility that pathological rhythmic discharges and tremor onset may be associated with loss
of the fractal pattern of nigrostriatal neurons. During paradoxical sleep and in the attentive
state, neurons in the mesencephalic reticular formation of unanaesthetized cats exhibit firing
patterns with 1/f spectral profile (Yamamoto et al, 1986). Kodama et al, 1988) extended this
observation to discharge properties of neurons in Hippocamus and ventrobasal thalamic
neurons, and suggest that 1/f structured patterns in discharge trains are indicative of spatial
and temporal summation of convergence. Variations in 1/f spectra in cortical and subcortical
brain structures of monkeys are apparently related to differences in emotional states (Andersen
et al., 2006).
In a very detailed and thorough study, Bianco et al (2007) identify the EEG time series
as a (non-ergodic) renewal non-Poisson process, reflecting strong deviation from exponential
decay. This startling claim is based on two premises: one, the comparison with the statistics of
an entirely different physical process, namely the fluorescence intermittency in blinking
quantum dots (Bianco et al, 2005); and, second, on the conjecture of the brain operating at or
near a self-organized critical state. The implication is that neuron synchronization can be
viewed as a kind of phase transition involving the close cooperation among many constituents
of a neuron set, each individual neuron in essence losing its identity. Furthermore, the absence
of exponential truncation would violate the ergodic condition (Bel and Barkai, 2005). The
authors then proceed to show that compositional music belongs to the same category of
processes. They finally claim that the effect of music on the human brain is in fact based on
the essential identity of their respective fractal dynamics, ensuing a kind of complexity
matching of the interacting brain-music systems. This aspect will be further pursued in section
5.1. Equally consequential are the inferences drawn by Allegrini et al (2008) from their EEG
data. The thrust of their analysis is on measuring the time distribution of reco rded events
occurring simultaneously at two or more electrodes (in their terminology: coincidences); they
find that the time interval between two consecutive coincidences has a waiting time
distribution corresponding to perfect 1/f noise. The theoretical analysis of this finding leads
these authors to infer that the coincidences are driven by a renewal process.
The electroencephalographic findings in support of 1/f scaling are supplemented by
observations with brain imaging: In 1997, Zarahn et al (1997) reported BOLD time series data
obtained from normal subjects at rest that exhibited a fractal power spectrum and self -similar
signal contributions, with disproportionate contribution of power in the spectrum for low
frequencies. The temporal variablility of brain activity in time series of fMRI data in combination
with a voxel-wise analysis of scaling exponents enabled Thurner et al (2004) to distinguish
different physiological states of the brain. In non-active brain regions, the voxel-profile activity
is described by a random walk model; in contrast, stimulus activated brain activity is
characterized as correlated fractional Browninan noise. The same group of investigators
(Shimizu et al, 2004) examined fMRI time series with a multifractal method to extract local
singularity (fractal) exponents: the range of Hoelder exponents in voxels with brain activation is
close to 1, whereas exponents in white matter and voxels in the absence of brain activation are
close to 0.5
Without further discussing at this point the far reaching implications of the non-
ergodicity claim (Tsallis, 2009; Tsallis, et al, 1995), I merely alert to two publications which
interpret human EEG signals in terms of a Tsallis Entropy measure (Capurro et al 1998, 1999).
The common theme of studies surveyed in the following is wavelet based
representations of functional magnetic imaging (fMRI) time series. Amongst others, Wornell
(1993) explicated in detail the role of wavelet based representations for the power law family
of processes. The remarkable feature of wavelet analysis is that it can be viewed as matching
self-similar processes since the wavelet coefficients exactly reproduce, from scale to scale, the
self-replicating statistical structure of such processes (Abry, 2003).
Publishing with various associates since 1994, Bullmore gathered extensive experience
with fractal analysis of human brain activity which led eventually to the suggesting that
wavelet-based f MRI time series estimates (Bullmore et al, 2001) can be viewed as realizations
of Fractional Brownian Motion , i.e. a class of fractals described by Mandelbrot & Ness (1986),
characterized by zero-mean, and non-stationary and non-differentiable time functions (see
Section 2). Extolling further the virtues of wavelet techniques for the purposes on hand,
Bullmore et al (2004) and Maxim et al, 2004) give a meticulous account of their use of the
’discrete wavelet transform’ approach to fMRI time series evaluation. In normal subjects at rest,
the time series is most parsimoniously described as Fractional Gaussian Noise, signifying a
persistent long-memory fractal processes of which the Hurst Exponent is a defining parameter.
Interestingly, the value of this parameter in Alzheimer subjects differs from the norm (Maxim et
al, 2004). Several results from the same laboratory contribute additional facets to the notion of
the active brain displaying fractal properties: Achard et al (2006, 2008) applied discrete wavelet
transform analysis to fMRI time series to estimate the frequency dependence of functional
connectivity between some ninety cortical and subcortical brain regions; the functional
networks is dominated by a neocortical core of highly connected hubs with an exponentially
truncated power law degree distribution. Dynamical analysis of brain at wavelet scales from 2 -
37 Hz show the emergence of long-range connections with execution of motor tasks (Bassett
et al, 2006). Under certain conditions (e.g. age, cognitive performance, certain pharmacologic
interventions) brain dynamics requires a more comprehensive description than is captured by
the monofractal analysis applied in the studies cited thus far (Suckling et al., 2008. In such
cases, a more comprehensive description must make allowance for scaling behavior that is
governed by several local scaling exponents. Multifractal analysis can then be characterized by
the histogram of the Holder. Following expenditure of cognitive effort, the brain’s fractal
oscillations require several minutes for returning to baseline activity, this time depending on
the task’s cognitive load; this is taken to signify the relevance of fractal scaling for adaptive
task processes, in addition to the role it plays for the “resting” brain (Barnes et al, 2009). The
substantial evidence for modular organization of brain networks is reviewed by Bullmore and
Sporns (2009), and was subsequently further refined by Meunier et al (2009), applying a
method for rapid, high-resolution modular decomposition of brain functional networks (Blodel
et al, 2008). Differences between low frequency BOLD signal spectral power in task and rest
periods also support the notion of fMRI reflecting meaningful brain states (Duff et al, 2008), as
do the emotional task dependent fractal fluctuations in fMRI of the cerebellar vermis
(Andersen et al, 2006). Brain imaging in Neuropathology has revealed significant differences
between patients suffering from unawareness of ownership of one arm (Asomatognosia) and
those with additional confabulations (Somatoparaphrenia): the latter patients display lesions in
the medial and orbitofrontal regions, in addition to the multiple large lesions including
temporo-parietal sectors which are common to both groups of patients (Feinberg et al, 2010).
For the distinction of stimuli related to the self (i.e. self -referential stimuli) from those not so
related, Northoff et al (2006) identify processes mediated by cortical midline structures.
Alternative approaches to the question of brain criticality have a distinguished
history: sudden transitions between stable states of motor behavior are well known since the
pioneering observations of Haken et al. 1985; Kelso, 1995). The transitions were interpreted as
manifestations of metastability in the self-organizing nonlinear dynamic system of the brain,
along the theoretical lines formulated in Synergetics (Haken, 1983). Criticality in brain and
behavior was first mentioned by Kelso (1984) in a brief note. Using a superconducting quantum
interference device (SQUID) sensor array, Kelso et al (1992) reported a few years later their
observations of spontaneous transitions in neuromagnetic field patterns which occur at a
critical value of a behavioral parameter: coherent states of both brain and behavior were
captured by the spatiotemporal pattern of phase relations among participating components.
This was considered evidence for the brain being a pattern forming system that can switch
flexibly from one coherent state to another. Chialvo credites also Varela (2001) with the vision
of brain large-scale dynamical properties. Locating cortical regions associated with such phase
transitions of motor behavior, Meyer-Lindenberg et al (2002) showed that TMS can induce
switches between two clearly defined and distinct motor behavior patterns. Additional new
evidence of aspects of critical brain behavior accrued in rapid sequence in the following years
in several forms: as the scale free connection pattern of cortical networks (Eguiluz et al, 2005;
Chialvo, 2004; 2008) ; as result of the avalanche analysis of Beggs and Plenz (2003,2004) at the
mesoscopic level; as coordination dynamics of large scale neural circuits subserving rhythmic
sensorimotor behavior (Jantzen et al, 2008); and finally from two fMRI studies at the
macroscopic level, which will be the subject of the following paragraphs.
Two virtually simultaneously published recent studies, using different experimental
strategies, deliver seemingly firm evidence for brain criticality. Kitzbichler et al (2009) based
their approach on the widely accepted view that many behavioral and cognitive states are
related to coherent or phase-locked oscillations in transient neuronal assemblies (for a recent
summary: Womelsdorf et al, 2007). The measures for determining phase synchronization
between component processes were in their study the phase lock intervals (estimating the
length of time a pair of bandpass filtered oscillations remain in phase synchronization), and the
lability of global synchronization (informally analogous to the previously discussed
avalanches). Applying these measures to functional MRI and MEG data recorded from normal
volunteers at resting state demonstrated power law scaling of both pair wise and global
synchronization. They then evaluated the performance of two models , both typically being
used in nonlinear dynamics: the Ising and the Kuramoto (1984) model. Observed and model
generated data were identical, provided the model system was in a critical state. Hence, the
authors conclude that the brain must be in a critical state. Moreover, the critical brain
dynamics obtained at frequency intervals ranging from 0.05-0.11 to 62.5-125 Hz, confirming
criticality of the human brain network organization across its functional bandwidth. They
consider therefore ‘Broadband Criticality’ as a characteristic property of the resting brain
network functional organization
Although also using the Ising model as reference point for determining brain criticality,
Fraiman et al (2009) followed an entirely different approach: the issue at stake in their study
was to determine whether and to what extent the dynamics of the paradigmatic two
dimensional Ising model at criticality displays features that correspond to patterns
encountered in the imaging of (resting) brain networks. However, unlike most prior studies of
brain dynamics cited in this Section, no prior assumptions on structural connectivity of brain
regions were made. Instead, network connectivity was extracted from voxel correlations: thus,
networks were here defined in terms of correlations among the activity at each location (voxel
in the case of the brain, and lattice site in the Ising model). Prior investigations showed that the
so called ‘resting state’ (absence of overt external stimulation) is subject to a Default Network
Dynamics, reflecting balanced positive and negative correlations between activity in
component brain regions (Fox and Raichle, 2007; Baliki et al, 2008); this is not the case under
certain abnormal conditions (Baliki et al, 2008) . The result was that the dynamics of the Ising
model at criticality, as captured by the correlation networks, exhibits average statistical
properties which are identical to those observed in the brain networks at resting condition.
Among several other network characteristics that match critical Ising dynamics with brain
dynamics was also the equality of the fraction of sites with positive and negative correlations ,
corroborating that the dynamics of the normally functioning brain at rest being near a critical
point. In any case, the unequivocal answer to the question the investigators set out to answer
was that networks derived from correlations of fMRI signals in human brains are
indistinguishable from networks extracted from Ising models at critical temperature.
In an important next step, Expert et al (2010) investigated the large-scale dynamical
properties of resting brain by examining more closely the character of the spatio-temporal
correlations: considering three successive steps in spatial coarse graining, two-point correlation
functions exhibit self-similarity; self-similarity in time was revealed by 1/f frequency behavior
of the power spectrum. The condition of long range correlations in space and time
presupposes a dynamical system at criticality; the strong correlations across large distances are
indicators of highly integrated cortical states, with nearby clusters functioning in synchrony.
Apart from this principal conclusion of this study, the authors of this study also alert to a
significant property of the brain networks which, as noted before, are extracted from the site-
to-site temporal voxel correlations: obviously, equally oriented spins in the Ising model coalesce
in large domains near the critical temperature where also nontrivial collective states emerge
in the Ising model’s otherwise regular lattice. Similarly, large regions of brain activate
concurrently with deactivation of other regions. How does the brain self-organize to negotiate
the dynamic balance between the extreme possibilities of total quiescence and explosive
massive excitation? The authors refer to a discussion of this stability problem which was
already noted by Abeles (1991). It motivates their question: is it necessary to confine brain
activation to structural connections linking brain regions, as is customary in most current
research ? (e.g.: Hagmann et al, 2008). Take the Ising model as example: there, a change of
temperature can lead to the emergence of functional collectives, in the absence of preexisting
structural connections. This leads Fraiman et al. to ask: might the brain, likewise, have this
capacity, as basis of a kind of adaptive coordination dynamics of the kind envisioned by Kelso
and Tognoli, (2007) and Tognoli and Kelso (2009) ?
Apparently, SOC, metastability and phase transitions constitute a nexus of intimately
interrelated dynamic processes of which fractals and self-similarity are pivotal aspects.
2.4.3: Significance of Brain Criticality
In statistical Physics, systems operating at the critical point of transition between
ordered and random behavior are metastable with respect to a set of control parameters, and
are capable of rapid qualitative change in response to fluctuations of external input. For
systems far from equilibrium most of the analytical and numerical methods of the ‘classical’
(equilibrium) theory appear to remain valid (Sornette, 2000 ). Moreover, dissipative (open)
Hamiltonian System, such as the brain, have the capacity to form “strange” attractors whose
boundaries and bases have fractal properties (Aguirre et al, 2009; see Section 4) . At or near the
point of phase transition, the systems exhibit complex patterns of fluctuations on all scales of
space and time, as one the indicators of an impending phase transition ; another is the slowing
down of relaxation processes, associated forming long range correlations for efficient functional
coupling among system components: both events are anticipatory signals of impending critical
transitions (Scheffer et al, 2009). Fractal clusters formed by phase transitions can be
characterized in terms of correlation length (Antoniou et al, 2000) which is associated with
fractal scaling of clusters of correlated elements on all scales; as a result, any intrinsic scale
before phase transition is de facto ‘forgotten’ (Stinchcombe, 1989). As a corollary, the system
presents at the critical transition qualitatively new properties, requ iring new macroscopic
descriptors. The important feature of the organization following the phase transition is to form
new objects with distinct properties. In physics, this is manifest as, for instance, the phase
transition from ferro- to para-magnetism, or from water to ice (Stanley, 1999). Typically, one
deals with a large collection of ‘microscopic’ constituents which, at phase transitions, arrange
to a macrostate which displays qualitatively novel features and properties. The macrostate’s
new properties have no referent at the microscopic level, and require new descriptors: by way
of illustration, think of hardness or liquidity in the ice -water example as descriptors of new
physical properties, originating de novo upon phase transition. The properties described in the
foregoing are universal in the sense that the apply irrespective of the system’s constituents at
the microscopic level.
One of the amazing features of phase transitions is that material systems of diverse
physical properties at their microscopic level form on phase transition but a small number of
Universality Classes which share identical macroscopic properties ( for a discussion in relation
to brain function: see Werner,2009 c). Based on a stochastic theory of neural activity, Buice
and Cowan (2007) developed field theoretic methods for nonequilibrium statistical processes;
their model exhibits a dynamical phase transition of the universality class of directed
percolation (see Section 2.4.2).
Critical Theory (Stanley, 1987; Marro and Dickman, 1999) considers reality as a hierarchy
of levels, each having its own scale, its own description and a theory that accounts for that
description. The scale on each level emerges from the scale on the next finer level by ignoring
some of the lower level details which become invisible at the higher level scale (Laughlin, 2005 ;
Sokal and Bricmont,2004)). The result is a drastic reduction of dimensionality. Coarse graining
(specifically renormalization group transformation ) (Fischer, 1998) unveils self-similarity at the
point of phase transition. The intimate relations between scaling, renormalization group, and
long-range correlations are addressed by Perez-Mercader (2004) and Penrose (1986), the latter
pointing out that the definition of fractal dimension depends primarily on the distribution of
widely separated sites, telling little on sites that are close together.
What is the significance of criticality ? Excitable systems at criticality exhibit an optimal
dynamical range for information processing (Kinouchi and Copelli, 2006; see Shew et al, 2009,
in Section 2.3) . Furthermore, a model that reproduces the typical features of systems at a
critical point learns and remembers complex logical rules: learning occurs by plastic adaptation
of synaptic strengths, and exhibits universal features in being independent of the specific task
assigned to the system (Arcangelis and Herrmann, 2010). Finally, Phase transition in critical
systems provide an universal mechanism for rapid switching between different cooperative
neuron collectives. These three attributes of criticality are the reason for its rapid ly moving into
center stage of current brain theory.
3. Psychological and Behavioral Processes
The following overview of psychological functions with power law scaling is predicated
on the notion that mental states may be viewed as macrostates emerging from EEG dynamics
(Allefeld, 2009), and neurophysiological processes generally. Classical Psychophysics of
Helmholtz, Fechner and Weber sought to establish dependencies of perceptual experience on
properties of physical stimuli impinging on sensory organs. In 1975, Stevens reported the
summary of the extensive work that led him to propose that this dependency is in many
sensory modalities a power function. In neurophysiological experiments, Werner and
Mountcastle (1963,1964) identified the power function scaling of responses in primary afferent
cutaneous nerve fibers to mechanical indentation of peripheral receptors. Neurons of primary
visual cortex (V1) exhibit a higher coding efficiency and information transmission rate for input
signals with natural long term (1/f) correlations (Yu et al, 2005). Copelli et al (2002) and
Kinouchi and Copelli (2006) claim that Stevens’ law (1957) for intensity of subjective sensory
experiences can be attributed to dynamics in a network of excitable elements constituting the
peripheral receptors, set at the edge of a phase transition, i.e.: of being in a state of crit icality.
For a discussion of this view, see Chialvo (2006).
Unlike dismissing the fluctuations in the performance of many psychophysical task as
“noise”, Gilden ( 1997,2001) attributes them to a memory process associated with active choice
and discrimination. This memory process is suggested to express itself as 1/f ‘noise’ in the three
major measurement paradigms in Psychophysics: speeded judgment, accuracy of discrimination
and production. The 1/f fluctuations are attributed to an intrinsic dynamics, associated with
the formation of representations, comparable to the kind of memory that arises in dynamical
systems as they flow forward in time, along principles outlined by Beran (1994). According to
this interpretations of the psychophysical observations, cognition would generate its dynamical
signature as a consequence of its own activity: this would entail a fundamental revision of what
is signal and what is noise in psychophysical data. Gilden points out that” the conventional
experimental design and data analysis using ANOVA does in fact bury “ one of the most
important signatures of what happens when the mind is working”.
Timing fluctuations in tasks requiring sensorimotor coordination display cycle -to-cycle
fluctuations which, analyzed as time series, show fractal scaling of power spectra . Ding et al
(2002) suggest that the reason for this lies in the multiple time scale activities of distributed
neural areas that contribute to the task performance. If asked to produce random series of
numbers from a given set, series with short and long range correlations are produced which in
most cases exhibit a power law spectrum (Morariu et al, 2001). Van Orden et al (2003)
interpret serial correlations in human cognition as evidence of self -organization. In their view,
self-organization coordinates the activities of the organism across a hierarchy of time scales,
producing correlated variation across time: variations in response times would then appear as a
natural fractal in which larger scale deviations nest within themselves smaller (self-similar) scale
deviations. Accordingly, 1/f noise is in this view not sufficient evidence for self -organized
criticality, but rather its necessary consequence. Similarly, Kello et al (2007) assembled reaction
time and response data which lead them to considering the 1/f scaling of their data as
expression of a coordinative, metastable basis of cognitive functions. This view is in effect an
extension of Van Orden’s et al (2003), shifting the genesis of 1/f scaling from self-organization
to metastability: the 1/f pervasiveness in the brain would be the signature of metastability
associated with cognitive functions. However, these claims are challenged by Wagenmakers et
al (2005) and contrasted with the alternative that long term serial dependence in data can be
explained in a number of ways , for instance by mixtures of a small number of short -range
processes) (Wagenmakers et al, 2004).
Applied to problem solving and insight, reasoning was viewed by Stephen and Dixon
(2009) as the self-organization of novel structures: taking a particular problem solving task as
example, the authors suggest that the problem solution can be viewed as a phase transition in a
self-organizing system whose dynamics would be reflected in power law behavior. Implications
for social psychology are reviewed by Correll (2008): cognitive effort to avoid bias in judgments
reduces the scaling exponents of response times relative to less challenging tasks. Grigolini et al
(2009) interpret Correll’s data to suggest that increasing the difficulty of cognitive tasks would
accelerate the transition from observed 1/f noise to white noise in decision making time series.
The temporal structure of many human-initiated activities can display a striking
regularity. Barabasi (2005) showed that a decision-based queuing process can account for the
dynamics of some human patterns of activity: when individuals execute tasks based on some
perceived priority, the timing of the tasks will indicate the signatures of fractal dynamics:
heavy-tailed distributions with initial fast bursts.
If patterns of expression in spoken language reflect in some way the organization of
brain processes, then Zipf’s law is of course the notable landmark that presages more recent
fascinating reports of fractal patterns and scale-invariant word transition probabilities in spoken
and written texts (Costa and Sigman, 2009; Altmann et al, 2009; Alvarex-Lacalle et al, 2007),
and their extension to music (Zanette, 2008). On the basis of EMG data, it appears that some
common features of patterning in language, music and syntax (Patel, 2003) can be attributed
to neural activity in Broca’s area and its right hemisphere homologue (Maess et al, 2001).
3.1 Symbol processing and fractals
The classical book “Language of Thought” (Fodor, 1975 ) epitomizes the framework of
computation-representation of the Computational Theory of Mind. However, with adopting a
dynamical perspective, it became appropriate to view ‘representation’ in terms of regions of
state space, and ‘computational rules’ as attractors (Elman, 1995); the dynamics is supplied by
Recurrent Neural Networks (RNN). Systems of this kind learn to recognize and generate
languages after being trained on suitable examples. Surprisingly, it turned out that the
induction of this ability occurs when small network parameter adjustments bring about a phase
transition in the neural network’s state space. Once in a certain state, machine states for
correct recognitions scale with an exponent of 1.4 (Pollack, 1991). Considering the RNN as a
dynamical system, it appears that its trajectories can locate regions in phase space which
support fractal dynamics: putting it in a graphical way, the phase space would seem ‘peppered’
with regions for fractal dynamics (i.e. attractors), which can be reached by the trajectories of
the complex system’s dynamics. Numerous additional sources point to a close, though often
not readily transparent relation between the dynamics of RNN and IFS: the principle is
consistent with the observations of Pollack (1991) inasmuch as fractal sets provide a method
for organizing recursive computation in a bounded state space (Tabor, 2000). Furthermore,
context-free grammar computation by connectionist networks using fractal sets can generate
spatial representations of symbolic sequences (Tino, 1999; Jeffrey ,1990) via IFS (Barnsley &
Demko, 1985; Barnsley et al, 1989). A class of associative reinforcement learning algorithms
was constructed by Bressloff and Stark (1992) as an extension of non-associative schemes in
stochastic automata theory; within the IFS framework, it suggested a possibly fractal nature of
the learning process. Tsuda and Kuroda (2004) recently elaborated this idea and developed a
mathematical model of Cantor Coding for the formation of episodic memory in the
hippocampus.
The intent is here merely to draw attention to a large segment of literature, o f which
the foregoing citations are but a small sample that implicates interrelations between state
space dynamics of RNN and IFS in the processing of symbolic information. For clarification of
this relationship, Kolen (1993, 1994) proposed that the RNN’s state dynamics itself is an IFS, as
a paradigmatic case of the synergism of fractal and complex system dynamics. Levy & Pollack
(2001) obtained supportive evidence in that every point in the hidden layer of the RNN is either
itself part of the fractal attractor of the IFS, or has an orbit that “ends on” the attractor in a
finite number of steps. Two tantalizing questions arise: one, wherein does the ‘computational’
power of a Fractal System lie? How does the self-similar structures of fractals unpack layers of
‘information” for guiding actions across many scales still eludes our comprehension. And,
second, what exactly is the nature of that apparent synergism between complex system and
fractal dynamics ? (see Section 5.2).
3.2 Motor Behavior and Allometric Control processes
During quiet standing, the human body sways in a seemingly erratic fashion. Collins and
De Luca (1994) determined that the pattern of this postural sway is exhibits intrinsic
correlations which can be modeled as a system of bounded correlated random walks. This
result suggested to the authors that the postural control system incorporates both open and
closed loop control mechanisms. The statistics of temporal patterns in spontaneous motor
activity of laboratory rodents can be replicated by the stochastic mechanism of Davidsen and
Schuster (2002, see Section 4) which generates power law distributions and 1 1/f power
spectrum over several decades (Anteneodo and Chialvo, 2009). The presence of long -time
correlations in the stride-interval time of normal humans suggested to Hausdorff et al (1995)
that the activity of walking may be a self-similar fractal. Acknowledging the complexity of
locomotor activity, the authors referred to its prerequisite of coordinating inputs from motor
cortex, basal ganglia and cerebellum, as well as feedback from vestibular, visual and
proprioceptive sources. In the Hausdorff et al. study, the gait cycle (synonymous with stride
interval) was defined as the time between consecutive heel strikes of the same foot.
West & Griffin (1998) and Griffin et al (2000) took a different and novel approach to the
analysis of gait patterns: using the time between consecutive maximal positive extensions of
the same knee for measuring the stride interval, their data analysis was based on determining
the long-time correlation properties of the stride interval time series . The Relative Dispersion
(given as the ratio of standard deviation to the mean) for different levels of data aggregation
captures the inter-relatedness of the data across multiple time scales. The method is described
in detail by Bassingwhite et al, 1994) and is designed to answer whether the correlations are
self-similar upon scaling (i.e.: identical between groups of neighbors at different time scales).
The result of the data analysis was that the fluctuations of the gait cycle were self-similar with
a fractal dimension of 1.25. In addition, stride-interval time series itself was in this study a
random fractal, consistent with the data of Hausdorff et al.
The importance of the West-Griffin results lies in showing that the correlation in their
data was an inverse power law of a form similar to the allometric scaling laws found in many
areas of Biology : typically, allometry establishes a relation between two properties of an
organism. Historically, the idea is based on Huxley’s (1931) definition of allometric growth,
describing that the different growth rates of two parts of an organism are proportional to one
another. In the West-Griffin studies, the allometric principle is reflected in the constancy of the
Relative Dispersion over the length of the stride interval time series which, in the present case,
has the non-integer fractal dimension of 1.25. Their data raise the issue of systematic control
of variability, which is generic of complex physiological processes. Unlike the familiar
homeostatic control that regulates system variables by negative feedback, an allometric
control system is conceived as regulating variability of a process involving multiple interactions
among sensors and effectors with intricate feedback arrangements, each with its own
characteristic set of frequencies and time scales. Their functions are reflected by the allometric
relation which captures the process’s long term memory with power law correlations, and by
the power law distributions of the system variable (West, 1999 b).
The significance of this principle is documented by West (1999 a, 2006) for the
numerous physiological processes which are identified as fractal, on the basis of their time
series behavior. Notable examples of fractal Physiology are heart rate, bronchial air ways and
body temperature variability, and integrated neural control networks. In these situations, the
regulatory mechanisms constitute coupled cascades of feedback loops in systems far from
equilibrium. Therapeutic interventions, commonly based on the homeostatic principle which
assumes the significant system variable to be normally distributed fails to take the regulatory
complexity into account and may be counterproductive. Instead, Allometric (fractional) Control
based on Fractional Calculus (Podlubny, 1999) provides the appropriate approach. Applications
of Fractional Calculus to modeling the interdependence and organization of complex system,
such as for instance the vestibulo-oculomotor system, are illustrated by Magin (2004).
Changing walking speed, using metronomically controlled walking, or aging and
pathological conditions introduce stress conditions to the neural control system which
requires expanding the theoretical framework. Based on the notion of a stochas tic model of
human gait dynamics (Ashkenazy et al, 2002), West and Scafetta (2003) tested the model of a
neural pattern generator on the data set obtained by Hausdorff et al, 1995) which they showed
to exhibit slightly multifractal fluctuations. Metronome timing breaks the long-time correlations
of the natural pace and generates a large fractal variability of the gait regime. The two essential
features of the model required for capturing the phenomenology of the data set were that the
dynamics of the system unfolds on an attractor in phase space, and that the natural frequency
of the attractor is replaced by a random walk over a restricted set of frequencies which leads to
the multifractal output for the dynamical model (Scafetta et al, 2009).
4. Processes that generate power law distributions
Antedating the modern theory of stochastic processes, Yule (1925) proposed a model of
speciation to explain the highly skewed distributions of abundances of biological genera. Thirty
years later, Simon (1955) derived several related stochastic processes from relatively general
probability assumptions that lead to Yule-type distributions. Their characteristic properties
distinguish them from the negative binomial and Fisher’s logarithmic series. Leaving open the
possibility of still other generative mechanisms for power law distributions, Simon suggests that
the frequency of occurrence of this empirical distribution should not come as surprise. The
preferential attachment scheme for network growth (Barabasi and Albert 1999) has stimulated
the recent interest in the Yule-Simon approach in as much as Bornholdt and Ebel (2001) could
show that they are closely related. The important step of introducing the notion of aging of
network nodes was taken by Dorogovtsev and Mendes (2000): the probability of being linked
to a newly added node is taken to be proportional to its current connectivity weighted by a
power law function of its age. This motivated Cattuto et al (2006) to propose a modified Yule -
Simon process that takes the full history of the system into account, applying a hyperbolic
memory kernel.
Simon’s conclusion that power law distributions can be derived from relatively general
assumption seems to be born out by the number of mathematical models that have been
proposed. A shot noise process, reviewed by Milotti (2002) is an example, as is the Reversible
Markov Chain Models (Erland and Greenwood, 2007), and the Clustering Poisson Point Process
(Grueneis,2001), the latter already introduced in Section 1.2. The simple stochastic mechanism
of Davidsen and Schuster (2002) generates pulse trains with power law distributions of pulse
intervals, and 1/f power spectra over several decades at low frequencies with an exponent
close to 1. Iterated function systems (IFS) are a unified approach for generating and classifying
a broad class of fractals with self-similarty (Barnsley and Demko , 1985). The Chaos Game is a
generalized form of this, designating a method for generating the attractor (fixed point) of any
IFS. Other Recurrence Models (Kaulakys et al, 1998, 2006) derive from a more specific frame of
reference insofar as they consider random walks in complex systems that display self -
organization. As alternative, Ruseckas and Kaulaskys (2010) generate 1/f noise with nonlinear
stochastic differential equations. Touboul and Destexhe (2009) followed a similar route when
developing their case against power law scaling of neural avalanches (see Section 2.3). Physical
systems whose observable properties exhibit values which randomly exceed certain critical
values are candidates for applying Extreme Value Theory: the aim of the classical form of this
theory is to quantify the properties of the extremes (large or small) occurring in random
sequences of independent numbers. Extremal dynamics may be applied to generate objects
with fractal structure (Miller et al, 1993); as Extremal Optimization, it successively eliminates
undesirable components of suboptimal problem solutions (Boettcher and Percus, 2000).
The various approaches discussed in the foregoing can essentially be viewed as ad hoc
(Milotti ,2002). In contrast, however, there are two types of conceptual anchors that ground
power law relations explicitly in larger foundational contexts. For one of the conceptual roots, I
turn to the theory of Random Walks and fractional difference equations. The continuum limit
of simple random walks is diffusion and, correspondingly, expressed in the mathematics of
differential equations. The simple random walk aggregates the random steps from a large
number of identically distributed random variables with finite variance. However, an extensive
range of investigations has made it abundantly clear that simple random walks with this
statistics do not capture the richness of biological data, and for that matter other fields of
investigation as well (for reviews see : West and Deehring, 1995; West 1999; Bassingthwaighte
et al, 1994). A decisive step beyond simple random walks was the introduction of the concept
of Continuous-Time-Random Walk (CTRW) by Montroll and Weiss (1965). Some forms of CTRW
are fundamentally different from the classical diffusion model by drawing the timing of steps
from waiting time distributions, or by taking steps of randomly varying length. This is for
instance the case when the waiting time distribution does not possess a characteristic time
scale (for instance, has a power law distribution) : in this situation, the mean square
displacement and the distribution of transition rates become fractal. Processes corresponding
to these and related random walk models are then referred to as fractal random walks,
corresponding to anomalous diffusion which occupies an important place for studying physical
processes such as transport in disordered media or non-exponential (anomalous) relaxation of,
for instance, glassy media. Along these lines, Montroll and West (1979), Hughes et al (1982) and
others examined a large repertoire of stochastic processes with unusual probability
distributions for the displacement per step. For certain parameters, these walks have infinite
spatial moments, generate fractal self-similar trajectories, have characteristic functions with
nonanalytic behavior, and lead to an analog of RNG transformations. In the continuum limit,
the fractal random walk leads to the Fractional Langevin Equation of motion describing
trajectories, and their ensemble densities, in phase space (West, 2006). Such processes are
viewed as fractional kinetics, and mathematically addressed in fractional calculus (Sokolov et al,
2002; Kleinz and Osler, 2000) and by Fractal Operators (West et al, 2003).
In an application to Neuroscience, Lundstrom et al (2008) showed that neocortical
pyramidal neurons’ firing rate is a fractional derivative of slowly varying stimulus parameters:
neuronal fractional differentiation effectively results in adaptation with many time scales (see
Section 5.2). Fractional order dynamics of brainstem vestibulo-oculomotor neurons was
demonstrated by Anastasio (1994) who also suggested that simulation of fractional-order
differentiators and integrators can be approximated by integer-order high- and low-pass filters,
respectively. Thus, fractional dynamics may possibly be applicable to motor control systems,
generally. This is also suggested by the stride-interval time series of human gait being a random
fractal, indicating the role of long-time correlations in walking (West and Griffin, 1999; see
Section 3.1). Mandelbrot and van Ness (1968) defined Fractional Brownian Mot ions as a family
of Gaussian random functions, parametrized according to the interdependence of successive
increments, with the parameter ranging from zero (Gaussian Fractional Random Walk) to
infinite in Fractional Brownian motion : the latter to account for the empirical studies of
random phenomena with interdependence of distant samples. The conceptual connections to
scaling invariance and to the theory of renormalization(Section 2.3.3) are discussed by Quian
(2003). Fractional reaction-diffusion in inhomogenous media stabilizes steady state solutions of
Turing patterns (Henry and Wearne, 2000).
In 1987, Shlesinger et al. introduced the Levy walk as a random walk with nonlocal
memory, coupling space and time in a scaling fashion. For the alpha-stable Levy Walks, the
transition probability varies with the size of the step (Montroll and West (1987). Anomalous
diffusion results from a Levy Flight which is a process where the time taken to complete a
transition depends on the length of the step (West et al, 1997). West et al (1994) also
identified dynamical generators of Levy Statistics . In an elegant step towards unifying various
classes of random walks, Zumofen and Klafter (1933) applied the framework of CTRW’s to
derive Levy stable processes. The interesting properties of Levy processes include their
satisfying a scaling law, self-similarity and possessing memory (Allegrini et al, 2002) . Levy
(1954) also generalized the Central Limit Theorem to include those phenomena for which the
second moment diverges. West and Deering (1995) and West (2006) assembled a large number
of data obtained from various biological systems that satisfy Levy walk statistics. In a motor
skill acquisition task, Cluff and Balasubramaniam (2009) report that probability distributions for
changes of fingertip speed in pole balancing are Levy distributed. In vitro recorded spontaneous
electrical activity of neuronal networks exhibits scale –invariant Levy distributions and long-
range correlations (Segev et al, 2002). This is thought to enable different size networks to self-
organize for adjusting their activities over many time scales. Among animal movement patterns
associated with random search behavior , Levy walks outperform fractional Brownian motion
(Reynolds, 2009), presumably evolved under section pressure (Bartumeus, 2007).
Physical process models to account for fractal heavy-tailed distributions of traffic
pattern of (information) packages in LAN’s (Local Area Networks) are based on renewal reward
processes, originally applied to commodity pricing (Taquu and Levy,1986). Applied to network
package traffic, the model takes into account the presence of long packet trains (“on periods”,
with packages arriving at regular intervals) and long inter-train pauses (“off periods”). The
superposition of many such packet trains displays on large time scales the self -similar behavior
LAN’s if the “on-off” distribution has infinite variance (Willinger et al, 1995, Willinger, 2000).
The second conceptual framework was already introduced in Section 2.3.3: power law
distributions are among the novelties that arise in the vicinity of or at the critical point of a
continuous phase transition, including criticality of the self -organized kind. This should not
come as surprise since scaling reflects long-time correlations in the underlying process,
analogous to the comparable re-ordering process at critical phase transitions (Wilson, 1979):
both cases address a class of phenomena where events at many scales make contributions of
equal importance. Significantly, the comprehensive review on Fractal structures in nonlinear
dynamics by Aguirre et al. (2009) begins with the sentence ”Fractal structures appear naturally
in nonlinear dynamics, in such a way that the two concepts are deeply related”. Their review
draws particular attention to numerous instances in nature where attractor basin boundaries in
dissipative (open) systems display fractal behavior. Giesinger’s (2001) comment is a propos: “ at
one point, Bak(1996) considered SOC as universal, with scaling as consequence; it appears,
however, that the balance of evidence shifted the question : why is there scale invariance in
Nature ? to the question: is Nature critical ? ”
For constructing theories that deal with problems that have multiple scales, the
renormalization group (RNG) offers a general method. In Physics, the most frequently studied
situation is ‘percolation transition’ for which Newman (2005) offers a detailed account of the
origin of power law scaling: the cumulative distribution of cluster sizes forms at the critical
point a power law distribution. Percolation transition is a special case under the closely
interconnected family of RNG and coarse graining that entails power law distributions as a
source of natural fractals (see Section 1.4). Coarse graining allows one to determine whether
the phenomenon under investigation has universality, apart from scaling: Universality implies
that macroscopic properties of a system are independent of the system’s particular microscopic
configuration. The particular values determined for a given instantiation of the system are then
not significant, apart from showing that the system scales: the theoretical foundations are
extensively discussed by Essam (1980) and by Stauffer and Aharony (1991/1994).
For Neuroscience, Kozma et al (2005) illustrated the potential relevance of percolation
for phase transitions in models of neural populations with mixed local and global interactions,
and (Werner,2009 b,c) proposed Renormalization Group Transformation as a general principle
to account for functional relations between levels of neural organization. Since fractals will in
both situations naturally arise, it is pertinent to ask what their ro le could be. West et al (2008)
and Allegrini et al (2006) attribute to them a complexity matching function which will be the
subject of review and comments in the next section 3.1.
Section 5: Fractals in Action.
Having reached the end of the largely phenomenological surveys of fractal scaling and
associated manifestations of fractality at the conventionally distinguished levels of organization
and function of nervous systems, it appears inescapable to recall the title of Barnsley’s (1993 )
book ‘Fractals everywhere”. The apparent prevalence in the nervous system is matched by
the numerous manifestation in physiological systems generally (West and Deering, 1995). Is
the ubiquity a sign of triviality, or the result of a generic and fundamental principle of Nature?
If that is the case in Biology at least, then Nature seems to adhere to it with remarkable
conservatism as, for instance, the monographs of Dewey (1999) for Molecular Biophysics and
of Seuront (2009) in Ecology attest.
As documented by Aguirre et al. (2009) , and is referenced in several sections of this
review: the intimate relation between fractals and nonlinear dynamics in dissipative systems is
apparent and well substantiated. Hence the seemingly disproportionate attention paid to
phase transitions and criticality in Sections 2.2 and 2.3. However, Section 4 lists a large number
of alternatives for generating fractals, some of which obviously qualifying as ‘natural’, as, for
instance Levy flights. Thus, the burden of proof of attributing observed fractals to nonlinear
dynamics lies on identifying the fractal boundaries or the critical phase transition that gave rise
to them. This subset of natural fractal must then be viewed as manifestation of, and a
consequence of, the particular dynamic regime from which they originated.
5.1: the Complexity Matching Effect (CME)
The issue under consideration is the communication among complex systems generating
fractal signatures. The starting point is the evidence presented in Section 1.4 that the EEG time
series can be identified as a (non-ergodic) non-Poisson renewal (NPR) process, reflecting strong
deviation from exponential decay. A brief account of CME will suffice at this point since a
comprehensive overview of the underlying principle of CME is available in West et al, (2008).
CME is concerned with the conditions under which one complex network responds to a
perturbation by a second complex network: Consider a NPR network with a power law index <
2 as measure of its complexity, and apply a random signal as perturbation: this is in essence
comparable to the condition of aperiodic Statistic Resonance (Gammaitoni et al, 1998).
Allegrini et al (2006a,b) then generalized the conditions by applying as perturbation another
complex network which also satisfies the NPR condition with power index < 2. Under these
conditions, it can be shown that the effect of the perturbation is maximal if the power law
indices of the interacting systems are equal. The claim is that CME, as illustrated in the
foregoing, applies to a large class of NPRs such as, for instance, return times for random walks,
either in regular lattices or in complex networks.
5.2: Linking actions across many scales ?
Even if Nature’s conservative adherence to fractals is a valid argument in support of
their functional significance, we are still in the dark as to what that function may be.
Emphasizing the feature of self-similarity, we can turn to types of functions which could benefit
from stacking extended ranges of space and/or time scales into one compact format; moreover,
there is no privileged time scale in power law dynamics. Sensory adaptation as a change over
time in the responsiveness of the sensory system to a constant stimulus is a situation of this
kind (Wark et al., 2007). Adaptation with power law dependence and multiple time scales has
been demonstrated in nervous systems under many different conditions. Examples come from
such diverse sources as electrosensory afferent nerve fibers in weakly electric fish (Xu et
al,1996), spider mechanoreceptor neurons (French and Torkkeli, 2008). The auditory sensory
memory which is thought to encode stimuli on multiple time scales (Ulanovsky et al., 2004).
Fairhall et al (2001, b) direct attention to the speed with which the dynamics of a neural code is
optimized, even when the statistical properties of the stimuli themselves evolve dynamically
over a wide range of timescales, from tens of milliseconds to minutes. The source of 1/f
fluctuations in human sensorimotor coordination tasks is presumably attributable to the
multiple time scale activities of neural centers (Ding et al, 2002). In visual psychophysics,
adaptation to contrast follows power law dynamics (Rose and Lowe, 1982) as does the tilt
aftereffect (Greenlee and Mangussen, 1987). For fractional order dynamics in adaptation , see:
Lundstrom,( 2008) and Anastasio (2004) in Section 4.
This fragmentary compilation of diverse observations is intended illustrate the
propensity of various neural structures to respond swiftly to a range of temporal and/or spatial
parameters: the image of a set of strings resonating to specific frequencies comes to mind.
Could self-similar fractal structures, having stored a repertoire of responses, each specific for a
particular range of temporal and/or spatial stimulus parameters, fulfil this function ? On
formal grounds, Thorson and Biederman-Thorson (1974) attributed sensory adaptation to
distributed relaxation processes, based on nonuniformities of local “efficacy” in the
transduction process at peripheral receptors. Might this “local nonuniformity” be the
expression of a self-similar functionality of receptors ?
Generically, models of adaptation integrate the response of a system and feed the
integrated signal back to curtail that response. The type of adaptation is determined by the
properties of the integrator. Applying this principle, Drew and Abbott (2006) examined a form
of power law integration with the result that the suppressive effect of repeated stimuli on
successive responses are accumulated with power law decline. In distinction from other forms
of integration (e.g.: exponential), power law integration has the notable feature of scale-
invariance and, in their simulations, replicated the published data of Xu et al (1996), cited in the
previous paragraph. On the same principle, these investigators also implemented power law
adaption within a standard spiking neuron model, except for approximating the time
dependence of the power-law adaptation integral for computational efficiency by a series
exponentials. This principle has been successfully applied by Hausdorff and Peng ( 1996) and is
the basis of Anderson’s (2001) assertion of the power law being an emergent function : it
amounts to linking the noninteracting exponential processes to a cascade which reproduces
the power law forgetting in power law integration. The adaptation obtained of the integrate-
and-fire model with adaptation current obtained by a cascade of exponential processes
matched perfectly that obtained by injecting adaptation currents to the model neurons.
The implications of this successful approach are considerable: the Dell and Abbott
results suggest that power law adaptation can be instantiated by a cascade of a large number of
processes with ordinary exponential dynamics , covering a wide range of time constants. The
cascaded model design lets the temporal stimulus dynamics set the appropriate adaptation
dynamics, in virtue of the numerous exponential processes with different time scales. They
present a telling argument in support of the biological significance of this mechanism of
adaptation: natural stimuli vary unpredictably over a wide of time scales; instead keeping the
recovery time after excitation constant, power-law adaptation allows the temporal statistics
of the stimuli themselves to determine the dynamics of adaptation. The work of Toib et al (
1998) and Gilboa et al ( 2005) referred to in Section 2.2.1 are additional examples of ‘multiscale
computing’ involving fractals.
Are we prepared to envision a general principle of self -organizing control structures for
multiscale behavior, extracting the statistics of an unpredictable environment by way of power
law integration ? The papers cited in the next paragraph speak to this question.
Fusi et al (2005) applied the principle of power-law forgetting of adaptation to a
cascade model for regulating the plasticity of synapses as the basis of stored memories. Note
that memory strength is here represented as synaptic plasticity, not as synaptic strength, as is
commonly the case. Each synapse model has two levels of synaptic strength, weak and strong.
Associated with each strength is a cascade of states (in one of their models, five). The cascades
introduce a range of probabilities for transition between weak and strong levels of synaptic
plasticity. A complicated heuristics is built into the model design that enables combinations of
states of plasticity, for instance: states with low probability are paired up with labile states, and
conversely, etc. The important point is that a high level of memory storage with long retention
times significantly outperforms other model designs. In a model of learning visuomotor
associations that are reversed unpredictably from time to time, synaptic modification occurring
on multiple time scales along the same principles can be the basis for flexible behavior; the
model predictions were validated with experimental data. (Fusi et al, 2007).
It appears then, that the principle of shifting between different scales on demand
manifests itself in many different forms, and on different levels of neural organization. Under
natural conditions, are the required exponential functions for power law integration models of
self-similar structures ?
6. Summary and final thoughts
The assembly of largely phenomenological data was presented to support the claim
that fractal processes and properties occur at many and diverse levels of neural organizations
and performance, and are functionally relevant. Several issues that must not glossed over
lightly needed discussion in several places : e.g. fractality as such is not an obligatory indicator
of SOC ; whether non-conservative systems may be limited to a state of ‘quasi-criticality
instead of being candidates for full criticality (Section 2.2); and what the origin of natural
fractals may be (Section 4). However, the close relation of fractals with critical phase
transitions is beyond dispute. Despite the evidence secured by Plenz and Chialvo in specially
targeted experiments, critical dynamics is still by some investigators called in question at the
mesoscopic level, but the evidence for its importance and essential role at the macroscopic
domain is uncontroversial and solid.
Among the virtues of brain criticality discussed in Section 2.4.3., there is one that has
attracted attention for the longest time: it is the ability for rapid changes of state. Phase
transition in critical systems provide an universal mechanism for rapid switching between
different cooperative neuron collectives. Less attention received the fact , well established in
Physics, that the system presents at the critical transition qualitatively new properties,
requiring new macroscopic descriptors. The important feature of the organization following
the phase transition is to form new objects with distinct properties: a new Ontology. In Physics,
transitions between ferro- and paramagnetism are the prototypical example. We need to ask:
What, if any, is the manifestation of ontological novelty in brain phase transition ? Customarily,
one tends to think in terms of integration and differentiation, brought about by the change in
correlation patterns of the system components, with nearby clusters functioning in synchrony.
However, this way of looking at the nature of the state change fails to meet the requirement
for qualitative (ontological) novelty, as the analogy with Physics would require. In the
paradigmatic cases in Physics, new aggregations are being formed on phase transition which
display novel properties at a macroscopic level of description. What can we assume to happen
in brain on phase transition ? The obvious answer is: partitioning into neural assemblies with
fractal properties. This is of course merely analogous to the fractal patterns formed at phase
transition in physical systems. But here is an essential difference between the systems studied
in Physics and the brain: Take the ferro-paramagnetic phase transition again as example, it
consist merely in a change of spin orientation of the elementary components. But in the case of
brain, the elementary components are reactive neurons which have the potential of entering
into aggregations with functional interactions. Model studies of Arcangelis and Herrman (2010)
with self-organizing neuronal networks show that avalanches formed at phase transition can
learn complex rules on the basis of a collective process. Under appropriate conditions, the
learning dynamics is universal inasmuch as even complex rules can be acquired. Recalling
Section 3.1, the relation of the state dynamics of these neural networks to Integrated Function
Systems and fractal attractors would constitute a notable case of the synergism of fractal and
complex system dynamics .
Re-visiting the work of Beggs and Plenz (Section 2.3) and putting it bluntly: phase
transition endows avalanches with novel capacities that are not analytically predictable from
the original state of the system. Generalizing from this, it is evident that pursuing these and
related directions in targeted studies of the aftermath of brain phase transitions is imperative
for seeking to gain a full appreciation of the functional novelties it may create. Whether such
studies will show a way of bridging the ultimate barrier of the epistemic cut (Pattee, 2001) that
separates the domain of integrable Physics from the domain of symbolic structures remains to
be seen: but this issue is, I submit, at the core of the apparent object-subject (brain-mind)
duality.
The limited sample of recent publication on network models cited in Section 2.4.1 is
but a fraction of the large number of combinations and permutations of network design
parameters that are conceivable: on the one hand, ranging from the of small -world to scale
free works with various degree distributions, on other hand each type being hierarchical,
modular with and without hubs, and fractal. Which one from among this collection is closest
to the natural brain networks? Suggestions abound, but tantalizing questions remain still open
to further inquiry. An architecture that does justice to self-similarity under renormalization
appears attractive as it might unite the dynamics of network topology in one common
mechanism, as I suggest in that Section.
Commenting on the wealth of existing data on anatomical and functional cortical
networks organization may seem like “carrying vocals to Newcastle”, notably in view of the
comprehensive review prepared by Bullmore and Sporns (2009). Nevertheless, a few general
considerations may be of some relevance. It is a commonplace observation that Nature loves
Hierarchies: in the classical “Architecture of Complexity”, Simon(1962) has given plausible
reasons for this apparent love affair. A few years later, in a discussion of the organization of
complex systems (Simon, 1973), observed that the term “hierarchy” has taken a somewhat
generalized meaning, divorced from its original denotation in human organizations of a vertical
authority structure. He there contended that in application to complex systems , “Hierarchy”
has come to denote a set of Chinese Boxes of a particular kind, usually consisting of their
recursion. But then, these Chinese Box hierarchies come in variants of one form or another of
ordering, e.g. in the partial ordering of trees. There is then the possibility for a variety (and
sometimes ambiguous) senses in which current investigators report their finding in
anatomical, fMRI and model studies when speaking of hierarchical order.
The network and graph-theoretical community has developed its own criteria:
modularity has been introduced a measure of a networks decomposability, for instance into
community structures ( Newman and Girvan, 2004), possibly at different hierarchical levels,
allowing one to “zoom in and out” for finding communities on different levels; this was
discussed in detail by Meunier et al, (l.c.). Mucha et al (2009) seek to address this same
objective in a new framework that encompasses coupling of individual networks via links
connecting nodes of one with nodes of another network at multiple scales, thus allowin g for
some nesting which escapes the Newman-Girvan method, but is accessible by the approach of
Sales-Pardo et al (2007). Keeping these issues in mind is, I suggest, of importance in the
definitive evaluation of many of the reports of fMRI data on cortical functional organization:
while describing modularity, they may not be able to capture nesting among modules, due to
methodological limitations. Yet, it is the nesting of the kind that Simon refers to as “Russian
doll” ( see also the Russian Matryoshka dolls of Agnati ,l.c.), where the pressing question of self-
similarity arises: It has some plausibility in view of Sporns ‘s (2006) suggestive evidence for
cortical connection patterns of fractal and self-similar nature. In different contexts, self-similar
community structures were noted by Guimera et al (2003) in a network of human interactions ,
and nesting of theta- and beta gamma oscillations was noted by Gireesh and Plenz (2008) in
neuronal avalanches of developing cortical layers 2/3.
The foregoing discussion was merely intended to draw attention to some of the
problems involved in attempts to characterize dynamic hierarchies of a system that consists of
multiple levels of organization, having dynamics within and between the entities described at
each of the different levels (Lenaerts et al., 2005). Hierarchical nesting would presumambly
qualify brain dynamics as a dynamic hierarchy. This is consonant with the recent findings of
Honey et al (2009) of indirect functional connectivity, not accounted for by structural
connections, and the variability of resting-state functional connectivity across scanning sessions
and model runs. As regards Criticality, it would then be imperative to examine the properties of
phase transitions in dynamically hierarchic systems. In one recent study of a kinetic Ising model,
it turned out that the universality class is consistent with the two-dimensional equilibrium Ising
model (Rikvold et al., 1999).
Sadly, despite the resourceful and imaginative studies undertaken by many investigators
during the past decade, we are still some distance away from understanding the structural and
functional basis of cortical dynamics; let alone the cortical-subcortical dynamics that, according
to the Global Workspace Hypothesis (Baars,1988) is now generally considered decisive for
cognitive functions.
Notwithstanding this disappointing state of affairs, one more speculations may perhaps
be permitted: recall the amazing discovery from Section 5.2 : there, power-law forgetting of
adaptation and synaptic modification on multiple time scales were discussed as principles for
flexible behavior. Recall also the evidence from Section 2.2.1 for multiscale computing as basis
cooperative fractal channel kinetics. The essential principle in these cases is the capacity of
having available the responsiveness, on demand, to a large range of unpredictable external
contingencies of a certain category, varying in one parameter: e.g. the capacity for selecting the
level of adaptation appropriate to a wide range of temporal patterns and stimulus intensity. In
the cases referred to, it amounted to scale shifting as required, involving fractality and self-
similarity of control structures. If Sporns’s evidence for cortical fractality and self-similarity
holds up, I can now re-visit the claim I made in the Introduction: “is there a natural capacity for
unpacking interactions between different levels of a fractal object or process, responsive to
circumstances and conditions, which eludes us ent irely? If it existed, fractals would surely be a
most extraordinary design principle for operational economy in complex systems.”
To which I now add: if this natural capacity embodies principles of power law integration
and cascading, similar to those identified for adaptation and fractal kinetics, we may come to
comprehend at least some aspects of the extraordinary functionality of the cortex.
7. References
Abbott, L.F., Rohrkemper, R. (2007). A simple growth model constructs critical avalanche networks. Progress in
Brain Research 165:13-19.
Abeles, M. (1991) Corticonics: Neural circuits of the cerebral cortex. Cambridge University Press, Cambridge.
Abry, P., (2003). Scaling and Wavelets: an introductory Walk. Lecture Notes in Physics 621:34-60.
Achard, S., Salvador R., Whicher B., Suckling J., and Bullmore E. (2006). A Resilient, Low -Frequency, Small-
World Human Brain Functional Network with Highly Connected Association Cortical Hubs. J Neurosci, 26(1):63-
72.
Achard, S., Bassett, D. S., Meyer-Lindenberg, A., and Bullmore, E. (2008). Fractal connectivity of long -memory
networks. Physical Review, E 77, 036104.
Agnati, L.F., Santarossa, L., Genedani, S., Canela, E.I., Leo, G., Franco, R., Woofs, A., Lluis, C., Ferre, S., Fuxe, K.
(2004) On the nested hierarchical organization of CNS: basic characteristics of neuronal molecular organization. In:
Cortical Dynamics LNCS 3146, edit. P. Erdi, pp. 24-54, Springer Berlin.
Agnati, L.F., Baluska, F., Barlow, P.W., Guidolin,D. (2009) Mosaic, self -similarity Logic, and Biological
Attraction. Commun. Integr. Biol. 2(6):552-563.
Aguirre, J., Viana, R.L., Sanjuan, M.A.F, (2009). Fractal structures in nonlinear dynamics. Reviews of Modern
Physics, 81:333-386.FF
Albert, R., Barabasi, A. (2002). Statistical Mechanics of complex Networks. Reviews of Modern Physics 74:47-97.
Allefeld, C., Atmanspacher, H., and Wackermann, J. (2009). Mental states as macrostates emerging from brain
electrical dynamics. Chaos, 19, 015102.
Allegrini, P., Bellazzini, J., Bramanti, G., Ignaccolo, M., Grigolini, P., and Yang., J. (2002). Scaling breakdown: A
signature of aging. Physical Review, E 66, 015101
Allegrini, P., Bologna, M., Grigolini, P., and West, B. J. (2006a). Response of Complex Systems to Complex
Perturbations: the Complexity Matching Effect. arXiv:cond-mat/0612303v1
Allegrini, P., Bologna, M., Grigolini, P., and Lukovic, M. (2006b). Response of Complex Systems to Complex
Perturbations: Complexity Matching. arXiv:cond -mat/0008341v1.
Allegrini, P., Menicucci, D., Bedini, R., Fronzoni, L., Gemignani, A., Grigolini, P., West, B. J., and Paradisi, P.
(2008). Spontaneous brain activity as a source of perfect 1/f noise.
Altmann, E. G., Pierrehumbert, J. B., and Motter, A. E. (2009). Beyond word frequency: Bursts, lulls, and scaling in
the temporal distributions of words. arXiv:0901.2349v1
Alvarez-Lacalle, E., Dorow, B., Eckmann, J.-P., and Moses, E. (2006). Hierarchical structures induce long -range
dynamical correlations in written texts. Proc.Natl.Acad.Sci USA, 103(21):7956-7961.
Anastasio, T.J. (1994). The fractional-order dynamics of brainstem vestiblo -oculomotor neurons. Biol. Cybern.,
72:69-79.
Andersen, C.M. (2000). From Molecules to Mind: how vertically convergent fractal time fluctuations unify
cognition and emotion. Consciousness & Emotion 1(2):193-226
Andersen, C.M., Lowen, S.B., and Renshaw, P.F. (2006). Emotional task -dependent low-frequency fluctuations and
methylphenidate: Wavelet scaling analysis of 1/f-type fluctuations in fMRI of the cerebellar vermis. J. Neurosci.
Methods, 151:52-61.
Anderson, R.B. (2001). The power law as an emergent property. Memory & Cognition 29(7): 1061-1068.
Anteneodo, C., Chialvo, D.R. (2009) Unraveling the fluctuations of animal motor activity. Chaos 19:1
Antoniou, N.G., Contoyiannis, Y. F., and Diakonos, F.K. (2000). Fractal geometry of critical systems. Physical
Review E, 62, 3.
Arenas, A., Diaz-Guilera, A., Kurths, J., Moreno, Y., and Zhou, C. (2008). Synchronization in complex networks.
Physics Reports, 469:93-153.
Arle, J. E., and Simon, R. H. (1990). An application of fractal dimension to the detection of transients in the
electroencephalogram. Electroencephalography and clinical Neurophysiology , 75:296-305.
Ashkenazy, Y., Hausdorff, J.M., Ivanov, P.Ch., Stanley, E. (2002). A stochastic model of human gait dynamics.
Physica A 316:662-670.
Baars, B.J., (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Babloyantz, A. (1986). Evidence of chaotic dynamics of brain activity during the sleep cycle. In: Dimensions and
Entropies in Chaotic Systems - quantification of complex behavior. ed. G. Mayer-Kress, Springer. pp. 114-122
Baddeley, R., Abbott, L.F., Booth, M. C. A., Sengpiel, F., Freeman, T., Wakeman, E.A., and Rolls, E. T. (1997).
Responses of neurons in primary and inferior temporal visual cortices to natural scenes. Proc. R. Soc. Lond. B,
264:1775-1783.
Bak,P. (1996). How Nature works: the Science of self-organized criticality. Springer, New York
Bak, P., Tang, C., and Wiesenfeld, K. (1987). Self-Organized Criticality: An Explanation of 1/f Noise. Physical
Review Letters, 59(4):381-384.
Bak, P., and Paczuski, M. (1995). Complexity, contingency, and criticality. Proc. Natl. Acad. Sci. USA, 92:6689-
6696.
Baliki, M.N., Geha, P.Y., Apkarian, A.V., Chialvo, D.R. (2008). Beyond feeling: chronic pain hurts the brain,
disrupting the default-mode network dynamics. J. Neuroscience 28(6):1398-1403.
Barabasi, A-L. (2005), The Origin of bursts and heavy tails in human dynamics. Nature 435:207-211.
Barnes, A., Bullmore, E.T., and Suckling, J. (2009). Endogenous Human Brain Dynamics Recover Slowly
Following Cognitive Effort. PLoS ONE, 4(8), e6626.
Barnsley, M.F. (1993). Fractals everywhere. Academic Press, N.Y.
Barnsley, M.F., and Demko, S. (1985). Iterated function systems and the global construction of fractals. Proc. R.
Soc. Lond. A, 399:243-275.
Barnsley, M.F., Elton, J.H., Hardin, D.P. (1989). Recurrent iterated function systems. Constr.Approx. 5:3-31.
Bartumeus, F. (2007) Levy processes in animal movement: an evolutionary hypothesis. Fractals 15(2): 151-162.
Bassett, D.S., Meyer-Lindenberg, A., Achard, S., Duke, T., and Bullmore, E. (2006). Adaptive reconfiguration of
fractal small-world human brain functional networks. PNAS, 103(51):19518-19523.
BassingwhiteJ.B., Liebovitch, L.S., West, B.J. (1994). Fractal Physiology, Oxford University Press, New York
Bedard, C., Kroeger, H., Destexhe, A. (2006a). Model of low-pass filter of local field potentials in brain tissue.
Physical Review E 73: 051911
Bedard,C., Kroeger H., Destexhe, A. (2006b). Does the 1/f frequency scaling of brain signals reflect self-organized
critical states? Physical Review Letters 97:118102.
Beggs, J.M. (2007) How to build a critical mind. Nature Physics 3:834-835.
Beggs, J.M. (2008) The criticality hypothesis: how local cortical networks might optimize information processing.
Phil. Trans. R.Soc. A 366:329-343.
Beggs, J. M., and Plenz, D. (2003). Neuronal Avalanches in Neocortical Circuits. J Neurosci, 23(35):11167-11177.
Beggs, J. M., and Plenz, D. (2004). Neuronal Avalanches Are Diverse and Precise Activity Patterns That Are Stable
for Many Hours in Cortical Slice Cultures. J Neurosci, 24(22):5216-5229.
Bel, G., and Barkai, E. (2005). Weak Ergodicity Breaking in the Continuous -Time Random Walk. Physical Review
Letters, 94, 240602.
Beran, J. (1994). Statistics for Long-Memory Processes. Chapman & Hall.
Bernard, F., Bossu, J-L., and Gaillard, S. (2001). Identification of living oligodendrocyte developmental stages by
fractal analysis of cell morphology. J. Neurosci. Res., 65(5):439-445.
Bhattacharya, J., and Petsche, H. (2001). Universality in the brain while listening to music. Proc. R. Soc. Lond. B,
268:2423-2433.
Bhattacharya, J., Edwards, J., Mamelak, A.N., Schuman, E.M. (2005). Long -range temporal correlations in the
spontaneous spiking of neurons in the Hippocampal-Amygdala complex of humans. Neuroscience 131:547-555.
Bieberich, E. (2002). Recurrent fractal neural networks: a strategy for the exchange of local and global information
processing in the brain. BioSystems, 66:145-164.
Bianco, S., Grigolini, P., and Paradisi, P. (2005). Fluorescence intermittency in blinking quantum dots: Renewal or
slow modulation? J. Chem. Physics, 123, 174704.
Bianco, S., Ignaccolo, M., Rider, M. S., Ross, M. J., Winsor, P., and Grigolini, P. (2007). Brain, music, and non -
Poisson renewal process. Physical Review E, 75, 061911.
Blondel, V.D., Guillame, J-L, Lambiotte, R., Lefebvre, E. (2008) Fast unfold ing of communities in large networks.
J. Stat.Mech.Theory E 1P10008
Bloom, J. L., (1969). Quantitative Aspects of neural activity in the cerebral cortex of the rabbit. Thesis, Amsterdam
University, Netherlands.
Boettcher, S., Percus, A. (2000) Nature‟s way of optimizing. Artificial Intelligence 119:275-286.
Bonachella, J.A., Munoz, M.A. (2009) Self-organization without conservation: true or just apparent scale invariance
? J.Stat.Mech. 1-37.
Boon, J. P., and Decroly, O. (1995). Dynamical systems theory for music dynamics. Chaos, 5(3):501-508.
Bornholdt, S., and Ebel, H. (2001). World Wide Web scaling exponent from Simon‟s 1955 model. Physical Review
E, 64, 0351004.
Breskin, I., Soriano, J., Moses, E., and Tlusty, T. (2006). Percolation in Living Neural Networks. Physical Review
Letters, 97, 188102.
Bressloff, P.C., Stark J. (1992). Analysis of associative reinforcement learning in neural netwo rks using iterated
function systems. IEEE Trans.Man,Cybern. 22:1348-1360.
Buiatti, M., Papo, D., Baudonniere, P.-M., and Vreeswijk, C. V. (2007). Feedback Modulates the Temporal Scale -
Free Dynamics of Brain Electrical Activity in a Hypothesis Testing Task. Neuroscience, 146:1400-1412.
Buice, M.A., Cowan, J.D. (2007). Field-theoretic approach to fluctuation effects in neural networks. Physical
Review E 75:051919.
Bullmore, E., Long, C., Suckling, J., Fadili, J., Calvert, G., Zelaya, F., Carpenter, T. A., and Brammer, M. (2001).
Colored Noise and Computational Inference in Neurophysiological (fMRI) Time Series Analysis: Resampling
Methods in Time and Wavelet Domains. Human Brain Mapping, 12:61-78.
Bullmore, E., Fadili, J., Maxim, V., Sendur, L., Whitcher, B., Suckling, J., Brammer, M., and Breakspear, M.
(2004). Wavelets and functional magnetic resonance imaging of the human brain. NeuroImage, 23:S234-S249.
Callaway, E.M. (2002). Cell type specificity of local cortical connections. J. Neurocytology 31:231 -237.
Capurro, A., Diambra, L., Lorenzo, D., Macadar, O., Martin, M. T., Mostaccio, C., Plastino, A., Rofman, E., Torres,
M. E., and Velluti, J. (1998). Tsallis entropy and cortical dynamics: the analysis of EEG signals. Physica A,
257:149-155.
Capurro, A., Diambra, L., Lorenzo, D., Macadar, O., Martin, M. T., Mostaccio, C., Plastino, A., Perez, J., Rofman,
E., Torres, M. E., and Velluti, J. (1999). Human brain dynamics: the analysis of EEG signals with Tsallis
information measure. Physica A, 265:235-254.
Caserta, F., Stanley, H. E., Eldred, W. D., Daccord, G., Hausman, R. E., and Nittman, J. (1990). Physical
Mechanisms Underlying Neurite Outgrowth: A Quantitative Analysis of Neuronal Shape. Physical Review Letters,
64(1):95-98.
Caserta, F., Eldred, W. D., Fernandez, E., Hausman, R. E., Stanford, L. R., Bulderev, S. V., Schwarzer, S., and
Stanley, H. E. (1995). Determination of fractal dimension of physiologically chara cterized neurons in two and three
dimensions. J. Neurosci. Methods, 56:133-144.
Cateau, H., and Reyes, A. D. (2006). Relation between Single Neuron and Population Spiking Statistics and Effects
on Network Activity. Physical Review Letters, 96, 058101.
Catutto, C., Loreto, V., and Servedio, V. D. P. (2006). A Yule -Simon process with memory. arXiv:cond-
mat/0608672v1.
Cessac, B. (2004). Some fractal aspects of Self-Organized Criticality. arXiv:nlin/0409004v1.
Changizi, M. A. (2001 a). Principles underlying mammalian neocortical scaling. Biol. Cybern., 84:207-215.
Changizi, M.A. (2001 b) Universal scaling laws for hierarchical complexity in languages, organisms, behaviors and
othjer combinatorial systems. J. Theoret.Biol. 211:277-295.
Changizi, M.A. (2003) The brain from 25,000 feet. Kluver Publisher, Dordrecht.
Chen, D., Wu, S., Guo, A., Yang, ZR. (1995). Self-organized criticality in a cellular automaton model of pulse -
coupled integrate-and-fire neurons. J. Physics A 28:5177-5182.
Chen, Y., Ding, M., and Kelso, J. A. S. (1997). Long Memory Processes (1/fα Type) in Human Coordination.
Physical Review Letters, 79(22):4501-4504.
Chialvo, D. R. (2004). Critical brain networks. Physica A, 340:756-765.
Chialvo, D. R. (2006). Are our senses critical? Nature Physics, 2:301-302.
Chialvo, D.R., Balenzuela, P., Fraiman, D. (2008). The brain: what is critical about it? Conf.Proc. American Institute
of Physics 1028:28-45.
Clauset, A., Shalizi, C. R., and Newman, M. E. J. (2009). Power -Law Distributions In Empirical Data.
arXiv:0706.1062v2.
Cluff, T., and Balasubramaniam, R. (2009). Motor Learning Characterized by Changing Levy Distributions. PLoS
ONE, 4(6), e5998.
Collins, J.J., De Luca, C.J. (1994) Random walking during quiet standing. Physical Review Letters 73(5):764-767.
Copelli, M., Roque, A. C., Oliviera, R. F., and Kinouchi, O. (2002). Physics of psychophysics: Stevens and Weber -
Fechner laws and transfer functions of excitable media. Physical Review E, 65, 060901.
Corner, M.A., van Pelt, J., Wolters P.S., Baker, R.E.,Nuytinck, R.H. (2002). Physiological effects of sustained
blockade of excitatory tynsotic 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.
Correll, J. (2008). 1/f Noise and Effort on Implicit Measures of Bias. J of Personality and Social Psych, 94(1):48-59.
Cosenza, M. G., and Kapral, R. (1992). Coupled maps on fractal lattices. Physical Review A., 46(4):1850-1858.
Costa, M. E., Sigman, M., and Bonomo, F. (2009). Scale -invariant transition probabilities in free word association
trajectories.
Da Silva, L., Papa, A. R. R., and de Souza, A. M. C. (1998). Criticality in a simple model for brain functioni ng.
Physics Letters A, 242:343-348.
Davidsen, L., Schuster, H.G. (2002). Simple model for 1/f noise. Physical Review E., 65:026120.
De Arcangelis, L., Perrone-Capano, C., Herrmann, H.J. (2006). Self-organized criticality for brain plasticity.
Physical Review Letters 96:028107,
De Arcangelis, L., Herrmann, H.J. (2010). Learning as a phenomenon occurring in a critical state.
arXiv:1003.1200v1 [q-biol.NC]
de Carvalho, J.X., Prado C.P.C. (2000) Self-organized criticality in the Olami-Feder-Christensen model. Physical
Review Letters 84:4006-4009.
DelCastillo, J., Katz, B. (1954). Quantal components of the end plate potential. J. Physiol. (London) 124:560-573.
Delignieres, D., Ramdani, S., Lemoine, L., Torre, K., Fortes, M., and Ninot, G. (2006). Fractal analyses for „short‟
time series: A re-assessment of classical methods. J of Mathematical Psych, 50:525-544.
De Los Rios, P., Zhang, Yi-Chen. (1999). Universal 1/f noise from dissipative self-organized criticality models.
Physical Review Letters 82 (3):472-475.
Destexhe, A., Contreras D., Steriade M. (1999). Spatiotemporal analysis of local field potentials and unit discharges
in cat cerebral cortex natural wake and sleep states. J. Neuroscience 19 (11):4595-4608.
Dewey, T.G. (1999). Fractals in Moleculal Biophysics. Oxford University Press, NY.
Ding, M., Chen, Y., and Kelso, J. A. S. (2001). Statistical Analysis of Timing Errors. Brain and Cognition, 48:98-
106.
Dorogovtsev,S., Mendes, J. (2003). Evolution of Networks: From Biological Nets to the Internet and WWW.
Cambridge University Press, N.Y.
Drew, P. J., and Abbott, L. F. (2006). Models and Properties of Power -Law Adaptation in Neural Systems. J
Neurophysiol, 96:826-833.
Duff, E. P., Johnston, L. A., Xiong, J., Fox, P. T., Mareels, I., and Egan, G. F. (2008). The Power of Spectral
Density Analysis for Mapping Endogenous BOLD Signal Fluctuations. Human Brain Mapping, 29:778-790.
Eguiluz, V. M., Chialvo, D. R., Cecchi, G. A, Baliki, M., and Apkarian, A. V. (2005). Scale-Free Brain Functional
Networks. Physical Review Letters, 94, 018102
Eke, A., Herman, P., Kocsis, L., and Kozak, L. R. (2002). Fractal characterization of complexity in temporal
physiological signals. Physiol. Meas., 23:R1-R38.
El Boustani, S., Destexhe, A. (2009) Does brain activity tem from high -dimensional chaotic dynamics ? Evidence
from the human electroencephalogram, cat cerebral cortex and artificial neuronal networks. arXiv:0904:4217v1
[nlin.CD]
El Boustani, S. E., Marre, O., Behuret, S., Baudot, P., Yger, P., Bal, T., Destexhe, A., and Fregnac, Y. (2009).
Network-State Modulation of Power-Law Frequency-Scaling in Visual Cortical Neurons. PLoS Computational
Biology, 5(9), e1000519.
Elman, J.L. (1995). Language as a dynamical system. In: Mind as Motion, Port, R.F. and Gelder,T., edits., MIT
Press, Cambridge, MA.
Erland, S., and Greenwood, P. E. (2007). Constructing 1/ω α noise from reversible Markov chains. Physical Reviews
E, 76, 031114.
Essam, J.W. (1980). Percolation Theory. Rep.Progr.Phys. 43:834-912.
Eurich, C.W., Herrmann, J.M., Ernst, U.A. (2002). Finite size effects of avalanche dynamics. Physical Reviews E
66:066137
Evarts, E. V. (1967). Unit activity in Sleep and Wakefulness. In:Neurosciences, eds. GC Quarton, T. Melnechuk,
F.O. Schmitt, Neurosciences Research program, pp.545-556.
Expert, P., Lambiotte, R., Chialvo, D.R., Christensen, K., Jensen, H.J., Sharp, D.J., Turkheimer, F. (2010) Self -
similar correlation function in brain resting state fMRI. arXiv:1003.3682v1 [q-biol.NC]
Fairhall, A. L., Lewen, G. D., Bialek, W., van Steveninck, R. R. (2001a). Multiple timescales of adaptation in a
neural code. Adv. Neural Inform. Proc. Syst. 13:124-130.
Fairhall,A.L.,Lewen, G.D., Bialek, W., de Ruyter, van Steveninck, R.R. (2001,b) Efficiency and ambiguity in an
adaptive neural code. Nature 412(6849):776-777.
Feinberg, T.E., Venneri, A., Simone A.M., Fan,Y., Northoff, G. (2010). The neuroanatomy of asomatognosia and
somatoparaphrenia. J. Neurol. Neursurg. Psychatry 81:276 -281.
Feng, J., Zhang, P. (2001). Behavior of integrate -and-fire and Hodgkin-Huxley models with correlated inputs.
Phys.Rev. E, 63: 051902.
Fisher, M. E. (1998). Renormalization group theory: Its basis and formulation in statistical physics. Reviews of
Modern Physics, 70:653-681.
Fodor, J. (1975). Language of Thought, Harvard U.P. , Cambridge, MA.
Fox,M.D., Raichle,M.E. (2007) Spontaneous fluctuations in brain activity observed with functional magnetic
imaging. Nat. Rev.Neurosci. 8:701-711
Fraiman, D., Balenzuela, P., Foss, J., and Chialvo, D. R. (2009). Ising -like dynamics in large-scale functional brain
networks. Physical Review E, 79, 061922..
Freeman, W.J. (2005). A field-theoretic approach to understanding scale-free neocortical dynamics. Biol.Cybern.
92:350-359.
Freeman, W. J., Holmes, M. D., Burke, B. C., and Vanhatalo, S. (2003). Spatial spectra of scalp EEG and EMG
from awake humans. Clinical Neurophys, 114:1053-1068.
French, A. S., and Torkkeli, P. H. (2008). The Power Law of Sensory Adaptation: Simulation by a Model of
Excitability in Spider Mechanoreceptor Neurons. Annals of Biomedical Engineering, 36(1):153-161.
Fusi, S., Drew, P. J., and Abbott, L. F. (2005). Cascade Models of Synaptically Stored Memories. Neuron, 45:599-
611.
Fusi, S., Asaad, W.F., Miller, E.K., Wang, X-J. (2007). A neural circuit model of flexible sensori-motor leraning and
forgetting on multiple time scales. Neuron 54:319-333.
Gallos, L.K., Song, C., Havlin, S., Makse, H.A. (2007), Scaling theory of transport in complex biological networks.
Proc.Nat.Acad.Sci. USA 104(19):7746-7751.
Gammaitoni, L., Hanggi, P., Jung, P., Marchesoni, F. (223). Stochastic resonance. Review of Modern Physics,
70(1):223-287.
Gefen, Y., Mandelbrot, B. B., and Aharony, A. (1980). Critical Phenomena on Fractal Lattices. Physical Review
Letters, 45(11):855-858.
Gerstein, G. L., and Mandelbrot, B. (1964). Random Walk models for the Spike Activity of a Single Neuron.
Biophysical Journal, 4:41-68.
Chialvo, D.R. (2004). Critical Brain Networks. Physica A 340:756-765.
Gilboa, G., Chen, R., and Brenner, N. (2005). History-Dependent Multiple-Time-Scale Dynamics in a Single-
Neuron Model. J Neurosci, 25(28):6479-6489.
Gilden, D.J. (1997) Fluctuations in the time required for elementary decisions. Psychological Science 8(4):296-301.
Gilden, D. L. (2001). Cognitive Emissions of 1/f Noise. Psychological Review, 108(1):33-56.
Gireesh, E.D., Plenz, D. (2008). Neuronal avalanches organize as nested theta - and beta/gamma oscillations during
development of cortical layer 2/3. Proc.Nat.Acad.Sci. USA 105:7576-7581.
Gisiger, T. (2001). Scale invariance in biology: coincidence or footp rint of a universal mechanism? Biol. Rev.,
76:161-209.
Giugliano,M., Darbon, P., Arsiero, M.,Luescher H.R., Streit, J. (2004). Single neuron discharge properties and
network activity in dissociated cultures of neocortex. J. Neurophysiology 92:977-996.
Gong, P., Nikolaev, A. R., Leeuwen, C. v. (2002). Scale -invariant fluctuations of the dynamical synchronization in
human brain electrical activity. Neurosci Letters, 336:33-36.
Goychuk, I., and Hanggi, P. (2002). Ion channel gating: A first -passage time analysis of the Kramers type.
Proc.Natl.Acad.Sci. USA, 99(6):3552-3556.
Griffin, L., West, DJ., West, BJ. (2000). Random stride Intervals with memory. J. Biol. Physics 26: 185-2000.
Grigolini, P., Aquino,G., Bologna, M., Lukovic, M., West, B.J. (2009). A theory of 1/f noise in human cognition.
Physics A 388:4192-4204.
Gruneis, F., Nakao, M., Mizutani, Y., Yamamoto, M, Meesmann M, Musha, T, (1993). Furher study on 1/f
fluctuations observed in central single neurons during REM sleep. Biol.Cybern. 68:193-198.
Gruneis, F. (2001). 1/f Noise, Intermittency and Clustering Poisson Process. Fluctuation and Noise Letters,
1(2):R119-R130.
Hagmann,P., Cammoun,L., Gigandet, X., Meuli,R., Honey,C., Van Wedeen,J., Spornsd,O. (2008). Mapping the
structural core of human cerebral cortex. PLoS Biology 6(7): e159.
Haken, H. (1983). Synergetics: an Introduction : nonequilibrium phase transitions and self -organization in Physics,
Chemistry and Biology. Springer, New York.
Haken, H., Kelso,J.A., Bunz, H. (1985) A theoretical model of phase transitions in human hand movements. Biol.
Cybern. 51: 347-356.
Harris, K.D. (2005) Neural signatures of cell assembly organization. Nature Revs.Neurosci 6:399-407.
Harris, T.E. (1989),The theory of branching processes. Dover, N.Y.
Harrison, K. H., Hof, P. R., and Wang, S. S.-H. (2002). Scaling laws in the mammalian neocortex: Does form
provide clues to function? J Neurocytology, 31:289-298.
Hausdorff, J.M., Peng, C.K., Ladin, Z., Wei, J.Y., Goldberger, A.L. (1995). Is walking a random event ? Evidence
for long-range c0rrelations in stride interval of human gait. J.Appl.Physiol. 78:349-358.
Hausdorff, J.M., Peng, C.K. (1996) Multiscaled randomness: a possible source of 1/f noise in Biology. Phys.Rev.E
54:2154-2157
Hennig, M.H., Adams, Ch., Willshaw D., Sernagor, E. (2009). Early-stage waves in the retinal network emerge
close to a critical phase transition between local and global functional connectivity. J. of Neuroscience 29(4):1077-
1086.
Henry, B.I., Wearne, S.L. (2000). Fractional reaction-diffusion. Physica A 276:448-455
Herz,A.V., Hopfield J.J. (1995). Earthquake cycles and neural reverberations: collective oscillations in systems with
pulse coupled threshold oscillators. Physical Review Letters 75: 1222-1225.
Hilgetag, C.C., Koetter, R., Stephan, K.E., Sporns, O.(2002). Computational methods for the Analysis of Brain
Connectivity. In: Computational Anatomy, edit. G.A. Ascoli, Humana Press, pp.295 -335.
Hilgetag, C. C., and Kaiser, M. (2004). Clustered Organization of Cortical Connectivity. Neuroinformatics, 2:353-
360.
Honey, C. J., Kotter, R., Breakspear, M., and Sporns, O. (2007). Network structure of the cerebral cortex shapes
functional connectivity on multiple time scales. PNAS, 104(24):10240-10245.
Honey, C.J., Sporns, O., Cammoun,L., Thiran, J.P., Meuli, R., Hagmann, P. (2009) Predicting human resting state
functional connectivity from structural connectivity. Proc.Nat.Acad.Sci. USA 106(6):2035-2040.
Hsu, K. J., and Hsu, A. (1991). Self-similarity of the “1/f noise” called music. Proc. Natl. Acad. Sci. USA, 88:3507-
3509.
Hughes, B. D., Montroll, E. W., and Shlesinger, M. F. (1982). Fractal Random Walks. J Statistical Phys, 28(1):111-
126.
Huett, M-T., Lesne, A. (2009). Interplay between topology and dynamics in excitation pattersn on hierarchical
graphs. NeuroInformatics 3:1-10.
Huxley, J.S. (1932). Problems of relative growth. The Dial Press, New York.
Iannacone, P.M., Khokha, M. (1995). Fractal Geometry in Biological Systems. CRC Press, Boca Raton.
Im, K., Lee, J.M., Kim, S.H., Lyttelton,)., Kim, S.H., Evans,A.C., K im, S.L., (2008) Brain size and cortical
structure. Cerebral Cortex 18:2181-2191.
Jantzen, K.J., Steinberg, F.L., Kelso,J.A.S. (2008). Coordination Dynamics of large scale neural circuitry underlying
rhythmic sensorimotor behavior. J. Cogn. Neurosci. 21(12):2420-2433.
Jeffrey, H.J. (1990). Chaos game representation of gene structure. Nucleic Acid Res. 18(8):2163-2170.
Jelinek, H. F., Elston,N., Zietsch,B. (1995). Fractal Analysis: Pitfalls and Revelations in Neuroscience. In: Fractals
in Biology and Medicine, eds. G.A. Losa, D., Merlini, T.F. Nonnenmacher, E,.R. Weibel. Birkhauser, Basel.
Jelinek, H. F. and Elston, G. N. (2001). Pyramidal Neurones in Macaque Visual Cortex: Interareal Phenotypic
Variation of Dendritic Branching Patterns. Fractals, 9(3):287-295.
Jones, C. L., and Jelinek, H. F. (2001). Wavelet Packet Fractal Analysis of Neuronal Morphology. DOI
10.1006/meth.2001.1205.
Juanico, D.E., Monterola, C., Saloma, C. (2007) Self-organized critical branching in systems that violate
conservation laws. New J. Physics 9:1-17
Kadanoff, L. P. (1990). Scaling and Universality in Statistical Physics. Physica A, 163:1-14.
Kadanoff, L.P., Nagel, S.R., Wu, L., Zhou, S. (1989). Scaling and universality on avalanches. Physical Review A 39
(12):6524-6537.
Kaiser, M., Goerner, M., Hilgetag, C.C. (2007) Criticality and spreading dynamics in hierarchical cluster networks
without inhibition. New J. Physics 9:110-123.
Kaiser, M. (2008). Brain architecture: a design for natural computation. arXiv:0802.4010v1 [q -biol. NC]
Kaiser, M., Hilgetag,C.C. (2010). Optimal hierarchical modular topologies for producing limited sustained
activation of neural networks. Front.Neuroinform. 4:8
Karperien, A. L., and Jelinek, H. F. (2008). Box-Counting Analysis of Microglia Form in Schizophrenia,
Alzheimer‟s Disease, and Affective Disorder. Fractals, 16(2):103-107.
Kass, R. E., and Ventura, V. (2001). A Spike -Train Probability Model. Neural Computation, 13:1713-1720.
Katsaloulis, P.,Verganelakis, D.A. (2009). Fractal Dimension and Lacunarity of Tractography Images of the human
Brain. Fractals 17 (2):181-189.
Kaulakys, B., and Meskauskas, T. (1998). Modeling 1/f noise. Physical Review E, 58(6):7013-7019.
Kaulakys, B., Ruscekas, J., Gontis, V., Alaburda, M. (2006). Nonlinear stochastic models of 1/f noise and power -
law distributions. Physica A, 365:217-221.
Kello, C. T., Beltz, B. C., Holden, J. G., and Van Order, G. C. (2007). The Emergent Coordination of Cognitive
Function. J Experimental Psych, 136(4):551-568.
Kelso, J.A.S. (1984) Phase transitions and critical behavior in human bimanual coordination. American Journal of
Physiology: Regulatory, Integrative and Comparative, 15, R1000-R1004.
Kelso, J.A.S., Bressler, S.L., Buchanan, S., DeGuzman, G.C., Ding, M., Fuchs, A. (1992).A phase transition in human
brain and behavior. Physics Letters A 169:134-144.
Kelso, J.A.S. (1995) Dynamic Patterns: The Self Organization of Brain and Behavior . Cambridge: MIT Press.
Kelso, J.A.S., Tognoli, E. (2007).Toward a complementary neuroscience: metastable coordination dynamics of the
brain. In: Kozma, R., Perlovsky, L.,(Eds). Neurodynamics of Cognition and Consciousness,, Springer, Perlin. Pp.39 -
59.
Kim, J. S., Goh, K. I., Kahng, B., and Kim, D. (2007). Fractality and self-similarity in scale-free networks. New
Journal of Physics, 9, 177.
Kinouchi, O., Prado, C.P.C. (1999) Robustness of scale invariance in mpodels with self -organized criticality.
Phys.Rev.E 59(5):4964-4969.
Kinouchi, O., Copelli, C. (2006). Optimal dynamical range of excitable networks at criticality. Nature Physics
2:348-352.
Kitzbichler, M. G., Smith, M. L., Christensen, S. R., and Bullmore, E. (2009). Broadband Criticality of Human
Brain Network Synchronization. PLoS Computational Biology, 5(3), e1000314.
Kleinz, M., and Osler, T. J. (2000). A Child‟s Garden of Fractional Derivatives. The College Mathematics Journal,
31:82-88.
Kniffki, K.-D., Pawlak, M., and Vahle-Hinz, C. (1993). Scaling Behavior of the Dendritic Branches of Thalamic
Neurons. Fractals, 1(2):171-178.
Kniffki, K.-D., Pawlak, M.,Vahle-Hinz,C. (1994). Fractal Dimension and Dendritic Branching of Neurons in the
soimatosensory Thalamus. In: Fractals in Biology and Medicine, edit.Nonnenmacher T.F. , Losa G., Weibel, E.R.,
Birkhauser Verlag, Basel, pp.221-229.
Kodama,T., Mushiake,H., Shima, K., Nakahama, H., Yamamoto, M. (1989). Slow fluctuations of single unit
activities of hippocampal and thalamic neurons in cats. I. Relation to natural sleep and alert states. Brain Research
487: 26-34.
Kolen, J.F. (1994) Fool‟s Gold: Extracting finet State Machines from recurrent network dynamics. Adsv.Neur .
Inform. Proc. Syst. 6:501-508.
Korn, H., Faure, P. (2003) Is there chaos in the brain ? II. Experimental evidence and related models. C.R.Biol. 326:
787-840.
Koulakov, A.A. (2010). On the scaling law for cortial magnification factor. arXiv:1002.4368v1[q -biol.NC]
Koutsoyiannis, D. (2002). The Hurst phenomenon and fractional Gaussian noise made easy. Hydrological Sciences,
47(4):573-595.
Kozma, R., Puljic, M., Ballister, P., Bollobas, B., Freeman, W. J. (2005).Phase transitions in the neuropercolation
model of neural populations with mixed local and non-local interactions. Biol. Cybern., 92:367-379.
Kuikka, J., Tiihonen, J, (1998). Fractal analysis – a new approach to receptor imaging. Annals of Medicine 30:242-
248.
Kulish, V., Sourin, A., Sourina, O. (2006). Human electroencephalograms seen as fractal time series: Mathematical
analysis and visualization. Computers in Biology and Medicine, 36:291-302.
Kuramoto, Y. (1984). Chemical oscillations, waves and turbulence. Springer, Berlin.
LaBarbera, M. (1989). Analyzing Body Size as a Factor in Ecology and Evolution. Annu. Rev. Ecol. Syst., 20:97-
117.
Laughlin, R.B. (2005). A different Universe:Reinventing Physics from the bottom down. Basic Books, New York.
Lee, J.-M., Hu, J., Gao, J., Crosson, B., Peck, K. K., Wierenga, C. E., McGregor, K., Zhao, Q., and White, K. D.
(2008). Discriminating brain activity from task-related artifacts in functional MRI: Fractal scaling analysis
simulation and application. NeuroImage, 40:197-212.
Lenaerts, T., Chu, D., Watson, R. (2005). Dynamical hierarchies. Artificial Life 11:403-405.
Levina,A., Herrmann, J.M., Geisel, T. (2007). Dynamical synapses causing self-organized criticality in neural
networks. Nature Physics 3:857-860.
Levina,A., Herrmann, J.M., Geisel, T. (2009). Phase transition towards criticality in a neural system with adaptive
interactions. Physical review Letters 102: 118110.
Levy, P., Theorie de l’addition des variables aleatoires. Gauthier-Villars, Paris
Levy, S.D., Pollack J.B. (2002). Logical computation on a fractal neural substrate.
http://www.demo.cs.btandeis.edu/papers/raam-ijcnn01.ps.gz
Lewis, C. D., Gebber, G. L., Larsen, P. D., and Barman, S. M. (2001). Long -Term Correlations in the Spike Trains
of Medullary Sympathetic Neurons. Downloaded from jn.physiology.org.
Liebovitch, L. S., Fischbarg, J., Koniarek, J. P., Todorova, I., and Wang, M. (198 7). Fractal model of ion-channel
kinetics. Biochimica et Biophysica Acta, 896:173-180.
Liebovitch, L. S., Scheurle, D., Rusek, M., and Zochowski, M. (2001). Fractal Methods to Analyze Ion Channel
Kinetics. Methods, 24:359-275.
Linkenkaer-Hansen, K., Nikouline, V. V., Palva, J. M., and Ilmoniemi, R. J. (2001). Long -Range Temporal
Correlations and Scaling Behavior in Human Brain Oscillations. J Neurosci, 21(4):1370-1377.
Linkenkaer-Hansen, K. (2003). Scaling and Criticality in Large-Scale Neuronal Activity. Lecture Notes in Physics,
621:324-338.
Linkenkaer-Hansen, K., Nikulin, V. V., Palva, J. M., Kaila, K., and Ilmoniemi, R. J. (2004). Stimulus -induces
change in long-range temporal correlations and scaling behaviour of sensorimoto r oscillations. European Journal of
Neuroscience, 19:203-211.
Lowen, S.B., Teich, M.C. (1993). Fractal auditory nerve firing patterns may derive from fractal switching in sensory
hair cell ion channels. AIP Conf. Proceedings 285 , eds. P.H. Handel, A.L. Chung, American Institute of Physics. Pp.
745-748.
Lowen, S. B., Cash, S. S., Poo, M.-m., and Teich, M. C. (1997). Quantal Neurotransmitter Secretion Rate Exhibits
Fractal Behavior. J Neurosci, 17(15):5666-5677.
Lowen, S. B., Liebovitch, L. S., and White, J. A. (1999). Fractal ion-channel behavior generates fractal firing
patterns in neuronal models. Physical Review E, 59(5):5970-5980.
Lowen, S. B., Teich, M.C. (2005). Fractal based Point Processes. Wiley, N.Y.
Lundstrom, B. N., Higgs, M.H., Spain, W.J., Fairhall, A. (2008). Fractional differentiation by neoircortical
pyramidal neurons. Nature Neuroscience doi: 10.1038/nn.2212
Maess, B., Koelsch, S., Gunter, T. C., and Friederici, A. D. (2001). Musical syntax is processed in Broca‟s area: an
MEG study. Nature Neuroscience, 4(5):540-545.
Magnussen,S., Greenlee, M.W. (1985). Marathon adaptation to spatial contrast: saturation in sight. Vision Res.
25:1409-1411.
Maimon, G., and Assad, J. A. (2009). Beyond Poisson: Increased Spike -Time Regularity across Primate Parietal
Cortex. Neuron, 62:426-440.
Mandelbrot, B. B., and Ness, J. W. V. (1968). Fractional Brownian Motions, Fractional Noises and Applications.
SIAM Review, 10(4):422-437.
Mandelbrot, B. B.(1977). The fractal Geometry of Nature.
Marro, J., Dickman R. (1999). Nonequilibrium phase transitions in Lattice Models. Cambridge University Press,
Cambridge, UK
Maxim, V., Sendur, L., Fadili, J., Suckling, J., Gould, R., Howard, R., and Bullmore, E. (2005). Fractional Gaussian
noise, functional MRI and Alzheimer‟s disease. NeuroImage, 25:141-158.
Mazzoni, A., Broccard, F. D., Garcia-Perez, E., Bonifazi, P., Ruaro, M. E., and Torre, V. (2007). On the Dynamics
of the Spontaneous Activity in Neuronal Networks. PLoS ONE, 5, e349.
McKenna, T. M., (1992). Single Neuron Computation. Elsevier.
Mayer-Kress, G., Layne, S.P. (1987). Dimensionality of the Human Electroencephalogram. Ann. N.Y. Acad. Sci.
504:62-87.
Meunier, D., Lambiotte, R., Fornito, A., Ersche, K.D., Bullmote, E.T. (2009) Hierarchical modularity in human
brain functional networks. Frontiers in Neuroinformatics 3: Article 37.
Meyer-Lindenberg, A., Ziemann, U., Hajak, G., Cohen, L., and Berman, K. F. (2002). Transition between dynamical
states of differing stability in the human brain. Proc.Natl.Acad.Sci. (USA), 99(17):10948-10953.
Miller, S.L., Miller W.M., McWhorter, P.J. (1993) Extremal Dynamics: a unifying physical explanation of fractals,
1/f noise and activated processes. J.Appl.Phys. 73(6):2617-2628.
Millhauser, G. L., Salpeter, E. E., and Oswald, R. E. (1988). Diffusion models of ion -channel gating and the origin
of power-law distributions from single-channel recording. Proc. Natl. Acad. Sci. USA, 85:1503-1507.
Milosevic, N. T., Ristanovic, D., Stankovic, J. B. (2005). Fractal analysis of laminar organization in spinal cord. J.
of Neuroscience Methods, doi: 10.1016/j.neumeth.2005.02.009 .
Milosevic, N. T., Ristanovic, D., Stankovic, J. B., and Gudovic, R. (2007). Fractal Analysis of Dendritic
Arborisation Patterns of Stalked and Islet Neurons in Substantia Gelatinosa of Different Species. Fractals, 15(1):1-
6.
Millotti, E. (2002). 1/f noise: a pedagogical review. arXiv:physics/0204033v1.
Mitzenmacher, M. (2003). A Brief History of Generative Models for Power Law and Lognormal Distributions.
Internet Mathematics, 1(2):226-251.
Montroll, E. W., Weiss, G. (1965). J. Math. Phys. 6:178.
Montroll, E.W., West, B.J. (1979): An enriched collection of stochastic processes. In Fluctuation Phenomena, eds.
E.W. Montroll, J. Lebowitz, North Holland.
Morariu, V. V., Cosa, A., Chis, M. A., Isvoran, A., Morariu, L. -C. (2001). Scaling in Cognition. Fractals, 9(4):379-
391.
Mucha, P.J., Richardson, T., Macon, K., Porter, M.A. (2009). Community structure in time -dependent, multiscale
and multiplex networks. arXiv:0911.1824v1 [physics.data -an]
Mueller-Linhow,M., Hilgetag, C.C., Huett, M-T. (2008). Organnization of excitable dynamics in hierarchical
biological networks. PLoS Comput.Biol. 4(9): e1000190.
Muzy,J.F., Bacry, E., Arneodo, A.(1993). Multifractal formalism for fractal signals: the structure-function approach
versus wavelet transform modulus maxima methods. Physical Review E 47:875-884.
Nakanishi, K., Kukita, F. (1998). Functional synapses in synchronized bursting of neocortical neurons in culture.
Brain Research 795:137-146.
Newman, M.E.J. (1996). Self-organized criticality, evolution and the fossil extinction record. Proc.R.Soc.London B
263:1605-1610.
Newman, M. E. J. (2003). The structure and function of complex networks. SIAM Review, 45:157-256.
Newman, M. E. J. (2005). Power laws, Pareto distributions and Zipf‟s law. Contemporary Physics, 46(5):323-351.
Newman, M.E.J., Girvan, M. (2004). Finding sand evaluating community structure in networks. Phys.Rev. E.
69:026113.
Nikulin, V. V., and Brismar, T. Long-range Temporal Correlations in Electroencephalographic Oscillations:
Relation to Topography, Frequency Band, Age and Gender. Neuroscience, 130:549-558.
Northoff, G., Heinzel, A., de Greck, M., Bermpohl, F., Dobrowolny,H., Panksepp, J. (2006). Self -referential
processing in our brain-- a meta-analysis of imaging studies on the self. NeuroImage 31:440-457,
Novikov, E., Novikov, A., Shannahoff-Khalsa, D., Schwartz, B., Wright, J. (1997). Scale-similar activity in the
brain. Physical Review E, 56(3), R2387.
Ohlshausen, B. A., Field, D.J. (1997). Sparse Coding with an overcomplete Basic Set: a strategy employed by V1?
Vision Res. 37:3311-3325.
Orer, H. S., Das, M., Barman, S. M., and Gebber, G. L. (2003). Fractal Activity Generated Independently by
Medullary Sympathetic Premotor and Preanglionic Sympathetic Neurons. J Neurophysiol, 90:47-54.
Paczuki, M., Maslov S., Bak, P. (1996). Avalanche dynamics in evolu tion, growth and depinning models. Physical
Review E 53(1):414-443.
Pajevik, S., Plenz D. (2009). Efficient network reconstruction from dynamical cascades identifies small world
topology of neuronal avalanches. PLoS ComputationalBiology 5(1): e1000271.
Papa, A. R. R., Silva, L. d. (1997). Earthquakes in the Brain. Theory Bioscienc., 116:321-327.
Paramanathan, P., Uthayakumar, R. (2008). Application of fractal theory in analysis of human
electroencephalographic signals. Computers in Biology and Medicine, 38:372-378.
Patel, A. D. (2003). Language, music, syntax, and the brain. Nature Neuroscience, 6(7):674-681.
Pattee, H.H. (2001) The physics of symbols: bridging the epistemic cut. BioSystems 60:5-21
Park, J. P., and Newman, M. E. J. (2004). Statistical mechanics of networks. Physics Review E, 70, 066117.
Pasquale, V. P., Massobrio, P., Bologna, L. L., Chiappalone, M. and Martinoia, S. (2008). Self -Organization and
Neuronal Avalanches in Networks of Dissociated Cortical Neurons. Neuroscience, 153:1354-1369.
Pellionisz, A.J. (1989). Neural Geometry: towards a fractal Model of Neurons. In: Models of Brain Function, edit.
R.M.L. Cotterill, Cambridge University Press.
Penrose, O. (1986). Phase Transitions on Fractal Lattices with Long-Range Interactions. J Statistical Physics,
45(1,2):69-88.
Perez-Mercader, J.(2004). Coarse-graining, Scaling and Hierarchies. In: Nonextensive Entropy-interdisciplinary
applications. Eds: M. Gell-Mann, C. Tsallis, Oxford Univ. Press. Pp. 357-376.
Perkel, D. H. and Feldman, M. W. (1979). Neurotransmitter Release Statistics: Moment Estimates for
Inhomogeneous Bernoulli Trials. J Math. Biology, 7:31-40.
Petermann, T., Thiagarajan, T. C., Lebedev, M. A., Nicolelis, M. A. L., Chial vo, D. R., and Plenz, D. (2009).
Spontaneous cortical activity in awake monkeys composed of neuronal avalanches. Proc. Natl. Acad. Sci. USA,
106(37):15921-15926.
Plenz, D., Aertsen A., (1996) Neuronal dynamics in cortex-striatum cultures II. Spatio-temporal characteristics of
neuronal activity. Neuroscience 70: 893-924.
Plenz, D., Thiagarajan, T. C. (2007). The organizing principles of neuronal avalanches: cell assemblies in the
cortex? TRENDS in Neurosciences, 30(3):101-110.
Plenz,D., Chialvo,D.R. (2010). Scaling properties of neuronal avalanches are consistent with critical dynamics.
arXiv:0912.5369v1 [q-bio.NC].
Podlubny, I. (1999) Fractional Differential Equations, Academic Press, San Diego
Poli, S.-S., Ooyen, A. v., Linkenkaer-Hansen, K. (2008). Avalanche Dynamics of Human Brain Oscillations:
Relation to Critical Branching Processes and Temporal Correlations. Human Brain Mapping, 29:770-777.
Pollack, J.B. (1991) Induction of dynamical recognizers. Machine Learning 7:227-252.Press, W. H. (1978). Flicker
Nosies in Astronomy and Elsewhere. Comments Astrophys., 7(4):103-119.
Quian, H. (2003).Fractional Brownian motion and Fractional Gaussian Noise. Lecture Notes in Physics 621:22-33
Reynolds, A.M. (2009) Scale-free animal movement patterns: Levy walks outperform fractional Brownian motions
and fractional Levy motions in random search scenarios. J. Phys. A 42:434006.
Richmond, B. J., Optican, L.M., Spitzer, H.(1990). Temporal encoding of two-dimensional patterns by single units
in Primary Visual Cortex I. Stimulus-Response relations. J. Neurophysiol. 64:351.
Rikvold, P.A., Kornoiss, G., White, C.J., Novotny, M.A., Sides,S.W. (1999). arXiv:cond-mat/9904028v2 [cond-
mat.stat.mech]
Robinson, H.P.C., Kawahara, M., Jimbo, Y., Torimitsu, K., Kuroda, Y., Kawana, A. (1993). Periodic synchronized
bursting and intracellular calcium transients elicited by low magnesium in cultured cortical neurons. J.
Neurophysiology 70 (4):1606-1616.
Rodriguez,M., Pereda, E., Gonzalez, J., Abdala,P., Obeso J. A.(2003). Neuronal activity in the Substantia Nigra in
the anesthetized rat has fractal characteristics: evidence for firing code patterns in the basal ganglia. Exp. Brain. Res.
151:167-172.
Roerig, B., Chan, B. (2002). Relationships of local inhibitory and excitatory circuits to orientation preference maps
in Ferret visual Cortex. Cerebral Cortex 12:187-198.
Roncaglia, R., Mannella, R., and Grigolini, P. (1994). Fractal Properties of Ion Channels and Diffusion.
Mathematical Biosciences, 123:77-101.
Rose, D., Lowe, I. (1982). Dynamics of adaptation to contrast. Perception 11:505-528.
Rotshenker, S., Rahamimoff, R. (1970). Neuromsucular Synapse: stochastic propertes of spontaneous release of
transmitter. Science 170:648-649.
Ruseckas, J., Kaulakys, B. (2010). 1/f noise from nonlinear stochastic differential equations . arXiv:1002.4316v1
[nlin.AO]
Sakai, Y., Funahashi, S., and Shinomoto, S. (1999). Temporally correlated inputs to leaky integrate -and-fire models
can reproduce spiking statistics of cortical neurons. Neural Networks, 12:1181-1190.
Salinas, E., and Sejnowski, T. J. (2002). Integrate-and-Fire Neurons Driven by Correlated Stochastic Input. Neural
Computation, 14:2111-2155.
Sales-Pardo, M.,Guimera, R., Moreira, A.A., Amaral, L.M.A (2007) Extracting the hierarchical organization of
complex systems. Proc.Nat.Acad.Sci. USA 104(39): 15224-15229
Salvadori, G., Biella G. (1994). Discriminating properties of wide dynamic range Neurons my means of universal
multifractals. In: Fractals in Biology and Medicine II, eds. G.A. Losa, D. Merlini, T.F. Nonnenmacher, E.R.
Weibel,. pp. 314-325.
Scafetta,N., Marchi,D., West, B.J. (2009). Understanding the complexity of human gait dynamics. Chaos
19:026108.
Scheffer, M., Bascompte, J., Brock, W.A., Brovkin, V., Carpenter S.R., Dakos, V., Held, H., van Nes, E.H.,
Rietkerk., M.D.,Sugihara, G. Early-warning signals for critical conditions. Nature 461:53-59.
Schierwagen, A. (2008). Neuronal Morphology: Shape Characteristics and Models. Neurophysiology, 40(4):310-
315.
Segev, R., Benveniste, M., Hulata, E., Cohen, N., Palevski, A., Kapon, E., Shapira, Y., and Ben -Jacob, E. (2002).
Long Term Behavior of Lithographically Prepared In Vitro Neuronal Networks. Physical Review Letters, 88(11),
118102.
Segev, R., Baruchi, I., Hulata, E., and Ben-Jacob, E. (2004). Hidden Neuronal Correlations in Cultured Networks.
Physical Review Letters, 92(11), 118102.
Seuront, L. (2010). Fractals and Multifractals in Ecology and aquatic science. CRC Press, Boca Raton,
Shadlen, M. N., and Newsome, W. T. (1998). The Variable Discharge of Cortical Neurons: Implications for
Connectivity, Computation, and Information Coding. J Neurosci, 18(10):3870-3896.
Shahverdian, A. Y., and Apkarian, A. V. (1999). On Irregular Behavior of Neuron Spike Trains. Fractals, 7(1):93-
103.
Shew,W.L., Yang,H., Petermann T., Roy, R., Plenz, D. (2009). Neuronal avalanches imply maximum dynamic
range in cortical networks at criticality. J. Neurosci. 29(49):15600-15595.
Shimizu, Y., Barth, M., Windischberger, C., Moser, E., and Thurner, S. (2004). Wavelet -based multifractal analysis
of fMRI time series. NeuroImage, 22:1195-1202.
Shimono, M., Owaki, T., Aman, K., Kitajo, K., and Takeda, T. (2007). Functional modulation of power -law
distribution in visual perception. Physical Review E, 75, 051902.
Shinomoto, S., Shima, K., Tanji, J. (2003). Differences in Spiking Patterns Among Cortical Neurons. Neural
Computation, 15:2823-2842.
Shlesinger, M. F., West, B. J., and Klafter, J. (1987). Levy Dynamics of Enhanced Diffusion: Application to
Turbulence. Physical Review Letters, 58(11):1100-1103.
Simon, H. A. (1955). On a Class of Skew Distribution Functions. Biometrika, 42:425-440.
Simon, H.A. (1962). The Architecture of Complexity. Proc. Amer. Philosoph. Soc. 106(6):467 -482.
Simon, H.A. (1973).The organization of complex systems. In: Hierarchy Theory, edit. H.H. Pattee, Brazillier, NY,
pp.3-27.
Smith Jr., T. G., Marks, W. B., Lange, G. D., Sheriff Jr., W. H., and Neale, E. A. (1989). A fractal analysis of cell
images. J Neurosci Methods, 27:173-180.
Smolders, F. D. J., Folgering H. Th. M., (1977). Actions and interactions of CO2 and O2 on the controlling system of
the lung ventilation. Thesis, Nijmegen University, Netherlands.
Sokal, A., Bricmont, J. (2004) Defense of a modest Scientific Realism. In: Carroer, M., Roggenhofer, J., Kueppers, G.,
Banchard ,P. (eds). Knowledge and the World: beyond the Science wars. Springer New York
Sokolov, I. M., Klafter, J., and Blumen, A. (2002). Fractional Kinetics. Physics Today, 48-54.
Soltys, Z., Ziaja, M., Pawlinski, R., Setkowicz, Z., and Janeczko, K. (2001). Morphology of Reactive Microglia in the
Injured Cerebral Cortex. Fractal Analysis and Complementary Quantitative Methods. J Neurosci. Res., 63:90-97.
Song, C., Havlin, S., and Makse, H. A. (2005). Self-similarity of complex networks. Nature, 433:392-395.
Song, C., Havlin, S., Makse, H.A. (2006). Orogins of gfractality in the growth of complex networks. Nature Physics
2:275-281.
Sornette, D., (2000), Critical Phenomena in Natural Sciences. Springer, Berlin.
Sporns, O., and Zwi, J. D. (2004). The Small World of the Cerebral Cortex. Neuroinformatics, 2:145-162.
Sporns, O., Koetter, R. (2004) Motifs in Brain networks. PLoS Biology 2(11): e369
Sporns,O., Chialvo, D.R., Kaiser, M., Hilgetag, C.C. (2004). Organization, development and function of complex brain
networks. Trends Cog. Sci. 8(9): 418-425.
Sporns, O. (2006). Small-world connectivity, motif composition, and complexity of fractal neuronal composition.
BioSystems, 85:55-64.
Sporns, O., Honey, C.J., Koetter, R. (2007). Identification and classification of hubs in brain networks. PLoS ONE 10:
e1049
Stam, C. J. (2004). Functional connectivity patterns of human magnetoencephalographic recordings: a ‘small -world’
network? Neurosci Letters, 355:25-28.
Stam, C. J., and Bruin, E. A. d. (2004). Scale-Free Dynamics of Global Functional Connectivity in the Human Brain.
Human Brain Mapping, 22:97-109.
Stam, C. J. (2005). Nonlinear dynamical analysis of EE and MEG: Review of an emerging field. Clinical Neurophys.,
116:2266-2301.
Stam, C. J., and Reijneveld, J. C. (2007). Graph theoretical analysis of complex network in the brain. Nonlinear
Biomedical Phys, 1:3.
Stanley, H.E. (1987). Introduction to phase transitions and critical phenomena. Oxford University Press, Oxford, UK
Stanley, H.E. (1999). Scaling, universality and renormalization: the three pillars of modern critical Phenomena. Rev.
Mod. Physiscs 71:S358-S366.
Stauffer,D., Aharony A. (1991/1994). Introduction to Percolation theory. CRC Press, Boca Raton.
Stephen,D.G., Dixon, J.A. (2009) The self-organization of insight: Entropy and power laws in problem solving. J. of
Problem Solving 2(1):72-101.
Stevens, S. S. (1957). On the psychophysical Law. Psychol.Rev.64 (3):153-181.
Stevens, C. F., and Zador, A. M. (1998). Input synchrony and the irregular firing of cortical neurons. Nature
Neuroscience, 1(3):210-217.
Stewart, C.V., Plenz, D. (2006) Inverted U profile of dopamine-NMDA mediated spontaneous avalanche recurrence
in superficial layers o rat prefrontal cortex. J. Neurosci. 23:8148-8159.
Stinchcombe, R. B. (1989). Fractals, phase transitions, and criticality. Proc. R. Soc. Lond. A, 423:17-33.
Stoop, R., Wagner, C. (2007) Neocortex’s architecture optimizes computation, information transfer and
synchronizability, at given total connection length. Int.J.Bifurc.Chaos 17(7): 2257-2279.
Suckling,J., Wink A.M., Bernard, F.A., Barnes, A., Bullmore, E. (2008) Endogenous multifractal brain dynamics are
modulated by age, cholinergic blockade and cognitive performance, J. Neuroscience Methods 174:292 -300.
Tabor, W. (2000). Fractal encoding of context-free grammars in connectionist networks. Expert Systems 17:41-56.
Takeda, T., Sakata, A., and Matsuoka, T. (1999). Fractal Dimensions in the Occurrence of Miniature End -Plate
Potential In a Vertebrate Neuromuscular Junction. Prog. Neuro-Psychopharmacol & Biol. Psychiat., 23:1157-1169.
Tateno, T., Kawana, A., Jimbo,Y. Analytical characterization of spontaneous firing in networks of developing rat
cultured cortical neurons. Physical Review E 65:051924.
Tebbens, S. F., and Burroughs, S. M. (2003). Self-Similar Criticality. Fractals, 11(3):221-231.
Teich, M. C., and Saleh, B. E. A (1981). Interevent-time statistics for shot-noise-driven self-exciting point processes
in photon detection. J. Opt. Soc. Am., 71(6):771-776.
Teich, M. C., Johnson, D. H., Kumar, A. R., and Turcott, R. G. (1990). Ra te fluctuations and fractional power-law
noise recorded from cells in the lower auditory pathway of the cat. Hearing Research, 46:41-52.
Teich, M. C., Heneghan, C., Lowen, S. B., Ozaki, T., and Kaplan, E. (1997). Fractal character of the neural spike train
in the visual system of the cat. . J. Opt. Soc. Am., 14(3):529-546.
Teramae, J.-n., and Fukai, T. (2007). Local cortical circuit model inferred from power-law distributed neuronal
avalanches. J Comput Neurosci, 22:301-312.
Thatcher, R. W., North, D. M., and Biver, C. J. (2009). Self-Organized Criticality and the Development of EEG Phase
Reset. Human Brain Mapping, 30:553-574.
Thiagarajan T.C., Lebedev, M.A., Nicolelis M.A., Plenz, D.(2010), Coherence potentials: Loss -less, all-or-none
network events in the cortex. PLoS Biology 8 (1):e1000278.
Thorson, J., Biederman-Thorson, M. (1974). Distributed relaxation processes in sensory adaptation. Science
183:161-183.
Thurner, S., Lowen, S. B., Feurstein, M. C., Heneghan, C., Feichtinger, H. G., and Teich, M. C. (1997). Analysis,
synthesis, and estimation of fractal-rate stochastic point processes. Fractals, 5(4):565-595.
Thurner, S., Windischberger, C., Moser, E., Walla, P., and Barth, M. (2003) Scaling laws and persistence in human
activity. Physica A, 326:511-521.
Tino, P. (1999). Spatial representation of symbolic sequences through iterative function systems. IEEE Trans. On
Systems, Man and Cybernetics 29(4):386-393.
Tognoli, E., Kelso, J.A.S. (2009). Brain Coordination Dynamics: True and false faces of phase synchrony and
metastability. Progr.Neurobiol. 87:31-40.
Toib, A., Lyakhov, V., and Marom, S. (1998). Interaction between Duration of Activity and Time Course of R ecovery
from Slow Inactivation in Mammalian Brain Na+ Channels. J Neurosci, 18(5):1893-1903.
Touboul, J., and Destexhe, A. (2009). Can power -law scaling and neuronal avalanches arise from stochastic
dynamics? arXiv:0910.0805v1.
Tsuda,I., Kuroda, S. (2004) A complex system approach to an interpretation of Dynamic Brain activity II. : does
Cantor coding provide a dynamical model for the formation of episodic memory ? in: Cortical Dynamics, LNCS
3146, edit: P. Erdi, pp. 129-139. Springer, Berlin.
Turcotte, d.L., (1999). Self-organized criticality. Rep.Prog.Phys. 62:1377-1429.
Tsallis, C., Levy, S. V. F., Souza, A. M. C., and Maynard, R. (1995). Statistical -Mechanical Foundation of the
Ubiquity of Levy Distributions in Nature. Physical Review Letters, 75(20):3589-3593.
Tsallis, C. (2009). Introduction to non-extensive Statistical Mechanics: approaching a complex World . Springer,
NY.
Uhlhaas,P.J., Pipa, G., Lima, B., Melloni, L., Neuenschwander, S., Nikolic, D., Singer, W. (2009). Neural
Synchrony in cortical networks : history, concept and current status. Frontiers in Integrative Neuroscience 3:1-19
Ulanovsky, N., Las, L., Farkas, D., Nelken, I. (2004). Multiple time scales of adaptation in auditory cortex neurons.
J. Neurosci. 17:10440-10453
Usher,M., Stemmler, M., Koch,C., Olami,Z. (1994). Network amplification of local fluctuations causes high spike
rate variability, fractal firing patterns and oscillatory local field potentials. Neural Computation 6:795-836.
Usher, M., and Stemmler, M. (1995). Dynamic Pattern Formation Leads to 1/f Noise in Neural Populations.
Physical Review Letters, 74(2):326-329.
Van den Heuvel, M. P., Stam, C. J., Boersma, M., and Pol, H. E. H (2008). Small -world and scale-free organization
of voxel-based resting-state functional connectivity in the human brain. NeuroImage, 43:528-539.
Van Order, G. C., Holden, J. G., and Turvey, M. T. (2003). Self -Organization of Cognitive Performance. J
Experimental Psych: General, 132(3):331-350.
Van Pelt, J., Coner, M.A., Wolters, P.S., Rutten, W.L.C., Ramakers, G.J.A. (2004). Longterm stability and
developmental changes in spontaneous network burst firing patterns in dissociated rat cerebral cortex cell cultures
on microelectrode arrays. Neuroscience Letters 361:86-89.
Van Vreeswijk, C. (2001). Informat ion transmission with renewal neurons. Neurocomputing, 38-40:417-422.
Varanda, W. A., Liebovitch, L. S., Figueiroa, J. N., and Nogueira, R. A. (2000). Hurst Analysis Applies to the Study
of Single Calcium-activated Potassium Channel Kinetics. J. theor. Biol., 206:343-353.
Varela F., Lachaux, J-P., Rodriguez,E., Martinerie,J. (2001), The Brainweb: phase synchronization and large scale
integration. Nature Reviews Neuroscience 2:229-239.
Voss, R. F., and Clarke, J. (1975). „1/f noise‟ in music and speech. Nature, 258:317-318.
Vrobel, S. (2007) Fractal time, observer perspective and levels of description in Nature.. Electronic J. of Theoretical
Physics 16(II):275-302.
Wagenaar, D.A., Pine, J., Potter, S.M. (2006 a) An extremely rich repertoire of bursting pa tterns of bursting patterns
during the development of cortical cultures. BMC Neuroscience 7: 7-11.
Wagenaar, D.A.,Nadasky, Z., Potter, S.M. (2006 b). Persistent dynamic attractors in activity patterns of cultured
neuronal networks. Physical Review E 73: 051907.
Wagenmakers, E.-J., Farrell, S., and Ratcliff, R. (2004). Estimation and interpretation of 1/f α noise in human
cognition. Psychonomic Bulletin & Review, 11(4):579-615.
Wagenmakers, E-J., Farrell, S. and Ratcliff, R. (2005). Human Cognition and a pile of sand: a discussion on serial
correlations and self-organized criticality. J.Exp.Psychol.Gen. 134(1): 108-116.
Wark, B., Lundstrom, B.N., Fairhall, A. (2008). Sensory Adaptation. Curr. Opin. Neurobiol. 17:423-429.
Watts, D., Strogatz S. (1998). Collective dynamics of Small World Networks. Nature 393:440-442.
Wen, Q., Stepanyants, A., Elston, G. N., Grosberg, A. Y., and Chklovskii, D. B. (2009). Maximization of the
connectivity repertoire as a statistical principle governing the shapes of dendritic arbors. Proc.Natl.Acad.Sci. USA,
106:12536-12541.
West, B.J., Griffin, L. (1999).Allometric control, Inverse power laws and human gait. Chaos, Solitons & Fractals
10(9):1519-1527.
Werner, G., and Mountcastle, V. B. (1963). The Variability of Central Neural Activity in a Sensory System, and its
Implications for the Central Reflection of Sensory Events. J Neurophysiol., 26:958-977.
Werner, G., and Mountcastle, V. B. (1964). Neural Activity in Mechanoreceptive Cutaneous Afferen ts; Stimulus-
Response Relations, Weber Functions, and Information Transmission. J. Neurophysiol., 28:359-394.
Werner, G. (2007a). Perspectives on the Neuroscience of Cognitions and Consciousness. BioSystems, 87:82-95.
Werner, G. (2007b). Metastability, criticality, and phase transitions in brain and its models. BioSystems, 90:496-508.
Werner, G. (2009a). Viewing brain processes and Critical State Transitions across levels of organization: Neural
events in Cognition and Consciousness, and general principles. BioSystems, 96:114-119.
Werner, G. (2009b). Consciousness related neural events viewed as brain state space transitions. Cogn. Neurodyn.,
3:83-95.
Werner, G. (2009c). On Critical State Transitions Between Different Levels in Neural Systems. New Math. And
Natural Computation, 5(1):185-196.
West, B.J. (1999 a). Physiology, Promiscuity and Prophecy at the Millenium: a tale of tails . Studies of Nonlinear
Phenomena in the Life Sciences, Vol. 7. World Scientific, Singapore.
West, B.J. (2006). Where Medicine went wrong: rediscovering the path to complexity. World Scientific,Singapore.
West, B.J. (2009 b). Control from an allometric perspective . Advances in Experimental Medicine and Biology
629:57-82.
West, B. J., Allegrini, P., Grigolini, P. (1994). Dynamical Generators of Levy Statistics in Biology. In: Fractals in
Biology and Medicine, Vol. II, eds. G.A. Losa, D. Merlini, T.F. Nonnenmacher, E.R. Weibel. Birkhauser, Basel.
West, B.J., Deering, B. (1995). The Lure of Modern Science. Studies in Nonlinear Phenomena in Life Sciences, Vol.
3. World Scientific.
West, B. J., Grigolini, P., Metzler, R., and Nonnenmacher, T. F. (1997). Fractional diffusion and Levy stable
processes. Physical Review E, 55(1):99-106.
West, B. J., and Griffin, L. (1999 a). Allometric Control, Inverse Power Laws and Human Gait. Chaos, Solitons, and
Fractals, 10(9):1519-1527.
.West, B. J., and Nonnenmacher, T. (2001). An ant in a gurge. Physics Letters A, 278:255-259.
West, B.J., Bologna, M., Grigolini, P. (2003). Physics of Fractal Operators. Springer.
West, B.J., Scafetta N. (2003) Nonlinear dynamical model of human gait. Physical Review E 67:051917.
West, B. J., Geneston, E. L., Grigolini, P.(2008). Maximizing Information Exchange between complex Networks.
Physics Reports, 468(1-3), 1-99.
Willinger,W., Taquu, M.S., Sherman, R., Wilson, D.V., (1995). Self-similarity through high variability: statistical
analysis of Ethernet LAN traffic at the source level. ACM SIGCOMM Computer Communication Review 25 (4):100-
113.
Willinger, W. (2000). The Discovery of Self-Similar Traffic. Lecture Notes in Computer Science 1769. Springer. pp
513-527.
Wilson, K. G. (1979). Problems in Physics with Many Scales of Length. Sci. Amer., 241:158-179.
Wise, M.E. (1981). Spike Interval Distributions for Neurons and Random Walks with Drift to a Fluctuating
Threshold. Statistical Distributions in Scientific Work, Taillie et al, eds., 6:211-231, Reidel Publ. Comp.
Witten Jr., T. A., and Sander, L. M. (1981). Diffusion-Limited Aggregation, a Kinetic Critical Phenomenon.
Physical Review Letters, 47(19):1400-1403.
Womelsdorf, T., Schoffelen,J.M., Oostenveld, R., Singer W., Desimone,R., Engel,A,K, Fries,P. (2007). Modulation
of neuronal interactions through neuronal synchronization. Science 316:1609-1612.
Wornell, G. W. (1993). Wavelet-Based Representations for the 1/f Family of Fractal Processes. Proceedings of the
IEEE, 81(10):1428-1450.
Xu, Z., Payne, J. R., and Nelson, M. E. (1996). Logarithmic Time Course of Sensory Adaptation in Electrosensory
Afferent Nerve Fibers in a Weakly Electric Fish. J Neurophysiol., 76(3):2020-2032.
Yamamoto, M., Nakahama, H., Shima, K., Kodama, T., Mushiake, H. (1986). Markov dependency and spectral
analyses on spike counts in mesencephalic reticular neurons during sleep and attentive states. Brain Research
366:279-289.
Yu, Y., Romero, R., Lee,T.S. (2005). Preference of sensory neural coding for 1/f signals. Phys. Rev. Lett.
94:108103.
Yule, G. U. (1925). A Mathematical Theory of Evolution, based on the Conclusions of Dr. J. C. Willis, F. R. S. Phil.
Trans. R. Soc. Lond. B, 213:21-87.
Zanette, D. H. (2008). Zipf‟s law and the creation of musical context. arXiv:cs/0406015v1.
Zapperi, S., Lauritsen, K.B., Stanley, H.E. (1995). Self-organized branching process: mean-field theory for
avalanches. Physical Review Letters 75(22): 4071-4074.
Zarahn, E., Aguirre, G. K., and D‟Esposito, M. (1997). Empirical Analyses of BOLD fMRI Statistics. NeuroImage,
5:179-197.
Zhang, L. (2006).Quantifying brain white matter structural changes in normal aging using fractal Dimension.
Doctoral Thesis, Case Western Reserve University. http://etd.ohiolink.edu/send-
pdf.cgi/Zhang%20Luduan.pdf?acc_num=case1126213038
Zietsch, B. and Elston, G. N. (2005). Fractal Analysis of Pyramidal Cells in the Visual Cortex of the Galago
(Otolemur Garnetti): Regional Variation in Dendritic branching Patterns Between Visual Areas. Fractals, 13(2):83-
90.
Zhou, C., Zemanova, L., Zamora-Lopez, G., C. C., and Kurths, J. (2007). Structure-function relationship in complex
brain networks expressed by hierarchical synchronization. New Journal of Physics, 9, 178.
Zumofen, G., and Klafter, J. (1993). Scale-invariant motion in intermittent chaotic systems. Physical Review E,
47(2):851-863.
|
1802.08279 | 4 | 1802 | 2019-03-20T16:04:03 | Implementing a Concept Network Model | [
"q-bio.NC"
] | The same concept can mean different things or be instantiated in different forms depending on context, suggesting a degree of flexibility within the conceptual system. We propose that a compositional network model can be used to capture and predict this flexibility. We modeled individual concepts (e.g., BANANA, BOTTLE) as graph-theoretical networks, in which properties (e.g., YELLOW, SWEET) were represented as nodes and their associations as edges. In this framework, networks capture the within-concept statistics that reflect how properties correlate with each other across instances of a concept. We ran a classification analysis using graph eigendecomposition to validate these models, and find that these models can successfully discriminate between object concepts. We then computed formal measures from these concept networks and explored their relationship to conceptual structure. We find that diversity coefficients and core-periphery structure can be interpreted as network-based measures of conceptual flexibility and stability, respectively. These results support the feasibility of a concept network framework and highlight its ability to formally capture important characteristics of the conceptual system. | q-bio.NC | q-bio | 1
The updated and peer-reviewed version of this
paper is now published. Please read and refer to
this final version:
Solomon, S.H., Medaglia, J.D., & Thompson-Schill, S.L.
(2019). Implementing a concept network model. Behavior
Research Methods. https://doi.org/10.3758/s13428-019-
01217-1
2
Implementing a Concept Network Model
Sarah H. Solomon1, John D. Medaglia2,3, and Sharon L. Thompson-Schill1
1 Department of Psychology, University of Pennsylvania
2 Department of Psychology, Drexel University
3 Department of Neurology, Perelman School of Medicine, University of
Pennsylvania
May 22, 2018
Abstract
The same concept can mean different things or be instantiated in different
forms depending on context, suggesting a degree of flexibility within the
conceptual system. We propose that a compositional network model can be
used to capture and predict this flexibility. We modeled individual concepts
(e.g., BANANA, BOTTLE) as graph-theoretical networks, in which properties
(e.g., YELLOW, SWEET) were represented as nodes and their associations as
edges. In this framework, networks capture the within-concept statistics that
reflect how properties correlate with each other across instances of a concept.
We ran a classification analysis using graph eigendecomposition to validate
these models, and find that these models can successfully discriminate
between object concepts. We then computed formal measures from these
concept networks and explored their relationship to conceptual structure. We
find that diversity coefficients and core-periphery structure can be interpreted
as network-based measures of conceptual
flexibility and stability,
respectively. These results support the feasibility of a concept network
framework and highlight
important
its ability
characteristics of the conceptual system.
Keywords:
conceptual knowledge, conceptual flexibility, network science
to formally capture
3
Introduction
The APPLE information evoked by "apple pie" is considerably different from
that evoked by "apple picking": the former is soft, warm, and wedge-shaped,
whereas the latter is firm, cool, and spherical. If you scour your conceptual
space for APPLE information, you will uncover the knowledge that apples can
be red, green, yellow, or brown when old; that they can be sweet or tart; that
they are crunchy when fresh and soft when baked; that they are naturally
round but can be cut into slices; that they are firm, but mushy if blended;
that they can be found in bowls, in jars, and on trees. Despite the complexity
of this conceptual knowledge, we can generate an appropriate APPLE
instance, with the appropriate features, based on the context we are in at the
time. In other words, the multi-faceted APPLE concept can be flexibly
adjusted in order to enable a near-infinite number of specific and appropriate
APPLE exemplars.
How is conceptual knowledge structured such that this generative and
essential flexibility is possible? In particular, we are interested in the
structure of individual concepts (e.g., APPLE, SNOW), rather than the structure
of super-ordinate categories (e.g., FRUIT, TOOLS) or the structure of semantic
space more broadly. This latter pursuit -- the modeling of semantic space
-- has already been approached from various theoretical orientations and
methodologies. In "compositional" cognitive theories, the meaning of a
concept can be decomposed into features and their relationships with each
other (e.g., Smith et al, 1974, McRae et al., 1997; Tyler & Moss, 2001). This
compositionality can be incorporated into the associated computational
models: researchers primarily use various forms of feature-based
connectionist models, in which concepts are represented as patterns of
activation over features, to simulate semantic behavior (e.g., Cree et al.,
1999; 2006; Randall et al., 2004). In "relational" frameworks, concepts are
defined in terms of how they relate to other concepts in semantic or lexical
space (e.g., Landauer & Dumais, 1997). Researchers within this framework
can use word co-occurrence or association statistics to create large semantic
networks, which can be analyzed using a rich set of network science tools
(e.g., Steyvers & Tenenbaum, 2005; Van Rensbergen et al., 2015; De Deyne
et al, 2016).
The compositional approach to conceptual knowledge generally represents
individual concepts as vectors of features. These features can span a range of
4
information-types (e.g., visual, functional, encyclopedic), consistent with a
distributed account of conceptual knowledge. The "conceptual structure"
account (Tyler & Moss, 2001) represents concepts as binary vectors
indicating the presence of absence of features, and argues that broad
semantic domains (e.g., ANIMALS, TOOLS) differ in their characteristic
properties and in their patterns of property-correlations (e.g., HAS-WINGS and
FLIES tend to co-occur within the ANIMAL domain). The "feature-correlation"
account (e.g., McRae et al., 1997; 1999; McRae, 2004) tweaks this model by
empirically deriving conceptual property statistics and by implementing this
framework in a type of connectionist model called an attractor network:
property statistics characterize the structure of conceptual space, and the
model can leverage these statistics to settle on an appropriate conceptual
representation given the current inputs (Cree et al., 1999; Cree et al., 2006).
These models contain a dynamic component in that the attractor networks
reveal how word comprehension may unfold over time, and is "flexible" in
the sense that the model may follow varied trajectories through conceptual
space in order to settle on a specific concept's representation. However, the
ability of this approach to capture conceptual flexibility -- in the sense
described above -- is not fully fleshed out. A feature-based framework is
valuable because the elements that can be adjusted during conceptual
processing are explicitly modeled, but a different set of tools may be helpful
in the pursuit of capturing how flexibility might emerge out of this
conceptual structure. Our initial steps to develop such tools are reported
here.
Another way to model conceptual knowledge is to capture statistical
relations between words or phrases in language. This approach is
"relational", rather than compositional, because a concept's meaning is
represented in terms of its relations to other concepts, rather than assuming
any kind of internal conceptual structure. Word co-occurrence statistics can
be extracted from text corpora and have been used to create probabilistic
models of word meanings (Griffiths et al., 2007), to represent semantic
similarity (Landauer & Dumais, 1997), and to characterize the structure of
the entire lexicon (e.g., WordNet; Miller & Fellbaum, 2007). In a similar
approach, word association data can be used to capture and analyze the
structure of semantic space (Steyvers & Tenenbaum, 2005; Van Rensbergen
et al., 2015; De Deyne et al., 2016). These data are generally modeled as
networks, which can be analyzed in formal ways.
5
The use of networks to model semantic knowledge has a well-established
history. The early "semantic network" models (Collins & Quillian, 1969;
Collins & Loftus, 1975) represent concepts as nodes in a network; links
between these nodes signify associations in semantic memory. These
networks capture the extent to which concepts are related to other concepts
and features, and can model the putatively hierarchical nature of conceptual
knowledge. Though these models are "network-based", they are so in a
rather informal way. On the other hand, network science, a mathematical
descendent of graph theory, has developed a rich set of tools to study
networks in a formal, quantitative framework (Barabási, 2016). There has
been a recent surge of research applying network science to neural and
linguistic data, providing new insights into these complex systems along
with new tools and measures we can use to study them.
Formal networks are composed of units (i.e., "nodes") and the links between
them (i.e., "edges"). Current network science approaches to semantic and
lexical knowledge use nodes to represent individual words, and edges to
represent their co-occurrence or association statistics. Once modeled in this
way, aspects of network structure can be quantitatively analyzed and
relationships between network structure and other phenomena can be
explored. For example, it has been suggested that human language exhibits
small-world properties (Steyvers & Tenenbaum, 2005; i Cancho & Sole,
2001), and that semantic networks exhibit an "assortative" structure,
meaning that semantic nodes tend to have connections to other semantic
nodes with similar characteristics (e.g., valence, arousal, concreteness; Van
Rensbergen et al., 2015). A spreading activation model applied to these
word-association networks makes accurate predictions of weak similarity
judgments (De Deyne et al., 2016). Further, Steyvers & Tenenbaum (2005)
report that a word's degree (i.e., how many links it has to other nodes)
predicts both age of acquisition and reaction times on lexical decision tasks.
The application of network science tools to semantic data enables
researchers to explore higher-level structure in semantic space, and to use
these structural characteristics to predict aspects of semantic and lexical
processing. However, this approach does not examine the internal structure
of concepts: individual concepts are characterized in terms of their statistical
co-occurrences with other concepts in language, and not in terms of their
6
Figure 1: Example images used to generate test data in classification analysis. Test data used in the
classification analysis were generated from participants who made property judgments on images of
conceptual exemplars. Yellow cross indicates object to be considered. Example images for grass (top)
and cookie (bottom).
unique, internal content. That is, these language-based approaches are non-
compositional, and we argue that compositionality is a key aspect of flexible
conceptual models. Modeling a concept's internal structure -- along with its
features and the ways those features interact -- enables us to flexibly adjust
which features are included, and to what degree, in a given conceptual
instance.
We believe that a compositional conceptual framework paired with network
science techniques provides a platform on which to model conceptual
flexibility. Unlike previous network approaches, we use networks to model
individual concepts, rather than the semantic system as a whole. In this case,
the nodes in each concept's network represent individual features, and the
edges of the network (i.e., the links between the nodes) represent the
statistical relationship between features within that concept. That is, edges
capture the extent to which certain properties tend to covary with each other
within a concept. The creation of such networks thus depends on our ability
to calculate within-concept statistics. These statistics provide the scaffolding
to build our networks, and also reveal how a concept's information may be
appropriately adjusted to form valid, yet varied, instances of that concept.
Our goal is to show that creation of such networks is possible, and that they
can be used to capture conceptual flexibility, and other phenomena, in a
formal way.
We chose 15 basic-level concepts (e.g., CHOCOLATE, TABLE, GRASS, KNIFE)
and defined a list of within-concept states for each one (e.g., DARK
7
CHOCOLATE, WHITE CHOCOLATE, CHOCOLATE SYRUP, CHOCOLATE CHIPS)
using a large sample of participants on Amazon Mechanical Turk (AMT).
We then compiled a large set of conceptual properties that could apply to
any of the 15 concepts (e.g., BROWN, GREEN, WOODEN, METAL, SHARP,
SWEET). A final sample of AMT participants then reported which of the
properties corresponded to each of the specific concept-states. This let us
know which properties applied to dark chocolate (e.g., BLACK, BITTER) and
which applied to white chocolate (e.g., WHITE, SWEET). These data enable us
to calculate the within-concept statistics necessary to construct our networks.
Each of the properties corresponds to a vector that denotes whether that
property is present or absent in each of the concept-states, and we can
correlate these property vectors with each other to determine the extent to
which each property covaries with every other property within that specific
concept. These correlation values are encoded as the edges in the concept's
network.
Once these networks were constructed, we had two objectives: First, we
intended to determine whether or not these networks contain concept-
specific information. We thus ran a classification analysis over these
networks to confirm that within-concept statistics can be used to
discriminate between concepts. We performed eigendecomposition on our
concept networks in a classification analysis, which provided a measure of
the extent to which a vector is consistent with an underlying network
structure (e.g., Medaglia et al., 2017, Huang et al., in press). In our case, our
test data were vectors of properties generated from photographs of
individual concept exemplars (Fig. 1). Second -- and most importantly --
we intended to extract useful and interpretable measures from our concept
networks. In particular, we aimed to quantify conceptual flexibility.
Many networks by their very nature permit flexibility, because a single
network can support different states, each characterized by different patterns
of activation across nodes. A node's contribution to network flexibility can
be understood in terms of its position in the context of the larger network.
Most natural systems exhibit "small-world" network structure (Bassett &
Bullmore, 2017), which means that there are clusters of nodes in a network
with strong connections between them (Amaral et al., 2000). These are
called "modules", and nodes can interact with these modules in different
ways. Some nodes may have links that are highly distributed across the
modules in a network, whereas other nodes may have links only in one
module. Each node in a network can be assigned a diversity coefficient, a
8
version of the participation coefficient calculated using normalized Shannon
entropy, which reflects this tendency. We interpreted network diversity as a
likely candidate for a formal flexibility measure, and pooled the diversity
coefficients across nodes in a concept's network to quantify that concept's
flexibility. We calculated this version of flexibility for our 15 basic-level
concepts, and predicted that it would correlate with a measure of "semantic
diversity" calculated separately using word co-occurrence statistics (SemD;
Hoffman et al., 2013). This would suggest that network-based measures can
successfully be used to quantify flexibility in a compositional conceptual
framework.
Another phenomenon of interest to cognitive scientists is the distinction
between context-independent and context-dependent conceptual properties
(Barsalou, 1982). Context-independent properties are those that are
automatically activated for a concept in all contexts, and are sometimes
referred to as "core" properties. On the other hand, context-dependent
properties are those that are only activated when the context renders them
relevant. Concepts are composed of both kinds of properties, such that some
properties are stable and are activated across all instances, and some are
more variable and are only activated some of the time. The distinction
between context-independent and -- dependent properties has been suggested
in reaction time differences in property-verification tasks (Barsalou, 1982),
but the classification of a property as one type or the other has been decided
upon by the experimenter rather than being calculated in a quantitative way.
One of our goals was to use our concept networks to extract this
information; that is, whether the structure of each concept does in fact
include such a core.
Network science provides techniques for assessing this core-periphery
structure (Borgatti & Everett, 2000; Bassett et al., 2013). In network terms, a
core is a set of nodes that are densely interconnected and therefore often co-
activated, whereas the periphery consists of nodes with sparser connections.
A measure can be extracted that represents the extent to which a given
network has a core-periphery structure; some networks might have more
prominent cores than others. We hypothesize that this construct of core-
periphery structure can provide a way to formally capture the notion of
context-dependent and context-independent conceptual properties. In other
words, perhaps a certain concept has a large set of context-independent
properties that are consistently activated across a large range of contexts:
this concept's network might have a strong core-periphery structure. It also
9
seems reasonable to suggest that those concepts with a stronger core might
be less flexible in the ways described above. If we interpret a core as a set of
properties whose activation patterns are stable across contexts, then there is
less room for variability in the expression of those properties, and therefore
less flexibility overall. On the other hand, more flexible concepts might have
a weaker core-periphery structure, reflecting the more variable patterns of
property activations. We therefore predict a negative relationship between
our network measures of flexibility and core-periphery structure.
Methods
General Methods
Network Construction In order to create our networks we first had to
define our nodes. Since our nodes represent individual conceptual properties,
we compiled a list of properties that could be applied to all of our target
concepts. Participants were recruited from Amazon Mechanical Turk (AMT)
and were asked to list all of the properties that must be true or can be true for
each concept. It was emphasized that the properties do not have to be true of
all types of the concept. Participants were required to report at least 10
properties per concept, but there was no limit on the number of responses
they could provide. Once these data were collected, we organized the data as
follows. For each concept, we collapsed across different forms of the same
property (e.g., "sugar", "sugary", "tastes sugary"), and removed responses
that were too general (e.g., "taste", "color"). For each concept, we only
included properties that were given by more than one participant. We then
combined properties across all concepts to create our final list of N
properties that will be represented as nodes in our concept networks.
The same AMT participants that provided conceptual properties also
provided concept-states for each of the target concepts. For each concept,
participants were asked to think about that object and all the different kinds,
forms, types, or states in which that object can be found. Participants were
required to make at least five responses, and could make up to 15 responses.
For each concept, we removed responses that we considered properties
rather than types (e.g., "sweet chocolate"), and responses that were too
specific (e.g., "Chiquita banana"). We only included responses that were
10
given by more than one participant, resulting in a set of K concept-states for
each concept.
A separate set of AMT participants was presented with one concept-state of
each of the target concepts in random order (e.g., "dark chocolate", "frozen
banana") and was asked to select the properties that are true of that specific
concept-state. The full list of N properties was displayed in a multiple-choice
format. For each concept-state, responses were combined across participants
and represented in a binary fashion. A property was only considered "true"
for a concept-state if more than one participant made that response. At this
point, each concept's data included a set of K concept-states, each of which
corresponds to a N-length vector that indicates the presence or absence of
each property. We could also view these data as a set of N conceptual
properties, each of which corresponded to a K-length vector that indicates its
presence or absence in each of the concept-states.
For each concept, we excluded properties that were not present in any of the
concept-states, resulting in a smaller set of NC properties. We created a
network by correlating the NC binary property-vectors with each other to
create a NC x NC symmetrical, weighted correlation matrix. The diagonal was
set to 0, and these networks were filtered using the triangulation filtering
method (Massara et al., 2016; Tumminello et al., 2005; Kenett et al., 2014).
This filtering approach generates a simpler subgraph that maximizes
information content while reducing the influence of noise. This method is
appropriate for graphs where edges are defined as correlations between
nodes, as is the case here. No parameter fitting is required. These final,
filtered concept networks were then analyzed using standard network
science methods.
Classification Analysis If our concept network models capture concept-
specific information, the networks should be able to successfully
discriminate between new concept exemplars. Exemplar data were generated
from sets of photographs for each concept; all concept-states were
represented. AMT participants were shown one image per concept, were
asked to imagine interacting with this object in the real world, and to
consider what properties it has. The full list of N properties was displayed in
multiple-choice format, and participants were asked to select the properties
that they believed applied to the object in the image. Individual participants'
responses to each concept-state were represented as N-length property
vectors and were used as test data in the classification analysis.
11
𝑥" = ∑ (𝑣(∙𝑥)+
,(-.
(1)
By performing eigendecomposition on each adjacency matrix (i.e., concept
network) we can assess the extent to which a vector is expected given an
underlying network structure (e.g., Medaglia et al., 2017; Huang et al., in
press). For each adjacency matrix A, V is the set of NC eigenvectors, ordered
by eigenvalue. M is the number of ordered eigenvectors to include in
analysis, and designates a subset of V. For each eigenvector v, we find the
dot product with signal vector x, which gives us the projection of x on that
dimension in the eigenspace of A. That is, it gives us an "alignment" value
for that particular signal and that particular eigenvector. We can include all
eigenvectors in M by taking the sum of squares of the dot products for each
eigenvector. The alignment value for each signal is defined as:
where 𝑥 is a property vector, M is the number of eigenvectors to include in
alignment (sorted by eigenvalue), 𝑣( is one of M eigenvectors of the
adjacency matrix, and 𝑥" is the scalar alignment value for signal x with
The concept network that resulted in the highest alignment value (𝑥") was
adjacency matrix A, given the eigenvectors 1-M. In our case, signal x is a
property vector corresponding to a particular exemplar image (e.g., Fig. 1),
which we align with each of the concept networks. Each exemplar was
restricted to the properties included in each concept model before
transformation; that is, exemplar data (x) were reduced to NC -- length vectors.
taken as the "guess" of the classifier; each exemplar was either classified
correctly (1), or incorrectly (0). We averaged these data across all exemplars
to calculate the average classifier accuracy.
To calculate a baseline measure of classification accuracy, we created
traditional vector models for each concept. These models were similar to
those used elsewhere in the literature (Tyler & Moss, 2001; McRae et al.
1997; 1999; 2004). For each concept, we averaged the K concept-state
vectors resulting in an NC -length vector containing mean property strength
values. Each concept's traditional vector model and network model
contained the same conceptual properties. We ran a separate classification
analysis using these traditional models and a correlational classifier. Each
exemplar property-vector was correlated with each of the traditional concept
vector models; the concept model that resulted in the highest correlation
value was taken as the guess of the classifier. We calculated average
12
measures of classifier performance using the same methods described above,
and also calculated classification accuracy within each concept.
Network Analysis We extracted network metrics from our concept
networks using the Brain Connectivity Toolbox (Rubinov & Sporns, 2010).
The set of nodes in each network is designated as N, and n is the number of
nodes. The set of links is L, and l is the number of links. The existence of a
link between nodes (i,j) is captured in 𝑎(0: 𝑎(0=1 if a link is present and
𝑎(0=0 if a link is absent. The weight of a link is represented as 𝑤(0, and is
normalized such that 0≤𝑤(0≤1. 𝑙6 is the sum of all weights in the
degree, modularity (𝑄), core-periphery structure, and diversity coefficients
connections to other nodes. Node degree (𝑘) is the number of connections
Sporns, 2010). In weighted (i.e., non-binary) networks, node strength (𝑘6) is
Nodes within a network differ in the number and strength of their
that each node has with other nodes in the network (Eq. 2; Rubinov &
calculated by summing the weights of the connections with other nodes (Eq.
3; Rubinov & Sporns, 2010). We separately averaged node strength and
node degree within each network to obtain mean strength and degree
measures for each concept network. We also counted the number of edges in
each network overall.
network. The network metrics we extracted included node strength, node
(Fig. 2).
𝑘( = ∑
𝑎(0
0∈:
𝑘(6 = ∑
𝑤(0
0∈:
We can attempt to partition a weighted network into sets of non-overlapping
nodes (i.e., modules) such that within-module connections are maximized
and between-module connections are minimized. Some networks exhibit
modularity for each weighted network (Eq. 4; Rubinov & Sporns, 2010),
which is defined as
Modularity (𝑄) is a metric that describes a network's community structure.
more of a modular structure than others; 𝑄6 is a quantitative measure of
𝑄6 = .;< ∑
(,0∈:
=𝑤(0− ?@<?A<;< B 𝛿D@,DA
(4)
(2)
(3)
13
Figure 2: Schematics of network structure. (A) Low-modularity network that contains nodes with
equal degree. (B) High-modularity network with nodes in either module 1 (red) or module 2 (blue). One
node (purple) participates in both modules; this is a high-diversity node. (C) Network with a strong core-
periphery structure; some nodes comprise a densely connected core (purple) and others a weakly
connected periphery (grey).
where 𝛿D@,DA=1 if nodes i,j are in the same module (m), 𝑤(0 is the specific
strength between nodes i,j, and ?@<?A<;< scales 𝑤(0 by the total strengths of
the set of modules (M): Nodes may have connections to many different
(ℎ(±) is a measure ascribed to individual nodes that reflects the diversity of
nodes i,j across the network. Given a network's community structure, we
can observe how individual nodes participate with each of the modules in
modules, or have very few such connections. The diversity coefficient
connections that each node has to modules in the network. This is a version
of the participation coefficient, and is calculated using normalized Shannon
entropy; we have previously used entropy to model property flexibility, and
so predicted that diversity would be a good candidate for a network-based
measure of conceptual flexibility. The diversity coefficient (Eq. 5; Rubinov
& Sporns, 2011) for each node is defined as
ℎ(± = − .HIJD∑
LÎ,
Q@± , 𝑠(±(𝑢) is the strength of node 𝑖 within module 𝑢, and
where 𝑝(±(𝑢) = Q@±(L)
𝑚 is the number of modules in modularity partition 𝑀. We averaged
(𝑢)log𝑝(±(𝑢),
diversity coefficients across nodes in a network to obtain a mean measure of
diversity for each concept network.
Core-periphery structure is another way to describe the structure of a
network. Here, we attempt to partition a network into two non-overlapping
𝑝(±
(5)
(6)
∑
(,0ÎV_
Z𝑤(0−𝛾V𝑤\]
'
sets of nodes such that connections within one set are maximized (i.e., the
"core") and connections in the other are minimized (i.e., the "periphery").
can be partitioned in this way (Eq. 6), and can be defined as
14
Core-periphery fit (𝑄V) is a quantitative measure of how well each network
𝑄V= .WX Y∑
(,0ÎV^
where 𝐶b is the set of all nodes in the core, 𝐶cis the set of nodes in the
periphery, 𝑤\ is the average edge weight, 𝛾Vis a parameter controlling the
size of the core, and 𝑣Vis a normalization constant (Rubinov et al., 2015).
Z𝑤(0−𝛾V𝑤\] −
Methods: Set 1
The 5 concepts used in Set 1 were CHOCOLATE, BANANA, BOTTLE, TABLE, and
PAPER. AMT participants (N=66) provided general properties for each
concept along with concept-states. Another group of AMT participants
(N=198) made property judgments on specific concept-states, and another
group of AMT participants (N=60) generated test data for the classification
analysis by making property judgments on individual images. The final
property list included 129 properties. The number of states for each concept
were as follows: chocolate=14, banana=15, bottle=11, table=14, paper=20.
The full list of states can be seen in the Appendix. In the classification
analysis, test data comprised a total of 300 property-vectors, with 60
exemplars/concept.
Methods: Set 2
The 10 concepts used in Set 2 were KEY, PUMPKIN, GRASS, COOKIE, PICKLE,
KNIFE, PILLOW, WOOD, PHONE, and CAR. AMT participants (N=60) provided
general properties for each concept along with concept-states. Another group
of AMT participants (N=108) made property judgments on specific concept-
states, and another group of AMT participants (N=30) generated test data for
the classification analysis by making property judgments on individual
images. The final property list included 276 properties. The number of states
for each concept were as follows: key=19, pumpkin=18, grass=16,
cookie=22, pickle=17, knife=15, pillow=16, wood=22, phone=16, car=20.
The full list of states can be seen in the Appendix. In the classification
analysis, test data comprised 300 property-vectors, with 30
exemplars/concept.
15
eigen-dimensions from our concept networks. Classification was successful using ≥ 7 dimensions in Set
1, and ≥1 dimension in Set 2. Classification performance increased as more dimensions were added, such
Figure 3: Classification results. We ran a range of classification analyses using different numbers of
that performance of the network-models approached performance of the vector-based models (single data
points). The sharp increase in performance in both sets is driven by eigenvectors with eigenvalues of 0,
suggesting that contribution of individual features, suggesting that the presence
Results
Classification Results
In order to determine whether our concept networks contained concept-
specific information, we ran a classification analysis using
eigendecomposition for both Set 1 and Set 2. We ran multiple analyses using
different ranges of eigenvectors, which were sorted by eigenvalue (positive
to negative). We started by only using the first eigenvector in each of the
concept networks and determined whether this dimension alone could be
used to classify the property vector. One dimension was enough to classify
exemplars in Set 2 (Mean Accuracy=0.27; SE=0.03; Chance=0.10) but not
Set 1 (Mean Accuracy=0.11; SE=0.02; Chance=0.20).
However, increasing the number of dimensions improved classification
performance for both sets (Fig. 3): for example, classification performance is
significantly above chance when only 10 dimensions are used in Set 1
(Mean Accuracy=0.38; SE=0.03; Chance=0.10) and Set 2 (Mean
Accuracy=0.38; SE=0.03; Chance=0.20). As more dimensions were
16
Table 2: Correlation results. We analyzed relationships between cognitive variables (SemS, SemD),
network variables (modularity, core fit, mean diversity, mean strength, mean degree, number of edges) and
vector-based classification results. *: p<0.05, **: p<0.01.
included in the analysis, classification performance significantly increases
and approaches the performance of the vector-based classifier, which was
successful at classifying exemplars in Set 1 (Mean Accuracy=0.85;
SD=0.06; Chance=.20) and Set 2 (Mean Accuracy=0.84; SD=0.10;
Chance=0.10).
The middle range on the x-axis of Fig. 3 contains eigenvectors with
eigenvalues of 0: this is a special case in which a single property is weighted
as 1. When one of these eigenvectors is multiplied by a signal (i.e., property
vector) that includes that particular property, the alignment value is driven
by the presence of that one property. In other words, the dramatic increase in
classification performance in both Set 1 and Set 2 is driven by the successive
contributions of individual features moving from left to right, suggesting that
the presence or absence of individual features is highly informative for
discriminating between concepts. Nevertheless, the significant classification
performance using eigenvectors representing multiple features does suggest
that our concept networks contain concept-specific information, motivating
us to look within a concept for structural elements that relate to conceptual
flexibility. It is this main goal that we pursue in the subsequent analyses.
Network Measures of Conceptual Structure
Networks across the two sets differed in node assignments, since they were
constructed using different properties. However, once classification and
network measures were extracted, we could pool the concepts together
(N=15) and examine relationships between these network-related measures
and other variables of interest.
17
Figure 4: A network-based measure of conceptual flexibility. Semantic diversity measures calculated
using word co-occurrence statistics (Hoffman et al., 2013) predict network diversity across 15 concepts;
network diversity is mean diversity coefficient across nodes.
We extracted network measures from the full concept networks and explored
how they relate to cognitive measures of conceptual flexibility and stability.
Hoffman et al. (2013) use word co-occurrence statistics to quantify the
context-dependent variations in word meanings found in language. The
authors provide a measure of semantic diversity (SemD) that captures this
variability, and we extracted SemD values for our 15 concepts. We also
extracted their reported mean cosine similarity of a word's contexts and used
this as a measure of semantic stability (which we refer to as SemS). As
expected, SemD negatively correlated with SemS across our 15 concepts
(r(15)=-0.96, p=<0.0001). The correlations between all measures of interest
are shown in Table 2.
One of our primary goals was to extract a network measure that reflects
conceptual flexibility. We used SemD (Hoffman et al., 2013) as a
benchmark for conceptual flexibility and determined whether our
hypothesized network measures of flexibility correlated with SemD across
our 15 concepts. A priori, we hypothesized that the mean diversity (i.e., the
average of a concept network's diversity coefficients across nodes) could
reflect conceptual flexibility. This network measure captures the extent to
which properties within a concept associate with different modules, or
property clusters. Another possible candidate measure was network
modularity, which reflects the extent to which a concept's network can be
18
partitioned into separate property clusters. Network modularity was not
significantly predicted by either SemD (r(15)=0. 22, p>0.4) or SemS
(r(15)=-0.19, p>0.5). On the other hand, mean diversity was positively
predicted by SemD (r(15)=0.56, p=0.03; Fig. 4) and negatively predicted by
SemS (r(15)=-0.60, p=0.02). Mean diversity was not significantly predicted
by mean node strength (r(15)=-0.08, p>0.7), mean node degree (r(15)=0.42,
p=0.12), or total number of network edges (r(15)=0.3, p>0.2). These results
suggest that the network measure of mean diversity is a strong candidate for
a quantitative measure of conceptual flexibility.
We also assessed the core-periphery structure for each concept network,
which determines how well a network can be divided into a densely
connected core and a sparsely connected periphery. If the core of a concept
network corresponds to the notion of a context-independent conceptual
"core", we would expect that more stable (i.e., less flexible) concepts would
have networks with a stronger core-periphery structure. Consistent with this
prediction, core fit was positively predicted by SemS (r(15)=0.54, p=0.04),
though the relationship with SemD was only marginally significant (r(15)=-
0.50, p=0.06). Furthermore, mean diversity and core fit were negatively
correlated (r(15)=-0.61, p=0.02), suggesting that these measures may be
used to capture conceptual flexibility and stability, respectively. We also
found that classification accuracy using the traditional vector model was
positively correlated with core fit (r(15)=0.56, p=0.03). This suggests that
standard cognitive models perform better on more stable concepts,
highlighting the need for a model that can adequately capture conceptual
flexibility.
Discussion
Here our goal was to model basic-level concepts using graph-theoretical
networks within a compositional conceptual framework. We argue that the
within-concept statistics encoded in these models capture useful, concept-
specific information. Using standard network science tools, we further reveal
the usefulness of these models by extracting formal metrics that relate to
cognitive notions of conceptual flexibility and stability.
Our concept network models capture the particular conceptual properties
that are associated with an individual concept along with those properties'
19
concept-specific covariation statistics. Individual concept networks are
distinct in that properties relate to each other in different ways across basic-
level concepts. For example, BLACK and SOFT may co-vary with each other
in BANANA, but BLACK and FIRM may co-vary with each other in
CHOCOLATE. We found that our network models could successfully
discriminate between new conceptual exemplars, suggesting that these
within-concept statistics differ reliably between basic-level concepts. These
results emerged out of a classification analysis based on eigendecomposition
of our concept networks. Eigendecomposition of graphs has previously been
used to assess the correspondences between anatomical brain network
structure and patterns of functional activation (Medaglia et al., 2017); here
we adapted this method to assess the correspondences between conceptual
structure and feature-patterns for individual conceptual exemplars. We found
that concept networks can simultaneously encode multi-property
relationships (i.e. within-concept statistics) and strong single-property
contributions, suggesting ways in which information might be organized
within the conceptual system and also establishing an exciting new direction
for further investigation.
A model structured using within-concept statistics provides a framework in
which varied yet appropriate instantiations of a concept may be flexibly
activated. An APPLE network may contain a strong connection between
CRUNCHY + FRESH and between SOFT + BAKED, enabling the conceptual
system to know what sets of properties should be activated in a particular
APPLE instance -- for example, in the representations evoked by "apple
picking" versus "apple pie." The property-covariation statistics for a given
concept will determine which sets of properties tend to be co-activated, and
how individual properties relate to those sets and to each other. We thus
sought to use our compositional concept network models, which contain
within-concept statistics, to extract quantitative measures of these
phenomena. We found that mean-diversity and core-periphery structure can
be interpreted as measures of conceptual flexibility and stability,
respectively: a concept network-model's mean-diversity positively predicts
semantic diversity (SemD; Hoffman et al., 2013), a network-model's core-
periphery fit positively predicts semantic stability (mean cosine similarity;
Hoffman et al., 2013), and these two network measures are negatively
related to each other across our concepts. Our results also suggest that
traditional property-vector models are better at capturing the representation
of stable versus flexible concepts, suggesting that a different kind of
conceptual model may be necessary to capture the intrinsic flexibility of the
20
conceptual system. We argue that a network-based model of basic-level
concepts is one such option.
Network mean-diversity was extracted by averaging over individual nodes'
diversity coefficients, a version of the participation coefficient calculated
using Shannon entropy. In network neuroscience, participation coefficients
have been related to the flexibility of functional network hubs. As discussed
above, networks can be partitioned into communities of modules, and nodes
can differentially participate in these modules: nodes that only have
connections to one module have low participation coefficients, whereas
nodes that have connections to many modules have high participation
coefficients. This measure is used to differentiate between different kinds of
network hubs. High-degree nodes that have low participation coefficients are
classified as "provincial" hubs, since they are only connected to one local
module. High-degree nodes that have high participation coefficients are
classified as "global" or "connector" hubs, since they have connections to
many modules across the network (van den Heuvel, 2013; Power et al.,
2013). Empirical studies have revealed that the fronto-parietal and cingulo-
opercular networks contain high-participation nodes, suggesting that these
regions contain functional network hubs (Sporns, 2014). It is further argued
that "flexible network hubs" are characterized by their ability to connect
with a diverse set of brain regions (Sporns, 2014).
Networks have varying degrees of flexibility with respect to how their states
can shift with time. On a higher level, networks can differ in how their
underlying structure changes with time. In order to study the fluctuations in
such dynamic neural networks, Bassett et al. (2011) quantify changes in
community structure over time. Instead of extracting a measure from a
single network, they define flexibility by analyzing multiple, consecutive
networks and determining whether individual nodes either maintain or
switch allegiances between communities. Given that networks can exhibit
flexibility at these different levels of analysis, we might expect some
features of static networks to support flexibility within dynamic networks.
We believe that node-participation -- and node-diversity, specifically --
captures a network characteristic of community allegiance similar to Bassett
et al. (2011). Furthermore, high-participation or high-diversity nodes (i.e.,
global hubs) should play a key role in flexibility because their dense,
variable connections mediate transitions between network states. We thus
argue that diversity might index flexibility across multiple levels of analysis,
and that concept network diversity can be reasonably interpreted as a
21
measure of conceptual flexibility. We do not rule out the possibility that
there are other candidate network measures that can also capture conceptual
flexibility, and intend to explore this moving forward.
Other frameworks have the potential to capture the flexibility of the
conceptual system; these include attractor networks (e.g., Cree et al., 1999;
2006; Rodd et al., 2004) and recent updates of the hub-and-spoke model
(Lambon Ralph et al., 2016). The concept network framework proposed here
is not in opposition with these other approaches; the development and
implementation of all of these methods will greatly benefit our
understanding of the semantic system. However, we do believe that a
network science approach to conceptual knowledge has its unique
advantages. Most broadly, the vast network science methodological toolkit
allows us to translate our analyses between cognitive and neural levels of
analysis. Given a sample of concept networks and extracted measures of
interest, we can observe the functional neural networks recruited for
conceptual processing and explore any correspondences between networks
across levels. Network neuroscientists have previously forged links to
cognitive processes such as motor-sequence learning (Bassett et al., 2011)
and cognitive control (Medaglia et al., 2018), setting a precedent for the
application of these methods to cognitive science.
The intersection of network science with control theory is another direction
that may prove useful for our current purposes. Network controllability
refers to the ability to move a network into different network states, and has
been applied to structural brain networks in order to shed insights into how
the brain may guide itself into easy- and difficult-to-reach functional states
(Gu et al., 2015). There have been additional attempts to link brain network
controllability to cognitive control (Medaglia, 2018). The application of
control theory to concept networks may provide an additional way to
quantify conceptual flexibility by identifying nodes that are well-positioned
to drive the brain into diverse, specific, or integrated states. Perhaps concept
networks that are more controllable overall -- that is, networks in which it is
easier to reach varied network states -- correspond to concepts that are more
cognitively flexible.
The concept network framework permits the application of spreading
activation models to assess how information flows through these networks.
De Deyne et al. (2016) paired a spreading activation model over relational
semantic networks and found that their network models could make accurate
22
predictions regarding semantic similarity judgments. In our case, we could
use spreading activation models to observe patterns of network activity as a
result of different inputs (i.e., contexts). For example, we can attempt to
model how conceptual information differs when presented in an adjective-
noun phrase: a vector representing adjectival information could be provided
as input to the networks, resulting in a specific pattern of activity across the
nodes. More generally, these techniques might provide a way to model
context-dependent conceptual meaning.
We acknowledge that our proposed concept network model framework has
some limitations. First, a large amount of data needs to be collected to
enable the calculation of within-concept statistics. However, now that we
have established the feasibility and usefulness of these models, it is possible
to develop online platforms that can streamline data collection. Second,
there is a sense in which some of the methodological decisions are arbitrary:
for example, the number of properties to model as network nodes, and the
number of concept-states, could influence the resulting networks and
extracted measures. These concerns are mitigated, however, by our findings
that results calculated from our concept models in Set 1 and Set 2 were
comparable, and that the number of concept-states did not significantly
predict our network measures of interest. Using standard network filtering
methods also reduces these concerns, since it decreases the number of
arbitrary parameter decisions required by the experimenter. It is also
imperative that the measures extracted from concept networks are
interpretable in the context of conceptual knowledge. Only by understanding
the potential correspondences between network structure and the structure of
conceptual knowledge will these ideas prove useful.
Here we have constructed concept network models, confirmed their ability
to capture concept-specific information, and extracted network measures that
relate to cognitive measures of conceptual flexibility and stability. We
believe the application of network science to conceptual knowledge will
provide a set of tools that will enable the intrinsic flexibility of the
conceptual system to be explored and quantified.
23
Acknowledgements
This work was supported by an NSF graduate research fellowship awarded
to SHS, DP5-OD021352 awarded to JDM, and National Institute of Health
grant R01 DC015359 awarded to STS.
Appendix
Concept-States (Set 1)
CHOCOLATE: bittersweet chocolate, caramel chocolate, chocolate bar, chocolate chips,
chocolate syrup, cocoa powder, dark chocolate, chocolate fudge, melted chocolate, milk
chocolate, nut chocolate, salted chocolate, white chocolate, hot chocolate
BANANA: banana chips, banana pudding, Cavendish banana, fried banana, frozen banana,
mashed banana, over-ripe banana, peeled banana, plantain, raw banana, red banana, rotten
banana, sliced banana, unripe banana, ripe banana
BOTTLE: baby bottle, beer bottle, broken bottle, juice bottle, liquor bottle, medicine bottle,
milk bottle, soda bottle, spray bottle, water bottle, wine bottle
TABLE: bedside table, changing table, coffee table, conference table, dining table, drafting
table, end table, folding table, kitchen table, play table, poker table, pool table, side table,
workbench
PAPER: butcher paper, cardboard, cardstock, construction paper, envelope, graph paper,
legal paper, newspaper, notebook paper, paper towel, papyrus, poster board, printer paper,
sandpaper, scrap paper, sketch paper, stationery, tissue paper, toilet paper, wrapping paper,
writing paper
Concept-States (Set 2)
KEY: car key, key to a city, door key, encryption key, garage key, key to my heart, house
key, key card, keyboard key, map key, master key, motorcycle key, office key, padlock
key, password, piano key, key to a safe, skeleton key
PUMPKIN: pumpkin bar, pumpkin bread, pumpkin candle, canned pumpkin, pumpkin
cookie, pumpkin in a field, Halloween pumpkin, Jack-O-Lantern, pumpkin latte, pumpkin
muffin, pumpkin pie, pumpkin puree, rotten pumpkin, pumpkin seeds, smashed pumpkin,
pumpkin soup, pumpkin spice, whole pumpkin, Thanksgiving pumpkin
24
GRASS: astroturf, bamboo, barley grass, bent grass, grass clippings, crab grass, dead grass,
hay, lawn grass, lemongrass, marijuana, oat grass, overgrown grass, grass seeds, sod,
wheatgrass
COOKIE: almond cookie, butter cookie, chocolate cookie, chocolate chip cookie, Christmas
cookie, cookie cake, cookie dough, ginger snap, Girl Scout cookie, lemon cookie, M&M
cookie, macadamia nut cookie, macaroon, mint cookie, no-bake cookie, oatmeal raisin
cookie, Oreo cookie, peanut butter cookie, shortbread, snickerdoodle, sugar cookie, wafer
cookie
PICKLE: bread and butter pickles, canned pickles, pickle chips, chopped pickles, cucumber
pickles, dill pickles, garlic pickles, gherkins, hamburger pickles, homemade pickles, jarred
pickles, kosher pickles, relish, sliced pickles, sandwich pickles, pickle spears, whole
pickles
PILLOW: airplane pillow, bed pillow, body pillow, cotton pillow, couch pillow, decorative
pillow, down pillow, feather pillow, foam pillow, hypo-allergenic pillow, memory foam
pillow, neck pillow, silk pillow, throw pillow, travel pillow
KNIFE: bread knife, butcher knife, butter knife, cheese knife, chef's knife, dagger, hunting
knife, jackknife, machete, paring knife, pocket knife, steak knife, switchblade, sword,
throwing knife, utility knife
WOOD: wood blocks, wood chips, chopped wood, wood fence, firewood, floor, wood
furniture, log, lumber, wood paneling, paper, planks, plywood, wood pulp, sticks, tree,
cedar wood, cherry wood, maple wood, oak wood, pine wood, walnut wood
PHONE: android phone, antique phone, broken phone, car phone, cell phone, emergency
phone, flip phone, home phone, iPhone, land line, pay phone, rotary phone, satellite phone,
smart phone, wall phone, wireless phone
CAR: broken down car, compact car, convertible, coupe, electric car, family car, hatchback,
hybrid car, jeep, luxury car, pickup truck, race car, rental car, sedan, sports car, station
wagon, SUV, toy car, truck, used car, van
References
Amaral, L. A. N., Scala, A., Barthelemy, M., & Stanley, H. E. (2000). Classes of small-
world networks. Proceedings of the national academy of sciences, 97(21), 11149-
11152.
Barabási, A. L. (2016). Network science. Cambridge university press.
25
concepts. Memory & Cognition, 10(1), 82-93.
Barsalou, L. W. (1982). Context-independent and context-dependent information in
Bassett, D. S., & Bullmore, E. T. (2017). Small-world brain networks revisited. The
Bassett, D. S., Wymbs, N. F., Porter, M. A., Mucha, P. J., Carlson, J. M., & Grafton, S.
Neuroscientist, 23(5), 499-516.
T. (2011). Dynamic reconfiguration of human brain networks during
learning. Proceedings of the National Academy of Sciences, 108(18), 7641-7646.
Bassett, D. S., Wymbs, N. F., Rombach, M. P., Porter, M. A., Mucha, P. J., & Grafton, S.
T. (2013). Task-based core-periphery organization of human brain dynamics. PLoS
computational biology, 9(9), e1003171.
networks, 21(4), 375-395.
Borgatti, S. P., & Everett, M. G. (2000). Models of core/periphery structures. Social
Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory of semantic
Collins, A. M., & Quillian, M. R. (1969). Retrieval time from semantic memory. Journal
Cree, G. S., McRae, K., & McNorgan, C. (1999). An attractor model of lexical
of verbal learning and verbal behavior, 8(2), 240-247.
processing. Psychological review, 82(6), 407.
conceptual processing: Simulating semantic priming. Cognitive Science, 23(3), 371-
414.
Cree, G. S., McNorgan, C., & McRae, K. (2006). Distinctive features hold a privileged
status in the computation of word meaning: Implications for theories of semantic
memory. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 32(4), 643.
De Deyne, S., Navarro, D. J., Perfors, A., & Storms, G. (2016). Structure at every scale:
A semantic network account of the similarities between unrelated concepts. Journal of
Experimental Psychology: General, 145(9), 1228.
Griffiths, T. L., Steyvers, M., & Tenenbaum, J. B. (2007). Topics in semantic
Gu, S., Pasqualetti, F., Cieslak, M., Telesford, Q. K., Alfred, B. Y., Kahn, A. E., ... &
representation. Psychological review, 114(2), 211.
Bassett, D. S. (2015). Controllability of structural brain networks. Nature
communications, 6, 8414.
Hoffman, P., Ralph, M. A. L., & Rogers, T. T. (2013). Semantic diversity: A measure of
semantic ambiguity based on variability in the contextual usage of words. Behavior
research methods, 45(3), 718-730.
26
Huang, W., Bolton, T.A.W., Medaglia, J.D., Bassett, D.S., Ribeiro, A., & Van De Ville,
D. (In Press). A Graph Signal Processing Perspective on Functional Brain Imaging.
IEEE.
of the Royal Society of London B: Biological Sciences, 268(1482), 2261-2265.
i Cancho, R. F., & Solé, R. V. (2001). The small world of human language. Proceedings
Kenett, Y. N., Anaki, D., & Faust, M. (2014). Investigating the structure of semantic
Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato's problem: The latent
networks in low and high creative persons. Frontiers in Human Neuroscience, 8, 407.
semantic analysis theory of acquisition, induction, and representation of
knowledge. Psychological review, 104(2), 211.
Massara, G. P., Di Matteo, T., & Aste, T. (2016). Network filtering for big data:
triangulated maximally filtered graph. Journal of complex Networks, 5(2), 161-178.
McRae, K. (2004). Semantic memory: Some insights from feature-based connectionist
attractor networks. The psychology of learning and motivation: Advances in research
and theory, 45, 41-86.
McRae, K., Cree, G. S., Westmacott, R., & Sa, V. R. D. (1999). Further evidence for
feature correlations in semantic memory. Canadian Journal of Experimental
Psychology/Revue canadienne de psychologie expérimentale, 53(4), 360.
McRae, K., De Sa, V. R., & Seidenberg, M. S. (1997). On the nature and scope of
featural representations of word meaning. Journal of Experimental Psychology:
General, 126(2), 99.
Medaglia, J. D., Huang, W., Karuza, E. A., Kelkar, A., Thompson-Schill, S. L., Ribeiro,
A., & Bassett, D. S. (2017). Functional alignment with anatomical networks is
associated with cognitive flexibility. Nature Human Behaviour, 1.
arXiv:1801.08806.
Evaluation, 41(2), 209-214.
Medaglia (2018). Clarifying cognitive control and the controllable connectome.
Miller, G. A., & Fellbaum, C. (2007). WordNet then and now. Language Resources and
Power, J. D., Schlaggar, B. L., Lessov-Schlaggar, C. N., & Petersen, S. E. (2013).
Evidence for hubs in human functional brain networks. Neuron, 79(4), 798-813.
Ralph, M. A. L., Jefferies, E., Patterson, K., & Rogers, T. T. (2017). The neural and
computational bases of semantic cognition. Nature Reviews Neuroscience, 18(1), 42.
27
Randall, B., Moss, H. E., Rodd, J. M., Greer, M., & Tyler, L. K. (2004). Distinctiveness
and correlation in conceptual structure: Behavioral and computational studies. Journal
of Experimental Psychology: Learning, Memory, and Cognition, 30(2), 393.
semantic ambiguity in word recognition. Cognitive Science, 28(1), 89-104.
Rodd, J. M., Gaskell, M. G., & Marslen-Wilson, W. D. (2004). Modelling the effects of
Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity:
Rubinov, M., & Sporns, O. (2011). Weight-conserving characterization of complex
Rubinov, M., Ypma, R. J., Watson, C., & Bullmore, E. T. (2015). Wiring cost and
functional brain networks. Neuroimage, 56(4), 2068-2079.
uses and interpretations. Neuroimage, 52(3), 1059-1069.
topological participation of the mouse brain connectome. Proceedings of the National
Academy of Sciences, 112(32), 10032-10037.
memory: A featural model for semantic decisions. Psychological review, 81(3), 214.
Smith, E. E., Shoben, E. J., & Rips, L. J. (1974). Structure and process in semantic
Sporns, O. (2014). Contributions and challenges for network models in cognitive
Steyvers, M., & Tenenbaum, J. B. (2005). The Large-scale structure of semantic
neuroscience. Nature neuroscience, 17(5), 652.
networks: Statistical analyses and a model of semantic growth. Cognitive
science, 29(1), 41-78.
Tumminello, M., Aste, T., Di Matteo, T., & Mantegna, R. N. (2005). A tool for filtering
information in complex systems. Proceedings of the National Academy of Sciences of
the United States of America, 102(30), 10421-10426.
knowledge. Trends in cognitive sciences, 5(6), 244-252.
Tyler, L. K., & Moss, H. E. (2001). Towards a distributed account of conceptual
van den Heuvel, M. P., & Sporns, O. (2013). Network hubs in the human brain. Trends in
Van Rensbergen, B., Storms, G., & De Deyne, S. (2015). Examining assortativity in the
cognitive sciences, 17(12), 683-696.
mental lexicon: Evidence from word associations. Psychonomic bulletin &
review, 22(6), 1717.
|
1905.02405 | 1 | 1905 | 2019-05-07T08:37:31 | Neuroplasticity in adult human visual cortex | [
"q-bio.NC"
] | Between 1 to 5 out of 100 people worldwide has never experienced normotypic vision due to a condition called amblyopia, and about 1 out of 4000 suffer from inherited retinal dystrophies that progressively lead them to blindness. While a wide range of technologies and therapies are being developed to restore vision, a fundamental question still remains unanswered: would the adult visual brain retain a sufficient plastic potential to learn how to see after a prolonged period of abnormal visual experience? In this review we summarize studies showing that the visual brain of sighted adults retains a type of developmental plasticity, called homeostatic plasticity, and this property has been recently exploited successfully for adult amblyopia recover. Next, we discuss how the brain circuits reorganizes when visual stimulation is partially restored by means of a bionic eye in late blinds with Retinitis Pigmentosa. The primary visual cortex in these patients slowly became activated by the artificial visual stimulation, indicating that sight restoration therapies can rely on a considerable degree of spared plasticity in adulthood. | q-bio.NC | q-bio | Neuroplasticity in adult human visual cortex
Elisa Castaldi1, Claudia Lunghi2 and Maria Concetta Morrone3,4
1 Department of Neuroscience, Psychology, Pharmacology and Child health, University of
Florence, Florence, Italy
2 Laboratoire des systèmes perceptifs, Département d'études cognitives, École normale
supérieure, PSL University, CNRS, 75005 Paris, France
3 Department of Translational Research and New technologies in Medicine and Surgery,
University of Pisa, Pisa, Italy
4 IRCCS Stella Maris, Calambrone (Pisa), Italy
Abstract
Between 1-5:100 people worldwide has never experienced normotypic vision due
to a condition called amblyopia, and about 1:4000 suffer from inherited retinal
dystrophies that progressively lead them to blindness. While a wide range of
technologies and therapies are being developed to restore vision, a fundamental
question still remains unanswered: would the adult visual brain retain a sufficient
plastic potential to learn how to 'see' after a prolonged period of abnormal visual
experience? In this review we summarize studies showing that the visual brain of
sighted adults retains a type of developmental plasticity, called homeostatic
plasticity, and this property has been recently exploited successfully for adult
amblyopia recover. Next, we discuss how the brain circuits reorganizes when visual
stimulation is partially restored by means of a 'bionic eye' in late blinds with Retinitis
Pigmentosa. The primary visual cortex in these patients slowly became activated
by the artificial visual stimulation, indicating that sight restoration therapies can rely
on a considerable degree of spared plasticity in adulthood.
Keywords
Keywords: 7T fMRI; amblyopia; binocular rivalry; bionic eye; blindness; cortical
excitability; critical period; cross-modal plasticity; retinal prosthesis; retinitis
pigmentosa; short-term monocular deprivation; visual restoration
Highlights
Short-term monocular deprivation triggers homeostatic plasticity in adult humans
Physical activity and inverse occlusion promote visual recovery in amblyopic
adults
Intra-modal and cross-modal brain reorganization occur in late-blind individuals
Sight restoration in late-blinds promotes early visual cortex re-organization
The adult visual system retains residual plasticity well after the critical period
Index
1. Introduction
2. Cortical plasticity in sighted adults revealed by short-term visual deprivation
2.1. Behavioral proxies for plasticity
2.2. Short-term visual deprivation alters visual neural responses
2.3. Neurochemical changes following short-term visual deprivation
2.4. Therapy for amblyopia
3. Cortical plasticity after vision loss following ophthalmological diseases
3.1. Retinotopic remapping of the visual cortex
3.2. Cross-modal plasticity in blind individuals
3.3. Cortical plasticity and sight restoration
4. Conclusions and further directions
1. Introduction
Neuroplasticity is the ability of the nervous system to adapt and optimize its limited
resources in response to physiological changes, injuries , new environmental demands
and sensory experiences (Pascual-Leone et al., 2005).
Early in the individual's development, during the so called critical period, the intrinsic
plastic potential of the nervous systems is maximal and sensory deprivation events can
cause profound morphological alterations of sensory cortices preventing the normal
development of sensory functions (Berardi et al., 2000; Wiesel and Hubel, 1963).
Classical studies in the visual system of kittens and non-human primates demonstrated
that even few days of abnormal visual experience during the period of high
susceptibility can cause severe visual impairment that cannot be recovered even after
prolonged periods of restored vision (Hubel and Wiesel, 1970; Hubel et al., 1977;
Wiesel and Hubel, 1963, 1965b). Visual cortical plasticity is typically assessed by using
the experimental paradigm of monocular deprivation: occluding one eye few weeks
after birth causes atrophy of the lateral geniculate nucleus (LGN) layers representing
the deprived eye, reorganization of primary visual cortex (V1) ocular dominance
columns, with ocular dominance shifting in favor of the open eye and with only a few
neurons responding to the deprived eye (Hubel and Wiesel, 1970; Hubel et al., 1977;
Wiesel and Hubel, 1963). At the functional level, this shift of ocular dominance is
reflected in lower visual acuity and reduced response to stimulation of the deprived
eye, a condition known as amblyopia (for a review see: Levi and Carkeet, 1993). In
adults, after the closure of the critical period, amblyopia cannot be induced or treated:
in adult cats monocular deprivation has only minor or no effects, suggesting that the
visual cortex retains very little or no experience-dependent plasticity in adulthood
(Wiesel and Hubel, 1963). Similarly, in humans early visual deprivation or suboptimal
visual experience (for example due to untreated congenital cataracts, astigmatism,
myopia, etc.) leads to plastic reorganization of the visual cortices and permanent visual
impairments that can hardly be recovered in adulthood even after corrections or visual
restoration (Braddick and Atkinson, 2011; Fine et al., 2002, 2003; Maurer et al., 2005,
2007). The ocular dominance imbalance induced by early altered monocular inputs
can be efficiently treated by occluding the dominant eye for prolonged periods of time
(occlusion therapy) but only within the critical period (Webber and Wood, 2005),
whereas only modest improvements of visual function are observed when therapy is
performed in adulthood (Fronius et al., 2014).
While neuroplasticity is preserved in adulthood for higher-level functions, endorsing
learning and memory (Fuchs and Flugge, 2014), the absence of experience-dependent
changes observed after the closure of the critical period led researchers to consider
the visual system, and in particular early visual cortex, as hard-wired and with no
spared plastic potential, especially for ocular dominance. However, several recent lines
of evidence have put this assumption into question: in animal models, ocular
dominance plasticity can be reactivated after the closure of the critical period by
manipulating the visual cortex excitability, either pharmacologically or through
environmental enrichment and physical activity (Baroncelli et al., 2010, 2012; Berardi
et al., 2015; Harauzov et al., 2010; He et al., 2006; Hensch and Fagiolini, 2005; Hensch
et al., 1998; Maya Vetencourt et al., 2008; Pizzorusso et al., 2002; Sale et al., 2007,
2014; Spolidoro et al., 2009). In adult humans, evidence of preserved visual plasticity
has been demonstrated by behavioral and neural changes associated with perceptual
learning (Beyeler et al., 2017; Dosher and Lu, 2017; Fiorentini and Berardi, 1980;
Watanabe and Sasaki, 2015), short term visual deprivation (Binda et al., 2018; Binda
and Lunghi, 2017; Lunghi et al., 2011, 2013, 2015a,b, 2018; Lunghi and Sale, 2015;
Zhou et al., 2013, 2014), progressive blinding pathologies and visual restoration
therapies (Aguirre et al., 2016; Baseler et al., 2011a; Burton, 2003; Castaldi et al.,
2016; Cunningham et al., 2015a,b; Dormal et al., 2015; Heimler et al., 2014).
feedback cycles might
The cellular mechanisms promoting neuroplasticity during development have been
extensively studied in animals (Smith et al., 2009). Cortical wiring in developing neural
systems is determined both by molecular cues, responsible for neural migration and
formation of synaptic contacts, as well as by activity-dependent mechanisms that fine
tune and optimize the number and strength of synaptic connections through Hebbian
plasticity (Levelt and Hubener, 2012). This is mediated by Long Term Potentiation and
Depression (LTP/LTD), mechanisms that have a crucial role also in learning and
memory (Malenka and Bear, 2004) and in defining the critical period (Berardi et al.,
2000; Hensch and Quinlan, 2018; Levelt and Hubener, 2012). In addition Hebbian
plasticity, another form of plasticity (named homeostatic plasticity) is acting to maintain
the overall balance of the network excitability (Maffei and Turrigiano, 2008a). This may
have an important role given that LTP/LTD themselves might lead to destabilization of
neural circuits: continuous
the progressive
strengthening/weakening of synaptic connections and to its consequent excessive
excitability/loss (Turrigiano and Nelson, 2000). Through mechanisms such as synaptic
scaling, synaptic redistribution and changes in neural excitability, the homeostatic
plasticity promotes stability in the neural circuits by readjusting the overall level of
network activity to optimize responses to sensory experiences and perturbations
(Abbott and Nelson, 2000; Maffei and Turrigiano, 2008a,b; Turrigiano, 2012; Turrigiano
and Nelson, 2000). After short-term deprivation and during the critical period
homeostatic plasticity is boosting the signals from the deprived eye, in the attempt to
contrast the deprivation effects. Recent evidence suggests that this homeostatic
synaptic rescaling may remain active even after the closure of the critical period and
support neuroplasticity throughout the life span. Interestingly, the visual cortex
maintains homeostasis by normalizing the individual neural responses with the overall
activity of a pool of neurons. This mechanisms is especially important to efficiently
control contrast gain (Carandini and Heeger, 2011; Heeger, 1992). It is possible that
homeostatic plasticity in adult humans may be implemented using similar cellular
mechanisms mediating contrast gain (see later Lunghi et al., 2011)
lead
to
Here, we review evidence for adult plasticity and we speculate on the most likely
mechanisms and principles supporting the residual ocular dominance plasticity in
adulthood. Deeply understanding the mechanisms regulating adult visual plasticity is
crucial for visual restoration in late-blind individuals, considering that the various
intervention (retinal prostheses but also pharmacological manipulations) are
implemented monocularly: in principle the restored monocular signals might be gated
at cortical level if ocular dominance plasticity cannot be endorsed.
We start reviewing recent studies reporting perceptual and neural changes induced
in healthy sighted individuals by temporary altering their visual experience and how
these findings have recently led to the development of a non-invasive training for adult
amblyopic patients. We discuss how these phenomena likely reflect homeostatic
plasticity in adulthood. In the second part we discuss the neural changes following
blinding diseases and in particular the plastic retinotopic remapping of the residual
visual input and the cross-modal reorganization of the visual brain when the visual
input is altered or interrupted. Finally, we describe the success and falls of the very
first attempts of restoring vision in adult blind patients.
2. Cortical plasticity in sighted adults revealed by short-
term visual deprivation
2.1 Behavioral proxies for plasticity
Recent studies have introduced new behavioral techniques to infer and estimate the
degree of residual plasticity for ocular dominance in adult sighted subjects, by
combining binocular rivalry (Lunghi et al., 2011) and, more recently, pupillometry
(Binda and Lunghi, 2017) with short periods of monocular deprivation. Binocular rivalry
is a form of bistable perception that is generated whenever two incompatible images
are separately projected to the eyes. In such condition the visual perception alternates
between the two monocular images which take turn in dominating visual awareness
(Blake and Logothetis, 2002; Levelt, 1965). Binocular rivalry is one of the most robust
psychophysical methods used to assess sensory eye dominance (Ooi and He, 2001):
under normal conditions, the average time in which the image presented to each eye
dominates the observer's perception is similar for the two eyes, reflecting balanced
ocular dominance. Lunghi et al. (2011) first observed that visually depriving healthy
adult individuals though monocular occlusion with a translucent patch for 150 minutes
profoundly altered ocular dominance measured with binocular rivalry. Surprisingly the
stimulus presented to deprived eye dominates twice as long as the one displayed in
the non-deprived eye after deprivation. The effect, although progressively attenuated,
lasted up to 90 or 180 minutes after patch removal (Lunghi et al., 2013). After patch
removal the apparent contrast increased for a short time, suggesting an up-regulation
of the contrast gain-control mechanisms in the occluded eye that boosted the neuronal
responses to compensate for the reduced incoming signal. The altered dynamics of
binocular rivalry and the contrast gain enhancement after monocular deprivation
suggested the existence of a spared plastic potential in adult individuals: these
phenomena most likely reflect a form of homeostatic plasticity attempting to maintain
the overall network activity stable and to optimize the individual's new visual
experience.
Even though the perceptual effects of short-term monocular deprivation might be in
principle interpreted as contrast adaptation, growing evidence indicates that they
reflect a genuine form of plasticity. For example, even though deprivation alters
apparent contrast, the 36% boost in apparent contrast is not sufficient to explain the
change in ocular dominance (Lunghi et al., 2011), and when chromatic vision is
specifically targeted (using iso-luminant visual stimuli), the effect can out-last the
duration of deprivation, lasting for up to 3h (Lunghi et al., 2013). Moreover, two recent
studies (Bai et al., 2017; Ramamurthy and Blaser, 2018) have shown that changes in
ocular dominance can be observed by altering the monocular input without reducing
monocular contrast, pointing to a crucial role of inter-ocular correlation in mediating
ocular dominance plasticity. This is consistent with the etiology of amblyopia which can
be produced by strabismus (Kiorpes et al., 1998) where both eye receive good vision
but with no spatial congruency. Another direct proof that the transient changes in ocular
dominance are an expression of homeostatic plasticity is given by the long term
improvement of vision in adult amblyopic patients that received a short-term monocular
deprivation of the amblyopic eye for 6 short sessions. The improvement in visual acuity
of about 2 lines lasted for up to one year, indicating long-term neuroplastic changes
(Lunghi et al., 2018).
Binda & Lunghi (2017) more recently tackled another biomarker of neuroplasticity in
adult humans by demonstrating that monocular deprivation affects spontaneous low
frequency oscillations of the pupil diameter at rest, a phenomenon called hippus
(Diamond, 2001). The authors measured pupillary oscillations before and after
monocular occlusion, following the same paradigm as Lunghi et al. (2011), and found
that hippus amplitude increased after visual deprivation and that participants with more
pronounced pupillary fluctuations also showed stronger ocular dominance changes in
binocular rivalry dynamics. The procedures to measure these two proxies of
neuroplasticity are completely non-invasive and suitable for application in clinical
populations.
2.2 Short-term visual deprivation alters visual neural responses
A substantial number of studies investigating the cortical effects of visual deprivation
in adult humans suggested that plasticity might be mediated by changes in the
excitation/inhibition balance of the visual cortex (Binda et al., 2018; Boroojerdi et al.,
2000; Fierro et al., 2005; Lou et al., 2011; Lunghi et al., 2015a; Pitskel et al., 2007).
The first studies (Boroojerdi et al., 2000; Fierro et al., 2005; Pitskel et al., 2007)
applied TMS pulses to the occipital cortex sufficient to elicit light perception
(phosphene) in absence of visual stimulation. The minimum intensity needed to elicit
phosphene perception (PT, phosphene threshold) is an indirect measure of cortical
excitability. Boroojeredi et al. (2000) showed that after 45 minutes of binocular light
deprivation PT was reduced in healthy adult subjects, suggesting increased cortical
excitability and this effect persisted for the entire deprivation period (180 minutes). Re-
exposure to light reverted the process and PT returned to pre-deprivation values within
120 minutes. Interestingly the same study provided also another measure of cortical
responsiveness, independent of the subjects' perception: neural activity in striate and
extrastriate cortices as measured by BOLD responses with functional magnetic
imaging (fMRI) was enhanced after 60 minutes of blindfolding and the increased fMRI
signal persisted for at least 30 minutes after re-exposure to light.
Monocular deprivation might cause different cortical effects with respect to
blindfolding, possibly due to inter-ocular competition which is absent in case of
binocular deprivation. In fact, evidence from animal studies has shown that, compared
with monocular deprivation, binocular deprivation performed during the critical period
induces a more modest reorganization of visual cortical circuits (Wiesel and Hubel,
1965a). In adult humans, monocular deprivation was shown to cause a decrease in
cortical excitability, rather than an increase, as measured by TMS PT (Lou et al., 2011).
However a recent study measuring pattern-onset visual evoked potentials (VEPs)
before and after monocular deprivation described a more complex pattern of results
with opposite effects for the two eyes (Lunghi et al., 2015a). The amplitude of the C1
component of the VEP responses, typically reflecting the earliest stage of visual
processing in V1 (Di Russo et al., 2002), and the peak in alpha band after monocular
deprivation were enhanced for the deprived eye, but reduced for the non-deprived eye.
Moreover the amplitude of later components, such as P1 and P2, were equally altered
by monocular deprivation, suggesting that the variations in cortical excitability
propagate to extrastriate areas and modulate feed-back projection to V1. Overall these
results, in line with the perceptual changes observed after short-term monocular
deprivation (described in section 2.1), suggest that the deprived and non-deprived eye
are respectively strengthened or weakened after monocular deprivation, reflecting
antagonistic homeostatic short-term plasticity in the two eyes.
Recently Binda et al. (2018) exploited the enhanced resolution and signal to noise
ratio provided by ultra-high field (7T) fMRI, to track ocular driven changes of BOLD
responses in V1 before and after 2h of monocular deprivation. During the scanning
participants' eyes were separately stimulated with either high contrast low- and high-
band-pass noise (optimized to differentially stimulate the magno- and parvocellular
pathways respectively) or with a luminance matched uniform background (Fig.1A).
BOLD responses to the high spatial frequency stimuli were strongly affected by
monocular deprivation, in opposite directions for the two eyes: the percentage BOLD
signal measured in V1 elicited by stimulation of the deprived eye and non-deprived eye
was respectively enhanced and reduced with respect to pre-deprivation values (Fig
1B). The authors further calculated for each voxel an index of Ocular Dominance
defined as the response difference to the deprived and non-deprived eye, which
reflects the eye preference of a given particular voxel (average biased signal), although
not strictly coinciding with the ocular dominance columns. Before deprivation, the
ocular dominance indices were symmetrically distributed around zero, reflecting
balanced V1 activity elicited by the two eyes. However, after deprivation the index
strongly shifted toward a stronger activation from the deprived eye. Interestingly the
voxels originally preferring the deprived eye did not change their preference (i.e. the
average signal measured in these voxels continued to be biased toward the same eye),
whereas the voxels originally preferring the non-deprived eye swapped their
preference and were most strongly activated by stimulation of the deprived eye.
Importantly the deprivation effect in BOLD responses correlated with the perceptual
effect of deprivation as measured by increased mean phase duration for the deprived
eye during binocular rivalry performed outside the scanner with the classic paradigm
(Fig. 1 C, D). Spatial frequency selectivity in V1 was also affected by deprivation: only
responses elicited by high-spatial frequencies stimulation of the deprived eye were
significantly reduced, while responses to low-spatial frequencies stimulation of the
same eye as well as to the whole frequencies range of the non-deprived eye were
unaffected. Population Spatial Frequency Tuning of V1 confirmed these effects and
the shift of selectivity toward higher spatial frequency in the deprived eye correlated
with the phase duration during binocular rivalry. Interestingly, the increase in BOLD
response to the high-spatial frequency stimulation of the deprived eye was strongest
in V1, V2, V3, attenuated but still significant in V4, while it was absent in V3a and
hMT+. The enhanced signals measured along the ventral but not along the dorsal
pathway suggested that plasticity acted more strongly on the parvocellular rather than
magnocellular pathway. This interpretation is also supported by the evidence that
deprivation-induced changes in binocular rivalry dynamics with chromatic equi-
luminant gratings resulted in much longer-lasting effects with respect to those
measured with luminance-modulated gratings (Lunghi et al., 2013).
In sum, these experiments showed that complex patterns of neuroplastic responses
can be induced in the visual cortex of sighted adults by short periods of visual
deprivation, which rapidly alter the visual cortex's excitability and responsiveness to
homeostatically adapt to the altered visual experience.
2.3 Neurochemical changes
following short-term visual
deprivation
The molecular mechanisms underlying experience-dependent plasticity have been
widely investigated in animals (Berardi et al., 2003; Heimel et al., 2011). The
maturation of intracortical inhibition has been proved to play a crucial role in regulating
the progression of critical period for ocular dominance and visual acuity (Berardi et al.,
2003; Fagiolini et al., 2004; Fagiolini and Hensch, 2000; Hensch et al., 1998; Huang
et al., 1999; Speed et al., 1991). For instance, increasing intracortical inhibition was
shown to anticipate the opening and closure of the critical period for monocular
deprivation in mice (Fagiolini and Hensch, 2000; Hanover et al., 1999; Huang et al.,
1999), and transgenic animals lacking a GABA-synthesizing enzyme showed deficient
ocular dominance plasticity to monocular deprivation which could be restored by
increasing inhibitory transmission with benzodiazepines (Hensch et al., 1998). In
adulthood, the increased inhibition may be a limiting factor for cortical plasticity and
reducing GABAergic inhibition was shown to partially restore ocular dominance
plasticity in adult rats and promote recovery from amblyopia (Harauzov et al., 2010;
Maya Vetencourt et al., 2008). Interestingly, environmental enrichment and physical
activity were recently found to be associated with reduced inhibitory tone in the rat's
visual cortex, providing a potential non-invasive strategy to promote recovery from
amblyopia (Baroncelli et al., 2010, 2012; Sale et al., 2007; Stryker, 2014).
In line with these results, one study recently showed that GABAergic inhibition plays
a key role in promoting ocular dominance plasticity also in adult humans (Lunghi et al.,
2015b). Participants underwent one psychophysical test measuring ocular dominance
by means of binocular rivalry and one 7T MR spectroscopy session before and after
150 min of monocular occlusion (same procedure as described in section 2.1). The
perceptual changes triggered by deprivation (resulting in the deprived eye dominating
over the non-deprived eye) were associated with decreased GABA concentration in
V1, at least when GABA concentration was assessed at rest, while participants kept
their eyes closed. Specifically, participants with greater decrease in resting GABA
concentration showed the greatest perceptual effects, i.e. perceptual boost and
dominance of the deprived eye. Interestingly, other studies showed that fMRI
responses in visual cortex inversely correlated with GABA concentration, potentially
suggesting a link between GABA level and cortical excitation-inhibition balance
(Donahue et al., 2010; Muthukumaraswamy et al., 2009).
Of course GABAergic circuits are not the only ones mediating plasticity in visual
cortex (Berardi et al., 2003). In animals, ocular dominance plasticity can be restored
also by enhancing excitatory neurotransmission systems such as serotoninergic (Maya
Vetencourt et al., 2008) and cholinergic (Morishita et al., 2010) systems. In humans
Boroojeredi et al. (2001) applied TMS over occipital cortex and measured phosphene
thresholds in adult sighted participants under the effect of various drugs interfering with
synaptic plasticity before and after a period of light deprivation. They found that the
rapid plastic changes typically triggered by light deprivation (here quantified as
phosphene detection thresholds) were blocked when participants were under the effect
of lorazepam (which enhance the functioning of GABAa receptors), dextromethorphan
(NMDA receptors antagonist) and scopolamine (a muscarinic receptor antagonist),
thus pointing at a key role of GABA, NMDA and cholinergic receptors in mediating
rapid plastic changes after visual deprivation.
Finally, in light of the recent results by Binda & Lunghi (2017) showing both
increased pupillary hippus at rest and enhanced eye-dominance of the deprived eye
during binocular rivalry after monocular deprivation, the role of the neurotransmitter
norepinephrine (NE) in mediating homeostatic plasticity should be further investigated
in adult humans. This neurotransmitter may indeed constitute the common source of
visual cortical excitability underlying these phenomena, given its known role in
regulating both pupil diameter modulation (Joshi et al., 2016) and visual cortical
plasticity (Kasamatsu et al., 1979, 1981).
Overall, these studies point at a spared plastic potential in the adult visual cortex
beyond the critical period which can be reactivated by altering the excitability level of
visual cortex either by pharmacologically targeting several neuromodulator systems or
by manipulating sensory and motor experience, such as being exposed to abnormal
visual experiences even for a short time period, or being exposed to an enriched
environment and physical activity.
2.4 Therapy for amblyopia
Inspired by studies on animal models showing that physical exercise triggers visual
cortical plasticity, modulates visual cortex excitability and increases neurotrophic
factors (Baroncelli et al., 2010, 2012; Sale et al., 2007), recent studies tested whether
similar effects could be obtained in adult humans (Lunghi and Sale, 2015; Lunghi et
al., 2018). Lunghi & Sale (2015) tested binocular rivalry dynamics in sighted
participants before and after monocular deprivation (still with the same paradigm
described in 2.1) while varying the level of physical activity performed by participants
during the deprivation period. In the 'inactive condition' participants were required to
watch a movie while sitting on a chair, while in the 'physical activity' condition they
watched a movie while intermittently cycling on a stationary bike. With respect to the
control inactive condition, the perceptual effect of deprivation on binocular rivalry
dynamic was much stronger when participant performed physical activity throughout
the 2h period tested after eye-patch removal, suggesting that physical activity had
further boosted homeostatic plasticity.
The beneficial effect of moderate physical activity for triggering neuroplasticity was
recently combined with short-term inverse occlusion to promote the visual recovery in
adult anisometropic amblyopes (Lunghi et al., 2018). Six 2h long training sessions over
a 4 weeks period consisting in simultaneous physical activity (intermitted cycling) and
occlusion of the amblyopic eye restored visual acuity in all patients and stereopsis in
six of them with improvement lasting up to 1 year. Although these results should be
replicated in a larger sample, this non-invasive training paradigm seemed to
successfully boost visual plasticity in adulthood and to constitute a valid approach to
treat amblyopia.
3. Cortical plasticity after vision
loss
following
ophthalmological diseases
3.1 Retinotopic remapping of the visual cortex
The blind brain certainly constitute a unique opportunity for studying plasticity in
adulthood. A highly studied phenomena, sometimes considered as an index of
neuroplasticity, is the ability of the visual cortex to remap the retinotopic organization
of the neuronal receptive fields following retinal lesions (for a recent review see:
Dumoulin and Knapen, 2018). The first evidence for this phenomenon was reported in
adult cats and monkeys (Gilbert and Wiesel, 1992): after retinal lesion V1 receptive
fields near the border of the lesion projection zone (LPZ) underwent an immediate
enlargement, and two months after the regions silenced by the lesion were found to
represent loci surrounding the scotoma. The receptive fields shift leading to complete
filling in of the scotoma was attributed to long-range horizontal connections within V1
rather than to spared geniculate afferent connections: none of the changes observed
in the recovered cortex where observed in the lateral geniculate nucleus, perhaps due
to its reduced plasticity with respect to higher level cortical regions (however see
section 3.3).
More recently these findings have been strongly questioned by a combined
neurophysiological and fMRI study that failed to record normal responsivity in adult
macaque V1 during 7.5 months of follow-up after retinal lesion (Smirnakis et al., 2005).
The authors found no change in the BOLD-defined LPZ border and suggested that
cortical reorganization is not needed to explain the apparent size of the LPZ (Smirnakis
et al., 2005). Incongruences with respect to previous results were attributed to several
factors, including sampling biases and differences in the recording methods (one
recording sub-sets of single neurons, the other reflecting average activation of
ensembles of cells) and originated an intense debate (for details on this debeat see:
Calford et al., 2005; Sereno, 2005; Smirnakis et al., 2005; Wandell and Smirnakis,
2009).
The ability of the adult visual cortex to reorganize has also been studied in retinal
dystrophies, namely macular degeneration (MD, both age-related MD and juvenile MD)
and retinitis pigmentosa (RP). Both diseases create scotomas in patients' central (MD)
and peripheral (RP) visual field that expand with illness progression. Some fMRI
studies described large-scale cortical reorganization of visual processing in response
to retinal disease by showing that the regions of V1 matching the patients' scotoma
are remapped to respond to stimuli outside it both in MD (Baker et al., 2005, 2008;
Dilks et al., 2009, 2014; Schumacher et al., 2008) and in RP patients (Ferreira et al.,
2017). However, others have found the LPZ to remain silent (Ritter et al., 2018;
Sunness et al., 2004) and found no evidence for large scale remapping of the visual
cortex following late blindness (Baseler et al., 2011b; Goesaert et al., 2014; Haak et
al., 2015). Masuda et al. (2008; 2010) found that the possibility to detect BOLD signals
in LPZ depends on task and might reflect the upregulation or unmasking of feed-back
projections from extrastriate areas to V1 that are normally suppressed in sighted
individuals. These authors proposed that the presence of activations in LPZ might not
need to reflect cortical reorganization. Similar conclusions were reached by studies
that simulated scotomas in sighted humans artificially by removing the visual stimulus
in a given location of a rich image. This studies found consistently altered receptive
fields properties around the simulated scotoma (Binda et al., 2013; Dumoulin and
Knapen, 2018; Haak et al., 2012).
In sum, studies in patients with retinal dystrophies have reported mixed results
regarding the presence of remapped visual activation in LPZ. Identifying the origin of
such activations remains matter of open debate: it can reflect cortical reorganization
leading to the formation of new connections or changes in the connections strength
between neurons or to unmasking of existing connections normally suppressed in
sighted individuals.
3.2 Cross-modal plasticity in blind individuals
Another widely studied phenomena, considered a marker of experience-dependent
neuroplasticity, is the cross-modal reorganization of the visual cortex in blind
individuals: several studies observed occipital activations elicited by non-visual
sensory stimulation in congenital and early-onset blinds (Amedi et al., 2010; Burton,
2003; Collignon et al., 2009, 2011b; Ptito et al., 2005). The mechanisms leading such
cross-modal reorganization are still under debate. Some authors proposed that cross-
modal activations in blinds reflect unmasking of 'latent' cross-modal connections that
are normally suppressed (Merabet et al., 2007, 2008; Pascual-Leone and Hamilton,
2001), while others pointed at an additive shifts of the cross-modal responses in early-
blind individuals rather than at rescaling or unmasking processes (Fine and Park, 2018;
Lewis et al., 2010). Interestingly the cross-modal reorganization of typically visual
areas is not random, but seems to rather reflect a supramodal functional organization
of the brain, encoding for an abstract representation of the perceived stimuli
independently from the sensory input (Pietrini et al., 2004; Ricciardi et al., 2014a,b;
Ricciardi and Pietrini, 2011). For example, responses in the typically visual motion area
MT+ can be elicited by motion-specific auditory and tactile stimulation in early-blinds
and sight recovered individuals (Ricciardi et al., 2007; Saenz et al., 2008). However
the overlap between cross-modal responses should be taken carefully and further
investigated to exclude artefactual co-localizations of cross-modal responses within
the same area (Jiang et al., 2015). Yet, signals reflecting functional selective cross-
modal plasticity have been reported by several studies across different cortical areas
and cognitive functions (such as object recognition, stimuli localization in the space,
reading) and seems to be a general reorganizational principle of the brain
independently from the deprived modality (Heimler et al., 2014): areas typically
involved in spoken language processing for example, are recruited by sign language
in deaf people (MacSweeney et al., 2008).
The presence of cross-modal responses in late-onsets blind individuals is more
controversial. For instance, Bedny et al. (2010) found activations elicited by auditory
motion stimuli in hMT+ only in congenital, but not in late-blind subjects. Similarly the
perceived direction of auditory moving stimuli could be classified in early, but not in
late-blind individuals (Jiang et al., 2014, 2016) using multi voxel pattern analysis.
However, there is also evidence for the cross-modal reorganization to take place when
blindness develops late in life: although reduced with respect to those observed in
early-blind patients, both auditory and tactile activations of the visual cortices have
been described in late-blinds (Burton, 2003; Cunningham et al., 2011, 2015a,b; Voss
et al., 2006), and one study showed that the extent and strength of tactile-evoked
responses in V1 correlate with vision loss in late-blind individuals affected with retinitis
pigmentosa (RP) (Cunningham et al., 2011, 2015b).
The underlying anatomical pathways supporting cross-modal responses may differ
between early and late-blind individuals (Collignon et al., 2009). For example,
Collingon et al. (2013) found that the auditory activity in the occipital cortex of
congenitally blind individuals was most likely conveyed by direct connections from A1
to V1, whereas in late -- blind individuals it appeared to be conveyed by feedback
projections from multisensory parietal areas. The specific mechanisms underlying the
rerouting of non-visual information are still unclear (Voss, 2019), however they might
deeply differ depending on the age of blindness onset: while non-visual connections to
the occipital cortex in congenital/early-blind individuals might be due to the lack of
pruning typically triggered by visual experience, cross-modal responses in late-blind
individuals might reflect the unmasking of non-visual pre-existing connections
(normally supporting multisensory integration in sighted individuals) which can be
progressively reinforced and result in permanent structural changes with new synapsis
formation (Merabet et al., 2008).
Whatever the cause or the precise mechanisms underlying cross-modal plasticity
might be, it can potentially interfere with the outcome of restoration techniques
(Collignon et al., 2011a; Merabet et al., 2005). This possibility has been clearly
demonstrated in deaf children who underwent cochlear implant: sustained and
prolonged periods of deafness induced stronger cross-modal reorganization of
acoustic cortex (as measured by glucose hypometabolism) and hampered recovery
after cochlear implant (Lee et al., 2001). Although the predictive link between the extent
of cross-modal reorganization and successful outcome of vision restoration still has to
be directly demonstrated, one potential prognostic predictor of treatment effectiveness
might be V1 cortical thickness: Aguirre et al (2016) showed that this parameter is
strictly related to the strengths of cross-modal responses, independently of the patient
age and blinding pathology.
Thus, cross-modal reorganization might pose a major challenge to vision restoration
therapies, and if we take into account the lower plasticity in adulthood, it seem even
more difficult to imagine successful visual recovery in late-blinds. On the other hand
however, it is possible that sight restoration might be more feasible in late than in early-
blind individuals, because visual cortex wiring successfully occurred before the disease
in late -- blind individuals. A recent study showed that functional connectivity between
visual areas is still retinotopically organized in blind patients affected with juvenile
macular degeneration, even after prolonged period of visual deprivation, directing
further hope on the possibility to have positive outcomes from sight restoration
techniques that can rely on a relatively intact visual system (Haak et al., 2016).
Yet, while these findings are encouraging, they do not guarantee that the adult visual
brain retains the plastic potential necessary to mask the non-visual inputs potentiated
or newly created during the blindness period and to boost the responses to the restored
visual input; nor is it obvious that the visual system would 'learn to see again' with an
artificial and monocular input.
3.3 Cortical plasticity after sight restoration
The first attempts to restore vision in adulthood focused on early-blind patients
and showed limited visual recovery in these patients (Dormal et al., 2015; Fine et al.,
2003; Gregory and Wallace, 1963; Saenz et al., 2008). Patient MM became blind at
the age of three years old, and once his sight was restored in his 40ies, he was only
able to recover visual abilities strictly related to the visual experience before blindness,
such as perception of simple forms, color and motion, while perception of more
complex 3D forms, objects and faces remained severely impaired even after several
years of restored vision (Fine et al., 2003; Huber et al., 2015). fMRI studies on patient
MM and on another early-blind patient whose vision was partially restored in adulthood,
showed that cortical plasticity was also limited although not completely absent: several
months after vision restoration, cross-modal auditory responses continued to coexist
with the restored visual activations in area V1 (Dormal et al., 2015) and MT (Saenz et
al., 2008). Interestingly however Dormal et al. (2015) observed that cross-modal
responses in extra striate areas decreased after surgery and vision improved
suggesting that cross-modal reorganization can be partially reversed in their early-blind
patient.
Very few studies tracked the cortical reorganization process in late-blind patients
after vision restoration. One study tracked the cortical responses in a 80 years old
woman with wet age-related AMD undergoing intravitreal antiangiogenic injections
(ranibizumab) over about 1 year period (Baseler et al., 2011a). Microperimetry showed
that the scotoma decreased over time, visual acuity, fixation stability and reading skills
improved as well. Interestingly after the first treatment, BOLD responses elicited by
full-field flickering lights showed a tendency to be located in more posterior occipital
regions, corresponding to the patient scotoma, however without filling it completely.
Cunningham et al (2015a) tested two late-blind patients affected by retinitis
pigmentosa (RP) implanted with Argus II Retinal Prosthesis and showed that the
strength of tactile-evoked responses in V1 depended from the time from surgery: cross-
modal activations were much reduced in the patient implanted for fifteen weeks before
the scanning with respect to the patient who had the implant only for six weeks (Fig
2A). No visual responses were detected in this study, potentially due to the relatively
short time period between the implantation and the scanning. A more recent study
tested RP patients implanted with the same system at least six months after surgery
and found that visual responses to flashes of lights increased in LGN and V1 after the
surgery (Fig. 2B) (Castaldi et al., 2016). Importantly visual recovery (quantified as the
behavioral performance on a challenging detection task) depended on the time from
surgery and practice with the device (Fig. 2C) and was mirrored by enhanced BOLD
responses to flashes after surgery, suggesting that the activity measured in visual
areas had a functional relevance. Interestingly some weak visual activations were
present already before the surgery, although participants never reported perceiving the
flashing lights during the scanning. In particular, responses in extra striate areas were
stronger before the surgery, whereas V1 showed stronger activations after the surgery.
It is possible that even before surgery some spared visual input reached patients' V1,
given the consistent albeit small BOLD response observed. However these visual
signals might have been actively suppressed by extra striate areas, probably because
the visual stimulation was not sufficiently reliable, aberrant and delayed. The
suppression could have cross-sensory or motor origin, but also visual given that direct
thalamic visual input to associative cortex have repeatedly observed in normal brain
(Ajina et al., 2015; Tamietto and Morrone, 2016). Animal models reported spontaneous
ganglion cells hyperactivity during photoreceptors degeneration process (Ivanova et
al., 2016; Trenholm and Awatramani, 2015; Zeck, 2016). Suppressing such
paroxysmal discharges might be beneficial
faster cross-modal
reorganization. This possibility is in line with the idea of a dysfunctional gating
mechanism in blinds which would allow feedback projections to freely interact with the
incoming input signal (Masuda et al., 2010), even when this is noisy and aberrant and
the nature of this interaction might be inhibitory.
to prompt
Taken together the results of these experiments suggest that when visual signals
are restored in late-blinds, not only cross-modal responses in V1 needs to be
attenuated or eliminated, but also the suppression of the incoming visual input to V1
should be released, and both these processes might take a long time. Yet, t he
decreased tactile-evoked responses (Cunningham et al., 2015a) and the increased
visual BOLD signal in V1 (Castaldi et al., 2016) in patients that used the prosthesis for
longer time, suggest that a spared plastic potential is retained by the adult visual brain
and encourage continuative research to overcome the major obstacles limiting the
expected outcome from vision restoration techniques. These obstacles include, among
others, the limited quality of visual percepts that can be obtained (at best) with the
current technology - for a review on advantages and limitations of different vision
restoration methodologies see: Fine and Boynton (2015). For example, simulation of
the likely perceived images with epiretinal devices showed that they might look
extremely distorted because the electrode array stimulate unspecifically axons of the
ganglion cells together with their cellular bodies (Fine and Boynton, 2015). It is thus
not surprising that patients find particularly difficult to recover complex aspects of visual
perception such as correctly discriminating the direction of drifting gratings (Castaldi et
al., 2016). Although the technology can certainly be further improved, several studies
have already shown that, after extensive training, RP patients implanted with retinal
prosthesis can learn to perform some easy task, which can be nevertheless important
for the patients' quality of life, such as moving independently in the space, locating
large sources of light and even read large highly contrasted letters (Barry et al., 2012;
Castaldi et al., 2016; Chader et al., 2009; da Cruz et al., 2013; Dorn et al., 2013; Rizzo
et al., 2014). Improvement of visual acuity and visual field perimetry were reported in
patients implanted with both epiretinal and subretinal implants (Chow 2013; Chow et
al., 2004, 2010; Rizzo et al., 2014, 2015). Interestingly visual field improvement
extended outside the retinotopic regions directly stimulated by the implant (Rizzo et al.,
2014) and even in the unoperated eye (Rizzo et al., 2015). The fact that the visual field
recovering is not strictly limited to the stimulation site might suggest that the
neuroplastic response acts 'peripherally' and originate from the release of retinal
trophic factors induced by the current injection which diffuse to non-stimulated regions
of the retina (Ciavatta et al., 2009). However it is also possible that the artificial vision
provided by retinal implants reopens visual plasticity at 'central level', which would
better explain the visual recovery of the fellow eye observed at least in one study (Rizzo
et al., 2015). Perhaps the fellow eye benefits from an anterograde effect mediated by
cross talk of neural discharges along the optic nerve after the chiasma or at the level
of LGN where the projections from the two eyes are closely interlayered. Interestingly
some increase of BOLD response have also been observed at LGN level after
prolonged use of the prosthetic devise (Castaldi et al., 2016).
In sum, despite long time and extensive training is needed to recover a functional
use of artificial vision and although the specific source of such neuroplastic responses
is currently unknown, the reviewed results suggest that vision restoration techniques
can rely on residual neuroplasticity retained by the adult brain and that, especially for
late-blinds, it might be possible to restore vision even after several years of blindness.
4. Conclusions
We reviewed behavioral and neural evidence suggesting that a considerable degree
of neuroplasticity is preserved well beyond the closure of the critical periods for vision.
Behavioral and neural evidence indicates that in sighted individual short periods of
monocular deprivation can trigger homeostatic plasticity and that this strategy can be
successfully used to improve visual perception in amblyopic adults. Evidence for cross-
modal reorganization of the visual brain of blind individuals is compelling and recent
findings suggest that this process can be reverted even after several years of
blindness. However it is important to acknowledge the fact that this neuroplasticity is a
very slow process and that the quality of the regained visual perception is limited. Yet,
the field of visual restoration techniques is still at the beginning and much improvement
can be expected in the upcoming years. Certainly neuroscientists can contribute to the
development of this field by exploring the properties, characteristics and mechanisms
of the spared plastic potential retained by the human visual brain to calibrate optimally
vision restoration therapies and maximize their expected outcome.
Perhaps, the most outstanding question to answer is whether we can boost
neuroplasticity to make visual recovery more effective and rapid. Although the answer
might depend on the age of blindness onset and on the cause determining the visual
deficit, identifying some common principles and mechanisms guiding neuroplasticity
during development and adulthood and across modalities mighty lead to develop
strategies effective under a wide range of circumstances and diseases. Hopefully in
the next decades we will have the complete approach to rescue visual function,
facilitating the adult brain to learn to "see" again, even after several years of blindness
or abnormal visual experience.
Acknowledgments
This research has received funding from Fondazione Roma under the Grants for
Biomedical Research: Retinitis Pigmentosa (RP)-Call for proposals 2013 (http://
www.fondazioneroma.it/it/index.html, http://wf-fondazioneroma.cbim.it/), project title:
Cortical Plasticity in Retinitis Pigmentosa: an Integrated Study from Animal Models to
Humans, from the European Research Council under the European Union's Seventh
Framework Programme (FPT/2007-2013) under grant agreement No. 338866
ECPLAIN (http://www.pisavisionlab.org/index.php/ projects/ecsplain) and under the
European Union's Horizon 2020 research and innovation programme (grant
agreement No 801715 -- PUPILTRAITS).
References
Abbott, L.F., Nelson, S.B., 2000. Synaptic plasticity: taming the beast. Nature neuroscience 3 Suppl,
1178-1183.
Aguirre, G.K., Datta, R., Benson, N.C., Prasad, S., Jacobson, S.G., Cideciyan, A.V., Bridge, H., Watkins,
K.E., Butt, O.H., Dain, A.S., Brandes, L., Gennatas, E.D., 2016. Patterns of Individual Variation in Visual
Pathway Structure and Function in the Sighted and Blind. Plos One 11, e0164677.
Ajina, S., Pestilli, F., Rokem, A., Kennard, C., Bridge, H., 2015. Human blindsight is mediated by an intact
geniculo-extrastriate pathway. eLife 4.
Amedi, A., Raz, N., Azulay, H., Malach, R., Zohary, E., 2010. Cortical activity during tactile exploration
of objects in blind and sighted humans. Restor Neurol Neurosci 28, 143-156.
Bai, J., Dong, X., He, S., Bao, M., 2017. Monocular deprivation of Fourier phase information boosts the
deprived eye's dominance during interocular competition but not interocular phase combination.
Neuroscience 352, 122-130.
Baker, C.I., Dilks, D.D., Peli, E., Kanwisher, N., 2008. Reorganization of visual processing in macular
degeneration: replication and clues about the role of foveal loss. Vision Res 48, 1910-1919.
Baker, C.I., Peli, E., Knouf, N., Kanwisher, N.G., 2005. Reorganization of visual processing in macular
degeneration. The Journal of neuroscience : the official journal of the Society for Neuroscience 25,
614-618.
Baroncelli, L., Bonaccorsi, J., Milanese, M., Bonifacino, T., Giribaldi, F., Manno, I., Cenni, M.C., Berardi,
N., Bonanno, G., Maffei, L., Sale, A., 2012. Enriched experience and recovery from amblyopia in adult
rats: impact of motor, social and sensory components. Neuropharmacology 62, 2388-2397.
Baroncelli, L., Sale, A., Viegi, A., Maya Vetencourt, J.F., De Pasquale, R., Baldini, S., Maffei, L., 2010.
Experience-dependent reactivation of ocular dominance plasticity in the adult visual cortex.
Experimental neurology 226, 100-109.
Barry, M.P., Dagnelie, G., Argus, I.I.S.G., 2012. Use of the Argus II retinal prosthesis to improve visual
guidance of fine hand movements. Invest Ophthalmol Vis Sci 53, 5095-5101.
Baseler, H.A., Gouws, A., Crossland, M.D., Leung, C., Tufail, A., Rubin, G.S., Morland, A.B., 2011a.
Objective visual assessment of antiangiogenic treatment for wet age-related macular degeneration.
Optometry and vision science : official publication of the American Academy of Optometry 88, 1255-
1261.
Baseler, H.A., Gouws, A., Haak, K.V., Racey, C., Crossland, M.D., Tufail, A., Rubin, G.S., Cornelissen, F.W.,
Morland, A.B., 2011b. Large-scale remapping of visual cortex is absent in adult humans with macular
degeneration. Nature neuroscience 14, 649-655.
Bedny, M., Konkle, T., Pelphrey, K., Saxe, R., Pascual-Leone, A., 2010. Sensitive period for a multimodal
response in human visual motion area MT/MST. Curr Biol 20, 1900-1906.
Berardi, N., Pizzorusso, T., Maffei, L., 2000. Critical periods during sensory development. Curr Opin
Neurobiol 10, 138-145.
Berardi, N., Pizzorusso, T., Ratto, G.M., Maffei, L., 2003. Molecular basis of plasticity in the visual cortex.
Trends in neurosciences 26, 369-378.
Berardi, N., Sale, A., Maffei, L., 2015. Brain structural and functional development: genetics and
experience. Developmental medicine and child neurology 57 Suppl 2, 4-9.
Beyeler, M., Rokem, A., Boynton, G.M., Fine, I., 2017. Learning to see again: biological constraints on
cortical plasticity and the implications for sight restoration technologies. Journal of neural engineering
14, 051003.
Binda, P., Kurzawski, J.W., Lunghi, C., Biagi, L., Tosetti, M., Morrone, M.C., 2018. Response to short-
term deprivation of the human adult visual cortex measured with 7T BOLD. eLife 7.
Binda, P., Lunghi, C., 2017. Short-Term Monocular Deprivation Enhances Physiological Pupillary
Oscillations. Neural plasticity 2017, 6724631.
Binda, P., Thomas, J.M., Boynton, G.M., Fine, I., 2013. Minimizing biases in estimating the
reorganization of human visual areas with BOLD retinotopic mapping. Journal of vision 13, 13.
Blake, R., Logothetis, N., 2002. Visual competition. Nat Rev Neurosci 3, 13-21.
Boroojerdi, B., Battaglia, F., Muellbacher, W., Cohen, L.G., 2001. Mechanisms underlying rapid
experience-dependent plasticity in the human visual cortex. Proceedings of the National Academy of
Sciences of the United States of America 98, 14698-14701.
Boroojerdi, B., Bushara, K.O., Corwell, B., Immisch, I., Battaglia, F., Muellbacher, W., Cohen, L.G., 2000.
Enhanced excitability of the human visual cortex induced by short-term light deprivation. Cerebral
cortex 10, 529-534.
Braddick, O., Atkinson, J., 2011. Development of human visual function. Vision Res 51, 1588-1609.
Burton, H., 2003. Visual cortex activity in early and late blind people. The Journal of neuroscience : the
official journal of the Society for Neuroscience 23, 4005-4011.
Calford, M.B., Chino, Y.M., Das, A., Eysel, U.T., Gilbert, C.D., Heinen, S.J., Kaas, J.H., Ullman, S., 2005.
Neuroscience: rewiring the adult brain. Nature 438, E3; discussion E3-4.
Carandini, M., Heeger, D.J., 2011. Normalization as a canonical neural computation. Nat Rev Neurosci
13, 51-62.
Castaldi, E., Cicchini, G.M., Cinelli, L., Biagi, L., Rizzo, S., Morrone, M.C., 2016. Visual BOLD Response in
Late Blind Subjects with Argus II Retinal Prosthesis. PLoS Biol 14, e1002569.
Chader, G.J., Weiland, J., Humayun, M.S., 2009. Artificial vision: needs, functioning, and testing of a
retinal electronic prosthesis. Progress in brain research 175, 317-332.
Chow, A.Y., 2013. Retinal Prostheses Development in Retinitis Pigmentosa Patients-Progress and
Comparison. Asia-Pacific journal of ophthalmology 2, 253-268.
Chow, A.Y., Bittner, A.K., Pardue, M.T., 2010. The artificial silicon retina in retinitis pigmentosa patients
(an American Ophthalmological Association thesis). Transactions of the American Ophthalmological
Society 108, 120-154.
Chow, A.Y., Chow, V.Y., Packo, K.H., Pollack, J.S., Peyman, G.A., Schuchard, R., 2004. The artificial silicon
retina microchip for the treatment of vision loss from retinitis pigmentosa. Archives of ophthalmology
122, 460-469.
Ciavatta, V.T., Kim, M., Wong, P., Nickerson, J.M., Shuler, R.K., Jr., McLean, G.Y., Pardue, M.T., 2009.
Retinal expression of Fgf2 in RCS rats with subretinal microphotodiode array. Invest Ophthalmol Vis Sci
50, 4523-4530.
Collignon, O., Champoux, F., Voss, P., Lepore, F., 2011a. Sensory rehabilitation in the plastic brain.
Progress in brain research 191, 211-231.
Collignon, O., Dormal, G., Albouy, G., Vandewalle, G., Voss, P., Phillips, C., Lepore, F., 2013. Impact of
blindness onset on the functional organization and the connectivity of the occipital cortex. Brain : a
journal of neurology 136, 2769-2783.
Collignon, O., Vandewalle, G., Voss, P., Albouy, G., Charbonneau, G., Lassonde, M., Lepore, F., 2011b.
Functional specialization for auditory-spatial processing in the occipital cortex of congenitally blind
humans. Proceedings of the National Academy of Sciences of the United States of America 108, 4435-
4440.
Collignon, O., Voss, P., Lassonde, M., Lepore, F., 2009. Cross-modal plasticity for the spatial processing
of sounds in visually deprived subjects. Experimental brain research 192, 343-358.
Cunningham, S.I., Shi, Y., Weiland, J.D., Falabella, P., Olmos de Koo, L.C., Zacks, D.N., Tjan, B.S., 2015a.
Feasibility of Structural and Functional MRI Acquisition with Unpowered Implants in Argus II Retinal
Prosthesis Patients: A Case Study. Transl Vis Sci Technol 4, 6.
Cunningham, S.I., Weiland, J.D., Bao, P., Lopez-Jaime, G.R., Tjan, B.S., 2015b. Correlation of vision loss
with tactile-evoked V1 responses in retinitis pigmentosa. Vision Res 111, 197-207.
Cunningham, S.I., Weiland, J.D., Bao, P., Tjan, B.S., 2011. Visual cortex activation induced by tactile
stimulation in late-blind individuals with retinitis pigmentosa. Conference proceedings : ... Annual
International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in
Medicine and Biology Society. Annual Conference 2011, 2841-2844.
da Cruz, L., Coley, B.F., Dorn, J., Merlini, F., Filley, E., Christopher, P., Chen, F.K., Wuyyuru, V., Sahel, J.,
Stanga, P., Humayun, M., Greenberg, R.J., Dagnelie, G., Argus, I.I.S.G., 2013. The Argus II epiretinal
prosthesis system allows letter and word reading and long-term function in patients with profound
vision loss. The British journal of ophthalmology 97, 632-636.
Di Russo, F., Martinez, A., Sereno, M.I., Pitzalis, S., Hillyard, S.A., 2002. Cortical sources of the early
components of the visual evoked potential. Human brain mapping 15, 95-111.
Diamond, J.P., 2001. The Pupil. Anatomy, Physiology and Clinical Applications. The British journal of
ophthalmology 85, 121E.
Dilks, D.D., Baker, C.I., Peli, E., Kanwisher, N., 2009. Reorganization of visual processing in macular
degeneration is not specific to the "preferred retinal locus". The Journal of neuroscience : the official
journal of the Society for Neuroscience 29, 2768-2773.
Dilks, D.D., Julian, J.B., Peli, E., Kanwisher, N., 2014. Reorganization of visual processing in age-related
macular degeneration depends on foveal loss. Optometry and vision science : official publication of
the American Academy of Optometry 91, e199-206.
Donahue, M.J., Near, J., Blicher, J.U., Jezzard, P., 2010. Baseline GABA concentration and fMRI
response. NeuroImage 53, 392-398.
Dormal, G., Lepore, F., Harissi-Dagher, M., Albouy, G., Bertone, A., Rossion, B., Collignon, O., 2015.
Tracking the evolution of crossmodal plasticity and visual functions before and after sight restoration.
J Neurophysiol 113, 1727-1742.
Dorn, J.D., Ahuja, A.K., Caspi, A., da Cruz, L., Dagnelie, G., Sahel, J.A., Greenberg, R.J., McMahon, M.J.,
gus, I.I.S.G., 2013. The Detection of Motion by Blind Subjects With the Epiretinal 60-Electrode (Argus
II) Retinal Prosthesis. JAMA ophthalmology 131, 183-189.
Dosher, B., Lu, Z.L., 2017. Visual Perceptual Learning and Models. Annu Rev Vis Sci 3, 343-363.
Dumoulin, S.O., Knapen, T., 2018. How Visual Cortical Organization Is Altered by Ophthalmologic and
Neurologic Disorders. Annu Rev Vis Sci 4, 357-379.
Fagiolini, M., Fritschy, J.M., Low, K., Mohler, H., Rudolph, U., Hensch, T.K., 2004. Specific GABAA circuits
for visual cortical plasticity. Science 303, 1681-1683.
Fagiolini, M., Hensch, T.K., 2000. Inhibitory threshold for critical-period activation in primary visual
cortex. Nature 404, 183-186.
Ferreira, S., Pereira, A.C., Quendera, B., Reis, A., Silva, E.D., Castelo-Branco, M., 2017. Primary visual
cortical remapping in patients with inherited peripheral retinal degeneration. Neuroimage Clin 13, 428-
438.
Fierro, B., Brighina, F., Vitello, G., Piazza, A., Scalia, S., Giglia, G., Daniele, O., Pascual-Leone, A., 2005.
Modulatory effects of low- and high-frequency repetitive transcranial magnetic stimulation on visual
cortex of healthy subjects undergoing light deprivation. The Journal of physiology 565, 659-665.
Fine, I., Boynton, G.M., 2015. Pulse trains to percepts: the challenge of creating a perceptually
intelligible world with sight recovery technologies. Philosophical transactions of the Royal Society of
London. Series B, Biological sciences 370, 20140208.
Fine, I., Park, J.M., 2018. Blindness and Human Brain Plasticity. Annu Rev Vis Sci 4, 337-356.
Fine, I., Smallman, H.S., Doyle, P., MacLeod, D.I., 2002. Visual function before and after the removal of
bilateral congenital cataracts in adulthood. Vision Res 42, 191-210.
Fine, I., Wade, A.R., Brewer, A.A., May, M.G., Goodman, D.F., Boynton, G.M., Wandell, B.A., MacLeod,
D.I., 2003. Long-term deprivation affects visual perception and cortex. Nature neuroscience 6, 915-
916.
Fiorentini, A., Berardi, N., 1980. Perceptual learning specific for orientation and spatial frequency.
Nature 287, 43-44.
Fronius, M., Cirina, L., Ackermann, H., Kohnen, T., Diehl, C.M., 2014. Efficiency of electronically
monitored amblyopia treatment between 5 and 16 years of age: new insight into declining
susceptibility of the visual system. Vision Res 103, 11-19.
Fuchs, E., Flugge, G., 2014. Adult neuroplasticity: more than 40 years of research. Neural plasticity
2014, 541870.
Gilbert, C.D., Wiesel, T.N., 1992. Receptive field dynamics in adult primary visual cortex. Nature 356,
150-152.
Goesaert, E., Van Baelen, M., Spileers, W., Wagemans, J., Op de Beeck, H.P., 2014. Visual Space and
Object Space in the Cerebral Cortex of Retinal Disease Patients. Plos One 9.
Gregory, R., Wallace, J., 1963. Recovery from early blindness: a case study. Heffer, Cambridge.
Haak, K.V., Cornelissen, F.W., Morland, A.B., 2012. Population receptive field dynamics in human visual
cortex. Plos One 7, e37686.
Haak, K.V., Morland, A.B., Engel, S.A., 2015. Plasticity, and Its Limits, in Adult Human Primary Visual
Cortex. Multisensory research 28, 297-307.
Haak, K.V., Morland, A.B., Rubin, G.S., Cornelissen, F.W., 2016. Preserved retinotopic brain connectivity
in macular degeneration. Ophthalmic & physiological optics : the journal of the British College of
Ophthalmic Opticians 36, 335-343.
Hanover, J.L., Huang, Z.J., Tonegawa, S., Stryker, M.P., 1999. Brain-derived neurotrophic factor
overexpression induces precocious critical period in mouse visual cortex. The Journal of neuroscience
: the official journal of the Society for Neuroscience 19, RC40.
Harauzov, A., Spolidoro, M., DiCristo, G., De Pasquale, R., Cancedda, L., Pizzorusso, T., Viegi, A., Berardi,
N., Maffei, L., 2010. Reducing intracortical inhibition in the adult visual cortex promotes ocular
dominance plasticity. The Journal of neuroscience : the official journal of the Society for Neuroscience
30, 361-371.
He, H.Y., Hodos, W., Quinlan, E.M., 2006. Visual deprivation reactivates rapid ocular dominance
plasticity in adult visual cortex. The Journal of neuroscience : the official journal of the Society for
Neuroscience 26, 2951-2955.
Heeger, D.J., 1992. Normalization of cell responses in cat striate cortex. Visual neuroscience 9, 181-
197.
Heimel, J.A., van Versendaal, D., Levelt, C.N., 2011. The role of GABAergic inhibition in ocular
dominance plasticity. Neural plasticity 2011, 391763.
Heimler, B., Weisz, N., Collignon, O., 2014. Revisiting the adaptive and maladaptive effects of
crossmodal plasticity. Neuroscience 283, 44-63.
Hensch, T.K., Fagiolini, M., 2005. Excitatory-inhibitory balance and critical period plasticity in
developing visual cortex. Progress in brain research 147, 115-124.
Hensch, T.K., Fagiolini, M., Mataga, N., Stryker, M.P., Baekkeskov, S., Kash, S.F., 1998. Local GABA
circuit control of experience-dependent plasticity in developing visual cortex. Science 282, 1504-1508.
Hensch, T.K., Quinlan, E.M., 2018. Critical periods in amblyopia. Visual neuroscience 35, E014.
Huang, Z.J., Kirkwood, A., Pizzorusso, T., Porciatti, V., Morales, B., Bear, M.F., Maffei, L., Tonegawa, S.,
1999. BDNF regulates the maturation of inhibition and the critical period of plasticity in mouse visual
cortex. Cell 98, 739-755.
Hubel, D.H., Wiesel, T.N., 1970. The period of susceptibility to the physiological effects of unilateral
eye closure in kittens. The Journal of physiology 206, 419-436.
Hubel, D.H., Wiesel, T.N., LeVay, S., 1977. Plasticity of ocular dominance columns in monkey striate
cortex. Philosophical transactions of the Royal Society of London. Series B, Biological sciences 278, 377-
409.
Huber, E., Webster, J.M., Brewer, A.A., MacLeod, D.I., Wandell, B.A., Boynton, G.M., Wade, A.R., Fine,
I., 2015. A lack of experience-dependent plasticity after more than a decade of recovered sight.
Psychological science 26, 393-401.
Ivanova, E., Yee, C.W., Baldoni, R., Sagdullaev, B.T., 2016. Aberrant activity in retinal degeneration
impairs central visual processing and relies on Cx36-containing gap junctions. Exp Eye Res 150, 81-89.
Jiang, F., Beauchamp, M.S., Fine, I., 2015. Re-examining overlap between tactile and visual motion
responses within hMT+ and STS. NeuroImage 119, 187-196.
Jiang, F., Stecker, G.C., Boynton, G.M., Fine, I., 2016. Early Blindness Results in Developmental Plasticity
for Auditory Motion Processing within Auditory and Occipital Cortex. Frontiers in human neuroscience
10, 324.
Jiang, F., Stecker, G.C., Fine, I., 2014. Auditory motion processing after early blindness. Journal of vision
14, 4.
Joshi, S., Li, Y., Kalwani, R.M., Gold, J.I., 2016. Relationships between Pupil Diameter and Neuronal
Activity in the Locus Coeruleus, Colliculi, and Cingulate Cortex. Neuron 89, 221-234.
Kasamatsu, T., Pettigrew, J.D., Ary, M., 1979. Restoration of visual cortical plasticity by local
microperfusion of norepinephrine. The Journal of comparative neurology 185, 163-181.
Kasamatsu, T., Pettigrew, J.D., Ary, M., 1981. Cortical recovery from effects of monocular deprivation:
acceleration with norepinephrine and suppression with 6-hydroxydopamine. J Neurophysiol 45, 254-
266.
Kiorpes, L., Kiper, D.C., O'Keefe, L.P., Cavanaugh, J.R., Movshon, J.A., 1998. Neuronal correlates of
amblyopia in the visual cortex of macaque monkeys with experimental strabismus and anisometropia.
The Journal of neuroscience : the official journal of the Society for Neuroscience 18, 6411-6424.
Lee, D.S., Lee, J.S., Oh, S.H., Kim, S.K., Kim, J.W., Chung, J.K., Lee, M.C., Kim, C.S., 2001. Cross-modal
plasticity and cochlear implants. Nature 409, 149-150.
Levelt, C.N., Hubener, M., 2012. Critical-period plasticity in the visual cortex. Annual review of
neuroscience 35, 309-330.
Levelt, W.J., 1965. On Binocular Rivalry Soesterberg: Netherlands: Institution for Perception.
Levi, D.M., Carkeet, A., 1993. Amblyopia: a consequence of abnormal visual development. , in: Simons,
K. (Ed.), Early Visual Development, Normal and Abnormal. Oxford University Press, New York, NY, pp.
391 -- 408.
Lewis, L.B., Saenz, M., Fine, I., 2010. Mechanisms of cross-modal plasticity in early-blind subjects. J
Neurophysiol 104, 2995-3008.
Lou, A.R., Madsen, K.H., Paulson, O.B., Julian, H.O., Prause, J.U., Siebner, H.R., Kjaer, T.W., 2011.
Monocular visual deprivation suppresses excitability in adult human visual cortex. Cerebral cortex 21,
2876-2882.
Lunghi, C., Berchicci, M., Morrone, M.C., Di Russo, F., 2015a. Short-term monocular deprivation alters
early components of visual evoked potentials. The Journal of physiology 593, 4361-4372.
Lunghi, C., Burr, D.C., Morrone, C., 2011. Brief periods of monocular deprivation disrupt ocular balance
in human adult visual cortex. Curr Biol 21, R538-539.
Lunghi, C., Burr, D.C., Morrone, M.C., 2013. Long-term effects of monocular deprivation revealed with
binocular rivalry gratings modulated in luminance and in color. Journal of vision 13.
Lunghi, C., Emir, U.E., Morrone, M.C., Bridge, H., 2015b. Short-term monocular deprivation alters GABA
in the adult human visual cortex. Curr Biol 25, 1496-1501.
Lunghi, C., Sale, A., 2015. A cycling lane for brain rewiring. Curr Biol 25, R1122-1123.
Lunghi, C., Sframeli, T., Lepri, A., Lepri, M., Lisi, D., Sale, A., Morrone, M.C., 2018. A new counterintuitive
training for adult amblyopia. Annals of Clinical and Translational Neurology.
MacSweeney, M., Capek, C.M., Campbell, R., Woll, B., 2008. The signing brain: the neurobiology of sign
language. Trends in cognitive sciences 12, 432-440.
Maffei, A., Turrigiano, G., 2008a. The age of plasticity: developmental regulation of synaptic plasticity
in neocortical microcircuits. Progress in brain research 169, 211-223.
Maffei, A., Turrigiano, G.G., 2008b. Multiple modes of network homeostasis in visual cortical layer 2/3.
The Journal of neuroscience : the official journal of the Society for Neuroscience 28, 4377-4384.
Malenka, R.C., Bear, M.F., 2004. LTP and LTD: an embarrassment of riches. Neuron 44, 5-21.
Masuda, Y., Dumoulin, S.O., Nakadomari, S., Wandell, B.A., 2008. V1 projection zone signals in human
macular degeneration depend on task, not stimulus. Cerebral cortex 18, 2483-2493.
Masuda, Y., Horiguchi, H., Dumoulin, S.O., Furuta, A., Miyauchi, S., Nakadomari, S., Wandell, B.A., 2010.
Task-dependent V1 responses in human retinitis pigmentosa. Invest Ophthalmol Vis Sci 51, 5356-5364.
Maurer, D., Lewis, T.L., Mondloch, C.J., 2005. Missing sights: consequences for visual cognitive
development. Trends in cognitive sciences 9, 144-151.
Maurer, D., Mondloch, C.J., Lewis, T.L., 2007. Effects of early visual deprivation on perceptual and
cognitive development. Progress in brain research 164, 87-104.
Maya Vetencourt, J.F., Sale, A., Viegi, A., Baroncelli, L., De Pasquale, R., O'Leary, O.F., Castren, E.,
Maffei, L., 2008. The antidepressant fluoxetine restores plasticity in the adult visual cortex. Science
320, 385-388.
Merabet, L.B., Hamilton, R., Schlaug, G., Swisher, J.D., Kiriakopoulos, E.T., Pitskel, N.B., Kauffman, T.,
Pascual-Leone, A., 2008. Rapid and reversible recruitment of early visual cortex for touch. Plos One 3,
e3046.
Merabet, L.B., Rizzo, J.F., Amedi, A., Somers, D.C., Pascual-Leone, A., 2005. What blindness can tell us
about seeing again: merging neuroplasticity and neuroprostheses. Nat Rev Neurosci 6, 71-77.
Merabet, L.B., Swisher, J.D., McMains, S.A., Halko, M.A., Amedi, A., Pascual-Leone, A., Somers, D.C.,
2007. Combined activation and deactivation of visual cortex during tactile sensory processing. J
Neurophysiol 97, 1633-1641.
Morishita, H., Miwa, J.M., Heintz, N., Hensch, T.K., 2010. Lynx1, a cholinergic brake, limits plasticity in
adult visual cortex. Science 330, 1238-1240.
Muthukumaraswamy, S.D., Edden, R.A., Jones, D.K., Swettenham, J.B., Singh, K.D., 2009. Resting GABA
concentration predicts peak gamma frequency and fMRI amplitude in response to visual stimulation
in humans. Proceedings of the National Academy of Sciences of the United States of America 106,
8356-8361.
Ooi, T.L., He, Z.J., 2001. Sensory eye dominance. Optometry 72, 168-178.
Pascual-Leone, A., Amedi, A., Fregni, F., Merabet, L.B., 2005. The plastic human brain cortex. Annual
review of neuroscience 28, 377-401.
Pascual-Leone, A., Hamilton, R., 2001. The metamodal organization of the brain. Progress in brain
research 134, 427-445.
Pietrini, P., Furey, M.L., Ricciardi, E., Gobbini, M.I., Wu, W.H., Cohen, L., Guazzelli, M., Haxby, J.V., 2004.
Beyond sensory images: Object-based representation in the human ventral pathway. Proceedings of
the National Academy of Sciences of the United States of America 101, 5658-5663.
Pitskel, N.B., Merabet, L.B., Ramos-Estebanez, C., Kauffman, T., Pascual-Leone, A., 2007. Time-
dependent changes in cortical excitability after prolonged visual deprivation. Neuroreport 18, 1703-
1707.
Pizzorusso, T., Medini, P., Berardi, N., Chierzi, S., Fawcett, J.W., Maffei, L., 2002. Reactivation of ocular
dominance plasticity in the adult visual cortex. Science 298, 1248-1251.
Ptito, M., Moesgaard, S.M., Gjedde, A., Kupers, R., 2005. Cross-modal plasticity revealed by
electrotactile stimulation of the tongue in the congenitally blind. Brain : a journal of neurology 128,
606-614.
Ramamurthy, M., Blaser, E., 2018. Assessing the kaleidoscope of monocular deprivation effects.
Journal of vision 18, 14.
Ricciardi, E., Bonino, D., Pellegrini, S., Pietrini, P., 2014a. Mind the blind brain to understand the sighted
one! Is there a supramodal cortical functional architecture? Neuroscience and biobehavioral reviews
41, 64-77.
Ricciardi, E., Pietrini, P., 2011. New light from the dark: what blindness can teach us about brain
function. Curr Opin Neurol 24, 357-363.
Ricciardi, E., Tozzi, L., Leo, A., Pietrini, P., 2014b. Modality dependent cross-modal functional
reorganization following congenital visual deprivation within occipital areas: a meta-analysis of tactile
and auditory studies. Multisensory research 27, 247-262.
Ricciardi, E., Vanello, N., Sani, L., Gentili, C., Scilingo, E.P., Landini, L., Guazzelli, M., Bicchi, A., Haxby,
J.V., Pietrini, P., 2007. The effect of visual experience on the development of functional architecture in
hMT+. Cerebral cortex 17, 2933-2939.
Ritter, M., Hummer, A., Ledolter, A., Holder, G., Windischberger, C., Schmidt-Erfurth, U., 2018.
Correspondence between
functional and
morphological assessment of retinal disease. Br J Ophthalmol.
Rizzo, S., Belting, C., Cinelli, L., Allegrini, L., 2015. Visual field changes following implantation of the
Argus II retinal prosthesis. Graefe's archive for clinical and experimental ophthalmology = Albrecht von
Graefes Archiv fur klinische und experimentelle Ophthalmologie 253, 323-325.
retinotopic cortical mapping and conventional
Rizzo, S., Belting, C., Cinelli, L., Allegrini, L., Genovesi-Ebert, F., Barca, F., di Bartolo, E., 2014. The Argus
II Retinal Prosthesis: 12-month outcomes from a single-study center. American journal of
ophthalmology 157, 1282-1290.
Saenz, M., Lewis, L.B., Huth, A.G., Fine, I., Koch, C., 2008. Visual Motion Area MT+/V5 Responds to
Auditory Motion in Human Sight-Recovery Subjects. The Journal of neuroscience : the official journal
of the Society for Neuroscience 28, 5141-5148.
Sale, A., Berardi, N., Maffei, L., 2014. Environment and brain plasticity: towards an endogenous
pharmacotherapy. Physiological reviews 94, 189-234.
Sale, A., Maya Vetencourt, J.F., Medini, P., Cenni, M.C., Baroncelli, L., De Pasquale, R., Maffei, L., 2007.
Environmental enrichment in adulthood promotes amblyopia recovery through a reduction of
intracortical inhibition. Nature neuroscience 10, 679-681.
Schumacher, E.H., Jacko, J.A., Primo, S.A., Main, K.L., Moloney, K.P., Kinzel, E.N., Ginn, J., 2008.
Reorganization of visual processing is related to eccentric viewing in patients with macular
degeneration. Restor Neurol Neurosci 26, 391-402.
Sereno, M.I., 2005. Neuroscience: plasticity and its limits. Nature 435, 288-289.
Smirnakis, S.M., Brewer, A.A., Schmid, M.C., Tolias, A.S., Schuz, A., Augath, M., Inhoffen, W., Wandell,
B.A., Logothetis, N.K., 2005. Lack of long-term cortical reorganization after macaque retinal lesions.
Nature 435, 300-307.
Smith, G.B., Heynen, A.J., Bear, M.F., 2009. Bidirectional synaptic mechanisms of ocular dominance
plasticity in visual cortex. Philosophical transactions of the Royal Society of London. Series B, Biological
sciences 364, 357-367.
Speed, H.D., Morrone, M.C., Burr, D.C., 1991. Effects of monocular deprivation on the development of
visual inhibitory interactions in kittens. Visual neuroscience 7, 335-343.
Spolidoro, M., Sale, A., Berardi, N., Maffei, L., 2009. Plasticity in the adult brain: lessons from the visual
system. Experimental brain research 192, 335-341.
Stryker, M.P., 2014. A Neural Circuit That Controls Cortical State, Plasticity, and the Gain of Sensory
Responses in Mouse. Cold Spring Harbor symposia on quantitative biology 79, 1-9.
Sunness, J.S., Liu, T., Yantis, S., 2004. Retinotopic mapping of the visual cortex using functional
magnetic resonance imaging in a patient with central scotomas from atrophic macular degeneration.
Ophthalmology 111, 1595-1598.
Tamietto, M., Morrone, M.C., 2016. Visual Plasticity: Blindsight Bridges Anatomy and Function in the
Visual System. Curr Biol 26, R70-R73.
Trenholm, S., Awatramani, G.B., 2015. Origins of spontaneous activity in the degenerating retina. Front
Cell Neurosci 9, 277.
Turrigiano, G., 2012. Homeostatic synaptic plasticity: local and global mechanisms for stabilizing
neuronal function. Cold Spring Harbor perspectives in biology 4, a005736.
Turrigiano, G., Nelson, S.B., 2000. Hebb and homeostasis in neuronal plasticity. Curr Opin Neurobiol
10, 358-364.
Voss, P., 2019. Brain (re)organization following visual loss. Wiley interdisciplinary reviews. Cognitive
science 10, e1468.
Voss, P., Gougoux, F., Lassonde, M., Zatorre, R.J., Lepore, F., 2006. A positron emission tomography
study during auditory localization by late-onset blind individuals. Neuroreport 17, 383-388.
Wandell, B.A., Smirnakis, S.M., 2009. Plasticity and stability of visual field maps in adult primary visual
cortex. Nat Rev Neurosci 10, 873-884.
Watanabe, T., Sasaki, Y., 2015. Perceptual learning: toward a comprehensive theory. Annual review of
psychology 66, 197-221.
Webber, A.L., Wood, J., 2005. Amblyopia: prevalence, natural history, functional effects and
treatment. Clin Exp Optom 88, 365-375.
Wiesel, T.N., Hubel, D.H., 1963. Effects of Visual Deprivation on Morphology and Physiology of Cells in
the Cats Lateral Geniculate Body. J Neurophysiol 26, 978-993.
Wiesel, T.N., Hubel, D.H., 1965a. Comparison of the effects of unilateral and bilateral eye closure on
cortical unit responses in kittens. J Neurophysiol 28, 1029-1040.
Wiesel, T.N., Hubel, D.H., 1965b. Extent of recovery from the effects of visual deprivation in kittens. J
Neurophysiol 28, 1060-1072.
Zeck, G., 2016. Aberrant Activity in Degenerated Retinas Revealed by Electrical Imaging. Front Cell
Neurosci 10, 25.
Zhou, J., Clavagnier, S., Hess, R.F., 2013. Short-term monocular deprivation strengthens the patched
eye's contribution to binocular combination. Journal of vision 13.
Zhou, J., Reynaud, A., Hess, R.F., 2014. Real-time modulation of perceptual eye dominance in humans.
Proceedings. Biological sciences 281.
Figure 1. Short-term visual deprivation induces functional reorganization of cortical
circuits in sighted adult humans.
(A) During 7T fMRI scanning participants' eyes were separately stimulated with band-
pass noise and BOLD responses were measured before and after 2h of monocular
deprivation. The flat map shows extensive BOLD response before deprivation elicited
by the noise stimuli in all visual areas. (B) The percentage BOLD signal measured in
V1 elicited by stimulation of the deprived eye and non-deprived eye was comparable
before deprivation, while respectively enhanced and reduced after 2h of monocular
deprivation. (C) The neural changes described are reflected at behavioral level: mean
phase duration during binocular rivalry are balanced across the two eyes before
deprivation, while mean phases are
the
deprived/non-deprived eye respectively. (D) These perceptual effects correlate with
the deprivation index measured with fMRI. (A-D) Reproduced from Figure 1 and Figure
3 (Binda et al., 2018), eLife, published under the Creative Commons Attribution 4.0
International
4.0;
https://creativecommons.org/licenses/by/4.0/)."
longer/shorter after deprivation
(CC
for
Public
License
BY
Figure 2. After sufficient time and practice the adult visual brain can learn to see again
with artificial vision.
(A) BOLD responses elicited by three different tactile tasks in two RP patients
implanted with Argus II retinal prosthesis (subjects A1 and A2) who underwent fMRI
scanning after 6 and 15 weeks after surgery. Participants were required to haptically
determine the symmetry of raised-line shapes (shape task), count the number of dots
in Braille letters (Braille-dot counting task), and evaluate the roughness of sandpaper
discs (sandpaper task). Independently of the task, strong tactile evoked responses can
be observed in occipital cortex after only 6 weeks after surgery (subject A1). The
occipital cross-modal activation is much reduced after 15 weeks from surgery (subject
A2). (B) BOLD responses elicited by visual stimuli (flashes of lights) before (blue) vs
after (red) surgery in a group of RP patients implanted with Argus II. After the surgery
visual BOLD signal is enhanced in V1. Crucially participants were scanned at least 6
months after the surgery. (C) RP patients were asked to choose the interval,
demarcated by sound, in which a large high contrast grating was presented. The
behavioral performance in a contrast detection task improves as a function of time from
surgery. Patients needed time and practice to learn how to interpret and use the
restored visual input. (A) Reproduced with permission from Figure 3 from Cunningham
SI, Shi Y, Weiland JD, et al. Feasibility of structural and functional MRI acquisition with
unpowered implants in Argus II retinal prosthesis patients: a case study. Trans Vis Sci
Tech. 2015;4(6):6., ARVO copyright holder. (B, C) Reproduced from Figure 2 and
Figure 5 from Castaldi et al. (2016) Visual BOLD Response in Late Blind Subjects with
Argus
e1002569.
https://doi.org/10.1371/journal.pbio.1002569, published under the Creative Commons
Attribution license (CC BY).
Prosthesis.
Biology
II
Retinal
PLOS
14(10):
|
1701.05080 | 1 | 1701 | 2017-01-17T12:43:17 | Individual versus collective cognition in social insects | [
"q-bio.NC",
"cs.DC"
] | The concerted responses of eusocial insects to environmental stimuli are often referred to as collective cognition on the level of the colony.To achieve collective cognitiona group can draw on two different sources: individual cognitionand the connectivity between individuals.Computation in neural-networks, for example,is attributedmore tosophisticated communication schemes than to the complexity of individual neurons. The case of social insects, however, can be expected to differ. This is since individual insects are cognitively capable units that are often able to process information that is directly relevant at the level of the colony.Furthermore, involved communication patterns seem difficult to implement in a group of insects since these lack clear network structure.This review discusses links between the cognition of an individual insect and that of the colony. We provide examples for collective cognition whose sources span the full spectrum between amplification of individual insect cognition and emergent group-level processes. | q-bio.NC | q-bio | Individual versus collective cognition in social insects
Ofer Feinermanᴥ, Amos Kormanˠ
ᴥ Department of Physics of Complex Systems, Weizmann Institute of Science, 7610001, Rehovot, Israel. Email:
[email protected]
ˠ Institut de Recherche en Informatique Fondamentale (IRIF), CNRS and University Paris Diderot, 75013, Paris,
France. Email: [email protected]
Abstract
The concerted responses of eusocial insects to environmental stimuli are often referred to as
collective cognition on the level of the colony.To achieve collective cognitiona group can
draw on two different sources:
individual cognitionand the connectivity between
individuals.Computation in neural-networks, for example,is attributedmore tosophisticated
communication schemes than to the complexity of individual neurons. The case of social
insects, however, can be expected to differ. This is since individual insects are cognitively
capable units that are often able to process information that is directly relevant at the level
of the colony.Furthermore, involved communication patterns seem difficult to implement in a
group of insects since these lack clear network structure.This review discusses links between
the cognition of an individual insect and that of the colony. We provide examples for
collective cognition whose sources span the full spectrum between amplification of individual
insect cognition and emergent group-level processes.
Introduction
The individuals that make up a social insect colony are so tightly knit that they are often
regarded as a single super-organism(Wilson and Hölldobler, 2009). This point of view seems
to go far beyond a simple metaphor(Gillooly et al., 2010)and encompasses aspects of the
colony that are analogous to cell differentiation(Emerson, 1939), metabolic rates(Hou et al.,
2010; Waters et al., 2010), nutrient regulation(Behmer, 2009),thermoregulation(Jones,
2004; Starks et al., 2000), gas exchange(King et al., 2015), and more.
It is tempting to push this analogy, one step further and attribute the superorganism with
collective cognition(Couzin, 2009; Franks, 1989; Seeley, 1996). In this respect, it is possible to
envision two extreme cases in which groups of insects may have evolved to exhibit cognition
on the scale of the entire colony. The first is reliance on the cognition of the individuals that
make up the group. Indeed, the cognitive abilities of a single ant or bee within the large
colony are far from being simple (Dornhaus and Franks, 2008). The group can benefit from
these capabilities, for example, by sharing and refining the knowledge of informed
individuals. The second extreme case is collective cognition derived from the interaction
between members. Manmade systems teach us that complex computation can be achieved
by the wiring together of very simple components such as logical gates (Lindgren and
Nordahl, 1990). Similarly, social insect colonies often display dense interaction networks
(Wilson and Hölldobler, 1988)and collective behaviors that appear to exceed the capacity of
the individuals of which they are comprised (Sumpter, 2006).
1
It is therefore of interest to trace the collective actions of the social insect colony to their
sources, be they the cognition of individuals or the communication network that connect
groups of such individuals. Mapping out the relations between these two organizational
scales is required if one is to understand and quantify collective cognition as well as learn
about its evolutionary origins.We hypothesize that individual-based collective behaviors will
be prevalent in cases where abilities, similar to those exhibited by solitary insects, suffice in
order to sense, grasp, and process knowledge that is relevant on the scale of the colony.
Deviations from this will tend to lead to group solutions that involve an increased emergent
component.
Outline
The outline of this review is as follows: First, we discuss the cognitive abilities of the
individuals that make up the social
insect colony andsome current knowledge of
communication networks in social insects. As mentioned above, these two components
provide the basis on which the colony could build its collective capabilities. Next, we present
a list of examples of collective cognition. These examples are ordered by the degree to
which collective behaviors rely on each of the two components, from individual-based to
connectivity-based. The examples are split into three categories: Individual-based collective
behaviors, collective behaviors that combine different individual perspectives, and, finally,
collective behaviors that display higher levels of emergence. Each of these categories is
divided into subcategories that further refine this division. Taken together, these examples
span a broad spectrum of relationships between individual and collective cognition. In the
final section we discuss the possible factors that may determine the degree of emergence in
a particular collective behavior.
Individual cognition
A good starting point for discussing the origins of cognition in social insect colonies is the
cognitive abilities of solitary insects. Insect brains have evolved hundreds of millions of
years(Farris and Schulmeister, 2010; Ma et al., 2012)prior to the appearance of
eusociality(Moreau, 2006). Despite the fact that their brains are relatively small(Chittka and
Niven, 2009),solitary insects exhibit high cognitive skills that include large behavioral
repertoires(Evans, 1966), complex forms of learning(Alloway, 1972; Blackiston et al., 2011),
and include navigational skills that often exceed those of humans(Brower, 1996). These
abilities aid the solitary insect, among other things, inforaging (O'Neill, 2001), finding or
constructing shelters(Raw, 1972), confronting predators(Schmidt, 1990), and identifying
appropriate mating partners(Dickson, 2008).
in this discussion
The next step
is the transitions to eusociality which happened
between100-150 million years ago(Brady et al., 2006; Engel et al., 2009). Eusociality is
characterized by reproductive division of labor that drastically lowers the level of conflict
between group members as they strive towards common goals (Crespi and Yanega, 1995). It
is first important to state the evident fact that, contrary to cells in a tissue or neurons in the
brain, insects within the colony superorganism maintain their individuality. They are able of
autonomous motion and decision making.Further, the brains of individuals within a colony
2
bear high resemblance to those of solitary insects(Strausfeld, 1976). One may therefore ask
how the cognitive capabilities a social insect compare to those of a solitary insect. To date, it
is not clear if once grouped into large groups evolution may work to increase or decrease the
cognitive complexity of individuals. On the one hand, it is known, mainly from vertebrate
groups that the communication requirements of group living may work to increase brain
complexity(Shultz and Dunbar, 2010). On the other hand, it has been suggested that, relying
on collective processes may ease the energetically expensive(Aiello and Wheeler, 1995)
maintenanceof brain tissue(Anderson and McShea, 2001; Feinerman and Traniello, 2015).
Whatever the exact comparison between the brains of a social and a solitary insect, it is
clear that thesocial insect is a cognitively capable individual. Individuals within the colony
possess the capacity for large behavioral repertoires(Chittka and Niven, 2009), for weighing
a large number of factors to reach individual decisions(Franks et al., 2003), andfor navigating
over large distances(Gathmann and Tscharntke, 2002; Wehner, 2003). Importantly, these
individual capabilities are relevant on the scale of the entire colony.
One aspectthat clearly differentiates the social insect from its solitary counterpart is the
capacity for communication. For example, eusocial insects display a huge diversification of
cuticular pheromones(van Wilgenburg et al., 2011) used to convey multiple signals that are
unique to colony life(Howard and Blomquist, 2005). Another famous example is the
honeybee waggle dance(Von Frisc, 1950). While solitary insects may have the motivation to
conceal a newly found item for personal consumption(Byrne et al., 2003), the bees have
evolved an elaborate communication scheme which allows them to share this location.These
and other interaction skills form the foundation of the insect society. In the next section we
discuss some of the properties of the communication networks via which social insects
coordinate their activities.
Interaction Networks
collection
is especially
important(Wilson
Group living animals combine personal and social information when deciding upon their next
action(Rieucau and Giraldeau, 2011). In eusocial insects – the social component of
information
and Hölldobler, 1988).
Correspondingly, the modalities of communication and richness of cues and signals is greatly
enhanced. Social insects use a variety of olfactory(Martin and Drijfhout, 2009; Morgan,
2009), tactile(Razin et al., 2013), visual, and vibrational(Delattre et al., 2015; Roces et al.,
1993)messages as well as multi-model combinations of these (Ramsden et al., 2009) in their
communication. Broadly speaking, these can be divided into several groups: Some messages
require direct contact between individuals and can thus be considered as local in both space
and time. Other signals are local in time but not space and are typically employed as alarm
signals (e.g. highly volatile pheromones(Blum, 1969)). Yet another group are signals that are
local in space but not in time. This group includes stigmergic, indirect communication
between insects in which one individual modifies the environment and a second individual
arriving at
its modified
surroundings(Theraulaz and Bonabeau, 1999). Mass recruitment pheromone trails (Jaffe and
Howse, 1979)and nest construction without a blueprint (Franks and Deneubourg, 1997)are
two
impressive examples of stigmergy. Note that, for the case of pheromonal
location at
the
same
some
later
time
reacts
to
3
communication, the time scales that characterize pheromonal communication are evolvable
as they depend on chemical evaporation times that,indeed, vary between species and
tasks(Morgan, 2009; Witte et al., 2007)(Holldobler and Wilson, 1990).
Quantifying the communication patterns requires descriptive frameworksfor the different
interaction types as described above. Contact dependent interactions can be described as
time-ordered (Blonder and Dornhaus, 2011)communication networks(Fewell, 2003; Moreau
et al., 2011). It has been shown that, in a laboratory setting, the high mobility of social
insectsdictates that, after a sufficient time window, interactions occur between practically
all possible pairs and the network become highly connected (Mersch et al., 2013). Stigmergic
communication has been described by using the
language of statistical mechanics
(Richardson et al., 2011), or by employing cellular-automata tools typically used to describe
self-organization processes (Khuong et al., 2016). These interactions are, inherently, one-to-
many signaling and have been shown to extend the connectivity induced by contact
dependent communication (Richardson and Gorochowski, 2015)..Adding
long range
communication such as that involving alarm pheromones, we obtain a picture of a system in
which, at least to first order and over long enough time-scales, interactions can be described
as well mixed. In other words, over time an insect receives signals from any other insect in
the colony. Not only are interactions mixed they are also, to a large extent anonymous. With
a few exceptions (Mallon and Franks, 2000; Tibbetts, 2002), it is reasonable to assume that
individuals do not recognize which of the hundreds to hundreds of thousands of other
individuals they are currently interacting with.
While the previous discussion seems to suggest that interactions are completely ergodic, it is
important to stress that the social insect colony is, by no means, devoid of structure. For
example, ant nests and bee hives are often concentrically arrangedsuch that young insects
reside in the deep center while older individuals occupy progressively occupy areas that are
closer to the boundaries or entrance (Beshers and Fewell, 2001). Even when structure is
initially lacking, self-organization and amplification of noise can work to create spatio-
temporal patterns over time (Richardson et al., 2011; Theraulaz et al., 2003). Moreover, ant
(Tschinkel, 2004) and termite (Noirot and Darlington, 2000) nests exhibit complex structures
of rooms and corridors and these further reflect on the spatial distribution of individuals
within the nest. It has been shown that different individuals tend to occupy specific locations
or chambers within the nest (Jandt and Dornhaus, 2009; Sendova-Franks and Franks, 1995)
and that this reflects on the probability that they interact with other individuals in other
parts of the nest (Mersch et al., 2013; Pinter-Wollman et al., 2011). Spatial locations
therefore inducea network structure that is composed of relatively stable clusters. Hence, in
a very broad sense, the communication patterns in a social insect colony can be viewed as
residing between a well-mixed(on the more local scale) and a fixed (on the global, cluster,
scale) network.
Collective cognition
Having described some of the basic "cognitive toolbox" available to the colony, we go on to
discuss its collective scale behavioral products. In what follows, we presenta non-
comprehensive
insect colonies.
list of examples for collective behaviors
in social
4
Theexamples where chosen while focusing on the different possible gaps between the
knowledge, actions, and capabilities of individuals and those of the entire group. They are
ordered by the extent of this gap and divided between individual based collective cognition
(small gaps) and emergent collective cognition that builds on the interaction between
insects (large gaps). This division is, by no means, strict.
Individual based collective cognition
In a eusocial colony the genetic conflict between individuals in the group is minimal(Queller
and Strassmann, 1998). This leads to an alignment of interests which implies that it is
generally advantageous for informed individuals to share their knowledge with other group
members. Utilizing this information is useful for the group as well. Since the number of
informed individuals may be small one could expect that their actions be too weak to elicit
any significanteffect or, alternatively, they be averaged out against opposing actions
performed by other,
losing this useful
information, it may be profitable for the group to amplify the actions of these focal
individuals. Anysuch amplification should be regulated to prevent runaway behavior in case
of mistakes. Such mistakes could arise from the informed individuals themselves: they may
hold only partial information or be plain wrong, or from communication: noisy interactions
may distort the original message.
less-informed, colony members. Instead of
Next, we discuss several examples ofamplification circuits that make the products of
individual cognitionavailable, effective, anduseful at the level of the group.
Unconditional amplification
The simplest example is the alarm response. When an individual ant senses danger she not
only directly reacts to it but further emits a volatile alarm pheromone(Blum, 1969). This
pheromone spreads around the ant eliciting similar responses from her neighboring nest-
mates. This positive feedback circuit quickly spreads the danger signal to affect a large
number of individuals(Jeanson and Deneubourg, 2009). This not only increases the group's
surveillance of its environment (the "many eyes principle") but also allows it to take
collective actions towards, for example, protection of the nest. Similar behaviors are
displayed by termiteswhere chemical communication is accompanied by vibrationalsignaling
(Delattre et al., 2015).
These collective positive feedback circuits provide an informed individual that senses danger
immediate and direct control over the actions of the group. In other words,the gap between
individual and collective cognition is, practically, nonexistent.The group forsakes regulation
and out-weights this crucial survival response over the possible price paid by false alarms.
Conditioned amplification
The mass recruitment foraging trail occurs as a single first ant locates a food source. This ant
then uses her navigational skills to return to the nest while laying a pheromone trail that
recruits others to the food such that foraging commences. While an emergent process may
work to straighten the trail and make it shorter (see below) the trail still follows the
qualitative solution as
initial
recruiter(Holldobler, 1971).
then communicated by
first discovered and
the
5
the group"double-checks"the
Importantly, in a large number of species, ants strengthen the initial trail only on the way
back from the target food source and only if they independently found it to be
profitable(Beckers et al., 1992a; Mailleux et al., 2003; Wilson, 1962). This regulationcan be
considered as "delayed" since it occurs only after an initial positive response to the initial
ant. This provides a mechanism by which
target
communicated by the initial ant before continuing to amplify her effect even further.
Itallows the colony to reduce its response to ants that may have outdated information or
are, for some reason, confused. Since pheromones are volatile and have a finite lifetime,
their concentration along the trail depends on the rate at which they are enhanced. A trail
which is not enhanced will eventually disappear. Thus, delayed regulation further supplies a
mechanism for calibrating the level of activity on the recruitment trail (Simon and Hefetz,
1992), discontinuingit once the food source is exhausted(Wilson, 1962). It further allows the
system to escape local minima by switching to foraging on more profitable food sources
when such are identified(Beekman and Dussutour, 2007).
Amplification with early regulation
Desert ants typically forage alone and display only a rudimentary form of recruitment(Amor
et al., 2010; Razin et al., 2013). The recruitment process occurs as ants that are informed
about a food source outside the nest attempt to alert their nest-mates using imperfect
communication. The interactions used are noisy in the sense that a recruitment interaction
may be ignored or, conversely, a non-recruitment interaction may induce an ant to leave the
nest. Therefore, amplification of the initial signal must be regulated so that recruitment
occurs only when a food source is present and runaway behavior that results from mere
interaction noise is avoided. It was shown that desert ants regulate recruitment early on in
the process, at the entrance chamber of their nest(Razin et al., 2013).
This regulation is the result of two behavioral components. The firstis the fact that
individuals "know that they know"(Greenwald et al., 2015). There is a clear difference in the
way directly and non-directly informed insects react to interactions with others(Razin et al.,
2013; Schultz et al., 2008; Stroeymeyt et al., 2011). For example, ants that have been to the
food areasimply disregard interactions and maintain high walking speed to increase the
number and effectiveness of their recruitment interactions(Razin et al., 2013). On the other
hand, ants with second hand knowledge react to interactions in a more cautious manner.
These ants can either upregulate or downregulate their propensity to be recruited
depending on the state of the individual they interact with(Razin et al., 2013). These rules
allow
regulate collective behavior with minimaldependence on
theirunreliablecommunication skills. Specifically, thisworks to decrease the chances that the
actions of non-informed ants have global consequences on the state of the nest and leaves
the stage, so to say, to the directly informed ants(Razin et al., 2013).
The second regulatory component is an early negative feedback. All else being equal, non-
informed ants tend to lower their propensity to exit with passing time. This leads to a
collective threshold that dissipates the effects of random or isolated interactions such that
their effect quickly dies away and this protects the system from noise. A persistent informed
the ants
to
6
ant with first-hand information can generate enough activity such that interaction rates
increase(Gordon and Mehdiabadi, 1999) and the system moves over this recruitment
threshold. This early feedback mechanism in which positive feedback occurs only if the
system passes a set threshold is similar to the generation of spikes in neurons(Razin et al.,
2013). Note that while delayed regulation (as described in the previous subsection) works to
regulate the amplitude of the collective response and terminate it when the stimulus ends,
early regulation works to prevent amplification in the first place.
Combining individual perspectives
Amplifying the optimal option
The social insect colony may do more than amplifying individual decisions – it can, in fact,
poll individuals to reach consensus choice regarding the best solution among several
alternative options. Examples for this come from house-hunting(Visscher, 2007) behaviors in
ants(Franks et al., 2002) and bees(Seeley et al., 2006). When assessing the quality of a
potential new nest site scout ants have been shown to incorporate an intricate, individually
based,evaluation scheme which combines the different attributes of this location (e.g. its
volume, the size of the door, and the level of light) in a non-trivial way(Franks et al.,
2003).There is evidence that this assessment results in a single grade given by the ant to the
new location(Robinson et al., 2011; Robinson et al., 2014). The group does not make its own
assessments of nest quality (a hypothetical example for this could be moving the whole
colony to occupy the alternative nests one at a time and using the resulting ant density
(Gordon et al., 1993)to accurately measure the area of each) but, rather, uses a quorum
sensing as a polling mechanism to compare the assessments of its individual scouts(Franks
et al., 2002; Seeley et al., 2006). With high probability, this leads to the colony choosing the
best among the alternatives with the accuracy of the decision growing with the size of the
group(Sasaki et al., 2013).
House-hunting provides another fascinating example of how the action of the group may
work to refine individual decisions: When comparing different nests with specific attributes
individual ants are prone to violate theregularity principle of rational decision making(Sasaki
and Pratt, 2011). This principle states that if option A is preferred over option B then this
should not change upon introducing a third option, C, that is inferior to both. This fallacy is
not specific to insects but, rather, affects many different animals including humans.
However, when a whole colony is presented with the choice between these nests it will tend
to make the rational choice(Edwards and Pratt, 2009; Sasaki and Pratt, 2011). This is because
the polling often terminates before individual ants have had the chance to fail the regularity
principle since this requires visiting multiple nests(Robinson et al., 2014).
We provided several examples (many more exist) of how the actions and decisions of
capable individuals reflect at the level of the group. The group does not create new solutions
but ratherworks to amplify, average, poll, and refine the actions of individual members. This
is done using collective communication circuits that involve positive and negative feedbacks
and certain non-linearities.
Amplification in dynamic settings
7
In dynamic, fast-evolving scenarios, Information can quickly become obsolete and the
individuals that carry useful information (Robson and Traniello, 2002) change over time
(Gelblum et al., 2015). This entails two inherent problems: First, the relevant individual at a
specific point in time has to be identified by the group. Second, when an individual that is
better updated appears the group must revert to following it instead.
This scenario is realized during cooperative transport by longhorn crazy ants (Czaczkes and
Ratnieks, 2013). When ants cooperatively transport a large food item they can often lose
orientation and become unknowledgeable regarding the correct way to the nest. To correct
their path, these ants rely on well-informed individuals that are in the vicinity of the load but
unattached to it (Gelblum et al., 2015).
Instead of identifying the ant that currently holds valuable navigational information, here
again, the group relies on the "know that you know" principle. In other words, an informed
ant acts in a manner that is different from the other carriers: She attaches to the object and,
without heeding to others, pulls it in the direction she knows to be correct. At the same time
the carrying ants "acknowledge that they don't know" and apply a different behavioral rule –
which is, in a sense, pull in the direction in which the load is currently moving. Together,
these different rules as applied by informed and non-informed ants allow the group to
optimally amplify the force of the informed leader (Gelblum et al., 2015).
Importantly, after a period of about 10 seconds, the newly attached leader loses her
orientation. This former leader then adapts the behavioral rules of an ordinary carrier and
may continue in this state for many minutes. The directionality of the carrying group is next
corrected by the attachment of a new leader ant that happened to be informed at that
particular time. The fast switching between leaders can be viewed as a mechanism that
enables the group to escape being trapped at local minima in which the group displays
coordinated motion but the direction is wrong.
Collective bootstrapping of individual solutions
Another form of colony level solutions that is based on the actions of a large number of
individuals is trail shortening. Ants are famous for the ability to gradually decrease the
length of their pheromone trail so that it finally draws a geodesic between the food source
and the nest (Feynmann, 1985). This process can occur by the accumulated effect of ants
that leave the trail (Deneubourg et al., 1983) and return to it a short distance away. Useful
detours, i.e. those that "cut a corner" and slightly decrease the trail's length, are then
amplified by the group while non-useful detours are abandoned (Deneubourg et al., 1983;
Goss et al., 1989; Reid et al., 2011).
Even though trail shortening utilizes segments that were offered by individual ants, it is
inherently different from the amplification schemes described in the previous section. This
difference is manifested in the fact that in the previous examples require that the informed
ant "know that she knows" and assess the quality of the information that she holds.
Conversely, during trail shortening ants that mark short-cuts are not required to hold any
knowledge about the quality of their solutions. Rather, it is the group that either amplifies or
8
eventually ignores this alternative trail segments through a pheromone based positive-
feedback mechanism.
Emergent collective cognition
So far, we focused on group level behaviors that gain their computational power by
amplifying the actions and decisions of individuals. This differs from the notion of emergent
cognition wherein collective scale processes allow the group to qualitatively transcend
individual capabilities.
An intuitive example for emergence is the different physical phases of matter. Here, minimal
changes in temperature or the coupling between microscopic particles leads to qualitatively
different macroscopic phases (e.g. the solid to liquid transition). It is therefore interesting to
ask whether grouping together, not simple inanimate particles, but rather cognitive
individuals with a memory and complex behavioral rules can be expected to display different
and perhaps higher forms of emergence. Specifically, what forms of cognitive emergence
occur in the case of social insects? In this section, we list several examples for emergent
collective actions. As before, the examples are loosely ordered according to the increasing
gaps between the individual and the group.
Weighted response to multiple stimuli
Since they are grouped into large ensembles individual insects are, inevitably, much smaller
than the size of the colony and its territory. As a consequence, individuals cannot have direct
access to large-scale environmental and internal colony conditions. Despite this, the colony
as a whole must react to the full set of stimuli and appropriately divide the work
force(Robinson, 1992).Models(Beshers and Fewell, 2001) suggestthat colony level division of
labor can result from single insects with different task thresholds(Bonabeau, 1996)that
resolve work demands which they locally experience(Franks and Tofts, 1994). Similar to the
house-hunting example described above this colony level phenomenon relies on the
cognitive assessments of individuals. The difference being that, in this case, the colony does
not form a consensus around the solution of a single individual but rather divides the work
force in a weighted manner according to information that is too spread out tobe available to
any one individual. Division of labor can include more complex mechanisms, such as
recruitment, that allow ants to employ not only personal but also social information in their
decisions(Robinson et al., 2009a).
Partial decoupling between individual and collective scales
Cooperative transport is the process in which a group of ants retrieves a food item much too
large for any of them to move on their own. During this process, the information available to
individuals may be plainly misleading and counterproductive for the group's collective goals.
This happens when trajectories to the nest as experienced by the small ants may be
inaccessible to the large loads and even take it towards dead-ends that are difficult to
escape. It was shown that to avoid such deadlocks, the macroscopic scale occasionally
decouples from the possibly misleading information available at the microscopic scale(Fonio
9
et al., 2016). This mechanism allows the group to utilize beneficial information while using
noise present at the group level to escape deadlocks (local minima) and avoid the potentially
devastating consequences of fully relying on misleading information. Importantly, such
decoupling does not require that any single individual detect at any point in time, whether
information is valuable or misleading.
Emergence in this system is evident as a separation between collective and individual
behavior. This is evident, as the carrying group (and the food item they transport) does not
follow a trajectory that was suggested by any single ant.
Collective response independent of individual actions
In some cases, the group can display effective reactions to stimuli that are not conceived by
any individual. Such can be the case in which a cooperatively carrying group hits an obstacle.
In such instances, instead of attempting to advance directly towards the nest, the group
decouples from the actions of its individual members and goes into a perpendicular motion
that takes the carried piece of food towards the edges of the obstacle(Gelblum et al., 2015).
It was shown that this change in collective motion does not require that any individual ant
be aware of the obstacle and individually change her behavior(Gelblum et al., 2016). Rather,
the physical constraint induced by the obstacle directly affects the group as a whole. As a
response, the group's mode of motion changes in a way that facilitates obstacle
circumvention.
Self-organization withouta blueprint
A final example for emergent behavior involves nest construction. Social insects construct
some of the most magnificent structures in the biological world(Theraulaz et al., 2003). This
is done through a stigmergic process(Theraulaz and Bonabeau, 1999) in which individuals
locally interact with features of the structure by adding (or removing, in the case of dug
nests) building material to them(Franks and Deneubourg, 1997). This induces indirect
communication as insects interact with the product of the action of their nest-mates. It was
further shown that, in some cases, building materials are combined with a volatile
construction pheromone(Khuong et al., 2016). This adds a temporal dimension to the
physical structure and expands the possibilities for local rules and the complexity of the
resulting structure(Khuong et al., 2016).
In contrast to trail formation, for example, where the result of the collective effort is a
refinement of almost complete structures suggested by individual ants, nest construction
creates structures which appear to be far from the capabilities of any individual. Indeed, in
this high form of emergence, it is unlikely that individuals have a blueprint of the desired
final product. Nevertheless, they follow local rules and such that their collective effort
results in the construction of intricate nests.
Discussion
The first three sections of this discussion follow the structure of the previous example
section. We raise and then discuss some hypotheses regarding the prevalence of collective-
scale behaviors which strongly rely on the cognitive capabilities of individuals.
10
Individual-based collective cognition
An important factor to note is that many of the immediate requirements of the social group
coincide with those of the individuals that comprise this group. Like the solitary insect, the
colony must also scan the environment for food (Gordon, 1995), locate shelters (Franks et
al., 2002), transport food (Gelblum et al., 2015), confront predators (Lamon and Topoff,
1981; Monographs, 2016), and care for brood (Siveter et al., 2014; Wang et al., 2015). As
noted above, the cognitive abilities of individuals in a colony do not seem to greatly differ
from those of their solitary counterparts. As such the cognition of individuals within a colony
hold direct advantages to the group as a whole.TheIndividual based collective
cognitionsection,above, providesexampleswherein the actions of a single individual suffice
for directing the entire colony.
Generally speaking, amplifying of the behavior of an individual requires two colony level
processes: First,the identification of the specific individual of interest and then the
amplification of its behavior. For example, as mentioned above, the first ant to find a food
source has the capacity to trace the complete path between the nest and the
food(Cammaerts and Cammaerts, 1980; Hölldobler, 1976; Wilson, 1962). This ant makes
itself identifiable by laying pheromone markings on the surface as it heads back to the
nest(Beckers et al., 1992a). Amplification occurs as ants that follow this trail enforce it with
further pheromonal markings. Finally, mass foraging develops along the trail drawn out by
the initial recruiting ant(Beckers et al., 1989). In many cases, solutions provided by different
individuals are in conflict such that one solution is amplified this must come at the expense
of others. In this case a third decision making process occurs alongside identification and
amplification. Some mechanisms by which such collective decisions occur include the
preferential reinforcement of preferred solutions, such as preferred food sources being
marked by higher pheromone concentrations (Beckers et al., 1992b; Jaffe and Howse, 1979;
Sumpter and Beekman, 2003)and cross inhibition between alternative emerging solutions,
as occurs during honeybee recruitment (Nieh, 2010).
Combining individual perspectives
It is not always the case that an individual insect holds the complete solution to the colony's
current needs. In fact, inherent constraints work to limit the value of individually held
information. A first constraint involves the size of the individual when compared to that of
its colony. Colonies reside over territories that are tens of meters (in the case of ants) or
even kilometers (for wasps and bees) across and include large elaborate nest structures on a
scale of several meters (Tschinkel, 2004). It is impossible for a single individual whose size is
on the order of one centimeter to individually monitor these large areas. Other constraints
are a single individual propensity to be badly informed, mistakenly wrong, or be limited by
its cognitive capacity (Sasaki and Pratt, 2012).
The second set of examples as above highlights behaviors in which the group holds useful
information about the environment but this information is distributed among a large
number of individuals. Using communication, these information fragments can be integrated
11
to yield collective decisions that take the "big picture" into account. Such integration can
often be classified using two general schemes defined for animal groups in general: The
"many eyes principle" (Ward et al., 2011) in which the group's capacity for surveillance
increases with the number of alert animals and the "many wrongs principle"(Biro et al.,
2006; Simons, 2004) in which averaging effects work to reduce "noise" at the scale of an
individual animal to yield accurate collective action. Integration of information happens over
a large number of contexts and species.Some recurring principles in this process include:
distributed integration, nonlinearities (Sumpter and Beekman, 2003), positive and negative
feedback loops (Franks et al., 2002; Nieh, 2010), use of interaction rates and cuing delays
(Camazine, 1993; Greene et al., 2013), higher influence and lower propensity to be
influenced by better informed individuals (Korman et al., 2014; Razin et al., 2013; Schultz et
al., 2008), and correctly weighing individual vs. social information (Robinson et al., 2009b;
Robinson et al., 2012).
This section also presents examples in which colonies have evolved to achieve a collective
task by amplifying the behaviors of individuals that work towards the small scale version of
this same task. For example, in the context of cooperative transport the actions of a leader
ant are exactly those she would take when individually transporting a small food item to the
nest. The group allows these actions to be the driving force behind the transport of a large
heavy item.
Emergent group level cognition
Last, some colony functions may fall outside of the solitary insect's behavioral or cognitive
repertoire. These include large scale behaviors such as assessing the relation between the
size of two objects when bothof these are much larger than the insect itself (Fonio et al.,
2016; Gelblum et al., 2016), and active food dissemination (Camazine, 1993; Greenwald et
al., 2015; Howard and Tschinek, 1981). In such cases, to achieve the group-level task, the
relevant capabilitiesmust arise either by newly evolved individual traits or through a
collective process that relies on the connectivity between individuals (a combination of
these two options is also possible). Cases in which group level processes bestow the group
with abilities that are qualitatively beyond those of its individual members can, by
definition(De Wolf and Holvoet, 2005), be considered as a form of emergence.
In the two examples of cooperative transport presented above, group problem solving
capabilities increase beyond those of individuals. In these cases, the small size of individual
insect may prevent it from grasping the relevant large-scale relationships between the
carried load and the obstacle. Therefore, in these cases the problem solving on the level of
the group must decouple to some extent from the array of, possibly misleading, solutions as
offered by individuals, even if these are highly adept navigators. The mechanisms that
enable such decoupling between the scales do not have to be complex. In these examples,
group-level noise in scent mark following (Fonio et al., 2016) and the persistence that results
from alignment of forces (Gelblum et al., 2016) suffice for efficient transport that interacts
with the environment on the relevant scale of the ant team and the large carried load.
12
Stigmergic nest construction holds the potential for higher levels of emergence. Here, nest
construction is carried out by individuals who follow simple local rules while constantly
reacting to the environmental product of their previous actions (Grassé, 1959). Following
these stigmergic principles (Theraulaz et al., 2003) may allow insects to construct elaborate
nest architectures (Tschinkel, 2004) without any blueprint. The degree of emergence in such
processes is difficult to define. On the one hand, by using a stigmergic process, a single
insect, a solitary queen for example, may construct a structure for which it holds no internal
representation (Camazine et al., 2001; Theraulaz and Bonabeau, 1999). A larger group in
which each individual follows the same rules may achieve faster construction but the quality
of emergence remains constant (Theraulaz and Bonabeau, 1999). On the other hand, it has
been shown that pheromone deposition onto the constructed nest is an essential part of
collective stigmergic nest construction. The addition of this time dependent component
(pheromones evaporate) has the potential of allowing the group to achieve more complex
forms of emergence not achievable by a single individual working alone (Khuong et al.,
2016). Indeed, such processes can be modeled using the mathematics of cellular automata a
theory which holds the potential of describing extreme forms of emergence(Wolfram, 2002).
Prevalence of individual-based mechanisms
Next, we raise somehypothesesin anattempt to provide rationale for the observed
imminence of individuals to colony scale processes.
As shown above, it is often the case that colony level behaviors are a consequence of the
direct of amplification of individual actionsthat rely on individual cognition. This didn't have
to be the case. One could imagine an evolutionary pathway where collective cognition is
constructed from a network of different individuals each manifesting different, possibly
more basic, capabilities.In this case, the actions of individuals don't directly coincide with the
actions of the group or even with each other. Rather, group performance emerges from the
coordination between individuals. An example for this are neurons in the brain. Neurons
have evolved to be cells whose actions are purely computational such that their spiking
activity bears no direct relations to the collective process at which they participate.
Hypothetically, there is no a-priori prevention that such structures arise in a social insect
colony to serve as the basis for highly efficient collective scale behaviors. Why, therefore, is
this not the general case?
We suggest that the first part of the answer to this question has to do with the availability
and high quality performances that individual-based group level cognition can achieve. As
stated above, by the time group cognition has evolved, individual insects were already
developed independent organisms(Farris and Schulmeister, 2010; Ma et al., 2012). As such,
they already possessed many of the cognitive resources that are required by the group.
Furthermore, these individual capabilities are, in no sense, simple. For example, individual
insects are highly adept navigators. To get from one place to another individual insects
employ a toolbox consisting of multiple tactics, often applied in parallel. Across different
insects, these include landmark navigation (Collett et al., 1993), dead reckoning (Collett and
Collett, 2000), backtracking (Wystrach et al., 2013), and cognitive map
(Gould,
1986)navigation by means of scents (Morgan, 2009), visual cues (Collett et al., 1993; Esch et
13
al., 2001; Wehner, 2003), temperature, and even magnetic fields. Collective cognition that
relies on such individual capabilities holds the advantage of utilizing these highly non-trivial
traits.
The fact that individual insects possess high cognitive capabilities does not, however, suffice
in explaining why these capabilities take a central role in many collective level processes.In
the context of the navigation example as in the previous paragraph,it may very well be the
case that a navigational toolbox that relies on the distributed actions of a large number of
cooperating insects not only exists but, also, outperforms other, individual-based, schemes.
Again, one may ask why this is not the general case and why the role of individuals has
remained so pronounced throughout evolution. We hypothesize that the answer to this
question has to do with the complexity of efficient distributed solutions given the inherent
constraints that apply for a colony of autonomous individuals. We hypothesize that in many
cases, simple distributed algorithms would not outperform what is readily achievable by
amplifying individual cognition. We further hypothesize that distributed algorithms that do
surpass simpler amplification processes may be expected to be highly complex and therefore
difficult to evolve. Hence, the system may be trapped within local minima within the fitness
landscape which utilize the individual cognitive components that have evolved at earlier
stages. The next sectiondiscusses some of the non-trivial limitations on emergent behavioral
solutions.
Computational constraints on emergent collective cognition
Emergence is often associated with the notion of a group that "exceeds the sum of its
parts". Theoretically speaking, achieving such highly effective cooperation typically requires
a substantial degree of coordination. An intuitive example for this is the parallel search
problem in which a group of non-communicating random walkers that start at a given
location aim to collectively cover an area and then share the profits of their findings. In this
context, to "exceed the sum of its parts" implies that the group of N searchers cover the
area more than N times faster than each individual, were it acting on its own. In fact, in
many natural topologies (Alon et al., 2011; Efremenko and Reingold, 2009), the situation is
very far from this. For example, in grid topologies when searchers do not communicate and
as long as N is not too small, multiple random walks that start from a single point, typically
achieve negligible speed-up in cover time when compared to a single searcher(Alon et al.,
2011) (i.e., the group takes almost as much time to cover an area as a single one of its
members). In this case, being part of a group becomes highly unbeneficial from an
individual's point of view: food must be shared but without the advantage of obtaining it at
a faster rate.
Biological systems often achieve emergence using involved communication. The most
celebrated example for this is that of the brain, viewed as a large ensemble of neurons. In
the case of brains, such coordination heavily relies on the fact that the neurons are
organized into stable networks such that the set of neighboring neurons of a given neuron
remains, relatively, fixed. Synaptic plasticity allows a pair of neighboring neurons to undergo
a mutual feedback process and fine-tune theirconnectivity(Hebb, 1949). This, as
conceptually demonstrated by the Hopfield model (Hopfield, 1982), may provide the system
14
with its immense computational power. Indeed, Hopfield networks have been shown to be
capable of universal computation in the Turing sense (Síma and Orponen, 2003).
What can one expect for social insect colonies? On the one hand,these are very far from
being non-communicating. On the other hand, interaction networks appear to be more
loosely defined than those that characterize a brain. Indeed, in the insect colony fixed
organization structures may be difficult to achieve for long periods of time, due to the
inherent mobility and anonymity constraints. This lack of structure leads to a lack of
knowledge that further constrains the system(Burgos and Polani, 2016). Namely, while
learning in the brain happens by strengthening the synaptic connections between pairs of
specific hard-wired (and therefore mutually identifiable) neurons, it is difficult imagine such
mechanisms for freely moving, anonymous insects. Further, many brain functions are known
to depend on precise interaction patterns that enable, for example, high levels of
synchrony(Abeles, 2010). Such preciseness is difficult to imagine in the case of autonomous
independent agents.
It appears that things become even worse if communication and information flows across
the systems are, themselves, limited or distorted (Feinerman and Korman, 2012; Feinerman
et al., 2014; Korman et al., 2014; Razin et al., 2013). The highly dynamic environment and
interchanging network structures that characterize the social insect colony make it difficult
to implement error correcting mechanisms such as repeatedly sending the same message in
orderto reduce distortion. As demonstrated in (Feinerman and Korman, 2012; Feinerman et
al., 2014; Korman et al., 2014; Razin et al., 2013) such circumstances make even basic
distributed tasks, such as rumor spreading, challenging. It is therefore reasonable to assume
that emergent phenomena would be even more difficult to implement in such conditions.
Summary
Useful information exists at the level of the individual insect. This often suffices for the
group's needs and indeed we have presented many examples in which the group follows
individuals by amplifying their effect. In other cases, the inherent scale gap between
individual and group may render information held by individuals partial, irrelevant, or even
misleading to the collective goals of the group. In such cases, emergent collective
phenomenon may kick in. Cognition that emerges at the level of the colony is subject to
multiple constraints that mainly result from the fact that individuals maintain their
autonomy as insects within the colony. Indeed, the most complex collective circuits
described to date(Nieh, 2010; Pratt et al., 2005; Seeley et al., 2011) may be defined as
simple when compared to the neural circuits that allow, for example, an ant to find her nest
by using vector integration(Ofstad et al., 2011; Wehner, 2003).
Interestingly, the ways in which collective cognition appears, even during a single behavior,
are not mutually exclusive.An example for thiscomes from cooperative transport in
Paratrechina longicornis ants. Indeed, during this behavior the ants simultaneously benefit
from both individual based(Gelblum et al., 2015) and collective based cognition(Fonio et al.,
15
2016; Gelblum et al., 2016). This balance between the organizational scales allows the
system to enjoy the best of both the macroscopic and the microscopic worlds.
Forming better connections between different collective behaviors and the levels of
emergence that characterize them requires further research. We suggest that such research
be focused on two avenues. The first is a computational study of the powers and limitations
of natural distributed algorithms(Feinerman et al., 2014; Greenwald et al., 2015).
Specifically, there is a need to better understand the complexity and evolvabilty of
distributed solutions of different qualities. The second avenue involves empirical studies that
attempt to trace the evolution of emergent, communication-based solutions. An example for
this direction can be a comparative study of nest structures across a large number of
phylogenetically related ant or termite species.
Acknowledgements
This work has received funding from the European Research Council (ERC) under the
European Union's Horizon 2020 research and innovation program (grant agreement No
648032).O.F., incumbent of the Louis and Ida Rich career development chair, supported by
Israel Science Foundation grant 833/15, would like to thank and the Clore Foundation for
their ongoing support. A.K. We would like to than Yossi Yovel for his comments on this
manuscript.
16
Bibliography
Abeles, M. (2010). Synfire Chains. In Encyclopedia of Neuroscience, pp. 829–832.
Aiello, L. C. and Wheeler, P. (1995). The Expensive-Tissue Hypothesis: The Brain and the
Digestive System in Human and Primate Evolution. Curr. Anthropol.36, 199.
Alloway, T. (1972). Learning and memory in insects. Annu. Rev. Entomol.17, 43–56.
Alon, N., Avin, C., Koucky, M., Kozma, G., Lotker, Z. and Tuttle, M. R. (2011). Many Random
Walks Are Faster Than One. Comb. Probab. Comput.20, 481–502.
Amor, F., Ortega, P., Cerdá, X. and Boulay, R. (2010). Cooperative prey-retrieving in the ant
Cataglyphis floricola: an unusual short-distance recruitment. Insectes Soc.57, 91–94.
Anderson, C. and McShea, D. W. (2001). Individual versus social complexity, with particular
reference to ant colonies. Biol. Rev. Camb. Philos. Soc.76, 211–37.
Beckers, R., Goss, S., Deneubourg, J. L. and Pasteels, J. (1989). Colony size, communication,
and ant foraging strategy. Psyche (Stuttg).96, 239–256.
Beckers, R., Deneubourg, J.-L. and Goss, S. (1992a). Trail Laying Behavior During Food
Recruitment In The Ant Lasius-Niger (L). Insectes Soc.39, 59–72.
Beckers, R., Deneubourg, J. L. and Goss, S. (1992b). Trails and U-turns in the selection of a
path by the ant Lasius niger. J. Theor. Biol.159, 397–415.
Beekman, M. and Dussutour, A. (2007). How to Tell Your Mates Recruitment Mechanisms.
105–124.
Behmer, S. T. (2009). Animal Behaviour: Feeding the Superorganism. Curr. Biol.19, R366–
R368.
Beshers, S. N. and Fewell, J. H. (2001). Models of division of labor in social insects. Annu.
Rev. Entomol.46, 413–40.
Biro, D., Sumpter, D. J. T., Meade, J. and Guilford, T. (2006). From compromise to
leadership in pigeon homing. Curr. Biol.16, 2123–8.
Blackiston, D., Briscoe, A. D. and Weiss, M. R. (2011). Color vision and learning in the
monarch butterfly, Danaus plexippus (Nymphalidae). J. Exp. Biol.214, 509–520.
Blonder, B. and Dornhaus, A. (2011). Time-ordered networks reveal limitations to
information flow in ant colonies. PLoS One6, e20298.
Blum, M. S. (1969). Alarm pheromones. Annu. Rev. Entomol.14, 57–80.
Bonabeau, E. (1996). Quantitative study of the fixed threshold model for the regulation of
division of labour in insect societies. Proc. R. Soc. B Biol. Sci.263, 1565–1569.
Brady, S. G., Schultz, T. R., Fisher, B. L. and Ward, P. S. (2006). Evaluating alternative
hypotheses for the early evolution and diversification of ants. Proc. Natl. Acad. Sci.103,
18172–18177.
Brower, L. (1996). Monarch butterfly orientation: missing pieces of a magnificent puzzle. J.
Exp. Biol.199, 93–103.
Burgos, A. C. and Polani, D. (2016). An Informational Study of the Evolution of Codes and of
Emerging Concepts in Populations of Agents. Artif. Life22, 196–210.
17
Byrne, M., Dacke, M., Nordstrom, P., Scholtz, C. and Warrant, E. (2003). Visual cues used by
ball-rolling dung beetles for orientation. J. Comp. Physiol. A Sensory, Neural, Behav.
Physiol.189, 411–418.
Camazine, S. (1993). The regulation of pollen foraging by honey bees: how foragers assess
the colony's need for pollen. Behav. Ecol. Sociobiol.32, 265–272.
Camazine, S., Deneubourg, J.-L., Franks, N. R., Sneyd, J., Theraulaz, G. and Bonabeau, E.
(2001). Self-organisation in biological systems.
Cammaerts, M.-C. and Cammaerts, R. (1980). Food recruitment strategies of the ants
Myrmica sabuleti and Myrmica ruginodis. Behav. Processes5, 251–270.
Chittka, L. and Niven, J. (2009). Are bigger brains better? Curr. Biol.19, R995–R1008.
Collett, M. and Collett, T. S. (2000). How do insects use path integration for their
navigation? Biol. Cybern.83, 245–59.
Collett, T. S., Fry, S. N. and Wehner, R. (1993). Sequence learning by honeybees. J. Comp.
Physiol. A172, 693–706.
Couzin, I. D. (2009). Collective cognition in animal groups. Trends Cogn. Sci.13, 36–43.
Crespi, B. J. and Yanega, D. (1995). The definition of eusociality. Behav. Ecol.6, 109–115.
Czaczkes, T. J. and Ratnieks, F. L. W. (2013). Cooperative transport in ants ( Hymenoptera :
Formicidae ) and elsewhere. Myrmecological News18, 1–11.
De Wolf, T. and Holvoet, T. (2005). Emergence versus self-organisation: Different concepts
but promising when combined. Lect. Notes Comput. Sci. (including Subser. Lect. Notes
Artif. Intell. Lect. Notes Bioinformatics)3464 LNAI, 1–15.
Delattre, O., Sillam-Dussès, D., Jandák, V., Brothánek, M., Rücker, K., Bourguignon, T.,
Vytisková, B., Cvačka, J., Jiříček, O. and Šobotník, J. (2015). Complex alarm strategy in
the most basal termite species. Behav. Ecol. Sociobiol. 1945–1955.
Deneubourg, J. L., Pasteels, J. M. and Verhaeghe, J. C. (1983). Probabilistic behaviour in
ants: A strategy of errors? J. Theor. Biol.105, 259–271.
Dickson, B. J. (2008). Wired for sex: the neurobiology of Drosophila mating decisions.
Science322, 904–909.
Dornhaus, A. and Franks, N. R. (2008). Individual and collective cognition in ants and other
insects (Hymenoptera: Formicidae). Myrmecological News11, 215–226.
Edwards, S. C. and Pratt, S. C. (2009). Rationality in collective decision-making by ant
colonies. Proc. Biol. Sci.276, 3655–61.
Efremenko, K. and Reingold, O. (2009). How Well Do Random Walks Parallelize? In
Approximation, Randomization, and Combinatorial Optimization. Algorithms and
Techniques, pp. 476–489.
Emerson, A. E. (1939). Social Coordination and the Superorganism. Am. Midl. Nat.21, 182–
209.
Engel, M. S., Grimaldi, D. a. and Krishna, K. (2009). Termites (Isoptera): Their Phylogeny,
Classification, and Rise to Ecological Dominance. Am. Museum Novit.3650, 1–27.
Esch, H. E., Zhang, S., Srinivasan, M. V and Tautz, J. (2001). Honeybee dances communicate
distances measured by optic flow. Nature411, 581–583.
18
Evans, H. E. (1966). The Behavior Patterns of Solitary Wasps. Annu. Rev. Entomol.11, 123–
154.
Farris, S. M. and Schulmeister, S. (2010). Parasitoidism, not sociality, is associated with the
evolution of elaborate mushroom bodies in the brains of hymenopteran insects. Proc.
Biol. Sci.278, 940–51.
Feinerman, O. and Korman, A. (2012). Memory lower bounds for randomized collaborative
search and implications for biology. In Proceedings of International Symposium on
Distributed COmputing (DISC), pp. 61–75.
Feinerman, O. and Traniello, J. F. A. (2015). Social complexity, diet, and brain evolution:
modeling the effects of colony size, worker size, brain size, and foraging behavior on
colony fitness in ants. Behav. Ecol. Sociobiol.70, 1063–1074.
Feinerman, O., Haeupler, B. and Korman, A. (2014). Breathe before speaking: efficient
information dissemination despite noisy, limited, and anonymous communication. In
Proceedings of the 2014 ACM symposium on Principles of distributed computing - PODC
'14, pp. 114–123.
Fewell, J. H. (2003). Social insect networks. Science301, 1867–70.
Feynmann, R. P. (1985). The Amateur Scientist. In Surely, you're joking Mr. Feynmann!, p.
79. Bantam Books.
Fonio, E., Heyman, Y., Boczkowski, L., Gelblum, A., Kosowski, A., Korman, A. and
Feinerman, O. (2016). A Locally-Blazed Ant Trail Achieves Efficient Collective
Navigation Despite Limited Information. Rev.
Franks, N. R. (1989). Army Ants: A Collective Intelligence. Am. Sci.77, 138–145.
Franks, N. and Deneubourg, J. (1997). Self-organizing nest construction in ants: individual
worker behaviour and the nest's dynamics. Anim. Behav.54, 779–96.
Franks, N. R. and Tofts, C. (1994). Foraging for work : how tasks allocate workers. Anim.
Behav.48, 470–472.
Franks, N. R., Pratt, S. C., Mallon, E. B., Britton, N. F. and Sumpter, D. J. T. (2002).
Information flow, opinion polling and collective intelligence in house-hunting social
insects. Philos. Trans. R. Soc. Lond. B. Biol. Sci.357, 1567–83.
Franks, N. R., Mallon, E. B., Bray, H. E., Hamilton, M. J. and Mischler, T. C. (2003). Strategies
for choosing between alternatives with different attributes: exemplified by house-
hunting ants. Anim. Behav.65, 215–223.
Gathmann, A. and Tscharntke, T. (2002). Foraging ranges of solitary bees. J. Anim. Ecol.71,
757–764.
Gelblum, A., Pinkoviezky, I., Fonio, E., Ghosh, A., Gov, N. and Feinerman, O. (2015). Ant
groups optimally amplify the effect of transiently informed individuals. Nat. Commun.6,
1–9.
Gelblum, A., Pinkoviezky, I., Fonio, E., Gov, N. S. and Feinerman, O. (2016). The Ant
Pendulum: problem solving through an emergent oscillatory phase. Proc. Natl. Acad.
Sci.Accepted f,.
Gillooly, J. F., Hou, C. and Kaspari, M. (2010). Eusocial insects as superorganisms Insights
from metabolic theory. Commun. Integr. Biol.3, 360–362.
Gordon, D. M. (1995). The expandable network of ant exploration. Anim. Behav.50, 995–
19
1007.
Gordon, D. M. and Mehdiabadi, N. J. (1999). Encounter rate and task allocation in harvester
ants. Behav. Ecol. Sociobiol.45, 370–377.
Gordon, D. M., Paul, R. E. and Thorpe, K. (1993). What is the function of encounter pattern
in ant colonies. Anim. Behav.45, 1083–1100.
Goss, S., Aron, S., Deneubourg, J. L. and Pasteels, J. M. (1989). Self-organized shortcuts in
the Argentine ant. Naturwissenschaften76, 579–581.
Gould, J. L. (1986). The locale map of honey bees: do insects have cognitive maps?
Science232, 861–863.
Grassé, P. P. (1959). La reconstruction du nid et les coordinations interindividuelles chez
Bellicositermes natalensis et Cubitermes sp. la th??orie de la stigmergie: Essai
d'interpr??tation du comportement des termites constructeurs. Insectes Soc.6, 41–80.
Greene, M. J., Pinter-Wollman, N. and Gordon, D. M. (2013). Interactions with combined
chemical cues inform harvester ant foragers' decisions to leave the nest in search of
food. PLoS One8, e52219.
Greenwald, E., Segre, E. and Feinerman, O. (2015). Ant trophallactic networks:
simultaneous measurement of interaction patterns and food dissemination. Sci. Rep.5,
12496.
Hebb, D. O. (1949). The organization of behavior: a neuropsychological theory.
Holldobler, B. (1971). Recruitment Behavior in Camponotus socius (Hym. Formicidae). Z. Vgl.
Physiol.75, 123–142.
Holldobler, B. and Wilson, E. O. (1990). The Ants. Harvard University Press.
Hölldobler, B. (1976). Recruitment behavior, home range orientation and territoriality in
harvester ants, Pogonomyrmex. Behav. Ecol. Sociobiol.1, 3–44.
Hou, C., Kaspari, M., Vander Zanden, H. B. and Gillooly, J. F. (2010). Energetic basis of
colonial living in social insects. Proc. Natl. Acad. Sci. U. S. A.107, 3634–3638.
Howard, R. W. and Blomquist, G. J. (2005). Ecological, behavioral, and biochemical aspects
of insect hydrocarbons. Annu. Rev. Entomol.50, 371–393.
Howard, D. F. and Tschinek, W. R. (1981). The flow of food in colonies of the fire ant,
Solenopsis invicta: a multifactorial study. Physiol. Entomol.6, 297–306.
Jaffe, K. and Howse, P. E. (1979). The mass recruitment system of the leaf cutting ant, Atta
cephalotes (L.). Anim. Behav.27, 930–939.
Jandt, J. M. and Dornhaus, A. (2009). Spatial organization and division of labour in the
bumblebee Bombus impatiens. Anim. Behav.77, 641–651.
Jeanson, R. and Deneubourg, J. L. (2009). Positive feedback, convergent collective patterns,
and social transitions in arthropods. In Organization of Insect Societies - From genome
to sociocomplexity, pp. 460–482.
Jones, J. C. (2004). Honey Bee Nest Thermoregulation: Diversity Promotes Stability. Science
(80-. ).305, 402–404.
Khuong, A., Gautrais, J., Perna, A., Sbaï, C., Combe, M., Kuntz, P., Jost, C. and Theraulaz, G.
information shape ant nest
(2016). Stigmergic construction and topochemical
architecture. Proc. Natl. Acad. Sci.113, 1303–1308.
20
King, H., Ocko, S. and Mahadevan, L. (2015). Termite mounds harness diurnal temperature
oscillations for ventilation. Proc. Natl. Acad. Sci.112, 11589–11593.
Korman, A., Greenwald, E. and Feinerman, O. (2014). Confidence Sharing: An Economic
Strategy for Efficient Information Flows in Animal Groups. PLoS Comput. Biol.10,
e1003862.
Lamon, B. and Topoff, H. (1981). Avoiding predation by army ants: Defensive behaviours of
three ant species of the genus Camponotus. Anim. Behav.29, 1070–1081.
Ma, X., Hou, X., Edgecombe, G. D. and Strausfeld, N. J. (2012). Complex brain and optic
lobes in an early Cambrian arthropod. Nature490, 258–261.
Mailleux, A.-C., Deneubourg, J.-L. and Detrain, C. (2003). Regulation of ants' foraging to
resource productivity. Proc. R. Soc. London. Ser. B Biol. Sci.270, 1609–1616.
Mallon, E. B. and Franks, N. R. (2000). Ants estimate area using Buffon's needle. Proc. Biol.
Sci.267, 765–70.
Martin, S. and Drijfhout, F. (2009). A review of ant cuticular hydrocarbons. J. Chem. Ecol.35,
1151–61.
Mersch, D. P., Crespi, A. and Keller, L. (2013). Tracking individuals shows spatial fidelity is a
key regulator of ant social organization. Science340, 1090–3.
Monographs, E. (2016). Colony Defense Strategies of the Honeybees in Thailand Author ( s ):
Thomas D . Seeley , Robin Hadlock Seeley and Pongthep Akratanakul Published by :
Wiley Stable URL : http://www.jstor.org/stable/2937344 REFERENCES
Linked
references are available on JSTO. 52, 43–63.
Moreau, C. S. (2006). Phylogeny of the Ants: Diversification in the Age of Angiosperms.
Science (80-. ).312, 101–104.
Moreau, M., Arrufat, P., Latil, G. and Jeanson, R. (2011). Use of radio-tagging to map spatial
organization and social interactions in insects. J. Exp. Biol.214, 17–21.
Morgan, D. E. (2009). Trail pheromones of ants. Physiol. Entomol.34, 1–17.
Nieh, J. C. (2010). A negative feedback signal that is triggered by peril curbs honey bee
recruitment. Curr. Biol.20, 310–5.
Noirot, C. and Darlington, J. P. E. C. (2000). Termite Nests: Architecture, Regulation, and
Defense. In Termites: Evolution, Sociality, Symbiosis Ecology, pp. 121–139. Springer.
O'Neill, K. (2001). Solitary Wasps: Behavior and Natural History.
Ofstad, T. A., Zuker, C. S. and Reiser, M. B. (2011). Visual place learning in Drosophila
melanogaster. Nature474, 204–7.
Pinter-Wollman, N., Wollman, R., Guetz, A., Holmes, S. and Gordon, D. M. (2011). The
effect of individual variation on the structure and function of interaction networks in
harvester ants. J. R. Soc. Interface8, 1562–73.
Pratt, S. C., Sumpter, D. J. T., Mallon, E. B. and Franks, N. R. (2005). An agent-based model
of collective nest choice by the ant Temnothorax albipennis. Anim. Behav.70, 1023–
1036.
Queller, D. C. and Strassmann, J. E. (1998). Kin Selection and Social Insects. Bioscience48,
165–175.
Ramsden, E., Adams, J., Holldobler, B., Shepherd, J. D., North, R. D., Jackson, C. W., Howse,
21
P. E., Bash, E., Millor, J., Am??, J. M., et al. (2009). Orientation cues used by ants.
Insectes Soc.37, 101–115.
Raw, A. (1972). The biology of the solitary bee Osmia rufa (L.) (Megachilidae). Trans. R.
Entomol. Soc. London124, 213–229.
Razin, N., Eckmann, J. and Feinerman, O. (2013). Desert ants achieve reliable recruitment
across noisy interactions. J. R. Soc. Interface10, 20130079.
Reid, C. R., Sumpter, D. J. T. and Beekman, M. (2011). Optimisation in a natural system:
Argentine ants solve the Towers of Hanoi. J. Exp. Biol.214, 50–8.
Richardson, T. O. and Gorochowski, T. E. (2015). Beyond contact-based transmission
networks : the role of spatial coincidence. J. R. Soc. Interface12, 20150705.
Richardson, T. O., Christensen, K., Franks, N. R., Jensen, H. J. and Sendova-Franks, A. B.
(2011). Ants in a labyrinth: A statistical mechanics approach to the division of labour.
PLoS One6,.
Rieucau, G. and Giraldeau, L.-A. (2011). Exploring the costs and benefits of social
information use: an appraisal of current experimental evidence. Philos. Trans. R. Soc.
Lond. B. Biol. Sci.366, 949–57.
Robinson, G. E. (1992). Regulation of division of labor in insect societies. Annu. Rev.
Entomol.37, 637–665.
Robinson, E. J. H., Feinerman, O. and Franks, N. R. (2009a). Flexible task allocation and the
organization of work in ants. Proc. Biol. Sci.276, 4373–80.
Robinson, E. J. H., Richardson, T. O., Sendova-Franks, A. B., Feinerman, O. and Franks, N. R.
(2009b). Radio tagging reveals the roles of corpulence, experience and social
information in ant decision making. Behav. Ecol. Sociobiol.63, 627–636.
Robinson, E. J. H., Franks, N. R., Ellis, S., Okuda, S. and Marshall, J. a R. (2011). A simple
threshold rule is sufficient to explain sophisticated collective decision-making. PLoS
One6, e19981.
Robinson, E. J. H., Feinerman, O. and Franks, N. R. (2012). Experience, corpulence and
decision making in ant foraging. J. Exp. Biol.215, 2653–9.
Robinson, E. J. H., Feinerman, O. and Franks, N. R. (2014). How collective comparisons
emerge without individual comparisons of the options. Proc. Biol. Sci.281,.
Robson, S. K. and Traniello, J. F. (2002). Transient division of labor and behavioral
specialization in the ant Formica schaufussi. Naturwissenschaften89, 128–131.
Roces, F., Tautz, J. and Hölldobler, B. (1993). Stridulation in leaf-cutting ants - Short-range
recruitment through plant-borne vibrations. Naturwissenschaften80, 521–524.
Sasaki, T. and Pratt, S. C. (2011). Emergence of group rationality from irrational individuals.
Behav. Ecol.22, 276–281.
Sasaki, T. and Pratt, S. C. (2012). Groups have a larger cognitive capacity than individuals.
Curr. Biol.22,.
Sasaki, T., Granovskiy, B., Mann, R. P., Sumpter, D. J. T. and Pratt, S. C. (2013). Ant colonies
outperform individuals when a sensory discrimination task is difficult but not when it is
easy. Proc. Natl. Acad. Sci.110, 13769–13773.
Schmidt, J. O. (1990). Insect defenses: adaptive mechanisms and strategies of prey and
22
predators. SUNY Press.
Schultz, K. M., Passino, K. M. and Seeley, T. D. (2008). The mechanism of flight guidance in
honeybee swarms: subtle guides or streaker bees? J. Exp. Biol.211, 3287–95.
Seeley, T. D. (1996). The Wisdom of the Hive.
Seeley, T. D., Visscher, P. K. and Passino, K. M. (2006). Group Decision Making in Honey Bee
Swarms. Am. Sci.94, 220–229.
Seeley, T. D., Visscher, P. K., Schlegel, T., Hogan, P. M., Franks, N. R. and Marshall, J. A. R.
(2011). Stop Signals Provide Cross Inhibition in Collective Decision-Making by Honeybee
Swarms. Science.
Sendova-Franks, a. B. and Franks, N. R. (1995). Spatial relationships within nests of the
antLeptothorax unifasciatus(Latr.) and their implications for the division of labour.
Anim. Behav.50, 121–136.
Shultz, S. and Dunbar, R. (2010). Encephalization is not a universal macroevolutionary
phenomenon in mammals but is associated with sociality. Proc. Natl. Acad. Sci. U. S.
A.107, 21582–21586.
Simon, T. and Hefetz, A. (1992). Dynamics of mass recruitment and its regulation in
Tapinoma simrothi. Biol. Evol. Soc. Insects 325–334.
Simons, A. M. (2004). Many wrongs: the advantage of group navigation. Trends Ecol.
Evol.19, 453–5.
Siveter, D. J., Tanaka, G., Farrell, U. C., Martin, M. J., Siveter, D. J. and Briggs, D. E. G.
(2014). Exceptionally preserved 450-million-year-old ordovician ostracods with brood
care. Curr. Biol.24, 801–806.
Starks, P. T., Blackie, C. A. and Thomas D Seeley, P. T. (2000). Fever in honeybee colonies.
Naturwissenschaften87, 229–231.
Strausfeld, N. J. (1976). Atlas of an insect brain. Springer Science & Business Media.
Stroeymeyt, N., Franks, N. R. and Giurfa, M. (2011). Knowledgeable individuals lead
collective decisions in ants. J. Exp. Biol.214, 3046–3054.
Sumpter, D. J. T. (2006). The principles of collective animal behaviour. Philos. Trans. R. Soc.
Lond. B. Biol. Sci.361, 5–22.
Sumpter, D. J. T. and Beekman, M. (2003). From nonlinearity to optimality: pheromone trail
foraging by ants. Anim. Behav.66, 273–280.
Theraulaz, G. and Bonabeau, E. (1999). A Brief History of Stigmergy. Artif. Life5, 97–116.
Theraulaz, G., Gautrais, J., Camazine, S. and Deneubourg, J.-L. (2003). The formation of
spatial patterns in social insects: from simple behaviours to complex structures. Philos.
Trans. A. Math. Phys. Eng. Sci.361, 1263–82.
Tibbetts, E. a (2002). Visual signals of individual identity in the wasp Polistes fuscatus. Proc.
Biol. Sci.269, 1423–1428.
Tschinkel, W. R. (2004). The nest architecture of the Florida harvester ant, Pogonomyrmex
badius. J. Insect Sci.4, 21.
van Wilgenburg, E., Symonds, M. R. E. and Elgar, M. A. (2011). Evolution of cuticular
hydrocarbon diversity in ants. J. Evol. Biol.24, 1188–1198.
23
Visscher, P. K. (2007). Group decision making in nest-site selection among social insects.
Annu. Rev. Entomol.52, 255–75.
Von Frisc, K. (1950). Bees: their vision, chemical senses, and language. Cornell University
Press.
Wang, B., Xia, F., Wappler, T., Simon, E., Zhang, H., Jarzembowski, E. A. and Szwedo, J.
(2015). Brood care in a 100-million-year-old scale insect. Elife2015, 4–11.
Ward, A. J. W., Herbert-read, J. E., Sumpter, D. J. T. and Krause, J. (2011). Fast and accurate
decisions through collective vigilance in fish shoals. Proc. Natl. Acad. Sci.108, E27–E27.
Waters, J. S., Holbrook, C. T., Fewell, J. H. and Harrison, J. F. (2010). Allometric scaling of
metabolism, growth, and activity in whole colonies of the seed-harvester ant
Pogonomyrmex californicus. Am. Nat.176, 501–10.
Wehner, R. (2003). Desert ant navigation: how miniature brains solve complex tasks. J.
Comp. Physiol. A. Neuroethol. Sens. Neural. Behav. Physiol.189, 579–88.
Wilson, E. O. (1962). Chemical communication among workers of the fire ant Solenopsis
saevissima (Fr. Smith) 1. The Organization of Mass-Foraging. Anim. Behav.10, 134–147.
Wilson, E. O. and Hölldobler, B. (1988). Dense heterarchies and mass communication as the
basis of organization in ant colonies. Trends Ecol. Evol. (Personal Ed.3, 65–8.
Wilson, E. and Hölldobler, B. (2009). The superorganism: The Beauty, Elegance, and
Strangeness of Insect Societies. W. W. Norton & Company.
Witte, V., Attygalle, A. B. and Meinwald, J. (2007). Complex chemical communication in the
Formicidae).
(Hymenoptera:
longicornis
Latreille
ant
crazy
Chemoecology17, 57–62.
Paratrechina
Wolfram, S. (2002). A New Kind Of Science. Wolfram Media.
Wystrach, A., Schwarz, S., Baniel, A. and Cheng, K. (2013). Backtracking behaviour in lost
ants: an additional strategy in their navigational toolkit. Proc. Biol. Sci.280, 20131677.
24
|
1907.13118 | 1 | 1907 | 2019-07-29T12:33:41 | Assessment and manipulation of the computational capacity of in vitro neuronal networks through criticality in neuronal avalanches | [
"q-bio.NC"
] | In this work, we report the preliminary analysis of the electrophysiological behavior of in vitro neuronal networks to identify when the networks are in a critical state based on the size distribution of network-wide avalanches of activity. The results presented here demonstrate the importance of selecting appropriate parameters in the evaluation of the size distribution and indicate that it is possible to perturb networks showing highly synchronized---or supercritical---behavior into the critical state by increasing the level of inhibition in the network. The 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. This study is part of a larger research project, the overarching aim of which is to develop computational models that are able to reproduce target behaviors observed in in vitro neuronal networks. These models will ultimately be used to aid in the realization of these behaviors in nanomagnet arrays to be used in novel computing hardwares. | q-bio.NC | q-bio | Assessment and manipulation of the computational
capacity of in vitro neuronal networks
through criticality in neuronal avalanches
Kristine Heiney∗†§, Ola Huse Ramstad‡, Ioanna Sandvig‡, Axel Sandvig‡, and Stefano Nichele∗
†Department of Computer Science, Norwegian University of Science and Technology, Trondheim, Norway
∗Department of Computer Science, Oslo Metropolitan University, Oslo, Norway
‡Department of Neuromedicine and Movement Science, Norwegian University of Science and Technology, Trondheim, Norway
§Email: [email protected]
Abstract -- In this work, we report the preliminary analysis of
the electrophysiological behavior of in vitro neuronal networks
to identify when the networks are in a critical state based on
the size distribution of network-wide avalanches of activity. The
results presented here demonstrate the importance of selecting
appropriate parameters in the evaluation of the size distribution
and indicate that it is possible to perturb networks showing
highly synchronized -- or supercritical -- behavior into the critical
state by increasing the level of inhibition in the network. The
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. This study is part of a larger research
project, the overarching aim of which is to develop computational
models that are able to reproduce target behaviors observed in
in vitro neuronal networks. These models will ultimately be used
to aid in the realization of these behaviors in nanomagnet arrays
to be used in novel computing hardwares.
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 scalable, capable of
learning, 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 -- 4]. 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
[5]. 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 [6], including
those geared toward computation. In addition, it has been
demonstrated that self-organizing substrates, such as magnetic
arrays and self-assembling molecules, can be used as com-
putational reservoirs -- untrained dynamical systems composed
of a collection of recurrently connected units -- by training a
readout layer to map the output of the physical system to a
target problem [7].
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 [8] 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.
9
1
0
2
l
u
J
9
2
]
.
C
N
o
i
b
-
q
[
1
v
8
1
1
3
1
.
7
0
9
1
:
v
i
X
r
a
This work is part of a project entitled Self-Organizing Com-
putational substRATES (SOCRATES) [9], 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
the computational capabilities and dynamics of the resulting
models will be 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 gives some background on the analysis method used in this
work to assess the criticality of in vitro neuronal networks
toward identifying networks that may be considered well-
suited for computation. Section III describes the methods of
preparing the dissociated primary cortical networks evaluated
in this study, which were cultured on a microelectrode array
(MEA) to enable observation of their electrical behavior, and
the analysis methods applied to the electrophysiological data
recorded from the networks. The results from the prelimi-
nary analysis of the recorded data are presented in Section
IV. These results include the observation of the course of
maturation of the networks and their response to chemical
perturbation to increase inhibition. The implications of the
presented results are discussed in Section V. A brief overview
of the long-term plan for this research project is then given in
Section VI, and section VII concludes the paper.
II. BACKGROUND: NEURONAL AVALANCHES
It has been theorized 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 works [10, 11]. A system in
the critical state rests at the boundary between two qualita-
tively 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].
To determine whether a neuronal network is in the crit-
ical state, we turn to the scaling behavior of network-wide
avalanches of activity. As first defined by Beggs and Plenz
[12], 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 no activity, as shown in
Fig. 1b. In their study, Beggs and Plenz [12] demonstrated
that the probability distributions of the size and duration of
neuronal avalanches follow a power law, indicating that the
propagation of activity in the cortex is in the critical state [13].
It has been further demonstrated that criticality is established
by a balance between excitation and inhibition and that cortical
networks at criticality show a larger dynamic range (sensitivity
to a wider range of inputs), higher information capacity
(number of output patterns generated in response to different
inputs), and greater information transmission (information
shared between two recording sites) than networks functioning
outside of criticality [14, 15]. 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, though not all networks settle
into this state and there is disagreement on when the networks
reach criticality [16 -- 18].
In the present study, an analysis method based on the size
distribution of neuronal avalanches was applied to the analysis
of in vitro primary cortical neuronal networks (Fig. 1a). This
method is based on previous analysis performed on cortical
networks (e.g., [16]). 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.
the present
As a long-term goal of
research project,
avalanche size distribution analysis is being applied to record-
ings obtained from different types of in vitro neuronal net-
works (e.g., human induced pluripotent stem cell (iPSC)-
derived dopaminergic networks, see [19]) to assess the criti-
cality of the networks. Emerging network dynamics in healthy
and perturbed conditions will be studied, characterized, and
classified. On the basis of the preliminary results obtained in
this work, we report the development of two cortical networks
from day in vitro (DIV) 7 to DIV 51 and their response to
chemical perturbation on DIV 51.
To the authors' knowledge, this work represents the first
time a network has been manipulated into the critical state
through chemical perturbation. Furthermore, this type of anal-
ysis has also not yet been applied to the characterization of the
neuronal network dynamics in in vitro disease models, which
is the focus of ongoing and future work.
III. METHODS
Neuronal networks were prepared using primary cortical
neurons. This section presents the methods for preparing and
recording from the neuronal networks, as well as the data
analysis methods applied to the recorded data.
A. Preparation and electrophysiology of neuronal networks
The neuronal networks assessed here were prepared as
follows. Primary rat cortical neurons (Thermo Fisher) were
seeded on a feeder layer of human astrocytes (Gibco, Thermo
2
Fisher) at a density of approximately 1 × 105 neurons per
MEA. The networks were left to mature for 7 DIV prior
to recording. 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 were taken
every second day for 15 min. Culture feedings occurred at the
start of each week immediately after recording.
γ-Aminobutyric acid (GABA, Sigma Aldrich) was added
to the cortical networks at rising concentrations to disrupt the
excitation-to-inhibition ratio by increasing network inhibition.
Prior to perturbation, a recording was taken to establish a
baseline. GABA was then added directly to the culture media
in microliter volumes, and recordings were taken immediately
after this perturbation. Different concentrations (10 and 50 µM
for Network A and 5, 10, and 25 µM for Network B) were
chosen to provide increasing degrees of perturbation; lower
concentrations were used for Network B because it showed
lower levels of activity. The 15-min recordings were performed
in succession with increasing concentrations. Following the
final perturbation, 90% of the culture media was replaced to
return the cultures to the baseline state.
B. Avalanche analysis
Avalanches were detected according to the method de-
scribed by Beggs and Plenz [12]. Spikes were detected using
a simple thresholding method [20] based on the standard
deviation of the noise of the signal after applying a bandpass
filter with a pass band of 300 Hz to 3 kHz1. The spike detection
threshold was set to 6, 7, and 8 standard deviations below the
median of the signal for the cortical networks, and the results
were compared to assess the effect of the detection threshold
on the classification results.
The spikes were 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
(Fig. 1b). 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 distribu-
tion data. 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 exponent of the fitted
power law. The exponent has been reported to take a value of
α = 1.5 for in vitro neuronal networks [12, 16]. The fitting
was performed using two different fitting methods, nonlinear
regression (NLR) and maximum likelihood estimation (MLE),
and the results were compared. The fit was applied over the
size range of smin = 2 to the maximum detected avalanche
1Code for spike detection is available at https://github.com/SocratesNFR/
MCSspikedetection.
(a)
(b)
Fig. 1: (a) Microscope images of the primary cortical networks
analyzed in this study. Network A (shown here at DIV 49)
showed activity on all electrodes, whereas Network B (shown
here at DIV 51) partially detached from the electrode area of
the MEA chip, resulting in approximately 12 -- 15 electrodes
of the 60 showing the vast majority of the recorded network
activity. (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.
3
size smax, with a cap at smax = 59 electrodes. The goodness
of fit was computed following Clauset et al. [21]. 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 greater than the empirical KS distance was less than
0.1 (p < 0.1)2.
IV. RESULTS
The avalanche size distributions of different in vitro neu-
ronal networks were observed as the networks matured, and
the networks were classified as being critical or not critical
at each recording time point based on the fitting results. In
this preliminary work, no rigorous analysis was yet applied
to classify networks as sub- or supercritical; rather, only
the goodness of fit of the size distribution to a power law
was evaluated to assess whether or not the network was in
a critical state during each analyzed recording. Preliminary
classification of non-critical cases as sub- or supercritical was
performed by visual inspection alone, where subcritical be-
havior is characterized by exponential decay and supercritical
by a bimodal distribution.
The primary cortical networks evaluated here (Networks A
and B; see Fig. 1a) were observed as they matured from DIV
7 to DIV 51, and then the change in the activity in response
to chemical perturbation by GABA on DIV 51 was evaluated.
These networks showed a large amount of activity with high-
amplitude spikes, providing richer results than iPSC-derived
dopaminergic networks investigated in a previous work [19]
and allowing for the comparison of the results across multiple
detection thresholds (6, 7, and 8 standard deviations below the
median) and fitting techniques (NLR and MLE), as described
in Section III-B.
This section will first present the observed course of matu-
ration for the two cortical networks based on their avalanche
size distributions and then discuss how the addition of GABA
affected the criticality of the networks. In both cases, the
avalanche size distributions and fitting results will be com-
pared for the different detection thresholds and fitting methods
considered in this work.
A. Spontaneous activity during maturation
Both networks showed an initial low-activity phase from
DIV 7 to DIV 12, in which the mean firing rate averaged
across all active electrodes was less than 0.1 s−1 and fewer
than approximately 1000 neuronal avalanches were detected
within the size range (2 -- 59 electrodes) considered for the
power law fitting. In this time period, there were considered
to be too few avalanches for the fitting results to be reliable.
In all subsequent recordings of spontaneous activity (DIVs
14 -- 51), both the NLR and MLE fitting methods produced fits
that did not meet the criterion (p > 0.1) for the networks to
be classified as critical, indicating that neither of the networks
2Code for avalanche detection and goodness of fit evaluation is available
at https://github.com/SocratesNFR/avalanche.
4
(a) DIV 21
(b) DIV 21
(c) DIV 44
(d) DIV 44
(e) DIV 37
(f) DIV 37
Fig. 2: Raster plots and size distributions for Network A.
(a) DIV 21. Raster plots for detection thresholds of 6 (top)
and 8 (bottom) standard deviations. The raster plots show 0.5-
s intervals of high network activity with the intervening silent
periods eliminated for the sake of visualization. (b) DIV 21.
Size distributions for detection thresholds of 6 and 8 standard
deviations. A power law function with an exponent of α = 1.5
is plotted for reference. (c,d) Same as (a,b) for DIV 44. (e,f)
Same as (a,b) for DIV 37.
entered the critical state during the observation period. All
subsequent discussion on classifying the networks as sub-
or supercritical over the course of their maturation is thus
based solely on visual inspection; as stated previously, the
implementation of a rigorous classification method in this
regard remains as a task for future work.
Figure 2 shows a selection of raster plots and avalanche
size probability distributions for Network A at three different
time points (DIVs 21, 44, and 37). The raster plots and size
distributions are shown for detection thresholds of 6 and 8
standard deviations at each time point. The raster plots show
the first three network bursts in each recording with the periods
of relative silence between the bursts removed for the sake of
visualization.
The first time point shown (DIV 21; Fig. 2a and 2b) is
representative of the type of activity seen from DIV 14 to 21.
During this period, the networks appeared to be in a subcritical
state, which is characterized by an exponential size distribution
showing a rapid drop in probability as the avalanche size
increases. As shown in the raster plots, this corresponds to
loose synchrony and relatively low levels of activity. The
second time point (DIV 44; Fig. 2c and 2d) is representative
of the type of activity seen from DIV 23 to 51. During
this period, the networks appeared to be in a supercritical
state, which is characterized by a bimodal size distribution
showing many very large avalanches and few avalanches of
intermediate size. The raster plots demonstrate that this type
of activity corresponds to tight synchronicity across most of
the electrodes and high activity levels.
The final time point shown in Fig. 2 (DIV 37; Fig. 2e and
2f) is shown to demonstrate the robustness of the classification
method based on the goodness of fit measure p against
false positives. The size distribution obtained for a detection
threshold of 8 standard deviations appears to follow a power
law, showing high linearity in log -- log space with a slope cor-
responding to the expected value of α = 1.5 (MLE: α = 1.62,
NLR: α = 1.53). However, in both cases, the criterion p > 0.1
was not met (MLE: p = 0.005, NLR: p = 0.0). The analysis
results at a threshold of 6 standard deviations more clearly
show the deviation of the size distribution from the linear fit,
displaying the same type of bimodal distribution seen at DIV
44 (Fig. 2d). This supports the reliability of the goodness of
fit measure.
B. Perturbation to increase inhibition
On DIV 51, both networks showed highly synchronized
activity and avalanche size distributions with peaks at large
avalanche sizes, behavior indicative of a supercritical state.
Figure 3a and 3e shows raster plots of the activity of Networks
A and B, respectively, on DIV 51 before chemical pertur-
bation, which demonstrate their high level of synchronicity.3
These plots were obtained with a detection threshold of 6
standard deviations. The corresponding avalanche size distri-
butions are shown in Fig. 3c and 3g, respectively. As stated
previously, Network B detached from some of the electrodes
(see Fig. 1a), causing approximately 12 -- 15 of the electrodes
to show the vast majority of the activity, especially during later
recordings. This is apparent in both the raster plot (Fig. 3e)
and the size distribution (Fig. 3g); there is a small peak in the
probability of avalanches of size 12, followed by a very sharp
drop off toward larger avalanches.
3Note that the raster plot for Network A shows 1 min of activity, whereas
that for Network B shows 2 min. Because Network B showed much less
activity than Network A both before and after perturbation, the type of activity
is better captured by showing a larger time window.
(a) Network A before perturbation
(b) Network A after perturbation
(c) Network A before perturbation
(d) Network A after perturbation
(e) Network B before perturbation
(f) Network B after perturbation
(g) Network B before perturbation
(h) Network B after perturbation
Fig. 3: Raster plots and avalanche size distributions before and
after chemical perturbation by GABA (DIV 51, threshold of
6 standard deviations). (a) Raster plot of Network A before
perturbation. (b) Raster plot of Network A after perturbation.
(c) Probability distribution of avalanche sizes for Network
A before perturbation. A power law with an exponent of
α = 1.5 is plotted for reference. (d) Probability distribution of
avalanches sizes for Network A after perturbation. The fitting
results obtained by NLR and MLE are plotted along with their
α and p values. (e -- h) Same as (a -- d) for Network B.
5
TABLE I: Power law fitting results for the two networks after
chemical perturbation on DIV 51. The results are given for
different spike detection thresholds and fitting methods. Bold
text indicates a good fit (p > 0.1). Italicized results indicate
that the number of avalanches N was considered insufficient
for reliable fitting results (N (cid:46) 1000).
Network
Threshold
Fitting method
A
(50 µM
GABA)
B
(25 µM
GABA)
8 st. dev.
7 st. dev.
6 st. dev.
8 st. dev.
7 st. dev.
6 st. dev.
NLR
MLE
NLR
MLE
NLR
MLE
NLR
MLE
NLR
MLE
NLR
MLE
α
1.32
1.54
1.42
1.54
1.40
1.54
1.95
1.67
1.72
1.69
1.60
1.64
p
0.28
0.02
0.72
0.00
0.14
0.00
0.54
0.48
0.16
0.06
0.05
0.31
Figure 3b and 3f shows the raster plots of Networks A
and B, respectively, after the addition of GABA to increase
inhibition. After the addition of GABA, the amount of activity
decreased, and the high level of synchronicity was broken. In
contrast to the behavior prior to perturbation, the perturbed
networks showed much less periodicity in their activity, with
different groupings of electrodes firing within close temporal
proximity. This is reflected in the avalanche size distributions
for the two networks shown in Fig. 3d, which both appear to
follow a power law. The fitting results obtained using the two
fitting methods are also given in the size distribution plots.
For Network A, NLR yielded a good power-law fit, whereas
MLE did not; for Network B, the converse was true.
The effect of the detection threshold and fitting method
was further evaluated by assessing the agreement among the
fitting results for the same recording under different evaluation
parameters. The fitting results for the two networks after per-
turbation on DIV 51 are given in Table I, with cases yielding a
good fit (indicating the network is in the critical state) reported
in bold. In cases where the number N of detected avalanches
was insufficient to produce a reliable fitting (N (cid:46) 1000), the
results are shown in italics. For Network A, the NLR fitting
results indicate the network is in the critical state, with reliable
results produced for thresholds of 6 and 7 standard deviations.
For Network B, only a threshold of 6 standard deviations
yielded enough activity for a reliable fitting, and the MLE
results indicate the network is in the critical state.
V. DISCUSSION
The two cortical networks observed in this study never self-
organized into the critical state; rather, they appeared to enter
into a subcritical state after an initial low-activity phase and
then settled into a stable highly synchronized supercritical
state. Pasquale et al. [16] have demonstrated that not all
dissociated cortical networks will reach the critical state during
their normal course of maturation. They also reported that
cultures tend to fall into a single preferred state (subcritical,
supercritical, or critical) by around DIV 21 and tend not
6
to diverge from that state. However, other previous studies
have demonstrated different courses of maturation for cortical
networks. For example, Yada et al. [18] have indicated that
cortical networks undergo early development from subcritical
to supercritical before finally reaching a critical state after
DIV 10. In contrast to this, Tetzlaff et al. [17] have described
activity traversing from supercritical to subcritical and finally
reaching criticality around DIV 58 (± 20 days). It is possible
that the present networks may have reached criticality if left to
mature further; however, it seems more likely that they would
have remained in the supercritical state, as would be predicted
by the type of trends observed by Pasquale et al. [16]. It will
be necessary to perform more rigorous fittings of exponential
curves to the size distribution to definitively demonstrate that
the cultures showed subcritical behavior. This remains a task
for future work.
An important achievement in this work was the successful
perturbation of the in vitro networks from supercritical to crit-
ical. The networks showed increasingly synchronized activity
accompanied by heightened activity levels as they matured
from DIV 23 onward. A balanced excitation-to-inhibition ratio
is widely considered to play an important role in enabling
the emergence of critical behaviors, and previous studies have
demonstrated the importance of this balance both through
modeling (e.g., [16, 22]) and experimentally (e.g., [14, 15]).
On the basis of this evidence, the networks were perturbed at
DIV 51, when they showed highly synchronized supercritical
behavior, by adding GABA;
the aim of this perturbation
was to artificially balance the excitation-to-inhibition ratio by
increasing inhibition, thereby pushing the networks into the
critical regime. The results indicate that the addition of GABA
(50 µM to Network A and 25 µM to Network B) successfully
broke the high level of synchronicity seen in the networks
and brought them into the critical state (see Fig. 3). As stated
in Section III, the GABA concentration was incrementally
raised for each network (10 and 50 µM for Network A; 5, 10,
and 25 µM for Network B). In both cases, the network only
reached criticality at the highest considered molarity; however,
in future work, smaller increments will be considered, as
GABA also has the undesirable effect of reducing activity.
The analysis of both the spontaneous and perturbed activity
also demonstrated that it is important to consider different
spike detection thresholds and fitting methods when conduct-
ing the analysis. The preliminary results reported here show
that the goodness of fit measure [21] was able to robustly
identify non-critical behavior even when the size distribution
showed apparent linearity in log -- log space and that lowering
the threshold in this case supported the identification of
supercritical activity (see Fig. 2f). When GABA was added,
there was insufficient activity at certain thresholds to achieve
a reliable fitting (see Table I), indicating the importance of an
appropriate threshold. Furthermore, although the results appear
to demonstrate critical behavior in Network B with a threshold
of 6 standard deviations, it is possible that this threshold was
too low and may have yielded false positives, as it is unlikely
that there was actual activity on the electrodes where neurons
had not adhered (see Fig. 1a). In the current experimental
setup, a threshold of 7 standard deviations seems to have been
most appropriate to balance these considerations. The results
also indicate that more work is needed to determine whether
NLR or MLE is best suited to yielding a good power law fit,
as the two fittings did not produce consistent values of α or
classification results.
VI. PLAN FOR FUTURE RESEARCH
This work represents a preliminary 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-
turbed networks to restore their dynamics to the unperturbed
state will then be explored.
7
VII. 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 results indicate
that the avalanche analysis method applied here is able to
successfully classify networks as critical or not critical and
that chemical perturbation is a feasible method of inducing
criticality when a network is in the supercritical state. In
future work, more networks will be analyzed with the hopes
of achieving self-organized criticality in some networks, and
further investigations of the effects of chemical perturbation
on both critical and supercritical networks will be conducted.
it can be determined if a
network is in a critical state, which gives an indication of
its suitability for use in computational tasks. Furthermore,
the results presented here indicate that it is possible to bring
cortical networks from the supercritical state to criticality by
increasing inhibition in the networks. This would be a useful
approach to manipulating networks to make them viable for
performing computational tasks. In addition to the computa-
tional 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.
With this type of analysis,
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] H. Broersma, J. F. Miller, and S. Nichele, "Computational
matter: Evolving computational functions in nanoscale
materials," in Advances in Unconventional Computing.
Springer, 2017, pp. 397 -- 428.
[3] Z. Konkoli, S. Nichele, M. Dale, and S. Stepney, "Reser-
voir computing with computational matter," in Compu-
tational Matter. Springer, 2018, pp. 269 -- 293.
[4] 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.
[5] 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.
[6] R. Doursat, H. Sayama, and O. Michel, "A review of mor-
phogenetic engineering," Natural Computing, vol. 12, pp.
517 -- 535, 2013.
[7] 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. p.
471-482 2007, 2007, pp. 471 -- 482.
[8] 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.
[9] (2018) SOCRATES: Self-organizing
computational
[Online]. Available: https://www.ntnu.edu/
substrates.
socrates
[10] W. L. Shew and D. Plenz, "The functional benefits of
criticality in the cortex," The Neuroscientist, vol. 19,
no. 1, pp. 88 -- 100, 2013.
[11] J. Hesse and T. Gross, "Self-organized criticality as a
fundamental propertie of neural systems," Frontiers in
Systems Neuroscience, vol. 8, 2014.
[12] 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.
[13] 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.
[14] 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.
[15] 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.
[16] 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.
[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] K. Heiney, V. Devold Valderhaug, I. Sand vig, A. Sand-
vig, G. Tufte, H. Lewi Hammer, and S. Nichele, "Eval-
uation of the criticality of in vitro neuronal networks:
Toward an assessment of computational capacity," arXiv
e-prints, p. arXiv:1907.02351, Jul 2019.
[20] K. Heiney, J. Mateus, C. D. F. Lopes, E. Neto,
M. Lamghari, and P. Aguiar, "µSpikeHunter: An
advanced computational tool for the analysis of neuronal
communication and action potential propagation in
microfluidic platforms," Scientific Reports, vol. 9,
no. 1, p. 5777, 2019.
[Online]. Available: https:
//doi.org/10.1038/s41598-019-42148-3
[21] 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.
[22] 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.
[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.
8
|
1803.03742 | 1 | 1803 | 2018-03-10T02:19:09 | K-shell decomposition reveals hierarchical cortical organization of the human brain | [
"q-bio.NC"
] | In recent years numerous attempts to understand the human brain were undertaken from a network point of view. A network framework takes into account the relationships between the different parts of the system and enables to examine how global and complex functions might emerge from network topology. Previous work revealed that the human brain features 'small world' characteristics and that cortical hubs tend to interconnect among themselves. However, in order to fully understand the topological structure of hubs one needs to go beyond the properties of a specific hub and examine the various structural layers of the network. To address this topic further, we applied an analysis known in statistical physics and network theory as k-shell decomposition analysis. The analysis was applied on a human cortical network, derived from MRI\DSI data of six participants. Such analysis enables us to portray a detailed account of cortical connectivity focusing on different neighborhoods of interconnected layers across the cortex. Our findings reveal that the human cortex is highly connected and efficient, and unlike the internet network contains no isolated nodes. The cortical network is comprised of a nucleus alongside shells of increasing connectivity that formed one connected giant component. All these components were further categorized into three hierarchies in accordance with their connectivity profile, with each hierarchy reflecting different functional roles. Such a model may explain an efficient flow of information from the lowest hierarchy to the highest one, with each step enabling increased data integration. At the top, the highest hierarchy (the nucleus) serves as a global interconnected collective and demonstrates high correlation with consciousness related regions, suggesting that the nucleus might serve as a platform for consciousness to emerge. | q-bio.NC | q-bio | K-shell decomposition reveals hierarchical cortical
organization of the human brain
Nir Lahav*1, Baruch Ksherim*1, Eti Ben-Simon2,3, Adi Maron-Katz2,3, Reuven
Cohen4, Shlomo Havlin1.
1. Dept. of Physics, Bar-Ilan University, Ramat Gan , Israel
2.Sackler Faculty of Medicine, Tel Aviv University, Tel-Aviv , Israel
3.Functional Brain Center, Wohl Institute for Advanced Imaging, Tel-Aviv Sourasky Medical Center,
Tel- Aviv, Israel
4.Dept. of Mathematics, Bar-Ilan University, Ramat Gan, Israel
*Authors contributed equally to this work
Corresponding author: Nir Lahav, Dept. of Physics, Bar-Ilan University,
Ramat Gan, Israel. E-mail: [email protected]
Abstract
In recent years numerous attempts to understand the human brain
were undertaken from a network point of view. A network framework takes
into account the relationships between the different parts of the system and
enables to examine how global and complex functions might emerge from
network topology. Previous work revealed that the human brain features
'small world' characteristics and that cortical hubs tend to interconnect
among themselves. However, in order to fully understand the topological
structure of hubs, and how their profile reflect the brain's global functional
organization, one needs to go beyond the properties of a specific hub and
examine the various structural layers that make up the network.
To address this topic further, we applied an analysis known in
statistical physics and network theory as k-shell decomposition analysis. The
analysis was applied on a human cortical network, derived from MRI\DSI
data of six participants. Such analysis enables us to portray a detailed account
of cortical connectivity focusing on different neighborhoods of inter-
connected layers across the cortex. Our findings reveal that the human cortex
is highly connected and efficient, and unlike the internet network contains no
isolated nodes. The cortical network is comprised of a nucleus alongside
shells of increasing connectivity that formed one connected giant component,
revealing the human brain's global functional organization. All these
components were further categorized into three hierarchies in accordance
with their connectivity profile, with each hierarchy reflecting different
functional roles. Such a model may explain an efficient flow of information
from the lowest hierarchy to the highest one, with each step enabling
increased data integration. At the top, the highest hierarchy (the nucleus)
serves as a global interconnected collective and demonstrates high correlation
with consciousness related regions, suggesting that the nucleus might serve as
a platform for consciousness to emerge.
Key words: Network theory, K-shell decomposition, cortical hubs, Graph
theory, functional hierarchies, consciousness
"..And you ask yourself, where is my mind?" The pixies (Where is my
mind)
Introduction
The human brain is one of the most complex systems in nature. In
recent years numerous attempts to understand such complex systems were
undertaken, in physics, from a network point of view (Carmi, 2007; Cohen
and Havlin, 2010; Newman, 2003; Colizza and Vespignani, 2007; Goh et al.,
2007). A network framework takes into account the relationships between the
different parts of the system and enables to examine how global and complex
functions might emerge from network topology. Previous work revealed that
the human brain features 'small world' characteristics (i.e. small average
distance and large clustering coefficient associated with a large number of
local structures (Achard et al., 2006; Bullmore and Sporns, 2009; He et al., 2007;
Ponten et al., 2007; Reijneveld et al., 2007; Sporns et al., 2004; Sporns and Zwi,
2004; Stam et al., 2007; Stam and Reijneveld, 2007; van den Heuvel et al.,
2008)), and that cortical hubs tend to interconnect and interact among
themselves (Achard et al., 2006; Buckner et al., 2009; Eguiluz et al., 2005; van
den Heuvel et al., 2008). For instance, van den Heuvel and Sporns
demonstrated that hubs tend to be more densely connected among
themselves than with nodes of lower degrees, creating a closed exclusive "rich
club" (Collin et al., 2014; Harriger et al., 2012; van den Heuvel and Sporns,
2011; van den Heuvel et al., 2013). These studies, however, mainly focused on
the individual degree (i.e. the number of edges that connect to a specific node)
of a given node, not taking into account how their neighbors' connectivity
profile might also influence their role or importance. In order to better
understand the topological structure of hubs, their relationship with other
nodes, and how their connectivity profile might reflect the brain's global
functional organization, one needs to go beyond the properties of a specific
hub and examine the various structural layers that make up the network.
In order to explore the relations between network topology and its
functional organization we applied a statistical physics analysis called k-shell
decomposition (Adler, 1991; Alvarez-Hamelin et al., 2005a; Alvarez-Hamelin
et al., 2005b; Carmi, 2007; Modha and Singh, 2010; Pittel et al., 1996; Garas et
al., 2010) on a human cortical network derived from MRI and DSI data. Unlike
regular degree analysis, k-shell decomposition does not only check a node's
degree but also considers the degree of the nodes connected to it. The k-shell
of a node reveals how central this node is in the network with respect to its
neighbors, meaning that a higher k-value signifies a more central node
belonging to a more connected neighborhood in the network. By removing
different degrees iteratively, the process enables to uncover the most
connected area of the network (i.e., the nucleus) as well as the connectivity
shells that surround it. Therefore, every shell defines a neighborhood of
nodes with similar connectivity (see Fig. 1). A few studies have already
applied this analysis in a preliminary way, focusing mainly on the network's
nucleus and its relevance to known functional networks (Hagmann et al.,
2008; van den Heuvel and Sporns, 2011). For instance, Hagmann et al.
revealed that the nucleus of the human cortical network is mostly comprised
of default mode network regions (Hagmann et al., 2008). However, when
examined more carefully, k-shell decomposition analysis, as shown here,
enables the creation of a topology model for the entire human cortex taking
into account the nucleus as well as the different connectivity shells ultimately
uncovering a reasonable picture of the global functional organization of the
cortical network. Furthermore, using previously published k-shell analysis of
internet network topology (Carmi, 2007) we were able to compare cortical
network topology with other types of networks.
We hypothesize that using k-shell decomposition would reveal that the
human cortical network exhibits a hierarchical structure reflected by shells of
higher connectivity, representing increasing levels of data processing and
integration all the way up to the nucleus. We further assume that different
groups of shells would reflect various cortical functions, with high order
functions associated with higher shells. In this way we aim to connect the
structural level with the functional level and to uncover how complex
behaviors might emerge from the network.
Figure 1: K-shell decomposition
process.
K-shell decomposition takes into
account the degree of the node as
well as the degree of the nodes
connected to it.
shows
example
This
the
difference of the K-shell method
compared with regular degree
count. Top panel: The whole
network. The yellow node is a
hub (k=5) and thus one might
think that it would be in the
nucleus. But on the first step of
the process (k=1), two of
its
neighbors will be removed to the
re-
first
(blue). When
computing the degree of
the
remaining nodes we notice there
are no more nodes with only one
link. The remaining network is
the 1st core. On step 2a (k=2),
another two of its neighbors will
be removed (black). Then, when
re-computing the degree of each
node (step 2b), the yellow node
has a low degree (k=1) and will
be removed to the second shell.
The process stops in k=3 when
the remaining nodes will be
removed and no node will remain
in
is
composed of
remaining
network in a given k step and the
nucleus is defined as the final k-
core in the process. The nucleus
of this network is thus the 2nd-
core,
last
remaining nodes (red). K-crust
includes the nodes that have been
removed until step k of the
process. This network has 5 nodes
in its 2nd-crust (blue, black and
yellow.
for more details see
methods).
the network. K-core
shell
the
the group of
the
Materials and methods
Imaging
The networks for our analysis were derived from two combined brain
imaging methods, MRI/DSI recorded by Patric Hagmann's group from
University of Lausanne (for all the functions and data sets, please refer to :
http://www.brain-connectivity-toolbox.net/). Using this data, clusters of
gray matter formed the nodes while fibers of white matter formed the edges
of the cortical network. In this technique, 998 cortical ROIs were used to
construct the nodes of each network and 14,865 edges were derived from
white matter fibers (for more specific details please see Hagmann P. et al.
(Hagmann et al., 2008)). Six structural human cortical networks were
transformed into six connection matrices by Patric Hagmann's group, derived
from five right handed subjects (first two networks were derived from the
same subject in different times). These connection matrices were utilized to
calculate the network's properties and to apply the k-shell decomposition
analysis. We used binary connection matrices ('1' –connected,
'0' –
disconnected) and not weighted connection matrices because of known
difficulties in determining the appropriate weights and how to normalize
them (Hagmann et al., 2007; Hagmann et al., 2003; Van Den Heuvel and Pol,
2010; van den Heuvel and Sporns, 2011). In order to connect between our
structural network and known functional networks the 998 nodes were
clustered into 66 known anatomical regions in accordance with Hagmann et
al. (Hagmann et al., 2008).
Network theory
Several network characteristics were used in our analysis (see supplementary
material 6 for further details):
Degree (k) of a node is the number of edges that connect to the node.
Hub is a node with degree above the average degree of the network.
Distance between nodes is the shortest path between node i and node j.
Average diameter (L) of the network is denoted by:
𝐿 =
1
𝑁(𝑁 − 1)
∑ 𝑑𝑖𝑗
𝑖≠𝑗
dij – distance between node i and node j ; N – total number of nodes in the
network
Local clustering coefficient (ci) of a node i reflects the probability that "my
friend's friend will also be my friend" (computed for each node). Clustering
coefficient is the average over all local ci and it provides estimation of the
amount of local structures in the network. Topologically it means that the
network will have a large quantity of triangles:
Small-world networks are networks that are significantly more clustered than
random networks, yet have approximately the same characteristic path length
as random networks (high clustering coefficient and low average distance).
Assortativity coefficient is the Pearson correlation coefficient of degree between
pairs of linked nodes. Positive values indicate a correlation between nodes of
similar degree, while negative values indicate relationships between nodes of
different degree. Assortativity coefficient lies between −1 and 1.
We also examined whether the cortical network exhibits a hierarchal
structure (not to be confused with the hierarchies derived from k-shell
decomposition analysis) in which hubs connect nodes which are otherwise
not directly connected. Networks with a hierarchal structure have a power
law clustering coefficient distribution- C~K-β which means that as the node
degree increases (k) the clustering coefficient (C) decreases. The presence of
hubs with low clustering coefficient means that the network has a hierarchal
structure (since hubs connect nodes which are not directly connected,
triangles with hubs are not frequent).
Module structures: the network's modular structure (community structure), is
revealed by subdividing the network into groups of nodes, with a maximally
iicNC1 possible number of within group links, and a minimally possible number of
between-group links.
K-shell decomposition method
In the k-shell decomposition method we revealed the network's
nucleus as well as the shells that surround it. The k-shell of a node indicates
the centrality of this node in the network with respect to its neighbors. The
method is an iterative process, starting from degree k=1 and in every step
raising the degree to remove nodes with lower or similar degree, until the
network's nucleus is revealed, along the following steps:
Step 1. Start with connectivity matrix M and degree k=1.
Step 2. Remove all nodes with degree ≤ k, resulting in a pruned connectivity
matrix M'.
Step 3. From the remaining set of nodes, compute the degree of each node. If
nodes have degree ≤ k, step 2 is repeated to obtain a new M'; otherwise, go
back to step 1 with degree k=k+1 and M=M'.
Stop when there are no more nodes in M' (M'=0).
The k-shell is composed of all the new removed nodes (along with their
edges) in a given k step. Accumulating the removed nodes of all previous
steps (i.e. all previous k-shells) is termed the k-crust. The k-core is composed
of the remaining network in a given k step and the nucleus is defined as the
final k-core in the process. In the end of every step a new k-shell, k-crust and
k-core are produced of the corresponding k degree. In the end of the process
the nucleus is revealed with the most central nodes of the network, and the
rest of the nodes are removed to the different shells (see Fig. 1). Typically, in
the process of revealing the nucleus, all removed nodes in the k-crust
eventually connect to each other forming one giant component.
The uniqueness of k-shell decomposition method is that it takes into account
both the degree of the node as well as the degree of the nodes connected to
that node. Thus, we can examine groups of nodes, every group with its own
unique connectivity pattern. In this way one can examine cortical anatomical
regions according to their connectivity neighborhood. For each node in the
network we determined its shell level (i.e. to which shell it belongs, or if it
survived the whole process, it belongs to the highest level – the nucleus). We
then calculated shell levels for every anatomical region, comprised of many
nodes, according to the weighted average shell level of its nodes.
Statistics and random networks
In order to evaluate the significant of the properties of the cortical network
each result was compared to that of a randomized network. The network was
randomized by keeping the degree distribution and sequence of the matrix
intact and only randomizing the edges between the nodes (Rubinov and
Sporns, 2010). For each cortical network several random networks were
computed with different amount of randomized edges (from 1% until 100% of
the edges). This process was repeated several times iteratively. K-shell
decomposition was applied for each of the randomized networks. Since the
results of the cortical network were resilient to small perturbations (1% of the
edges randomized) we raise the amount of randomization. For greater
amount of randomization the results were fixed around an average value
after 5 iterations (or more) using 100% random edges. Thus we took the
random networks to be with 100% randomized edges and 5 iterations.
To assess statistical significance of our results across networks, permutation
testing was used (Van Den Heuvel and Pol, 2010). Matrix Correlations across
6 networks were computed and compared with correlations obtained from
1000 random networks. These random network correlations yielded a null
distribution comprised of correlations between any two networks obtained
from the random topologies. Next, we tested whether the real correlations
significantly exceeded the random correlations, validated by a p-value< 0.01.
Moreover, the significance of the observed connectivity within and between
hierarchies was evaluated using a random permutation test. In this test, each
node was randomly assigned with a hierarchy, while preserving the
connectivity structure of the graph as well as hierarchy sizes. This process
was repeated 10,000 times, and in each repetition, the number of connections
within each hierarchy and between each pair of hierarchies was recorded. For
each pair of hierarchies, a connectivity p-value was calculated using the
fraction of the permutations in which the number of connections linking them
was equal or higher than this number in the real data. Resulting p-values
were corrected for multiple comparisons using the false discovery rate (FDR)
procedure thresholded at 0.05.
Results
Cortex network topology
The results of the K-shell decomposition process revealed that the
human cortex topology model has an "egg-like" shape (see Fig. 2). In the
"middle", 22% (± 12%) of the networks' nodes formed the nucleus ("the yolk"
in the egg analogy) and "surrounding" the nucleus about 77% (± 12%) of the
removed nodes formed the shells. These removed nodes did not reach the
nucleus and connected to each other to form one giant component. The nucleus
has on average 217 nodes (± 117) and the giant component has on average 770
nodes (± 121). The rest of the nodes are isolated nodes. These removed nodes
did not connect to the giant component, and essentially connect to the rest of
the network solely through the nucleus (some nodes are not connect to any
other node in the network and thus were removed; on average 9± 6 nodes per
cortical network).
Over all 6 networks, the average k-core of the nucleus was 19(±1),
which means that during the iterative process the nucleus was revealed after
the removal of 19(±1) shells. Thus, the minimum degree in the nucleus is 20
and the average degree of the nodes in the nucleus is 45 (±4). In comparison,
the average degree across the entire cortical network is 29 (±1), demonstrating
that the nucleus contains hubs with significantly higher degree than that of
the average network. In addition, the nucleus had considerably lower average
distance compared to the average distance of the entire cortical network (2
±0.2 vs. 3±0.1 , respectively). This finding means that it takes 2 steps, on
average, to get from one node to any other node in the 217 nodes of the
nucleus.
Fig. 2: Topology of the cortical network.
Topology of the cortical network (middle) compared with the internet topology, after
Carmi et al(Carmi, 2007) (left) and random cortex network (right). In the cortical
network the nucleus consists of 20% of the nodes while the remaining 80% compose
a one giant component from all the removed nodes in the different shells. Note, a
much bigger nucleus in the random cortical network and contrary to the cortical
network larger amount of isolated nodes in both random and internet topologies.
The giant component is formed in a process similar to a first order
phase transition with several critical points, as for the internet (Carmi, 2007;
Pittel et al., 1996). In the beginning of the process islands of removed nodes
were forming and growing, but at some stage all of these islands connect
together to form the giant component (see Fig. S1 for more details). This
abrupt phase transition occurred, on average, in k-crust 15 ±1 (i.e. big islands
of removed nodes were formed in crust 14, comprised of all previous shells
including shell 14, but in crust 15 all of these islands disappear and a single
giant component is formed). There is no significant difference between the
number of removed nodes that were added to crust 15 compared to crust 14,
yet a phase transition had occurred, suggesting that the difference is in the
amount of the removed hubs. In crust 15, for the first time, enough hubs
(which connect to lower degree nodes) were removed at once and connect all
the islands to form the giant component. Later, another critical point is
observed. On average in crust 18 (±1), a very large amount of nodes are
removed at once to join the giant component (on average 282 nodes
comprising 37% ±10% of all the nodes in the giant component (see Fig. S1).
This may suggest that the process reached yet another group of higher hubs
which have been removed along with their connections. These hubs connect
to significantly more nodes than the previous hubs leading to a massive
removal of nodes. We also note that the giant component features small world
characteristics similar to the entire network (C=0.4 for both giant component
and the whole network, average distance is 3.6 ±0.5 for the giant component,
slightly higher than that of the whole network (3±0.1), see Fig. S2).
Cortex network topology in comparison to other networks
The cortical network topology is found to be very different from the
topologies of a randomized cortex or the internet network (at the autonomous
systems level) which displayed a "medusa-like" shape (Carmi, 2007) (see Fig.
2). In addition to the nucleus and the giant component both random and
internet topologies have a large amount of isolated nodes, forming the
"medusa legs" in the medusa shape (on average 17% in the randomized
cortical networks and 25% in the internet network, unlike close to 0.3% ±0.3%
in the cortical network).
In addition, the average nucleus size of the randomized cortex is nearly
three times bigger than the average nucleus of the human cortex (56% vs.
20%). The cortical nucleus contains only 50% of the hubs, the rest fall on
average in the last 4-5 shells before the nucleus, while in the random cortex
100% of all hubs reached the nucleus (see Fig. S3). A network that displays a
significant amount of hubs on several levels and not just in the nucleus could
support a hierarchical structure that enables modular integration, as evident
in cortical function (Bassett et al., 2008; Christoff and Gabrieli, 2000; Gray et al.,
2002; Northoff and Bermpohl, 2004; Northoff et al., 2006). Note that in the
cortical network the hubs outside the nucleus start on average at shell 14-15
which supports the hypothesis that the first phase transition (shell 15±1 ) is
due to the removal of those hubs (as mentioned above).
Correlation between topology and known brain functions
In the k-shell decomposition analysis the connections of a node as well
as its neighborhood determine at which shell that node will be removed.
Neighborhood of High degree will be removed in a higher shell, or might
survive the entire process and be part of the nucleus. Therefore, the giant
component is comprised of different shells which represent different
neighborhood densities of connectivity. These shells, corresponding to known
cortical networks, enable an effective examination of cortical hierarchical
organization.
We, therefore, examined the functional attributes of the nodes found
in the nucleus and in all shells, by checking the shell level of every anatomical
region (mapping how many nodes from the anatomical region have been
removed to the different shells). Subsequently, we were able to score each
anatomical region in accordance with its place in the network's hierarchy
represented by its shell level. This characterization is demonstrated to be
more accurate than just analyzing the average degree of each anatomical
region (see Fig. S4 and supplementary material 1 for further details).
Furthermore, we examined the nucleus and revealed known functional
areas that are always found in the nucleus across all 6 networks (see Fig. 3).
These areas comprise the entire bilateral midline region and overlap with
five major functional networks: motor and motor planning, the default
network, executive control network, high order visual areas and the salience
network (see Table 1 for full details). In contrast, several known functional
areas were never in the nucleus across all 6 networks. These areas include
most of the right temporal lobe (e.g. the fusiform gyrus, A1, V5), right Broca
and Wernicke homologues and right inferior parietal cortex. Interestingly, all
the areas that never appear in the nucleus are from the right hemisphere.
Furthermore, 70% of all the lowest shells are from the right hemisphere while
60% of the areas that are always in the nucleus belong to the left hemisphere
(see supplementary material 2 for more details ).
Fig. 3: Anatomical regions and the network nucleus. Brain maps displaying
anatomical regions that are always in the nucleus (red) and never in the nucleus
(blue). Note that all the regions that never reach the nucleus are from the right
hemisphere.
Next, we used the critical points that were observed during the giant
component formation (see supplementary material 3 for more details) in
order to detect and establish different hierarchies of shells. Briefly, the
creation of the giant component corresponded to the shell threshold of a
middle hierarchy and the creation of the nucleus corresponded to the
threshold of a high hierarchy. This analysis resulted in three major hierarchal
groups (low, middle and high) as portrayed in Figure 4.
Fig. 4: Anatomical regions according to their hierarchies. Brain maps displaying
cortical anatomical regions according to their hierarchies. Red – low hierarchy, green
– middle hierarchy, blue – high hierarchy. One can divide the cortex to low hierarchy
regions found in the lateral bottom part of the cortex, middle hierarchy in the lateral
middle part of the cortex, and high hierarchy in the lateral top and midline part of
the cortex. RH - Right hemisphere, LH - Left hemisphere
The first hierarchal group consists of regions found in the lowest shells
(average shell level 8.8, number of nodes/edges: 99/730 respectively). The
removed nodes of this group are distributed across the shells with relatively
high standard deviation (4.42, e.g. fusiform gyrus, entorhinal cortex,
parahippocampal cortex. See Table 1 and Fig. 5 for full details). Notably, in
this hierarchal group 75% of the regions were bilateral and 50% of the regions
were never in the nucleus. The second hierarchy is a middle group which
includes nodes found in the highest shells, but still not in the nucleus (number
of nodes/edges: 335/4377 respectively). This group can be further subdivided
to two subgroups, distributed middle and localized middle according to their
average shell level and standard deviation. The average shell level of the
distributed middle group is 14.5 (±3.07). This subgroup includes regions like
right A1, right V5 and right Broca's homologue (for full details see table 1 and
Fig. 5d). The average shell level of the localized middle group is 16.67 (±1.13).
This subgroup includes regions like right wernicke homologue and right
middle frontal gyrus. In the middle hierarchy 56% of the regions are bilateral
and 40% of the regions are from the right hemisphere (in localized middle
88% right). 48% of the regions in this hierarchy were never found in the
nucleus (for full details see Table 1 and Fig. 5c).
The third group is the highest hierarchy which contains regions
predominantly found in the nucleus (number of nodes/edges: 561/8430
respectively). This group can also be subdivided to two subgroups, distributed
high and localized high according to their average shell level and standard
deviation. Average shell level of distributed high is 16.92 ( ± 2.82). This
subgroup includes the superior frontal gyrus, left Wernicke, left Broca and left
V5. The average shell level of the localized high group is 19.30 (±0.97) and
includes the precuneus and the cingulate cortex (for full details see Table 1
and Fig. 5). In this hierarchal group 69% of the regions were bilateral while
28% of the regions belonged to the left hemisphere. 44% of the regions in this
hierarchy were always in the nucleus (66% in localized high). Altogether, all
the regions that are always in the nucleus are from the high hierarchy while
the regions that never reached the nucleus are from lower hierarchies.
Fig. 5: Hierarchies of the cortical network. Top left panel: average shell level of the
hierarchies. X-axis: hierarchy, Y-axis: shell level. Top right panel: an example of a
single anatomical region representing each hierarchy (derived from average cortical
network over all 6 networks. For exact data see supplementary data 1 and Fig. S5).
Right Precuneus as an example of localized high hierarchy regions (blue). Notably,
this area always reached the nucleus. Right caudal middle frontal as an example of
localized middle hierarchy regions (green). Notably, this area never reached the
nucleus. Right fusiform gyrus as an example of low hierarchy regions (red). Note the
high standard deviation of the shell distribution. This region never reached the
nucleus. X-axis: k-shell number; Y-axis: number of nodes. Dashed line: nucleus.
Bottom left: Right lateral occipital cortex as an example of distributed middle
hierarchy regions (striped green, localized middle hierarchy as above). X-axis: k-shell
number; Y-axis: number of nodes. Dashed line: nucleus. Bottom right: Left precentral
gyrus as an example of distributed high hierarchy regions (striped blue, localized
high hierarchy as above). This area always reached the nucleus. X-axis: k-shell
number; Y-axis: number of nodes, dashed line: nucleus.
Using the shell score we could further estimate the average shell level
of known functional regions or networks (see Table S2). Interestingly, average
shell level often reflected known functional lateralization as detailed in Table
2. For instance, while Broca's area is found in the nucleus, its right
homologues never reached the nucleus. In a similar way, Wernicke's area is
found in the high hierarchy and its right homologue in the middle hierarchy,
again never reaching the nucleus. Right primary motor region and right TPJ
are found in the middle hierarchy (and also never reached the nucleus)
whereas their left counterparts are found in the high hierarchy (and left
primary motor region always reached the nucleus). The functional network
with the highest average shell level was the default mode network (DMN)
with a score of 18.1. 81% of its regions were found in the high hierarchy with
70% always reaching the nucleus. Following the DMN, the salience and the
sensorimotor networks also demonstrate high average shell level (17.3 and
17.5, respectfully) reflecting their high functional relevance. These results are
detailed in Table 2 and in supplementary material 2.
Connections between hierarchies
In order to examine the connections between the different hierarchies,
we compared the number of connections within each hierarchy to the number
of connections with other hierarchies (calculated as a percentage of its total
connections). Within the lowest hierarchy it was found that only 22% ±6.33%
were self-connections and the rest were distributed between the middle group
(30% ±3.36%) and the highest group (48% ±4.24%). In the middle hierarchy
approximately half of the connections (52% ±2.6%) were self-connections and
41.5% ±2.6% were linked to the highest group. Interestingly, only 7% ±0.77%
of the connections from the middle hierarchy were linked to the lowest
hierarchy. The highest hierarchy exhibited the highest levels of self-
connections (72% ±1.6%). Only 22.5% ±1.5% of its connections were linked to
the middle hierarchy and 6% ±0.6% to the lowest hierarchy (for more details
see table S1). These findings suggest a flow of information from the lowest to
the highest hierarchy with each step enabling greater local processing,
possibly supporting increased data integration.
We further tried to distinguish the differences between localized and
distributed hierarchies. Distributed hierarchies have high standard deviation
of the shell distribution and localized hierarchies have small standard
deviation of the shell distribution (see fig. 5). Notably, while most of the edges
of the localized hierarchies were mainly self-connections or connections to
their distributed partner in the same hierarchy (e.g. distributed to localized
middle), the distributed hierarchies displayed more connections to other
hierarchies (~15% in distributed subgroups compared to only ~8% in
localized subgroups) supporting their role in cross-hierarchy data integration.
Moreover, many of these connections were also across similar categories (e.g.
distribute middle with distribute high, app. 25%). Furthermore, the
distributed and localized subgroups within the same hierarchy displayed a
large amount of connections between themselves (~33% of their connections),
supporting the fact that they originate from the same hierarchy. The
significance of the observed connectivity within and between hierarchies was
evaluated using a random permutation test. The results showed that
connectivity within each hierarchy is significantly higher (FDR q<0.0005) and
that connectivity between all hierarchies was significantly lower (FDR
q<0.0005) than expected according the size of the hierarchies (see Figure 6).
Fig. 6: Connections between hierarchies. The size of the hierarchies represents total
amount of intra hierarchy connections. Connections within any hierarchy is found to
be significantly higher (arrows) and connections between hierarchies was
significantly smaller (dash arrows) than expected when taking into account the size
of the hierarchies, supporting the modularity nature of every hierarchy. Note the
increased self-connections as the hierarchies increase (percent connections are
normalized by the total amount of connections in each hierarchy).
Discussion
In the current study we applied the k-shell decomposition analysis to
reveal the global functional organization of the human cortical network.
Using this analysis we managed to build a model of cortex topology and
connect the structural with the functional level. Our findings indicate that the
human cortex is highly connected and efficient, compared to other networks,
comprised of a nucleus and a giant component with virtually no isolated
nodes. The giant component consists of different degree shells which
represent different neighborhoods of connectivity, revealing the global
properties of the cortical network. Together with the nucleus, these
connectivity shells were categorized into three hierarchies representing an
increasing number of regional connections, possibly supporting an increase in
data processing and integration within each hierarchy. In accordance, the
highest hierarchy was predominantly comprised of left and midline cortical
regions (including regions of the default mode network) known to be
associated with high-order functions (Northoff et al., 2006). Lastly, this
collective of interconnected regions, integrating information throughout the
cortex, might allow global properties such as consciousness to emerge.
Network properties
Our findings demonstrate, in accordance with previous work (Achard
et al., 2006; Cohen and Havlin, 2010; Ekman et al., 2012) that the cortical
network is resilient to small perturbations, highly organized, interconnected
and much more efficient compared with a random cortical network or the
internet network. K-shell decomposition analysis further proved to be more
accurate and provide better resolution of network properties compared to
standard methods (e.g. counting degrees, for full details see supplementary
material 1).
The two main components of the cortical network, the nucleus and the
giant component, both have small world properties though they might serve
different roles. A higher clustering coefficient of the giant component
alongside short average distance of the nucleus suggest that the majority of
local processing takes place within the giant component while the nucleus
mainly adds shortcuts and global structures to the network. Indeed, although
the nucleus is highly connected, it includes only 50% of all hubs unlike the
random nucleus which includes all network hubs (see Fig. S2, S3). These
'peripheral' hubs were located in the giant component and, as previously
suggested (Achard et al., 2006), might enable efficient data integration and
local information processing. Hubs outside the nucleus might therefore, serve
as local processors integrating information from lower shells and transfer it
forward to a higher hierarchy, eventually reaching the nucleus (for more
information see supplementary material 4).
Network hierarchies and data integration
K-shell decomposition analysis reveals that the creation of the giant
component entails several critical points. From these critical points we could
characterize three major neighborhoods of connectivity or three hierarchies
(for more details see supplementary material 3). The regions in the lowest
hierarchy appeared to be mostly involved in localized sensory perception
(e.g. the fusiform face area and visual "what" stream (Goodale and Milner,
1992)). The different nodes within this hierarchy broadly distributed along the
shells which might enable efficient data transfer and processing before
sending it to higher hierarchies.
The middle hierarchy is found to be composed of high shells with high
degree nodes, though half of them never reached the nucleus, a property that
separates these regions from the high hierarchy. Functional regions found in
this hierarchy appeared to be involved in high cognitive functions and data
integration. For instance, most of the auditory network and regions involved
in the integration of audio and visual perception were found in the middle
hierarchy. In addition 40% of the executive control network (including right
dorsolateral PFC, a crucial region in executive control and working memory
(Raz and Buhle, 2006)) and the dorsal visual stream (where\what stream
(Goodale and Milner, 1992)) are found in this hierarchy. Broca's area was also
located in the middle hierarchy as well as other homologue regions related to
language such as Broca and Wernicke homologues.
The high hierarchy contained regions predominantly found in the
nucleus. All regions that reached the nucleus across all cortical networks are
found in this hierarchy. Unlike other hierarchies, this unique hierarchy is a
single, highly interconnected component, which enables high levels of data
integration and processing, probably involved in the highest cognitive
functions. In accordance, the high hierarchy exhibited the highest amount of
self-connections across hierarchies suggesting that it processes data mostly
within itself (see Fig. 6). The nucleus (represented by the high hierarchy) has
a very strong overlap with the default mode network (81%), in accordance
with the result of Hagmann et al (Hagmann et al., 2008), and also with the
visual cortex (75%), sensorimotor network (75%) and salience network (71%).
The visual dorsal stream and the executive control network also display 60%
overlap with the nucleus. Interestingly, all the regions that never appear in
the nucleus (across all 6 networks) belong to the right hemisphere, while a
strong tendency towards the left hemisphere appeared when examining the
nucleus. As mentioned above, all the regions that reached the nucleus are
mostly midline or left hemisphere regions. Roughly speaking, the left
hemisphere is comprised of high hierarchy regions and the right hemisphere
is comprised of middle hierarchy regions (see Fig. 4 and supplementary
material 3 and 2).
Looking across hierarchies it's evident that the lowest hierarchy has the
smallest amount of connections to other hierarchies and within itself; the
middle hierarchy has more connections, almost equally distributed between
itself and others; and the high hierarchy has the largest amount of
connections, most of them within itself (see Fig. 6). Interestingly, self-
connections within each hierarchy were significantly higher and between
hierarchies significantly smaller, than expected in random control according
the size of the hierarchies. This finding suggests that every hierarchy can be
seen as a different module mostly involved in self-processing and only then
transfers information to other hierarchies (Bullmore and Sporns, 2009;
Hagmann et al., 2008; van den Heuvel and Sporns, 2011). Regarding cross
hierarchy connections, it is important to note that most of the connections
between middle and high hierarchies occur in their distributed subgroups.
This finding suggests that in every hierarchy distributed regions are more
involved in data transfer and integration across hierarchies, while localized
regions deal more with data processing.
Assuming that data integration requires cross hierarchy connections
(the amount of data that a hierarchy receives from other hierarchies – the
centrality of the hierarchy (Rubinov and Sporns, 2010)) and data processing
depend on interconnected regions (the amount of calculations taking place
inside the hierarchy – specialized processing within densely interconnected
module (Rubinov and Sporns, 2010)), then data integration and processing
seem to increase as we step up in the hierarchies. These findings could
therefore suggest a flow of information from the lowest to the highest
hierarchy with every hierarchy integrating more data and executing further
processing, in line with previous studies and theoretical work (Christoff and
Gabrieli, 2000; Gray et al., 2002; Northoff et al., 2006; Damasio, 2000). The low
hierarchy receives information, performs specific calculations with its small
amount of intra connections and passes the information to the higher
hierarchies. The middle hierarchy is further able to integrate more data and
locally process more information. At the top, the nucleus receives the most
information from all other hierarchies and executes further processing using
its dense interconnections, suggesting its vital involvement in data integration
within the cortical network.
The Nucleus as a platform for consciousness
The regions in the nucleus form one component and constitute the
most connected neighborhoods in the cortical network with the highest
degrees. In contrast to the giant component, which mostly exhibits local
structures (i.e. high clustering coefficient), all the regions in the nucleus form
global structures (see supplementary material 4) and densely connect within
themselves creating a unique interconnected collective module all over the
brain. The regions and profile of this collective module are consistent with
previous work (Collin et al., 2014; Hagmann et al., 2008; van den Heuvel and
Sporns, 2011), mostly comprised of posterior medial and parietal regions.
Furthermore, in Hagmann et al's structural cortical core, 70% of the core's
edges were self-connections, similar to our findings within the high hierarchy
(72%). In addition, this structural core forms one module and connected with
connector hubs to all other modules in the network, reflecting our results that
the nucleus is a single interconnected module with increased global
structures. These findings further suggest that the distributed high hierarchy
is composed of such connector hubs, in charge of connecting other hierarchies
with the nucleus.
A strong inter-connected nucleus has also been demonstrated by
Sporns et al suggesting a rich club organization of the human connectome
(Collin et al., 2014; van den Heuvel and Sporns, 2011; van den Heuvel et al.,
2013). Their results revealed a group of "12 strongly interconnected
bihemispheric hub regions, comprising, in the cortex, the precuneus, superior
frontal and superior parietal cortex". These six cortical regions were part of
our more detailed interconnected nucleus which further includes more
regions of the high hierarchy (see Table 1). This interconnected collective
module creates one global structure, involving regions from all over the
cortex, which may create one global function. Given recent theories that
explain consciousness as a complex process of global data integration
(Balduzzi and Tononi, 2008; Damasio, 2000; Dehaene and Naccache, 2001;
Tononi and Edelman, 1998; Godwin et al., 2015), in particular Global Work
space Theory and integrated information theory (Balduzzi and Tononi, 2008;
Dehaene and Naccache, 2001; Tononi and Edelman, 1998), one can postulate
that such global function could be related to conscious abilities. We therefore
suggest that the global interconnected collective module of the nucleus can
serve as a platform for consciousness to emerge. Indeed, all of the regions in
the nucleus have been previously correlated to consciousness activities
(Achard et al., 2006; Godwin et al., 2015; Goodale and Milner, 1992; Gray et al.,
2002; Northoff and Bermpohl, 2004; Northoff et al., 2006; Christoff et al., 2009),
especially midline and fronto-parietal regions. The nucleus, receiving the
most information from all other hierarchies and integrating it to a unified
global function, is therefore a perfect candidate to be the high integrative,
global work space region in which consciousness can emerge (for more
information see supplementary material 5).
Study limitations
Some limitation issues have to be taken into account when interpreting
the current results. First, our network is limited only to the cortex; future
studies should examine the entire brain network and its influence on the
profile of the hierarchies or nucleus. It is possible, for instance, that regions
within the low hierarchy (e.g. the fusiform gyrus) might belong to higher
hierarchies and are affected by lack of subcortical regions (such as the
hippocampus). Lastly, the structural connections of our network were
mapped with DSI followed by computational tractography (Hagmann et al.,
2008; Hagmann et al., 2007; Hagmann et al., 2003; Schmahmann et al., 2007).
Although DSI has been shown to be especially sensitive with regard to
detecting fiber crossings (Hagmann et al., 2008; Hagmann et al., 2007;
Hagmann et al., 2003; Schmahmann et al., 2007), it must be noted that this
method may be influenced by errors in fiber reconstruction, and systematic
detection biases.
Conclusions
The current study used k-shell decomposition analysis in order to
reveal the global functional organization of the human cortical network.
Consequently, we built a model of human cortex topology and revealed the
hierarchical structure of the cortical network. In addition, this analysis proved
to be more accurate than standard methods in the characterization of cortical
regions and hierarchies. Our findings indicate that the human cortex is highly
connected and efficient, compared to other networks, comprised of a nucleus
and a giant component with virtually no isolated nodes. The giant component
consists of different connectivity shells, which we categorized into three
hierarchies representing an increasing number of regional connections. Such a
topological model could support an efficient flow of information from the
lowest hierarchy to the highest one, with each step enabling more data
integration and data processing. At the top, the highest hierarchy (the global
interconnected collective module) receives information from all previous
hierarchies, integrates it into one global function and thus might serve as a
platform for consciousness to emerge.
Acknowledgments
We would like to thank Mr. Kobi Flax for his crucial role in data analysis,
without him this work would not be accomplished! We would also like to
thank Dr. Itay Hurvitz for his invaluable comments. We thank the European
MULTIPLEX (EU-FET project 317532) project, the Israel Science Foundation,
ONR and DTRA for financial support.
Tables:
Table 1: Cortical anatomical regions according to hierarchies
Anatomical Region
Localized High
Side
Function
Paracentral lobule
Mid
SMA - sensorimotor network (Always)
Caudal anterior cingulate cortex
Caudal anterior cingulate cortex
Inferior parietal cortex
L
R
L
Salience network (Always)
Salience\Executive control network (Always)
DMN, Sensorimotor network, Visual dorsal stream
(Always)
Posterior cingulate cortex
Mid
DMN (Always)
Rostral anterior cingulate cortex
Mid
Salience\Executive control network, DMN (Always)
Precuneus
Mid
Isthmus of the cingulate cortex
Pericalcarine cortex
Postcentral gyrus
Superior parietal cortex
Supramarginal gyrus
Bank of the Superior temporal
sulcus
Cuneus
Distributed high
Superior frontal cortex
Precentral gyrus
R
R
L
L
L
L
R
L
L
DMN (Always)
DMN (Always)
Primary visual area
Primary somatosensory cortex - Sensorimotor
network
Executive control, Sensory integration, Sensorimotor
network, Visual dorsal stream
Wernicke area, TPJ
Visual dorsal stream
Visual
DMN\ Executive\ Salience, Sensorimotor network
(Always)
Primary motor cortex - sensorimotor network
(Always)
Superior temporal cortex
Pericalcarine cortex
Pars orbitalis
Middle temporal cortex
Lateral occipital cortex
Isthmus of the cingulate cortex
Cuneus
Rostral middle frontal cortex
Superior parietal cortex
Superior frontal cortex
Postcentral gyrus
Lingual gyrus
Localized middle
Inferior parietal cortex
Caudal middle frontal cortex
Bank of the superior temporal
sulcus
Supramarginal gyrus
Superior temporal cortex
Frontal pole
Frontal pole
Medial orbitofrontal cortex
Distributed middle
Pars triangularis
Pars triangularis
L
L
L
L
L
L
L
L
R
R
R
R
R
R
R
R
R
R
L
R
R
L
Wernicke ,TPJ, Visual dorsal stream
Primary visual
Executive control network
V5 (Visual dorsal stream), DMN
Primary visual, Visual ventral stream
DMN
Visual
Executive control network, DMN
Executive, sensory integration, Sensorimotor
network, Visual dorsal stream
DMN\ Executive\ Salience\ Sensorimotor network
Primary somatosensory cortex - Sensorimotor
network
Visual
DMN, Sensorimotor network, Visual dorsal stream
(Never)
Executive control network, Sensorimotor network
(Never)
Visual dorsal stream (Never)
Wernicke homologue, TPJ (Never)
Wernicke homologue, TPJ, Visual dorsal stream
(Never)
Executive control network
Salience and executive control networks
Stimulus-reward associations
Broca homologue (Never)
Broca
Middle temporal cortex
Pars opercularis
Pars opercularis
Inferior temporal cortex
Inferior temporal cortex
Rostral middle frontal cortex
Pars orbitalis
Transverse temporal cortex
Temporal pole
R
R
L
R
L
R
R
R
L
V5 (Visual dorsal stream), DMN (Never)
Broca homologue (Never)
Broca
Visual association, Visual ventral stream (Never)
Visual association, Visual ventral stream
Salience and executive control networks (Never)
Salience and executive control networks (Never)
Primary auditory cortex (Never)
Salience network
Lateral orbitofrontal cortex
L+R
Stimulus-reward associations
Medial orbitofrontal cortex
Precentral gyrus
Caudal middle frontal cortex
Lateral occipital cortex
Low
Temporal pole
Parahippocampal cortex
Parahippocampal cortex
Fusiform gyrus
Fusiform gyrus
Entorhinal cortex
Entorhinal cortex
Lingual gyrus
L
R
L
R
R
R
L
R
L
R
L
L
Stimulus-reward associations
Primary motor cortex - Sensorimotor network
Executive control network, DMN, Sensorimotor
network
Primary visual, Visual ventral stream
Salience network (Never)
Hippocampal support, Visual ventral stream (Never)
Hippocampal support, Visual ventral stream
Face recognition, Visual ventral stream (Never)
Face recognition, Visual ventral stream
Hippocampal support, Visual ventral stream (Never)
Hippocampal support, Visual ventral stream
Visual association
DMN= Default mode network, TPJ=temporal parietal junction. Always= region that always reaches
the nucleus for all networks, Never= region that never reaches the nucleus for all networks.
Table 2: Laterality effects
Anatomical region
Left
Precentral gyrus (primary motor cortex)
High (always)
Right
Middle
Inferior parietal
High (always)
Middle (never)
Supra marginal gyrus (Wernicke area,TPJ)
Superior temporal (Wernicke area ,TPJ)
Lateral occipital cortex (primary visual)
Lingual gyrus (visual association)
Bank of the superior temporal sulcus (vision)
Pars Orbitalis (executive control network)
Middle temporal (V5, DMN)
Rostral middle frontal cortex (executive control
network, DMN)
High
High
High
Low
High
High
High
High
Middle (never)
Middle (never)
Middle
High
Middle (never)
Middle (never)
Middle (never)
Middle (never)
Superior frontal cortex
High (always)
High
Caudal middle frontal cortex (executive control
network, DMN)
Inferior temporal cortex (visual association)
Pars triangularis (Broca homologue)
Pars opercularis (Broca homologue)
Temporal pole (salience network)
Parahippocampal cortex
Fusiform gyrus
Entorhinal cortex
Functional Networks
Middle
Middle
Middle
Middle
Middle
Low
Low
Low
Middle (never)
Middle (never)
Middle (never)
Middle (never)
Low (never)
Low (never)
Low (never)
Low (never)
Dorsal stream (where stream)
100% high
80% middle (80% never)
Ventral stream (what stream)
60% low
60% low (80% never)
Auditory network
100% high
100% middle (100% never)
Executive control network
77% high
55% middle
Default mode network
Salience network
Sensorimotor network
89% high (55% always)
71% high (57% always)
60% (always)
40% (always)
83% (always)
17% (always)
DMN= Default mode network, TPJ=temporal parietal junction. Always= region that always reaches
the nucleus for all networks, Never= region that never reaches the nucleus for all networks.
References:
Achard S, Salvador R, Whitcher B, Suckling J and Bullmore E 2006 A resilient, low-frequency,
small-world human brain functional network with highly connected association
cortical hubs The Journal of neuroscience 26 63-72
Adler J 1991 Bootstrap percolation Physica A: Statistical Mechanics and its Applications 171
453-70
Alvarez-Hamelin J I, Dall'Asta L, Barrat A and Vespignani A 2005a k-core decomposition of
Internet graphs: hierarchies, self-similarity and measurement biases arXiv preprint
cs/0511007
Alvarez-Hamelin J I, Dall'Asta L, Barrat A and Vespignani A 2005b Large scale networks
fingerprinting and visualization using the k-core decomposition. In: Advances in
neural information processing systems, pp 41-50
Balduzzi D and Tononi G 2008 Integrated information in discrete dynamical systems:
motivation and theoretical framework PLoS computational biology 4 e1000091
Bassett D S, Bullmore E, Verchinski B A, Mattay V S, Weinberger D R and Meyer-Lindenberg A
2008 Hierarchical organization of human cortical networks in health and
schizophrenia The Journal of Neuroscience 28 9239-48
Buckner R L, Sepulcre J, Talukdar T, Krienen F M, Liu H, Hedden T, Andrews-Hanna J R,
Sperling R A and Johnson K A 2009 Cortical hubs revealed by intrinsic functional
connectivity: mapping, assessment of stability, and relation to Alzheimer's disease
The Journal of Neuroscience 29 1860-73
Bullmore E and Sporns O 2009 Complex brain networks: graph theoretical analysis of
structural and functional systems Nature Reviews Neuroscience 10 186-98
Carmi S 2007 A model of internet topology using K-shell decomposition PNAS 104 11150-4
Christoff K and Gabrieli J D 2000 The frontopolar cortex and human cognition: Evidence for a
rostrocaudal hierarchical organization within the human prefrontal cortex
Psychobiology 28 168-86
Christoff K, Gordon A M, Smallwood J, Smith R and Schooler J W 2009 Experience sampling
during fMRI reveals default network and executive system contributions to mind
wandering Proceedings of the National Academy of Sciences 106 8719-24
Cohen R and Havlin S 2010 Complex networks: structure, robustness and function:
Cambridge University Press)
Cohen R e a 2003 Scale free networks are ultra small Phys Rev Lett 90
Colizza V and Vespignani A 2007 Invasion threshold in heterogeneous metapopulation
networks Physical Review Letters 99 148701
Collin G, Sporns O, Mandl R C and van den Heuvel M P 2014 Structural and functional
aspects relating to cost and benefit of rich club organization in the human cerebral
cortex Cerebral Cortex 24 2258-67
Damasio A R 2000 The feeling of what happens: Body and emotion in the making of
consciousness: Random House)
Dehaene S and Naccache L 2001 Towards a cognitive neuroscience of consciousness: basic
evidence and a workspace framework Cognition 79 1-37
Eguiluz V M, Chialvo D R, Cecchi G A, Baliki M and Apkarian A V 2005 Scale-free brain
functional networks Phys Rev Lett 94 018102
Ekman M, Derrfuss J, Tittgemeyer M and Fiebach C J 2012 Predicting errors from
reconfiguration patterns in human brain networks Proceedings of the National
Academy of Sciences 109 16714-9
Garas A, Argyrakis P, Rozenblat C, Tomassini M and Havlin S 2010 Worldwide spreading of
economic crisis New journal of Physics 12 113043
Godwin D, Barry R L and Marois R 2015 Breakdown of the brain's functional network
modularity with awareness Proceedings of the National Academy of Sciences
201414466
Goh K-I, Cusick M E, Valle D, Childs B, Vidal M and Barabási A-L 2007 The human disease
network Proceedings of the National Academy of Sciences 104 8685-90
Goodale M A and Milner A D 1992 Separate visual pathways for perception and action
Trends in neurosciences 15 20-5
Gray J R, Braver T S and Raichle M E 2002 Integration of emotion and cognition in the lateral
prefrontal cortex Proceedings of the National Academy of Sciences 99 4115-20
Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey C J, Wedeen V J and Sporns O 2008
Mapping the structural core of human cerebral cortex PLoS Biol 6 e159
Hagmann P, Kurant M, Gigandet X, Thiran P, Wedeen V J, Meuli R and Thiran J P 2007
Mapping human whole-brain structural networks with diffusion MRI PLoS ONE 2
e597
Hagmann P, Thiran J P, Jonasson L, Vandergheynst P, Clarke S, Maeder P and Meuli R 2003
DTI mapping of human brain connectivity: statistical fibre tracking and virtual
dissection Neuroimage 19 545-54
Harriger L, Van Den Heuvel M P and Sporns O 2012 Rich club organization of macaque
cerebral cortex and its role in network communication PloS one 7 e46497
He Y, Chen Z J and Evans A C 2007 Small-world anatomical networks in the human brain
revealed by cortical thickness from MRI Cerebral cortex 17 2407-19
Modha D S and Singh R 2010 Network architecture of the long-distance pathways in the
macaque brain Proceedings of the National Academy of Sciences 107 13485-90
Newman M E J 2003 The structure and function of complex networks SIAM review 45 167-
256
Northoff G and Bermpohl F 2004 Cortical midline structures and the self Trends in cognitive
sciences 8 102-7
Northoff G, Heinzel A, de Greck M, Bermpohl F, Dobrowolny H and Panksepp J 2006 Self-
referential processing in our brain-a meta-analysis of imaging studies on the self
Neuroimage 31 440-57
Pittel B, Spencer J and Wormald N 1996 Sudden emergence of a giantk-core in a random
graph Journal of Combinatorial Theory, Series B 67 111-51
Ponten S C, Bartolomei F and Stam C J 2007 Small-world networks and epilepsy: graph
theoretical analysis of intracerebrally recorded mesial temporal lobe seizures Clin
Neurophysiol 118 918-27
Raz A and Buhle J 2006 Typologies of attentional networks Nature Reviews Neuroscience 7
367-79
Reijneveld J C, Ponten S C, Berendse H W and Stam C J 2007 The application of graph
theoretical analysis to complex networks in the brain Clin Neurophysiol 118 2317-31
Rubinov M and Sporns O 2010 Complex network measures of brain connectivity: uses and
interpretations Neuroimage 52 1059-69
Schmahmann J D, Pandya D N, Wang R, Dai G, D'Arceuil H E, de Crespigny A J and Wedeen V
J 2007 Association fibre pathways of the brain: parallel observations from diffusion
spectrum imaging and autoradiography Brain 130 630-53
Sporns O, Chialvo D R, Kaiser M and Hilgetag C C 2004 Organization, development and
function of complex brain networks Trends Cogn Sci 8 418-25
Sporns O and Zwi J D 2004 The small world of the cerebral cortex Neuroinformatics 2 145-62
Stam C J, Jones B F, Nolte G, Breakspear M and Scheltens P 2007 Small-world networks and
functional connectivity in Alzheimer's disease Cereb Cortex 17 92-9
Stam C J and Reijneveld J C 2007 Graph theoretical analysis of complex networks in the brain
Nonlinear Biomed Phys 1 3
Tononi G and Edelman G M 1998 Consciousness and complexity science 282 1846-51
Van Den Heuvel M P and Pol H E H 2010 Exploring the brain network: a review on resting-
state fMRI functional connectivity European Neuropsychopharmacology 20 519-34
van den Heuvel M P and Sporns O 2011 Rich-club organization of the human connectome
The Journal of neuroscience 31 15775-86
van den Heuvel M P, Sporns O, Collin G, Scheewe T, Mandl R C, Cahn W, Goñi J, Pol H E H
and Kahn R S 2013 Abnormal rich club organization and functional brain dynamics in
schizophrenia JAMA psychiatry 70 783-92
van den Heuvel M P, Stam C J, Boersma M and Pol H H 2008 Small-world and scale-free
organization of voxel-based resting-state functional connectivity in the human brain
Neuroimage 43 528-39
|
1812.00598 | 3 | 1812 | 2019-08-21T04:07:18 | A Static Distributed-parameter Circuit Model Explains Electrical Stimulation on the Neuromuscular System | [
"q-bio.NC"
] | Finite Element Modeling (FEM) has been widely used to model the electric field distribution, to study the interaction between stimulation electrodes and neural tissue. However, due to the insufficient computational capability to represent neural tissue down to an atom-level, the existing FEM fails to model the real electric field that is perpendicular to neuron membrane to initiate an action potential. Thus, to reveal the real electrode-tissue interactions, we developed a circuit to model transmembrane voltage waveforms. Here, we show a distributed-parameter circuit model to systematically study how electrode-tissue interaction is affected by electrode position, input current waveform, and biological structures in the neuromuscular system. Our model explains and predicts various phenomena in neuromuscular stimulation, guides new stimulation electrode and method design, and more importantly, facilitates a fundamental understanding of the physical process during electrode-tissue interaction. In our model, myelin is assumed to be inductive. The voltage waveform resonance caused by this inductive myelin accounts for the much lower stimulation threshold to activate motoneurons than muscle fibers, which is observed with in vivo measurements. These findings confirmed the feasibility of studying electrode-tissue interaction using a proper distributed-parameter circuit. Our current application on the neuromuscular system also raises the possibility that this distributed-parameter circuit model could potentially be applied to study other neural tissues, including the Peripheral Nervous System (PNS) and the Central Nervous System (CNS). | q-bio.NC | q-bio | 117576, Singapore
117456, Singapore
Singapore
Abstract
The existing Finite Element Modeling (FEM) fails to model the real electric field that activates a
neuron, because the computation capability is insufficient to represent the neural tissue down to an
atom-level. Thus, to reveal the real electrode-tissue interactions, we adopt the idea of using a circuit to
simulate voltage waveforms, as proposed in the Circuit-Probability theory (C-P theory). Here we show a
distributed-parameter circuit model to systematically study how the interaction between electrode and
tissue is affected by the electrode position, input current waveform, and the biological structures in the
muscle. Our model explains and predicts various phenomena in muscle stimulation, guides new
importantly, facilitates a fundamental
stimulation electrode and method design, and more
understanding of the physical process during electrode-tissue
interaction. With some proper
modifications, other neural tissues, including the Peripheral Nervous System (PNS) and the Central
Nervous System (CNS), can also be studied with this distributed-parameter circuit model.
A Static Distributed-parameter Circuit Model to Study Electrical
Jiahui Wang1,2,3, Hao Wang1,2,3*, Xin Yuan Thow2, Nitish V. Thakor1,2, Shih-Cheng
Stimulation on Muscle Tissue
Yen1,2, Chengkuo Lee1,2,3*
1 Department of Electrical & Computer Engineering, National University of Singapore, 4 Engineering Drive 3,
2 Singapore Institute for Neurotechnology (SINAPSE), National University of Singapore, 28 Medical Drive, #05-COR,
3 Center for Intelligent Sensors and MEMS, National University of Singapore, 4 Engineering Drive 3, 117576,
Electroceuticals, where electrical stimulation is delivered to the nervous system and muscles, are
becoming widespread therapeutic solutions to people with neurological disabilities. For example, people
with spinal cord injury (SCI) above the sixth cervical vertebra are unable to control extant limbs due to
interruption to the motor pathway[1]. Functional electrical stimulation (FES) could benefit these people
by restoring functional actions, like voluntary grasp, via the electrical stimulation of specific muscles.
However, despite its application and medical promise, the electrophysiology of electrical stimulation is
neither precise nor well understood. The history of electrophysiology dates back to 1770, when Luigi
Galvani first discovered bioelectricity as he made a frog muscle twitch by accidentally creating a battery
from surgical instruments[2]. The mechanism of electrical stimulation remained elusive until 1952, when
Alan Hodgkin and Andrew Huxley proposed a quantitative description of membrane current on the
unmyelinated squid giant axon[3]. In 1976, Donald McNeal first applied the established nerve axon
models to explain excitation of nerve tissue, by bringing in the concept of shared electric field between
the stimulation electrodes and the excitable nerve tissue[4]. With increased computing power, modern
computational models further improved McNeal's method, by accounting for additional parameters,
such as the anisotropic extracellular conductivity in electric field simulation and non-linear response of
neuronal cells and axons. Such modern computational models include the field-neuron model, which
has been applied to analyze electrical stimulation on the peripheral nervous system[5][6][7] and the central
nervous system[8][9].
Despite the development of these computational models, there are large discrepancies between
these computational models and experimental observations. Firstly, the tissue should not be considered
as purely resistive[5-9], as recent indirect[10][11][12] and direct[13] evidence suggests that the extracellular
medium consists of non-resistive components. Second, the deterministic gating property assumed in the
computational models[5-9] contradicts the probabilistic gating property as observed in single ion channel
recording[14-18]. In addition, there lacks a complete model that can systematically study how the
interaction between the electrode and tissue is affected by the electrode position, input current
waveforms, and the biological structures of the neural tissue. In this paper, we aim to achieve such a
complete model that explains and predicts all various phenomena in muscle stimulation.
In the circuit-probability theory (C-P theory) proposed in our previous submission[19], the composite
biological structures in the tissue are neglected for circuit simplification. In this study, by considering
motoneurons, muscle fibers and extracellular medium, the lumped-parameter circuit adopted in the C-P
theory is now expanded to a distributed-parameter circuit, enabling a more detailed investigation of the
interaction between stimulating electrode sites and excitable tissues. This muscle model can
qualitatively explain many phenomena:
1. With different locations of the stimulation electrode sites relative to the target motoneuron
(motoneuron-electrode position), the stimulation efficiency can either increase or decrease with respect
to the spacing between the two electrode sites.
2. The motoneuron-electrode position determines the polarity of the cross-membrane voltage
waveform, which is an unsolved issue using the lumped-parameter circuit in the C-P theory.
3. The force mapping curves measured with increasing current amplitude may form a certain shape
with multiple curvatures, which has been observed in our previous experiments as well. A conventional
explanation for this phenomenon is that multiple groups of motoneurons are sequentially recruited
when the current amplitude is increased[20][21]. But in our model, this phenomenon can be theoretically
derived and numerically calculated.
4. Instead of forming a force mapping curve (a measured curve of the generated force with respect
to input current amplitude or pulse width) with multiple curvatures, we observe that the recruitment of
multiple motoneuron groups will induce an unstable output force at the transition current, showing an
abnormally high error bar in the force mapping curve. This phenomenon is reflected in the
measurement data. It helps us to tell when an additional group of motoneurons is recruited during the
experiment.
Based on a comprehensive understanding of muscle stimulation, we propose an effective design of
the stimulation electrode and a corresponding stimulation method to reduce muscle fatigue and control
stimulation efficiency. Firstly, a double-side multiple-channel polyimide electrode design is proposed to
reduce stimulation fatigue by alternatively activating the motoneurons on the frontside and backside of
the electrode. Secondly, a comprehensive calibration of the force mapping curve using different
combinations of the electrode sites is necessary because the stimulation efficiency and linearity of force
control using a specific electrode pair is determined by the motoneuron-electrode location, meaning
that a universal parametric force control method does not exist.
In addition, we propose new evidence to support the idea of inductive myelin sheath on
motoneuron axons, which is claimed in the C-P theory. In this study, we provide a theoretical and
experimental exploration of the excitability differences between the myelinated motoneurons and
unmyelinated muscle fibers.
Circuit-Probability simulation concept with distributed-parameter circuit
In the previous C-P theory, the target motoneuron is modeled as a parallel RLC circuit. The
extracellular medium is simplified as parallel impedance and integrated within the leakage resistor of
the parallel RLC circuit. In this paper, the entire muscle tissue, including the target motoneurons and
extracellular tissue, is modeled as a distributed-parameter circuit network (Fig. 1). The detailed circuit
configuration of each block in this network is determined by its biological features. The myelinated
motoneurons are modeled as parallel RLC components (green block). The unmyelinated muscle fibers
are modeled as RC components (blue block). The extracellular medium is represented with pure resistive
components (𝑹𝑹𝒔𝒔) connecting each block.
This circuit network can be further expanded by adding more functional components (e.g. fat tissue
and skin as the peripheral circuits) and revised when studying different experimental scenarios (e.g.
different configurations and implantation positions of the stimulating electrodes). In this study, a home-
made double-side multiple-channel polyimide electrode was implanted in the muscle belly, transversal
to the muscle fibers of the Tibialis Anterior (TA) muscle in all experiments, and the force generated by
stimulation was measured by a force gauge tied to the ankle (Fig. 2a). Thus, the corresponding circuit
network is revised as Fig. 2b. Each electrode site (e1 to e6 refer to the frontside; e1' and e6' refer to the
backside) is connected to a node (E1 and E6 refer to the frontside; E1' and E6' refer to the backside). The
non-conducting polyimide layer is modeled as broken connections between the frontside and backside.
By connecting a current source to an arbitrary electrode sites and a voltage meter to the capacitor in an
arbitrary block, the effective voltage waveform at any position in the tissue can be calculated. This
voltage waveform will then be used for a probability calculation with a set of parameters, 𝜶𝜶, 𝜷𝜷, and
𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻 (parameters to calculate the probability of initiating an action potential), which is the same
as the C-P theory.
Results
Influence of motoneuron-electrode position on voltage waveforms. Apparently, the exact voltage
waveform on each block in Fig. 2b is determined by the block position and the electrode sites selected
to deliver current. This motoneuron-electrode position will affect the amplitude, shape and the polarity
of the voltage waveform on each block. Since there are many motoneuron-electrode combinations, two
simplified situations are simulated (Fig. 3a and Fig. 3b), to qualitatively investigate the effect of this
motoneuron-electrode position upon the voltage waveform.
In Fig. 3a, fixed electrode sites (positive electrode site between P3 and P4; negative electrode site
between P9 and P10) are selected to deliver negative-first biphasic square current input. Voltage
In another situation (Fig. 3b), two blocks are fixed (P1 and P2) while the spacing between the
electrode sites is increased (positive electrode site between P1 and P2 is fixed; negative electrode site
changes from E1 to E8). The voltage waveforms on P1 and P2 (Fig. 3d and Fig. 3d') show an opposite
changing trend. The peak amplitude of the block outside the electrode sites (P1) and between the
electrode sites (P2) increases and decreases with the increasing electrode spacing, respectively.
The changing trends of the peak amplitude and polarity (green lines in Fig. 3c, Fig. 3d and Fig. 3d')
agree with the results using conventional method of electric field distribution simulation. However, our
model can also reveal information in the time domain: voltage waveform, which is critical in determining
neuron activation under the C-P theory, but unavailable in previous methods. This voltage waveform can
help us understand more phenomena, which will be discussed in the next session.
Influence of voltage waveform on force mapping curve: An explanation of the multi-curvature
phenomenon. For muscle stimulation using square wave current, it is widely observed that the force
mapping curves measured with increasing current amplitude may show multiple curvatures[20][21]. A
conventional explanation for this phenomenon is that multiple groups of motoneurons are sequentially
recruited when current amplitude is increased. Alternatively, this phenomenon can be quantitatively
derived from the C-P theory. Using the same circuit configuration as Fig. 3a, the voltage waveforms of
the voltage oscillation induced by RLC components, multiple effective voltage areas (indicated as A and
block P1 to P12 in Fig. 3c are shown in Fig. 4a and Fig. 4b, along with an estimated 𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻. Due to
B in Fig. 4a) will sequentially exceed 𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻. when current amplitude is increased, resulting in
multiple curvatures in the probability mapping curves (Fig. 4a'). But this phenomenon will not always
happen. For the voltage waveforms with only one area exceeding voltage threshold (indicated as C in Fig.
4b), their corresponding probability mapping curves will show a single curvature (Fig. 4b'). These two
types of force mapping curves, with and without multiple curvatures, were observed in experiments.
waveforms upon different blocks (P1 to P12) are modeled and shown in Fig. 3c. For these different
blocks, the voltage amplitude changes, and the polarity also switches. The blocks which are next to the
electrode sites (P3 and P4; P9 and P10) have the largest voltage amplitude. Meanwhile, the voltage
polarity gradually changes from negative-first to positive-first when the position changes from P1 to P12.
Two conclusions can be summarized from the above modeling results and validated by force
mapping results in two in vivo experiments (Fig. 4c, Fig. 4c'). First, due to the unknown motoneuron-
electrode position after electrode implantation, the stimulating efficiency using the same electrode pair
will vary in different experimental trials. In Fig. 4c, the electrode sites with the largest spacing (e1and e6)
show minimum stimulation efficiency. However, in another experiment (Fig. 4c'), electrode sites of e1
and e6 show medium stimulation efficiency. Second, the multi-curvature phenomenon will not always
happen in the force mapping curves. Whether the curve will have multiple curvatures is determined by
the shape of the voltage waveform. The blue curve in Fig. 4c and purple curve in Fig. 4c' clearly show
multi-curvature pattern, while other curves don't show such pattern.
Activation of multiple groups of motoneurons. As explained above, in contrast to the conventional
explanation of multi-curvature force mapping curve as a sequential activation of multiple groups of
motoneurons, the activation of a single group of motoneurons can already account for this phenomenon
in our theory. We find that the unstable force output during transition periods of additional
motorneuron group activation better characterize the recruitment of motornueron groups due to
attenuation of voltage over a distance.
Two measured force mapping curves (e2e5, orange and e3e4, purple) are selected from Fig. 4c' and
plotted in Fig. 5c (blue and orange). There is a sudden force increment accompanied with an abnormally
large error bar for the blue curve. We believe that this error bar is a sign of an additional group of
motoneurons starting to be recruited, due to the differing voltage acting on spaced out groups of
motorneurons. In our experiments (Fig. 1a), the force measured is the summed-up force generated by
multiple groups of activated motoneurons. When an additional group of motoneurons starts to be
recruited in the stimulation, while the total force will be higher than activating one group alone, the
force generated by this second group of motoneurons is still low at this stage. Furthermore,
motorneurons that require higher stimulation current to be recruited are necessarily further away and
the stimulation strength (voltage upon the motoneurons) will be low, resulting in unstable activation
and force generated. Hence, the measured force curve will show a large error bar at this current. Then,
when the current further increases, the stimulation strength of this second group of motoneurons will
be higher to generate a stable force output, making the error bar recover to a normal level. This explains
why this abnormal high error bar only occurs within a certain current range. Fig. 5d shows the force
profile measured at this transition range (1000 µA to 1200 µA) of the blue curve in Fig. 5c. When current
is high enough (1200 µA), the force profile will recover to a stable condition.
Force mapping curves measured in two in-vivo experiments are consistent with the simulation
results (Fig. 6c, Fig. 6c'). One force mapping curve was measured at small pulse width range (100-300 µs),
and the other is measured at large pulse width range (100-800 µs). Two electrode sites (e1 and e6) of
the furthest distance on our polyimide electrode were used. The force mapping curves are the same
when e1 delivers positive current or e6 delivers negative current (blue and orange curves, with Fig. 6b'
type activation). The recruitment curves are also the same when e1 delivers negative current or e6
delivers positive current (purple and green curves, with Fig. 6b type activation). In each measurement,
the four recruitment curves form two groups of different changing trend with increasing pulse width.
The flipping of voltage waveform polarity (Fig. 6b and Fig. 6b') causes the voltage waveforms to exceed
the threshold in different patterns, so that the two groups have totally different changing trend with
increasing pulse width.
In addition to the measured force mapping curves, a good curve fitting with distributed-parameter
circuit modeling can help us to understand the spatial distribution of these motoneurons within the
muscle. Here a simple modeling case is demonstrated. The corresponding electrode pairs of these two
curves are modeled in the circuit network (e3e4 in Fig.2b as E1+E1- in Fig. 5a; e2e5 in Fig. 2b as E2+E2-
in Fig. 5a). The locations of the two target motoneurons (P1 and P2) are captured to fit the force
mapping curves. The probability curves (the simulated curves of activation probability under the C-P
theory with respect to input current amplitude or pulse width) are shown in Fig. 5b and Fig. 5b'. At low
stimulation current, both E1 and E2 electrode pairs can only stimulate P1. However, when current
increases to around 1600 µA, E2 starts to stimulate another motoneuron P2, while E1 has no influence
on P2. From this modeling, the relative locations of the two motoneurons with respect to the electrode
sites can be roughly estimated. This simple case demonstration shows the potential to apply this model
for an in-depth investigation of the biological structure in the future.
Influence of input current waveform polarity on motoneuron activation. As shown in Fig. 3c, the
polarity of the voltage waveforms at different positions will not be the same. This polarity is not only
determined by its position and the polarity of the applied current, but also determined by which
electrode is connected to the positive terminal of the current source. It means if the polarity of the input
current and the connection of the electrode sites to current source are both reversed, the voltage
waveform measured on the specific block will remain the same. One case by modeling is shown in Fig.6.
Two electrode sites (E1 and E2) are connected to the current source, which can deliver monophasic
square wave current pulses of different pulse widths. The voltage waveform by applying positive pulses
from E1 and negative pulses from E2 are the same (Fig. 6b). Similarly, the voltage waveform by applying
negative pulses from E1 and positive pulses from E2 are the same. Although the voltage waveforms in
Fig. 6b and Fig. 6b' are just of the opposite polarity, the changing trend of the effective voltage area by
increasing the pulse width is completely different.
Thus, for stimulations using two-electrode configurations, it is important to point out the polarity of
current input together with which electrode site is connected to the positive terminal, as even when the
same current input is used, results will be different when positive electrode is changed. This issue may
not be so prominent in muscle stimulation, because the difference of using different positive electrode
only shows in the stimulation efficiency. However, when it comes to the sciatic nerve stimulation using
cuff electrode, different positive electrode may stimulate nerve fibers innervating different muscles with
the same current input. This will lead to totally different results of activating different muscle groups.
Thus, to get consistent results in the experiments, it is important to pay attention to both the polarity of
current input and the chosen positive electrode.
Independent activation of motoneurons using two-side polyimide electrode. In Fig. 7a, due to the
non-conducting polyimide layer, there will be broken connections between the frontside and backside
of the electrode. Thus, when current is applied from the frontside electrode sites, motoneurons on the
backside won't be activated. In other words, the electrode sites on one side can only cause extremely
limited motoneuron activations on the other side of the electrode.
A modeling demonstration is shown in Fig. 7a-c. Two electrode sites (E+ and E-) on the frontside of
the electrode are selected for current delivery. F1, F2, F3 are three motoneuron positions on the
frontside, while B1, B2, B3 are three motoneuron positions on the backside of the electrode. The overall
voltage amplitude on the frontside motoneurons is much larger than the backside motoneurons (Fig. 7b,
Fig. 7b'). The probability curves show that only with much larger current, the backside motoneurons can
be activated by the frontside electrode sites (Fig. 7c).
In-vivo experiments proved the independent activation of motoneurons using two-side polyimide
electrode. At the beginning of the experiment, force mapping curves for both frontside and backside
electrode sites were measured (blue curves in Fig. 7d and Fig. 7d'). Then, the frontside electrode sites
were used for a long-time electrical stimulation to induce fatigue. The stimulation lasted for 2.5 min, and
the force profile can be found in Fig. 7e. This force profile showed signs of muscle fatigue at the end of
the 2.5 min stimulation, as the force dropped to half of the initial value. Right after this 2.5 min
stimulation, the force mapping curve using frontside electrode sites was measured (orange curve in Fig.
7d). At that moment, because of the muscle fatigue on the frontside, the force mapping curve was much
lower than the one measured in the beginning. Then, the force mapping curve using backside electrode
sites was also measured (orange curve in Fig. 7d'). It was only slightly lower than the one measured in
the beginning. Lastly, after 5 min resting, the force mapping curve using the frontside electrode sites
was measured again (purple curve in Fig. 7d), which showed recovery of muscle fatigue but was still
much lower than the measurement in the beginning. Thus, frontside electrode sites can stimulate
frontside motoneurons, but can barely stimulate backside motoneurons. It proves our prediction of
independent activation of motoneurons using two-side polyimide electrode sites.
A theoretical explanation of the difference in excitability of motoneurons and muscle fibers. There
are two types of excitable cells in the muscle tissue: motoneurons and muscle fibers. They show
distinctive excitability properties in electrical stimulation. In our model, these differences can be
theoretically explained and proved by in vivo experiments. The experiments targeting motoneurons and
muscle fibers are conducted on healthy muscle and denervated muscle, respectively.
Here, two distributed circuits are built to model the healthy muscle and denervated muscle. Since
the myelin sheath is proposed as an inductor, the equivalent circuit of a myelinated motoneuron is
modeled as a RLC component (Fig. 8a), while an unmyelinated muscle fiber is modeled as a RC
component (Fig. 8b). To make a fair comparison, the block representing the target motoneuron or
muscle fiber is placed at the same position in the circuit network. The modeling parameters are set
based on two reasonable considerations as follows:
1. Due to the lack of myelin sheath, muscle fibers have a larger exposed cell membrane surface and
more leakage channels, which can be modeled as a larger capacitor (𝑪𝑪𝟐𝟐>>𝑪𝑪𝟏𝟏) and a lower leakage
resistor (𝑹𝑹𝑹𝑹𝟐𝟐<<𝑹𝑹𝑹𝑹𝟏𝟏), respectively.
among which the most important is the same threshold voltage, 𝑽𝑽𝑇𝑇ℎ𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑜𝑜𝑜𝑜𝑜𝑜.
2. It has been reported that the conductance and gating properties of the sodium channels in
motoneuron and muscle fiber are nearly the same[22], and muscle fiber sodium channels only require a
slightly more negative potential to activate than motoneuron sodium channels[23][24]. Therefore, in our
modeling, the motoneurons and muscle fibers share the same parameters for probability calculus,
Firstly, the voltage waveforms are compared when the same negative monophasic current pulse
(100 µs pulse width) is applied (Fig. 8c1). The voltage waveform of the RC component shows a typical
charging and discharging curve as a capacitor, while the voltage waveform of the RLC component shows
a voltage oscillation, resulting in a much higher peak amplitude. Thus, to exceed the threshold voltage
and activate the muscle, the threshold current required for RLC component is much lower than RC
component. The probability mapping curves by changing current amplitude in Fig. 8c2 show the
difference in threshold current. The probability mapping curves with increasing pulse width in Fig. 8c3
show that RC circuit requires a much higher current to achieve the same force. This phenomenon is also
observed in previous report[25] and confirmed by our in vivo experiments (Fig. 8c4). To achieve the
similar force output, healthy muscle requires only 1 mA while the denervated muscle requires 10 mA.
Then the changing trend of the voltage waveforms with increasing pulse width is compared in Fig.
8d1 and Fig. 8d2. The voltage waveforms of the RC component show a simple charging and discharging
process as a capacitor. Although the slope of the curve, which represents the charging rate, will vary
with the circuit parameters, the general shape will not change. Thus, the effective voltage area, which is
the part exceeding 𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻, will always have a monotonically increasing trend. As a result, the
corresponding probability mapping curves (Fig. 8d3) and measured force mapping curves (Fig. 8d4) also
increase with respect to pulse width monotonically. However, the changing trend of the effective
voltage area of the RLC component is quite complex. Many factors, such as current amplitude, circuit
parameter and extracellular environment, can induce nonlinear effect upon the effective voltage area
and finally result in various probability mapping patterns, which is investigated in details in our previous
C-P theory[19].Fig. 8d2 just shows a typical voltage waveform of RLC component. With different current
amplitudes, the effective voltage waveform will be completely different.
In summary, due to the lack of myelin sheath, muscle fiber is not very easily stimulated being pure
RC, but because of the branching motoneurons coordinating the impulse it is able to be activated with
lower current in healthy muscles. Meanwhile, the force mapping curves by changing pulse width in
muscle fiber stimulation always follow a monotonically increasing trend while the force mapping curves
of motoneuron stimulation will have a complex pattern[19].
Discussion
A static system-level distributed-parameter circuit to study electrical stimulation on muscle tissue.
In this paper, instead of treating the muscle tissue as a black box and trying to get some simplified
descriptions about the relationships between input and output, we thoroughly studied the interaction
between the stimulation electrodes and muscle tissue. This interaction is highly affected by all
parameters in the system, including the motoneuron-electrode position, input current waveforms
(amplitude, pulse width, frequency, and polarity), circuit properties of the tissue components
(motoneuron, muscle fiber, and extracellular medium), and the ion channel properties (𝜶𝜶, 𝜷𝜷, and
𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻). Therefore, a simplified general description of the relationship between input and output
does not exist. No matter what kind of simplified conclusions are summarized from the experiment
observations, exceptions will always exist.
Here, we demonstrate a modified distributed-parameter circuit to study a specific electrode
implantation: intramuscular implantation of a double-side multiple-channel polyimide electrode. There
are many electrode designs for muscle stimulation, including epimysial electrode and skin surface
electrode. These electrodes also come in different structures, like flexible strip electrodes, wire
electrodes and large-area patch electrodes. By including circuit components representing additional
tissue layers (fat layer, skin) and properly assigning the motoneuron-electrode positions, modified
distributed-parameter circuit can be applied to study all these implantations.
A potential solution to achieve sustained force output with reduced muscle fatigue in electrical
stimulation. Our distributed-parameter circuit can provide guidance to solve practical issues in electrical
stimulation. One example is to use double-side electrode stimulation for less muscle fatigue. The goal of
Functional Electrical Stimulation (FES) is to achieve meaningful functions using precisely controlled
electrical stimulation. However, electrical stimulation causes muscle fatigue much more easily than the
voluntary controlled movements. The fast kick-in of muscle fatigue is an obstacle to achieve some
meaningful functions, such as holding and gripping. In this paper, we have proved the independent
recruitment of motoneurons using double-side polyimide electrode. Even when the frontside
motoneurons are severely fatigued with the frontside electrodes delivering electrical stimulation, the
backside motoneurons are barely affected and can still generate large force with the backside
electrodes. The independent recruitment of motoneurons results in independent fatigability within a
muscle. If we expand this concept to a meaningful function, such as a holding movement, the frontside
and backside electrodes can be controlled to alternatively deliver the electrical stimulation (Fig. 9). By
alternating the side of electrode sites during a long stimulation, the muscle fibers and motoneurons on
one side can rest and recover when the other side is recruited for force output.
Potential applications of distributed-parameter circuit model to study electrical stimulation on
other neural tissue. Apart from muscles, this distributed-parameter circuit can also be applied to other
excitable tissues, including the Peripheral Nervous System (PNS) and the Central Nervous System (CNS).
In the distributed-parameter circuit description for the PNS and CNS, the muscle fibers will be removed,
and much more neuron distribution will be added. Further study using our distributed-parameter circuit
will help to elucidate important issues in the PNS and CNS electrical stimulation. An example is to use
high frequency localized electrical stimulation for selective activation. Selective stimulation of different
nerve fascicles on the sciatic nerve can enable functional activation of various muscle groups. In addition,
localized stimulation in the spinal cord can eliminate unwanted disturbance on other functional areas.
In the previous research, without this system-level model, researchers tended to use experiment
observations to judge the efficiency of electrical stimulation. Now we know the stimulation effect under
a specific condition is affected by so many parameters in the system. Since in the way of experiment
observations, only a certain range of parameters can be applied, this may give rise to a segmented
conclusion about this specific stimulation. This segmented conclusion may neglect the complicated
interaction between the stimulation electrodes and excitable tissue. Here, we take the study of tripolar
stimulation on the sciatic nerve as an example, to demonstrate how segmented conclusions can arise
when only parameters of a certain range are applied during the experiment observations. In Fig. 10, the
two outer electrode sites are connected to the positive terminal of the current source, and the middle
electrode site is connected to the negative terminal. Positive monophasic input current is delivered
using this tripolar structure. With the knowledge about voltage waveform polarity in Fig. 6, we now
know that it is the middle electrode site which is delivering negative input current that stimulates the
nerve axons to generate action potentials. When these action potentials travel to the electrode site
located at the lower stream of the nerve axons, they will be distorted by the external stimulation at that
location. Whether these action potentials can successfully pass through the lower-stream electrode site
depends on the external stimulation at that location: they can pass through when the amplitude of the
external stimulation is smaller than the action potential (Fig. 10a), and they will be blocked when the
amplitude of the external stimulation is larger (Fig. 10b). However, it is also possible that when the
action potentials arrive at the lower-stream electrode site, the external stimulation of a short pulse
width has already finished. Then, in this case, the action potentials will pass through at both small and
large amplitude (Fig. 10c, Fig. 10d). Thus, if we use force measurement as a quantification of the action
potentials passing through the lower-stream electrode site, the force curve with respect to the current
amplitude will show totally different patterns when different pulse width is used.
A new explanation of cathodal make, cathodal break, anodal make, and anodal break stimulation
phenomena. It has been observed that cardiac tissue can be electrically stimulated with the onset (make)
or termination (break) of an input current that is delivered with either a negative (cathodal) or positive
(anodal) electrode[26][27]. These phenomena can be easily explained with our model. In the case of
unipolar stimulation configuration, where only one electrode delivers input current, the effectiveness of
cathodal make can be intuitively understood, as it will depolarize the tissue surrounding the electrode.
However, considering the dramatic resonance caused by the RLC component, cathodal break, anodal
make, and anodal break may also generate voltage waveforms with a certain area sufficiently negative
to exceed the voltage threshold.
The gap between static and dynamic distributed-parameter circuit model. Inherited and expanded
from the C-P theory, we still assigned static values to parameters of both the circuit components
(resistance, inductance, and capacitance) and probability calculation (𝜶𝜶, 𝜷𝜷 and 𝑽𝑽𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝒔𝒔𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻). By using
static parameters of circuit components, we indeed assumed that the electrical properties of the muscle
tissue remain unchanged during applied electrical stimulation. Similarly, by using static parameters of
probability calculation, we assumed that external electrical stimulation won't affect the resting potential
and threshold voltage of motoneurons. This adoption of static parameters limits our capability to study
some specific issues. One example is the quantitative description of muscle fatigue during electrical
stimulation, which is characterized as change in force output. During muscle fatigue, the same current
input is applied, but the output force is different. It means something must have changed, either the
generated voltage waveform affected by circuit parameters, or the ion channel response to the voltage
waveform which is affected by probability calculation parameters.
Another limitation caused by the adoption of static parameters is that we cannot quantitatively
study how an external electrical stimulation interacts with a generated action potential. With the static
circuit parameters, we can only study how the external stimulation initiates an action potential. Unlike
the natural action potentials, the action potential initiated by the external electrical stimulation may be
distorted by this external stimulation. For the external stimulation that lingers on after the action
potential is already initiated, it will interact with the movements of local ions across the neuron
membrane and result in a distorted action potential.
Thus, to quantitatively study these dynamic processes in electrical stimulation, further studies are
required to establish a bridge between the parameters in our model and the dynamic change of muscle
tissue.
Reason for using voltage waveforms simulated with a distributed-parameter circuit, instead of
electric field distribution simulated with the existing Finite Element Modeling (FEM). The cross-
membrane electric field opens the ion channels. Considering the neuron membrane is partially
permeable to ions, charged ions will accumulate on both sides of the neuron membrane. Thus, the
electric field generated by these local ions is perpendicular to the membrane. To model the effect of this
perpendicular cross-membrane electric field (𝑬𝑬) on the neuron, we calculate the voltage across the
the capacitor can be used to characterize the electric field perpendicular to the membrane, which really
opens an ion channel. However, the electric field simulated with the existing FEM is only determined by
the physical boundary condition and electrode positions, which is irrelevant to the electric field
perpendicular to the neuron membrane.
capacitor (𝑽𝑽) in our model. This 𝑽𝑽 is the path integral of the electric field 𝑬𝑬 in the direction
perpendicular to the capacitor plates (neuron membrane) as described by the equation ∆𝑽𝑽=𝑬𝑬∗∆𝑻𝑻
(∆𝑽𝑽 is a change of 𝑽𝑽, ∆𝑻𝑻 is a change of membrane thickness). Thus, in our model, the voltage across
nonlinear relationship between 𝑽𝑽(𝒕𝒕) (cross-membrane voltage at time point t) and 𝑰𝑰(𝒕𝒕) (input current
at time point t). In response to an external input current 𝑰𝑰(𝒕𝒕) , the local ions on both sides of the neuron
assuming a linear relationship between 𝑽𝑽(𝒕𝒕) and 𝑰𝑰(𝒕𝒕) .
membrane will move. This local ion movement is not only affected by the charging and discharging
process of the neuron membrane, but also affected by the surrounding tissue environment. That's why
we build a distributed-parameter circuit network to account for the influence of the other parts of the
tissue. However, in the existing FEM, the tissue is modeled in a medium. In this way, the existing FEM is
In our model, by calculating the voltage across the capacitor (neuron membrane), we can model the
Methods
Animal statement. All experiments were conducted according to protocols approved by the
Institutional Animal Care and Use Committee at the National University of Singapore.
Preparation of denervated muscles. Sprague-Dawley rats (around 450 g) were used for the sciatic
nerve transaction. Anesthesia was induced with isoflurane. Buprenorphine was injected for pain relief
before the surgery. After the rat was anesthetized, a shaver was used to gently remove the fur on the
left leg. Then, the skin was disinfected with 70% ethanol wipes, and an incision was made to expose the
bicep femoris muscle. An incision was then made on this bicep femoris muscle to expose the sciatic
nerve. The sciatic nerve was transacted on the site before it branched into three smaller nerves. Then
the bicep femoris muscle and the skin were both sutured back. Right after the surgery, enrofloxacin was
injected. During the first five days after the nerve transaction surgery, buprenorphine was injected twice
a day and enrofloxacin was injected once a day.
Interface implantation. The homemade flexible interface was folded to form a loop on the tip. A
suture was threaded through this loop. Then, this suture was threaded through the center of the
exposed TA muscle belly. Pulling by the suture, the interface was also threaded into the muscle belly.
The interface was sutured to the muscle surface for fixation. Then, the skin was sutured back.
Electrical stimulation. A-M SYSTEMS model 4100 isolated high-power stimulator was used for
electrical stimulation. Every one second, a train of 10 negative-positive biphasic pulses of 60 Hz was
applied. In the experiment of comparing four waveforms, positive-negative biphasic pulses, negative-
positive biphasic pulses, positive monophasic pulses, and negative monophasic pulses were applied.
Force data collection and analysis. The anesthetized rat was fixed on a stand, and the ankle of the
left leg was connected to a dual-range force sensor (Vernier). This force sensor was connected to a
laptop through a data acquisition device (National Instruments). LabView (National Instruments) was
used for on-site result visualization during the measurements. After the measurements, MATLAB
(MathWorks) was used for data analysis. The recorded forces exceeding a certain amplitude threshold
were used for assessing the strength of the excitability. If the forces were below this amplitude
threshold, then the muscle was considered as not activated.
Distributed-parameter circuit simulation. The simulation was performed on MATLAB (MathWorks).
Firstly, a circuit description was performed in Simulink (MathWorks). Then, current inputs of different
waveforms were recursively fed to the circuit model, and the voltage responses of the targeted RC or
RLC component were collected. Lastly, these voltage responses were fed into the probability equation to
Acknowledgements
calculate the probability of excitation under these current inputs. The circuit parameters and probability
calculation parameters are shown in Table. 1 and Table. 2.
This work was supported by the following grant from the National Research Foundation:
Competitive Research Project 'Peripheral Nerve Prostheses: A Paradigm Shift in Restoring Dexterous
Limb Function' (NRF-CRP10-2012-01), and the grant from the HIFES Seed Funding: 'Hybrid Integration of
Flexible Power Source and Pressure Sensors' (R-263-501-012-133).
We would like to thank for the experiment setup support from Han Wu, Li Jing Ong and Wendy Yen
Xian Peh. We also would like to thank for the animal experiment support from Gammad Gil Gerald
Lasam.
[1] Anderson, K. D. Targeting recovery: priorities of the spinal cord-injured population. J. Neurotrauma.
21(10), 1371-1383 (2004).
[2] Galvani, L. De viribus electricitatis in motu musculari commentarius. De Bononiensi Scientiarum et
Artium Instituto atque Academia Commentarii, 7, 363 -- 418 (1791).
[3] Hodgkin, A. L., and Andrew F. H. A quantitative description of membrane current and its application
to conduction and excitation in nerve. J. Physiol. 117, no. 4: 500-544 (1952).
[4] McNeal, D. R. Analysis of a model for excitation of myelinated nerve. IEEE Trans Biomed Eng. 4, 329-
337 (1976).
[5] Raspopovic, S., Capogrosso, M., & Micera, S. A computational model for the stimulation of rat sciatic
nerve using a transverse intrafascicular multichannel electrode. IEEE Trans Neural Syst Rehabil Eng 19(4),
333-344 (2011).
[6] Raspopovic, S., Petrini, F. M., Zelechowski, M., & Valle, G. Framework for the development of
neuroprostheses: from basic understanding by sciatic and median nerves models to bionic legs and
hands. Proc. IEEE. 105(1), 34-49 (2017).
[7] Pelot, N. A., Behrend, C. E., & Grill, W. M. Modeling the response of small myelinated axons in a
compound nerve to kilohertz frequency signals. J Neural Eng. 14(4), 046022 (2017).
[8] Capogrosso, M. et al. A computational model for epidural electrical stimulation of spinal
sensorimotor circuits. J. Neurosci. 33(49), 19326-19340 (2013).
[9] McIntyre, C. C., & Grill, W. M. Extracellular stimulation of central neurons: influence of stimulus
waveform and frequency on neuronal output. J. Neurophysiol. 88(4), 1592-1604 (2002).
[10] Dehghani, N., Bédard, C., Cash, S. S., Halgren, E., & Destexhe, A. Comparative power spectral
analysis of simultaneous elecroencephalographic and magnetoencephalographic recordings in humans
suggests non-resistive extracellular media. J Comput Neurosci. 29(3), 405-421 (2010).
[11] Bédard, C., Rodrigues, S., Roy, N., Contreras, D., & Destexhe, A. Evidence for frequency-dependent
extracellular impedance from the transfer function between extracellular and intracellular potentials. J
Comput Neurosci. 29(3), 389-403 (2010).
[12] Reimann, M. W. et al. A biophysically detailed model of neocortical local field potentials predicts
the critical role of active membrane currents. Neuron, 79(2), 375-390 (2013).
[13] Gomes, J. M. et al. Intracellular impedance measurements reveal non-ohmic properties of the
extracellular medium around neurons. Biophys. J. 110(1), 234-246 (2016).
[14] Aldrich, R. W., & Stevens, C. F. Voltage-dependent gating of single sodium channels from
mammalian neuroblastoma cells. J. Neurosci. 7(2), 418-431 (1987).
[15] Zagotta, W. N., & Aldrich, R. W. Voltage-dependent gating of Shaker A-type potassium channels in
Drosophila muscle. J. Gen. Physiol. 95(1), 29-60 (1990).
[16] Zagotta, W. N., Hoshi, T., & Aldrich, R. W. Gating of single Shaker potassium channels in Drosophila
muscle and in Xenopus oocytes injected with Shaker mRNA. PNAS, 86(18), 7243-7247 (1989).
[17] White, J. A., Rubinstein, J. T., & Kay, A. R. Channel noise in neurons. Trends Neurosci. 23(3), 131-137
(2000).
[18] Mino, H., Rubinstein, J. T., & White, J. A. Comparison of algorithms for the simulation of action
potentials with stochastic sodium channels. Ann Biomed Eng, 30(4), 578-587 (2002).
[19] Hao, W. et al. Unveiling Stimulation Secrets of Electrical Excitation of Neural Tissue Using a Circuit
Probability Theory. Submitted to Nature.
[20] Crago, P. E., Peckham, P. H., & Thrope, G. B. Modulation of muscle force by recruitment during
intramuscular stimulation. IEEE Trans Biomed Eng. 12, 679-684 (1980).
[21] Popovic, D., Baker, L. L., & Loeb, G. E. Recruitment and comfort of BION implanted electrical
stimulation: implications for FES applications. IEEE Trans Neural Syst Rehabil Eng. 15(4), 577-586 (2007).
[22] Campbell, D. T., & Hille, B. E. R. T. I. L. Kinetic and pharmacological properties of the sodium channel
of frog skeletal muscle. J. Gen. Physiol., 67(3), 309-323 (1976).
[23] Pappone, P. A. Voltage‐clamp experiments in normal and denervated mammalian skeletal muscle
fibres. J. Physiol. 306(1), 377-410 (1980).
[24] Weiss, R. E., & Horn, R. Functional differences between two classes of sodium channels in
developing rat skeletal muscle. Science, 233(4761), 361-364 (1986).
[25] Peckham, P. H., & Knutson, J. S. Functional electrical stimulation
applications. Annu. Rev. Biomed. Eng., 7, 327-360 (2005).
[26] Dekker, E., Direct current make and break thresholds for pacemaker electrodes on the canine
ventricle. Circ. Res. 27(5), pp.811-823 (1970).
[27] Wikswo Jr, J. P., Lin, S. F., & Abbas, R. A. Virtual electrodes in cardiac tissue: a common mechanism
for anodal and cathodal stimulation. Biophys. J. 69(6), 2195-2210 (1995).
for neuromuscular
Figure 1 Distributed-parameter circuit simulation concept. Rs represents resistor of extracellular
media. RLC components (green blocks) represent motoneurons distributed in the healthy innervated
muscle tissue. RC components (blue blocks) represent muscle fibers.
Figure 2 In vivo measurement setup and the corresponding distributed-parameter circuit concept. (a)
Our home-made double-side multiple-channel polyimide electrode is sutured in the muscle belly,
transversal to muscle fibers of the Tibialis Anterior (TA) muscle. When electrical stimulation is delivered
to the TA muscle, the leg can freely kick forward. Force is measured from the ankle. (b) The
corresponding distributed-parameter circuit with broken connections between two sides of electrode
sites representing the insulating polyimide substrate, and RLC components (green blocks) representing
distributed motoneurons.
Figure 3 Influence of motoneuron-electrode position on voltage waveform. (a) Distributed-parameter
circuit of fixed electrode site location and changing targeted motoneuron positions (green blocks P1-
P12). (b) Distributed-parameter circuit of fixed positive electrode site location and changing negative
electrode site locations (E1-E8). Two targeted motoneuron positions (P1, P2) are studied, P2 is between
the electrode sites, and P1 is outside. (c) Voltage waveforms on motoneuron positions in (a). (d, d')
Voltage waveforms on P1 (d) and P2 (d') with negative electrode positions in (b).
Figure 4 Force mapping curves with multiple curvatures can be explained with the voltage
waveforms. (a, a') Voltage waveforms (a) and corresponding probability curves (a') of P1-P6 in Fig. 3c.
(b, b') Voltage waveforms (b) and corresponding probability curves (b') of P7-P12 in Fig. 3c. For voltage
waveforms with two areas exceeding the threshold (area A and area B), the probability curves have two
corresponding curvatures. For voltage waveforms with only one area exceeding the threshold (area C),
the probability curves also only show one curvature. (c, c') Quantification of force at different current
measured in two in-vivo experiments. Data are means ± s.d. (n=15 per group).
Figure 5 Activation of multiple groups of target motoneurons. (a) Distributed-parameter circuit with
two groups of electrode sites (E1+ and E1-, E2+ and E2-) and two targeted motoneuron positions (P1,
P2). (b, b') Simulation results. Probability curves of two groups of electrode sites on P1 and P2
separately (b). Probability curves of the summed stimulation effects on P1 and P2 using two groups of
electrode sites (b'). (c) Quantification of force at different current. The two curves using different groups
of electrodes have a crossing point. (d) Force profile at four current corresponding to the transition
points in (c). Data are means ± s.d. (n=15 per group).
Figure 6 Influence of input current waveform polarity on motoneuron activation. (a) Distributed-
parameter circuit with two electrode sites (E1 and E2) and a targeted motoneuron (green block). (b, b')
Simulated voltage waveforms of different pulse width. E1 delivers positive monophasic current or E2
delivers negative monophasic current (b). E1 delivers negative monophasic current or E2 delivers
positive monophasic current (b'). (c) Quantification of force at different pulse width in two experiments.
Data are means ± s.d. (n=15 per group).
Figure 7 Independent activation of motoneurons using double-side polyimide electrode. (a)
Distributed-parameter circuit with broken connections to represent the non-conducting polyimide
substrate. E- and E+ are electrode sites on the frontside of the polyimide electrode. F1, F2, F3 are
motoneurons on the frontside, and B1, B2, B3 are motoneurons on the backside of the polyimide
electrode. (b, b') Voltage waveforms on F1-F3 (b) and B1-B3 (b') when current input is delivered from
frontside electrode sites. The scale of the front motoneuron waveforms are much larger than the back
motoneuron waveforms. (c) The corresponding probability curves. (d, d') In-vivo measurement results.
Quantification of force at different current during frontside stimulation (d) and backside stimulation (d').
(e) Force profile of 2.5 min electrical stimulation to induce fatigue on frontside. Data are means ± s.d.
(n=15 per group).
Figure 8 The inductive myelin sheath on motoneurons accounts for the huge difference in
excitability of motoneurons and muscle fibers. (a) Distributed-parameter circuit with RLC component
(green block) representing the target motoneuron in healthy muscle stimulation. (b) Distributed-
parameter circuit with RC component (blue block) representing the target muscle fiber in denervated
muscle stimulation. (c1, c2) Simulated voltage waveforms (c1) and probability curves (c2) on RLC
component and RC component (c3, c4) Simulated probability curves (c3) and quantification of force (c4)
at small pulse width. (d1, d2) Simulated voltage waveforms of RC (d1) and RLC (d2) component at
different pulse width (d3, d4) Simulated probability curve (d3) and quantification of force of a
denervated muscle (d4) at large pulse width. Data are means ± s.d. (n=15 per group).
Figure 9 Illustration of a long-duration current waveform using double-side electrode stimulation for
reduced muscle fatigue. The frontside and backside electrode sites are controlled to alternatively
deliver the stimulation current.
Figure 10 Illustration of tripolar stimulation on the sciatic nerve. The two outer electrode sites are
connected to the positive terminal of the current source, and the middle electrode site is connected to
the negative terminal. A positive monophasic input current is applied using this tripolar structure. (a, b)
Large pulse width. Action potentials generated by the negative electrode site can pass through the
lower-stream positive electrode when the current amplitude is small (a). Action potentials cannot pass
through when the current amplitude is large (b). (c, d) Small pulse width. Action potentials can pass
through when the current amplitude is small (c) and large (d).
Figure 11 Concept of conventional computational model and our model to study excitability of
tissue.
Table 1 Parameter table of components used in the distributed-parameter circuit.
Table 2 Parameter table used in distributed-parameter circuit simulation and probability calculation.
Extended Figure 1 Influence of input biphasic current polarity on motoneuron activation. (a)
Distributed-parameter circuit with two electrode sites (E1 and E2) and a targeted motoneuron (green
block). (b) Simulated voltage waveforms of different pulse width. E1 delivering positive-negative
biphasic current and E2 delivering negative-positive biphasic current share the same voltage waveform.
E1 delivering negative-positive biphasic current and E2 delivering positive-negative biphasic current
share the same voltage waveform. (c) Quantification of force at different pulse width in two
experiments. According to the simulation in (b), the four measured recruitment curves should come in
two groups (e1 positive-negative and e6 negative-positive in one group, e1 negative-positive and e6
positive-negative in another group). However, the two measurements show a single group. (d) To
investigate the reason for showing a single group in recruitment curves, instead of two, simulation was
carried out to see how the two groups of possibility curves change with stimulation current. The
probability curve shows that the voltage waveform in response to positive-negative, and negative-
positive current input have a slightly different changing trend with increasing current, and there is a
crossing point. (e) Recruitment curves of e1 delivering positive-negative and negative-positive current.
The two recruitment curves overlap in a large range of current. This explains why only a single group of
recruitment curves is observed in (c).
Extended Figure 2 Shift of force mapping curves due to muscle fatigue. Quantification of force at
different current in the first round of recruitment curve measurement (first panel). Then, the
recruitment force using each electrode site combination was measured again. For each electrode site
combination, the recruitment curve shifted towards larger current and smaller force output, due to
accumulated muscle fatigue. Data are means ± s.d. (n=15 per group).
Extended Figure 3 Shift of recruitment curves due to muscle fatigue. In both testing, three rounds of
recruitment measurement were carried out. In each round, the current was changed from large to small,
and then small back to large. In both testing, the recruitment curves shift towards smaller force output,
due to accumulated muscle fatigue.
Extended Figure 4 Recruitment curves of denervated muscles using 10 mA current. The recruitment
curves were measured on denervated muscle, using both sides of the polyimide electrode.
Extended Figure 5 Electric potential distribution generated by two point charges. Electric potential is
calculated for positions in the line of the two charges. The amplitude of the generated electric potential
is the largest near the two charges. The polarity of the electric potential changes at the center of the
two charges. This distribution can help to understand the change of voltage waveforms generated at
different location in our model.
|
1501.05174 | 2 | 1501 | 2015-04-19T14:13:12 | Characterizing the neural correlates of reasoning | [
"q-bio.NC"
] | The brain did not develop a dedicated device for reasoning. This fact bears dramatic consequences. While for perceptuo-motor functions neural activity is shaped by the input's statistical properties, and processing is carried out at high speed in hardwired spatially segregated modules, in reasoning, neural activity is driven by internal dynamics, and processing times, stages, and functional brain geometry are largely unconstrained a priori. Here, it is shown that the complex properties of spontaneous activity, which can be ignored in a short-lived event-related world, become prominent at the long time scales of certain forms of reasoning which stretch over sufficiently long periods of time. It is argued that the neural correlates of reasoning should in fact be defined in terms of non-trivial generic properties of spontaneous brain activity, and that this implies resorting to concepts, analytical tools, and ways of designing experiments that are as yet non-standard in cognitive neuroscience. The implications in terms of models of brain activity, shape of the neural correlates, methods of data analysis, observability of the phenomenon and experimental designs are discussed. | q-bio.NC | q-bio | Characterizing the neural correlates of reasoning
David Papo1,2*
1
Laboratory of Biological Networks, Center for Biomedical Technology, Universidad Politécnica de Madrid
2 LPC Institute, Madrid
*Correspondence:
David Papo, Laboratory of Biological
Networks, Center for Biomedical
Technology, Universidad Politécnica de
Madrid, Calle Ramiro de Maeztu, 7, 28040
Madrid, Spain
e-mail: [email protected]
www: https://davidpapo.wordpress.com/
The brain did not develop a dedicated device for reasoning. This fact bears dramatic consequences. While for
perceptuo-motor functions neural activity is shaped by the input’s statistical properties, and processing is carried
out at high speed in hardwired spatially segregated modules, in reasoning, neural activity is driven by internal
dynamics, and processing times, stages, and functional brain geometry are largely unconstrained a priori. Here, it is
shown that the complex properties of spontaneous activity, which can be ignored in a short-lived event-related
world, become prominent at the long time scales of certain forms of reasoning which stretch over sufficiently long
periods of time. It is argued that the neural correlates of reasoning should in fact be defined in terms of non-trivial
generic properties of spontaneous brain activity, and that this implies resorting to concepts, analytical tools, and
ways of designing experiments that are as yet non-standard in cognitive neuroscience. The implications in terms of
models of brain activity, shape of the neural correlates, methods of data analysis, observability of the phenomenon
and experimental designs are discussed.
Keywords: cognitive neuroscience, reasoning, scaling, non-stationarity, non-ergodicity, characteristic scales,
observation time, resting brain activity.
INTRODUCTION
Consider an individual trying to solve a problem, and
reasoning for ten minutes before attaining a solution.
Take the middle five minutes. Clearly, though containing
no behaviourally salient event, these
five minutes
represent a genuine, indeed rather general, instance of
reasoning. What do we know about the brain regime far
from its conclusion? Can we use this regime to predict a
solution, and a solution to retrodict this regime?
Here, I concentrate on a form of reasoning, of which
the above scenario constitutes an example, which can
broadly be defined as "thinking in which there is a
conscious intent to reach a conclusion and in which
methods are used that are logically justified" [1], with no
a priori assumption on the type of reasoning process that
may take place during it. It is argued that finding the
generic properties of this form of reasoning entails
addressing the following fundamental issues: What are
reasoning's temporal and spatial scales? When is a given
observation time sufficient? How should we integrate the
information contained in various reasoning episodes?
A MINI LITERATURE REVIEW
The neural correlates of reasoning have traditionally been
expressed in terms of brain spatial coordinates. Early
neuropsychological work viewed reasoning as emerging
from global brain processing [2], consistent with
evidence indicating that it is negatively affected by diffuse
brain damage [3]. Neuroimaging studies have framed the
neural correlates of reasoning
local
functionally specialized brain activity, either by taking a
normative
or
by fractionating it into sub-component processes [11-14].
The results often lack specificity to reasoning [15]. Most
terms of
reasoning
approach
[4-10],
in
to
importantly though, these investigations provide a static
characterization of reasoning.
The neuroimaging literature mostly focused on short-
term and normative forms of reasoning [9,16-18]. This
minimizes variability in reasoning episode length and
allows segmenting reasoning episodes into separable
chunks, but does that at the price of limitations in the
phenomenology and ecologic value of
its stimuli.
Some neuroimaging [19,20] and electrophysiological [21-
29] studies examined more ecological forms of reasoning,
viz.
even
electrophysiological studies, despite optimal temporal
resolution, adopted an event-related perspective,
concentrating on activity occurring few seconds before
insight emergence , which only documents the outcome of
the reasoning process, not the process itself.
However,
problems
insight
[30].
Event-related neural activity associated with the
solution of riddles with insight was found to be related to
properties of preceding resting activity [26,27]. These
studies had the remarkable merit of using spontaneous
brain activity to characterize reasoning, but in essence
provided a comparative statics description. Although
some behavioural studies
as a
dynamical process [31], a comparable neurophysiological
characterization
incomplete. Altogether, the
research accomplished so far has generally not looked at
reasoning as a dynamical process, and produced either
time-averaged frames or discrete maps in partial time
cuts.
treated reasoning
is still
THE PROBLEM(S) WITH REASONING
The generalized form of reasoning considered in this
study comes in episodes offering scant behaviourally
salient events with no characteristic temporal length.
Each episode is a non-reproducible instance, as reasoning
task can be carry out in multiple ways. Brain activity
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
associated with reasoning is not event-related, and many
neurophysiological processes interact in a wide range of
spatial and temporal scales.
These phenomena can all be traced back to a basic
fact: the brain did not develop a dedicated device for
reasoning. Hardwired partially segregated modules
ensure that perceptuo-motor functions are carried out at
great speed, with stereotyped duration and time-varying
profile, and identifiable stages, largely determined by
input statistical properties. Reasoning, on the contrary, is
an internally-driven dynamics: processing times, stages,
and functional brain geometry are largely unconstrained.
Considering these extraordinary challenges, can we
still find general reasoning properties, over and above
specific task demands and individual differences? What
sort of process is reasoning in its general form? Is it a
series of simpler reasoning cycles? Can we segment it into
stages? What are the best neural variables and tools to
make these properties observable?
CHARACTERIZING THE REASONING PROCESS
tools accordingly. For
Robust characterizations of reasoning should incorporate
stylized facts, i.e. properties consistently appearing on
different subjects and in different periods of time, and
select analytical
instance,
perceptual response sensitivity to
incoming signals,
stability against noise, and minimal dependence on initial
conditions favour tools capturing transient dynamics,
which naturally reproduce
these properties under
appropriate conditions, over tools handling asymptotic
activity, which fail to do so [32].
from
instability and
and protection
Reasoning's relative
inefficiency
that optimal circuitry may need constant
suggest
reconstruction
interference,
summoning protracted support of energetically costly
long-range communications. Thus, reasoning may be a
sort of resonant regime, where functional efficiency
would be achieved with specific, though unstable, spatio-
temporal patterns, and should be studied with tools
capable of
spatially-extended dynamic
transients.
extracting
REASONING DYNAMICS
Each cognitive process can be translated in dynamical
terms and corresponding aspects of neural activity.
are
processes
Perceptual
quasi-
stereotyped short duration processes. The brain can
prima facie be modelled as an excitable
medium:
perturbations above a threshold induce a dynamical cycle,
before the system reverts to its initial silent state.
relaxational,
the brain
Learning too is a relaxational process: following a
gradient dynamics,
the
environment's statistical relationships, by representing
them in terms of its functional connectivity. Cycles can be
of much longer duration and non-trivial shape than
perceptual ones. The dynamics
is dominated by
fluctuations much shorter than the whole process.
incorporates
Reasoning may not be purely relaxational. The
corresponding neural activity is a fluctuation-dominated
endogenously modulated spontaneous brain activity, with
no clear gradient, and no single instant summarizing the
entire process. The reasoning scientist's world
is
considerably more complex than the event-related short
time scale one of perception. It is through the generic
properties of spontaneous activity's long time scales that
robust reasoning properties should be formulated [33].
THE STARTING POINT: SPONTANEOUS BRAIN ACTIVITY
Spontaneous activity can be thought of as a data bank of
cortical states, continuously reedited across the cortex
[34]. This re-editing process contains rich non-random
spatio-temporal structure [35-40].
The building blocks of this structure are fluctuations
which the brain, as all dissipative out-of-equilibrium
systems, generates even for fixed control parameter
values and in the absence of external stimuli, and which
constitute the trademark of its functioning. The dynamics
is intermittent, with alternating laminar and turbulent
phases [41,42], weakly non-ergodic, i.e. some phase space
regions take extremely long times to be visited, and
shows aging, i.e. temporal correlations depend on the
observation time [43]. Various aspects of spontaneous
activity display similar properties at all temporal and
spatial scales [44-51]. Self-similarity break-down [52,53],
with Gaussian low-frequency and non-Gaussian high-
frequency fluctuations [54] were also reported.
Understanding fluctuations
We can imagine brain activity as the motion of a random
walker, making steps of a length taken from some
distribution, at times taken from some other distribution
or, equivalently, of a macroscopic particle diffusing in a
liquid, subject to viscous friction and to an additive
random force [55].
from a Gaussian distribution
The relationship between these two forces' time scales
determines how microscopic
fluctuations produce
observable macroscopic properties. In the equilibrium
world of perceptual scientists, the brain makes steps
taken
and whose
correlations, much faster than the friction time scale, have
no macroscopical effect. Reasoning scientists live far from
equilibrium: random fluctuations are no faster than the
friction term and show long-lasting correlations, which
renormalize, becoming macroscopically detectable [56].
Complex fluctuations reveal the particle's navigation
'style', e.g. how travelled distances and the times to reach
a given target scale with time [57]. They also allow
deducing a system's characteristic temporal scales. For an
equilibrium
temporal
autocorrelation decay, the correlation length, i.e. the value
ߦ that makes the autocorrelation ܥሺݐ = ߦሻ = 0, or the
correlation time ߬ = ܥሺݐሻ݀ݐ
, endow the process with a
unique temporal scale. However, at long time scales, the
presence of scaling indicates that the brain relaxes more
system with
exponential
క
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
slowly than an exponential and fluctuates at all scales [58-
60]. Both ߦand ߬ may diverge, and a characteristic time
ceases to exist. Temporal scales are characterized by
some relationship between them, which can be treated as
a dynamical system, relating the behaviour of an observed
variable over a series of nested scales [61,62].
The brain's functional and corresponding dynamical
heterogeneity produces a spatial distribution of time
scales. While only some of these are usually considered of
interest, global temporal scales may not coincide with any
of the local dynamics, and may emerge from transient
connectivity patterns created and destroyed by rewiring
processes [63]. Because spatial scales also have non-
trivial topological properties [64], global dynamics is a
field endowed with arbitrarily real and phase-space
complex topologies [65], where scales replace metric
distances, fractal geometry the Euclidean one, and
scale invariance Galilean invariance.
By resorting to nonlinear analysis, algebraic and
differential topology, renormalization group methods etc.
[66,67], brain activity can then be described in terms of
universal properties, i.e. regimes with robust macroscopic
behaviour with respect to the nature of microscopic
interactions and sharing the same symmetries. These
descriptions partition the phase space, identify dynamical
pathways leading to specific regions of this space, and
allow relating descriptions of the same brain at different
scales, and grouping descriptions of different brains
exhibiting the same large-scale behaviour [66].
Chunking
Correlated noise and cross-scale relationships produce
temporally ordered structures which can help segmenting
reasoning episodes into chunks.
This can be done by defining quasi-stationary
segments boundaries [68]. The scaling properties of
quasi-stationary segments' durations may help clarifying
whether reasoning in its general form is merely a
repetition of simple cycles seen in more controlled forms
of reasoning, or is of a qualitatively different nature, and
in any case, determining the time scales at which simpler
cycles are reedited.
The waiting-time distribution between steps defines
an internal operational time, which may grow sub- or
super-linearly with physical time [69]. Multiplicative
cross-scale interactions bias the waiting-time distribution
so that operational and physical times no longer coincide,
and local probability densities become time-dependent
and intermittent [70].
FROM SPONTANEOUS ACTIVITY TO REASONING
Cognitive processes can be thought of as selections and
orchestrations of cortical states already present
in
spontaneous activity [71,72]. Each process corresponds
to a specific phase space cut, with its own topological
properties and symmetries, and characteristic kinematics,
memory, ageing properties, degree of ergodicity, and
internal clock [33].
Reasoning may modulate not brain activity's
frequency or amplitude but its functional form [33], e.g.
by pushing
it towards the basin of attraction of
advantageous probability distributions: good reasoning
could be tantamount to designing a driving noise function
forcing the system’s stationary distribution to equal a
target one.
Cognitive demands may generally change
the
symmetries of brain activity. For instance, scaling regime
modulations observed in associations with a reasoning
task [73] may correspond to cross-overs between
universality classes, reflecting dynamical transitions in
the system’s behaviour [74].
APPRAISING REASONING
interpreted
fluctuations can be
Brain
in various
interrelated ways that help evaluating the quality of
reasoning, by using models of the function it fulfils and
quantifying
faces while
performing it.
the constraints
the brain
Metaphors for reasoning
search process
[77]. This
in
Reasoning, as other cognitive processes, e.g. memory
recall [75,76], can be represented as a search process
similar to that of animals foraging in an unknown
can be
environment
characterized
terms of random walks [78-81].
Importantly, random walk types can quantify the extent to
which a given trajectory optimizes search, given the
characteristics of the explored space and the resources
available to the individual [81]. Such a characterisation
would allow assessing in a context-specific way the
quality of both the reasoning and the 'reasoner'. That
behavioural aspects of human cognition [75,76] and brain
activity
heavy-tailed
distributions might indicate search optimality [80,82].
However, because these properties are generic
in
spontaneous activity, reasoning's quality can only be
described in terms of its modulations. Furthermore, since
local dynamics may be nowhere Lévy-like, finding the
neural property and spatial scale showing scaling are the
crucial steps.
non-Gaussian,
show
both
The reasoning regime could also be represented as a
network traffic regulation problem, where phenomena
such as overload or jamming may be quantified in terms
of information creation, erasure and transmission rates,
by regarding simple fluctuations as letters of an alphabet
and fluctuation complexes as words, and quantifying the
amount of information in the system. Characterizing
traffic regulation may involve understanding the interplay
between the underlying network's topology, burstiness of
information packets and
fluctuation
distributions [83-85]. Although only causal information
[86] may directly serve reasoning purposes, the total
the shape of
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
information encoded in the network may describe the
noise-control mechanisms indirectly optimizing it.
The sudden onset of insight may be thought of as an
extreme event comparable to earthquakes, financial
crashes or epileptic seizures [87,88], e.g. as a rupture
phenomenon, and the route to it as a long charging
process, with nested hierarchical "earthquakes", and try
predicting its occurrence. For such phenomena, the
coupling strengths distribution and topology constitute
the relevant field [89], which may be described using
complex network theory [64]. It is tempting to conjecture
that insight onset may be predicted by monitoring e.g.
anomalous diffusion parameters [88], variations
in
Gaussianity
fractal spectrum
complexity [91,92].
[90], or changes
in
From dynamics to thermodynamics
statistical aspects of brain
Differences
in reasoning abilities can produce not
only differences in processing times of some orders
of magnitude, but also qualitative ones in dynamical
fluctuations, and
and
topology.
in
A
is
"How efficiently does a given brain carry
then:
out reasoning?".
corresponding brain
natural question when studying
topography and
reasoning
There are various ways to assess efficiency of a given
device. If reasoning was the output of an engine, e.g. a
Carnot's engine, it would be natural to understand not
only how the engine produces it, but also how efficiently
it performs it.
While functional network reconstruction can provide
an indirect characterization of brain activity's energetic
aspects [93,94], the brain's thermodynamics can directly
be deduced from its dynamics [95], which can be
interpreted as the walker's diffusion on the surface of the
entropy production rate in the space of kinetic variables
[96] .
autocorrelation
function
in
Temperature represents a good example of how
thermodynamical variables can be used to describe brain
activity [97]. For a system at equilibrium, temperature is
proportional to the ratio between the response to an
external field conjugate to some observable and the
corresponding
the
unperturbed system [98]. In the brain [99], equilibrium
temperature must be substituted by an effective
temperature
thermometer
responding on the time scale at which the system reverts
to equilibrium would measure [101]. Each scales can have
its own effective temperature even within the same
spatial region, corresponding to different distances from
equilibrium and reflecting qualitatively different diffusion
processes [102]. Thus, measuring effective temperature
at various scales allows understanding the extent to
which each
from
equilibrium, produces entropy, etc..
[100] reflecting what a
spatio-temporal
scale deviates
Thermodynamic functions can be used to directly
describe brain activity, but also as control parameters, i.e.
one can monitor different aspects of brain activity as one
measured thermodynamical variable varies in time. For
instance, one may observe temperature variations
during
a reasoning task, but also possible phase
transitions in some other property of neural activity, as
temperature is varied [97].
FROM THEORY TO EXPERIMENT
OBSERVING REASONING
Reasoning is a difficult phenomenon to observe: tasks can
be executed in more than one way, each possibly
corresponding to a neural phase space with convoluted
geometry and the processes involved in reasoning may
evolve over time-scales exceeding those typical of
laboratory testing.
Proper observation of a given process requires that
the observation time be much larger than any scale in the
system. A process is observable if it has a finite ratio
between the characteristic time of the independent
variable and the length of the available time series [103].
Factors including long-term memory, aging and weak
ergodicity breaking may result in a diverging ratio [104].
The observation time should also be much larger than the
time needed to visit the neural phase space. However,
deciding whether a reasoning episode, or even an
ensemble of episodes, sufficiently sample a subject's
repertoire is non-trivial.
Cognitive neuroscientists observe phenomena through
experiments where subjects typically carry out given
tasks a large number of times, assumed to be independent
realizations of the same observable, and to adequately
sample the phase space of task-related brain activity.
However, in the presence of complex fluctuations, trials
may not self-average, i.e. dispersion would not vanish
even for an infinite number of trials [105].
Furthermore, the time needed to explore this space
may far exceed the typical reasoning episode duration,
and reasoning episodes may explore different aspects of
the space of available strategies. Thus, trials may improve
phase space exploration rather than the signal-to-noise
ratio [106].
EXPERIMENTAL IMPLICATIONS
its
Reasoning's characteristics, particularly
lack of
characteristic temporal duration, have implications at
various levels. First, episodes cannot be compared in an
event-related fashion. Second, defining reliable neural
correlates of reasoning requires defining its characteristic
temporal scales. Third, measures of brain activity should
be invariant with respect to overall duration. Scaling
exponents, data collapse and universality of fluctuations
statistics [107-109], or explicit evolution equations for
the particle's momenta and for the cross-scale fluctuation
probabilities [62] can be retrieved from data and applied
to unevenly lengthen trials. Thermodynamic quantities
such as free energy or temperature can also be estimated
for stochastic trajectories over finite time durations
[97,110-113]. In all cases, the reconstruction of the
Reasoning presents a dilemma between ensuring
complete phase space exploration, which may require
extremely long trials, and signal stationarity, which is
guaranteed only for time scales much shorter than the
reasoning episodes' duration. At fast time scales, the
window in which relevant quantities are calculated
should not introduce spurious time scales, filtering out
genuine ones. Altogether, reasoning's inherently unstable
nature suggests that describing it may boil down to
characterizing non-stationarities and their aetiologies.
Reasoning tasks may be so difficult that only few
participants manage to produce solutions within a
reasonable time. This represents a shortcoming when
trials are considered as independent and identically
distributed, as the signal-to-noise ratio improves with the
square root of the number of trials. Smoothing response
times is a frequent strategy to obviate this problem, but
limits or distorts the reasoning process. Furthermore,
however many, short trials may insufficiently explore the
phase space. Designs with few long trials may express
richer spatiotemporal brain dynamics than many short
ones of equivalent overall length.
Finally, while observed scaling properties may help
understanding whether insight is predictable, i.e. whether
it is an outlier or it is generated by the same distribution
producing anonymous events, predicting insight onset in
real data appears a challenging task, as reasoning
episodes are various orders of magnitude shorter than
earthquake, financial or epilepsy time series [114].
CONCLUSIONS
Reasoning elicits an exceptionally rich repertoire of
otherwise unexpressed neural properties. Its neural
correlates
to
neuroscientists, whom it compels to consider hitherto
neglected brain properties, as they are to psychologists
striving to understand its underlying processes.
as much helpful
therefore
are
tools
(borrowed
Defining general and robust mechanistic properties of
healthy and dysfunctional reasoning will require as yet
non-standard brain metrics, experimental designs, and
analytical
rarely
made available to psychologists). This may shed light
fundamental mechanistic properties of both
on
and dysfunctional reasoning, and ultimately
healthy
help
cognitive
used
and
as
brain
to achieve
desired states [115].
understanding
pharmacological
enhancers and
interventions
them
targeting
actions
fields
from
the
of
ACKNOWLEDGEMENT
author
The
(FIS201238949-C03-01).
acknowledges
1.
2.
3.
4.
5.
6.
7.
8.
9.
Moshman, D. (1995). Reasoning as self-constrained
thinking. Hum. Devel. 38, 53–64.
Gloning, K., and Hoff, H. (1969). Cerebral localization of
disorders of higher nervous activity. In Vincken and Bruyn
(Eds.), Handbook of clinical neurology (Vol. 3, Disorders of
higher nervous activity). New York: Wiley.
Lezak, M.D. (1995). Neuropsychological assessment. 3rd
Edition. Oxford: Oxford University Press.
Goel, V., Gold, B., Kapur, S., and Houle S. (1997). The seats
of reason? An imaging study of deductive and inductive
reasoning. Neuroreport 8, 1305–1310.
Goel, V., Gold, B., Kapur, S., and Houle, S. (1998).
Neuroanatomical correlates of human reasoning. J. Cogn.
Neurosci. 10, 293–302.
Noveck, I.A., Goel, V., and Smith, K.W. (2004). The neural
basis of conditional reasoning with arbitrary content.
Cortex 40, 613–622.
Osherson, D., Perani, D., Cappa, S., Schnur, T., Grassi, F., and
Fazio, F. (1998). Distinct brain foci in deductive versus
probabilistic reasoning. Neuropsychologia 36, 369–376.
Parsons, L.M., and Osherson, D. (2001). New evidence for
distinct right and left brain systems for deductive versus
probabilistic reasoning. Cereb. Cortex 11, 954–965.
Prado, J., Spotorno, N., Koun, E., Hewitt, E., Van der Henst,
J.B., Sperber, D., and Noveck, I.A. (2014). Neural interaction
between logical reasoning and pragmatic processing
in narrative discourse. J. Cogn. Neurosci. 16, 1–13.
10. Prado, J., Chadha, A., and Booth, J.R. (2011). The brain
network for deductive reasoning: a quantitative meta-
analysis of 28 neuroimaging studies. J. Cogn. Neurosci. 23,
3483–3497.
12.
11. Acuna, B.D., Eliassen, J.C., Donoghue, J.P., and Sanes, J.N.
(2002). Frontal and parietal
lobe activation during
transitive inference in humans. Cereb. Cortex 12, 1312–
1321.
Houdé, O., Zago, L., Crivello, F., Moutier, S., Pineau, A.,
Mazoyer, B., and Tzourio- Mazoyer, N. (2001). Access to
deductive logic depends on a right ventromedial prefrontal
area devoted to emotion and feeling: evidence from a
training paradigm. Neuroimage 14, 1486– 1492.
Kroger, J.K., Sabb, F.W., Fales, C.L., Bookheimer, S.Y., Cohen,
M.S., and Holyoak, K.J. (2002). Recruitment of anterior
dorsolateral prefrontal cortex
in human reasoning:
A parametric study of relational complexity. Cereb. Cortex
12, 477–485.
13.
14. Reverberi, C., Bonatti, L.L., Frackowiak, R.S., Paulesu, E.,
Cherubini, P., and Macaluso, E. (2012). Large scale brain
activations predict reasoning profiles. Neuroimage 59,
1752–1764.
P.-M.
(2007).
Baudonnière,
15. Papo, D., Douiri, A., Bouchet, F., Bourzeix, J.-C., Caverni, J.-P.,
and
Time-frequency
intracranial source localization of feedback-related EEG
activity in hypothesis testing. Cereb. Cortex 17, 1314–1322.
Bonnefond, M., Noveck I., Baillet S., Cheylus, A., Delpuech,
C., Bertrand, O., Fourneret, P. and Van Der Henst, J.B.
(2013). What MEG can reveal about reasoning: the case of
if…then sentences. Hum. Brain Mapp. 34, 684–697.
16.
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
underlying dynamics
device's resolution.
improves with
the recording
REFERENCES
the
support
of MINECO
17. Bonnefond, M., Castelain, T., Cheylus, A., and Van der Henst,
J.B. (2014). Reasoning from transitive premises: an EEG
study. Brain Cogn. 90, 100– 108.
18. Bonnefond, M., Kaliuzhna, M., Van der Henst, J.B., and De
Neys, W. (2014). Disabling conditional inferences: an EEG
study. Neuropsychologia 56, 255–262.
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
19. Luo, J., Niki, K., and Phillips, S. (2004). Neural correlates of
20.
21.
22.
23.
24.
the ‘Aha! reaction’. Neuroreport 15, 2013–2017.
Subramaniam, K., Kounios, J., Parrish, T.B., and Jung-
Beeman, M. (2008). A brain mechanism for facilitation of
insight by positive affect. J. Cogn. Neurosci. 21, 415–432.
Bowden, E.M., and Jung-Beeman, M. (2007). Methods for
investigating the neural components of insight. Methods 42,
87–99.
Jung-Beeman, M., Bowden, E.M., Haberman., J., Frymiare,
J.L., Arambel-Liu, S., Greenblatt, R., Reber, P.J., and Kounios,
J. (2004). Neural activity when people solve
verbal
problems with insight. PLoS Biol. 2, E97.
Mai, X-Q., Luo, J., Wu, J-H., and Luo, Y-J. (2004). “Aha!”
Effects in a guessing riddle task: An event-related potential
study. Hum. Brain Mapp. 22, 261–270.
Lang, S., Kanngieser, N., Jaśkowski, P., Haider, H., Rose, M.,
and Verleger, R. (2006). Precursors of insight in event-
related brain potentials. J. Cogn. Neurosci. 18, 2052–2066.
39. Dragoi, G., and Tonegawa, S. (2011 ). Preplay of future
place cell sequences by hippocampal cellular assemblies.
Nature 469, 397–401.
40. Betzel, R.F., Erickson, M.A., Abell, M., O'Donnell, B.F.,
Hetrick, W.P., and Sporns, O. (2012). Synchronization
dynamics and evidence for a repertoire of network states
in resting EEG. Front. Comput. Neurosci. 6, 74.
41. Gong, P., Nikolaev, A.R., and van Leeuwen, C. (2007).
Intermittent dynamics underlying the intrinsic fluctuations
of the collective synchronization patterns in electrocortical
activity. Phys. Rev. E 76, 011904.
42. Allegrini, P., Menicucci, D., Paradisi, P., and Gemignani, A.
(2010). Fractal
spontaneous EEG
metastable-state transitions: new vistas on integrated
neural dynamics. Front. Physio. 1, 128.
complexity
in
43. Bianco, S., Ignaccolo, M., Rider, M.S., Ross, M.J., Winsor, P.,
and Grigolini, P. (2007). Brain, music, and non-Poisson
renewal processes. Phys. Rev. E 75, 061911.
25. Qiu, J., Li, H., Yang, D., Luo, Y., Li, Y., Wu, Z., and Zhang, Q.
(2008). The neural basis of insight problem solving: an
event-related potential study. Brain Cogn. 68, 100–106.
44. Novikov, E., Novikov, A., Shannahoff-Khalsa, D., Schwartz,
B., and Wright, J. (1997). Scale-similar activity in the brain.
Phys. Rev. E 56, R2387–R2389.
27.
26. Kounios, J., Fleck, J.I., Green, D.L., Payne, L., Stevenson, J.L.,
Bowden, E.M., and Jung- Beeman, M. (2008). The origins of
insight in resting-state brain activity. Neuropsychologia 46,
281–291.
Kounios, J., Frymiare, J.L., Bowden, E.M., Fleck, J.I.,
Subramaniam, K., Parrish, T.B., and
Jung-Beeman, M.
(2006). The prepared mind: neural activity prior to
problem presentation predicts subsequent solution by
sudden insight. Psychol. Sci. 17, 882–890.
28. Sheth, B.R., Sandkühler, S., and Bhattacharya, J. (2008).
Posterior beta and anterior gamma oscillations predict
cognitive insight. J. Cogn. Neurosci. 21, 1269–1279.
29. Sandkühler, S., and Bhattacharya, J. (2008). Deconstructing
insight: EEG correlates of insightful problem solving. PLoS
ONE 3, e1459.
30. Knoblich, G., Ohlsson, S., Haider, H., and Rhenius, D. (1999).
Constraint relaxation and chunk decomposition in insight
problem solving. J. Exp. Psychol. Learn. 25, 1534–1556.
31. Stephen, D.G., Boncoddo, R. A., Magnuson, J. S., & Dixon, J. A.
(2009). The dynamics of insight: mathematical discovery
as a phase transition. Mem. Cognition 37, 1132–1149.
32. Rabinovich, M., Huerta, R., and Laurent, G. (2008).
Transient dynamics for neural processing. Science 321,
48–50.
33. Papo
(2014). Functional
complex
in brain activity: from resting state to
significance of
fluctuations
cognitive neuroscience. Front. Syst. Neurosci. 8, 112.
34. Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A., and
Arieli, A.
(2003). Spontaneously emerging cortical
representations of visual attributes. Nature 425, 954–956.
35. Cossart, R., Aronov, D., and Yuste, R. (2003). Attractor
dynamics of network UP states in the neocortex. Nature
423, 283–288.
Ikegaya, Y., Aaron, G., Cossart, R., Aronov, D., Lampl, I.,
Ferster, D. and Yuste, R. (2004). Synfire chains and cortical
songs: temporal modules of cortical activity. Science 304,
559–564.
36.
37. Beggs, J.M., and Plenz, D. (2003). Neuronal avalanches in
neocortical circuits. J. Neurosci. 23, 11167–11177.
38. Beggs, J.M., and 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.
45. Linkenkaer-Hansen, K., Nikouline, V.V., Palva, J.M., and
Ilmoniemi, R. (2001). Long-range temporal correlations
and scaling behavior in human oscillations. J. Neurosci. 15,
1370–1377.
46. Gong, P., Nikolaev, A.R., and van Leeuwen, C. (2002). Scale-
invariant fluctuations of the dynamical synchronization in
human brain electrical activity. Neurosci. Lett. 336, 33–36.
47. Stam, C.J., and de Bruin, E.A. (2004). Scale-free dynamics of
global functional connectivity in the human brain. Hum.
Brain Mapp. 22, 97–109.
48. Freeman, W.J., Holmes, M.D., Burke, B.C., and Vanhatalo, S.
(2003). Spatial spectra of scalp EEG and EMG from awake
humans. Clin. Neurophysiol. 114, 1053–1068.
49. van de Ville, D., Britz, J., and Michel, C.M. (2010). EEG
microstate sequences in healthy humans at rest reveal
scale-free dynamics. Proc. Natl. Acad. Sci. USA 107, 18179–
18184.
50. Expert, P., Lambiotte, R., Chialvo, D.R., Christensen, K.,
Jensen, H.J., Sharp, D.J., and Turkheimer, F. (2010). Self-
similar correlation
resting-state
functional magnetic resonance imaging. J. R. Soc. Interface
8, 472–479.
in brain
function
52.
51. Fraiman, D., and Chialvo D.R. (2012). What kind of noise is
brain noise: anomalous scaling behavior of the resting
brain activity fluctuations. Front. Physiol. 3, 307.
Suckling, J., Wink, A.M., Bernard, F.A., Barnes, A., and
Bullmore, E. (2009). Endogenous multifractal brain
dynamics are modulated by age, cholinergic blockade
and cognitive performance. J. Neurosci. Methods 174, 292–
300.
53. Zilber, N., Ciuciu, P., Abry, P., and van Wassenhove, V.
(2012). Modulation of scale-free properties of brain
activity in MEG. IEEE I. S. Biomed. Imaging (Barcelona),
1531–1534.
54. Freyer, F., Aquino, K., Robinson, P.A., Ritter, P., and
Breakspear, M. (2009). Bistability and Non-Gaussian
fluctuations in spontaneous cortical activity. J. Neurosci. 29,
8512–8524.
55. Hsu, D., and Hsu, M. (2009). Zwanzig-Mori projection
operators and EEG dynamics : Deriving a simple equation
of motion. PMC Biophysics 2, 6.
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
56. Grigolini, P., Rocco, A., and West, B.J. (1999). Fractional
calculus as a macroscopic manifestation of randomness.
Phys. Rev. E 59, 2603–2613.
76. Baronchelli, A., and Radicchi, F. (2013). Lévy flights in
human behaviour and cognition. Chaos Soliton. Fract. 56,
101–105.
57. Papo, D. (2013). Time scales in cognitive neuroscience.
Front. Physio. 4, 86.
58. Fairhall, A.L., Lewen, G.D., Bialek, W., and de Ruyter van
Steveninck, R. (2001). Multiple timescales of adaptation in
a neural
in Neural Information
Processing Systems 13. ( Leen TK, Dietterich TG, Tresp V.
Eds.), MIT Press, pp. 124–130.
In Advances
code.
59. Gilboa, G., Chen, R., and Brenner N. (2005). History-
dependent multiple-timescale dynamics in a single-neuron
model. J. Neurosci. 25, 6479–6489.
60. Lundstrom, B.N., Higgs, M.H., Spain, W.J., and Fairhall, A.L.
neocortical
(2008).
pyramidal neurons. Nat. Neurosci. 11, 1335–13342.
differentiation
Fractional
by
61. Bacry, E., Delour, J., and Muzy, J.F. (2001). Multifractal
random walk. Phys. Rev. E 64, 026103.
62. Friedrich, R., Peinke, J., Sahimi, M., and Reza Rahimi Tabar,
M. (2011). Approaching complexity by stochastic methods:
from biological systems to turbulence. Phys. Rep. 506, 87–
162.
63. Bianco, S., Geneston, E., Grigolini, P., and Ignaccolo, M.
(2007). Renewal aging as emerging property of phase
synchronization. Physica A 387, 1387–1392.
77. Viswanathan, G.M., da Luz, M.G.E., Raposo, E.P., and Stanley,
H.E. (2011). The physics of foraging: an introduction to
random searches and biological encounters. Cambridge
University Press, Cambridge.
78. Shlesinger, M., Zaslavsky, G., and Klafter, J. (1993). Strange
kinetics. Nature 363, 31–37.
79. Codling, E.A., Plank, M.J., and Benhamou, S. (2008).
Random walk models in biology. J. R. Soc. Interface 5, 813-
834.
in
82.
intermittent
search processes
80. Lomholt, M., Tal, K., Metzler, R., and Joseph, K. (2008). Lévy
strategies
are
advantageous. Proc. Natl. Acad. Sci. USA 105, 11055–11059.
81. Bénichou, O., Loverdo, C., Moreau, M., and Voituriez, R.
(2011). Intermittent search strategies. Rev. Mod. Phys. 83,
81.
Humphries, N.E., Weimerskirch, H., Queiroz, N., Southall,
E.J., and Sims, D.W. (2012). Foraging success of biological
Lévy flights recorded in situ. Proc. Natl. Acad. Sci. USA 109,
7169–7174.
DeDeo, S., and Krakauer, D.C.
and
Soc. Interface 9, 2131–2144.
(2012). Dynamics
in finite self-similar networks. J. R.
processing
83.
64. Bullmore, E.T., and Sporns, O. (2009). Complex brain
networks: graph theoretical analysis of structural and
functional systems. Nat. Rev. Neurosci. 10, 186–198.
84. Delvenne, J-C., Lambiotte, R., and Rocha, L.E.C. (2013).
Bottlenecks, burstiness, and fat tails regulate mixing times
of non-Poissonian random walks. arXiv:1309.4155.
65. Zaslavsky, GM. (2002). Chaos, fractional kinetics, and
anomalous transport. Phys. Rep. 371, 461–580.
66. Lesne, A. (2008). Regularization, renormalization, and
renormalization groups: relationships and epistemological
aspects. In. Vision of Oneness. (I. Licata and A. Sakaji Eds),
Aracne, Roma, pp. 121–154.
67. Petri, G., Expert, P., Turkheimer, F., Carhart-Harris, R., Nutt,
D., Hellyer, J., and Vaccarino, F. (2014). Homological
scaffolds of brain functional networks. J. R. Soc. Interface
11, 20140873.
68. Kaplan, A.Y., Fingelkurts, A.A., Fingelkurts, A.A., Borisov,
B.S., and Darkhovsky, B.S. (2005). Nonstationary nature of
the brain activity as revealed by EEG/MEG: methodological,
practical and conceptual challenges. Signal Process. 85,
2190– 2212.
Sokolov, I.M., and Klafter, J. (2005). From diffusion to
anomalous diffusion: a century after Einstein’s Brownian
motion. Chaos 15, 026103.
69.
70. Crisanti, A., and Ritort, F. (2003). Violation of the
fluctuation–dissipation theorem in glassy systems: basic
notions and the numerical evidence. J. Phys. A-Math. Gen.
36, R181–R290.
71. Fiser, J., Chiu, C., and Weliky, M. (2004). Small modulation
of ongoing cortical dynamics by sensory input during
natural vision. Nature 431, 573–578.
72. Luczak, A., Barthó, P., and Harris, K.D. (2009). Spontaneous
events outline the realm of possible sensory responses in
neocortical populations. Neuron 62, 413–425.
73. Buiatti, M., Papo, D., Baudonnière, P.M., and van Vreeswijk,
C. (2007). Feedback modulates the temporal scale-free
dynamics of brain electrical activity in a hypothesis testing
task. Neuroscience 146, 1400–1412.
74. Burov, S., and Barkai, E. (2008). Critical exponent of the
fractional Langevin equation. Phys. Rev. Lett. 100, 070601.
75. Rhodes, T., and Turvey, M.T. (2007). Human memory
retrieval as Lévy foraging. Physica A 385, 255–260.
85. Lambiotte, R., Tabourier, L., and Delvenne, J-C. (2013).
Burstiness and spreading on temporal networks. Eur. Phys.
J. B 86, 320.
86. Shalizi, C.R., and Moore, C. (2003). What is a macrostate?
objective dynamics.
and
Subjective
arXiv:cond-mat/0303625v1.
observations
87. Osorio, I., Frei, MG, Sornette, D, Milton, J, and Lai YC.
(2010 ). Epileptic seizures: quakes of the brain? Phys. Rev. E
82, 021919.
88. Contoyiannis Y.F., and Eftaxias, K.A. (2008 ). Tsallis and
Levy statistics in the preparation of an earthquake. Nonlin.
Processes Geophys. 15, 379–388.
89. Sornette, D.,
and Osorio,
I.
(2010 ). Prediction.
arXiv:1007.2420v1.
90. Manshour, P., Saberi, S., Sahimi, M. Peinke, J., Pacheco, A.F.,
and Rahimi Tabar, M.R. (2009). Turbulencelike behavior
of seismic time series. Phys. Rev. Lett. 102, 014101.
91. de Arcangelis, L., and Herrmann, H.J. (1989). Scaling and
multiscaling laws in random fuse networks. Phys. Rev. B 39,
2678.
92. Kapiris, P.G., Eftaxias, K.A., and Chelidze, T.L. (2004).
Electromagnetic signature of prefracture criticality in
heterogeneous media. Phys. Rev. Lett. 92, 065702.
93. Bullmore, E.T., and Sporns, O. (2012). The economy of
brain network organization. Nat. Rev. Neurosci. 13, 336-
349.
94. Papo, D., Zanin, M., Pineda, J.A., Boccaletti, S., & Buldú, J.M.
(2014). Brain networks: great expectations, hard times,
and the big leap forward. Phil. Trans. R. Soc. B 369,
20130525.
95. Sekimoto,
and
thermodynamics. Prog. Theor. Phys. Supplement 130, 17–
27.
Langevin
equation
(1998).
K.
96. Kirkaldy, J.S. (1964 ). Thermodynamics of the human brain.
Biophys. J. 5, 981–986.
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neural correlates of reasoning
l correlates of reasoning
Papo Characterizing the neura
Papo Characterizing the neura
l correlates of reasoning
97.
Papo, D. (2013). Brain temperature: what it means and
what
neuroscientists.
arXiv:1310.2906v1.
(cognitive)
can do
for
it
98. Kubo, R. (1966). The fluctuation-dissipation theorem. Rep.
99.
Progr. Phys. 29, 255–284.
Papo, D. (2013). Why should cognitive neuroscientists
study the brain's resting state? Front. Hum. Neurosci. 7, 45.
doi: 10.3389/fnhum.2013.00045
100. Cugliandolo, L. F. (2011). The effective temperature. J. Phys.
A Math. Theor. 44, 483001.
101. Kurchan, J. (2005). In and out of equilibrium. Nature 433,
222–225.
102. Papo, D. (2014). Measuring brain temperature without a
thermometer. Front. Physio. 5, 124.
103. Reiner, M. (1964). The Deborah number. Phys. Today 17,
62.
104. Rebenshtok, A., and Barkai, E. (2007). Distribution of time-
averaged observables for weak ergodicity breaking. Phys.
Rev. Lett. 99, 210601.
105. Aharony, A., and Harris, A.B. (1996). Absence of self-
averaging and universal fluctuations in random systems
near critical points. Phys. Rev. Lett. 77, 3700–3703.
106. Ghosh, A., Rho, Y., McIntosh, A.R., Kötter, R., and Jirsa, V.K.
(2007). Noise during rest enables the exploration of the
brain's dynamic repertoire. PLoS Comput. Biol. 4,
e1000196.
107. Bramwell, S.T., Holdsworth, P.C.W., and Pinton, J.-F. (1998).
in turbulence and
Universality of rare
critical phenomena. Nature 396, 552–554.
fluctuations
108. Bhattacharya, J. (2009). Increase of universality in human
brain during mental imagery from visual perception. PLoS
One 4, e4121.
109. Friedman, N., Ito, S., Brinkman, B.A., Shimono, M., Deville,
R.E., Dahmen, K.A., Beggs, J.M., and Butler, T.C. (2012).
Universal critical dynamics in high resolution neuronal
avalanche data. Phys. Rev. Lett. 108, 208102.
Thermodynamic
110. Ruelle,
formalism.
(1978).
D.
Addison Wesley Publ. Co., Reading, MA.
111. Beck, C., and Schlögl, F. (1997). Thermodynamics of chaotic
systems: an introduction. Cambridge University press,
Cambridge.
112. Canessa, E. (2000). Multifractality in time series. J. Phys.
A: Math. Gen. 33, 3637–3651.
113. Olemskoi, A., and Kokhan, S. (2006). Effective temperature
of self-similar time series: analytical and numerical
developments. Physica A 360, 37–58.
114. Sornette, D. (2002). Predictability of catastrophic events:
Material rupture, earthquakes,
financial
crashes, and human birth. Proc. Natl. Acad. Sci. USA 99,
2522–2529.
turbulence,
115. Gutiérrez R, Sendiña-Nadal I, Zanin M, Papo D, Boccaletti S.
2012 Targeting the dynamics of complex networks. Sci.
Rep. 2, 396.
|
1705.04739 | 1 | 1705 | 2017-05-12T20:02:11 | Face recognition assessments used in the study of super-recognisers | [
"q-bio.NC"
] | The purpose of this paper is to provide a brief overview of nine assessments of face processing skills. These tests have been used commonly in recent years to gauge the skills of perspective 'super-recognisers' with respect to the general population. In the literature, a person has been considered to be a 'super-recogniser' based on superior scores on one or more of these tests (cf., Noyes, Phillips & O'Toole, in press). The paper provides a supplement to a recent review of super-recognisers aimed at readers who are unfamiliar with these tests. That review provides a complete summary of the super-recongiser literature to date (2017). It also provides a theory and a set of action points directed at answering the question "What is a super-recogniser?" | q-bio.NC | q-bio | FACE RECOGNITION ASSESSMENTS USED IN
THE STUDY OF SUPER-RECOGNISERS
Eilidh Noyes*,1 and Alice J. O'Toole1
1 School of Behavioural and Brain Sciences, The University of Texas at
Dallas, Richardson, USA
The purpose of this paper is to provide a brief overview of nine assessments of
face processing skills. These tests have been used commonly in recent years to
gauge the skills of perspective 'super-recognisers' with respect to the general
population. In the literature, a person has been considered to be a 'super-
recogniser' based on superior scores on one or more of these tests (cf., Noyes,
Phillips, & O'Toole, in press). The paper provides a supplement to a recent review
of super-recognisers aimed at readers who are unfamiliar with these tests. That
review provides a complete summary of the super-recoginser literature to date
(2017). It also provides a theory and a set of action points directed at answering
the question "What is a super-recogniser?"
Before They Were Famous Test (BTWFT)
The Before They Were Famous Task (BTWFT) was designed to test accuracy
in recognising famous faces across age-related changes. The task uses 56 images
of famous identities, taken at a time before each celebrity was famous (usually a
photograph of the celebrity that had been taken during childhood). Participants are
asked to identify each of the faces. A correct identification is accepted if the
participant correctly names the face, or if they provide a correct uniquely
identifying fact for the celebrity. Results for this task are calculated in terms of the
percentage of correct responses.
The Cambridge Face Memory Task Long Form (CFMT Long-Form)
(Russell, Duchaine, & Nakayama, 2009)
The CFMT long-form is an extension of the CFMT (Duchaine &
Nakayama, 2006), a task originally used to distinguish prosopagnosics from
controls. The original CFMT task requires participants to learn six unfamiliar
male faces by viewing them from different viewpoints during a learning phase.
Participants are tested subsequently in three different test conditions, each
* Corresponding Author address
Email: [email protected]
2
Eilidh Noyes and Alice J. O'Toole
progressively more difficult than the previous one. First, participants are tested on
their ability to recognise the exact image that they learned of each identity. Next,
they are tested on whether they can recognise the faces in new images. In the third
test, they must recognise the identities in novel images to which noise has been
added. Across the three tests, there are a total of 72 recognition trials.
After testing just one super-recogniser on the task, it was clear to Russell
et al. (2009) that the CFMT was not suitable to test super-recognisers, due to
ceiling levels of performance. Therefore, a fourth section was added to create the
CFMT long-form. This section consisted of very challenging trials-the test
images of previously learned faces varied drastically from the images that were
learned. The images used at training were all of good quality, and had been
cropped to include only the face. Test images in the fourth section were not
cropped and had added extra noise (more than in section three), or were profile
images with added extra noise. To make the task even more challenging,
distractor test images (image of faces that had not been learned) were repeated
more often than any of the previous conditions. This increased familiarity with the
distractor items made them harder to distinguish from other known (i.e.,
previously learned) identities. Comparison of performance (percentage accuracy)
across the sections of the test confirmed that section four of CFMT long form was
the most difficult of the sections (Russell et al., 2009).
Cambridge Face Perception Test (CFPT) (Duchaine, Germine, & Nakayama,
2007)
The CFPT was designed to assess face perception ability – i.e., identification
of faces without a memory component. Participants are presented with a target
face image (in ¾ view) and frontal images of morphs between the target and six
other faces. These images were created by morphing the target face with the faces
of six other individuals. The morph proportion between the individual and target
faces retains differing degrees of similarity (e.g., 88%, 76%, 64%, 52% 28%) to
the target face. Participants rank the six faces from one to six, in order of
similarity with the target face. The correct rank order would place the face with
the greatest proportion of the target image as most similar to the target face,
followed by the image with the next greatest percentage of the target face and so
on.
Participants are tested with both upright and inverted images (8 trial types of
each of these conditions). This test design makes it possible to measure face
matching through similarity ratings. The inclusion of upside-down faces also
allows for testing of an inversion effect.
What is a Super-Recogniser?
3
The Glasgow Face Matching Task (short version) (Burton, White, & McNeill
2010)
The Glasgow Face Matching Task (GFMT) (short version) is a
standardised test of unfamiliar face matching with well-established general
population norms. The task consists of the most difficult 40 trials from the
Glasgow Face Matching Task (Burton, White, & McNeill, 2010). When
completing the GFMT short form, participants view face image pairs and are
tasked with deciding whether the images in each pair are two different images of
the same identity or are instead images of two different people (see Figure 1).
Results for this test are calculated in terms of the percentage of correct responses
made.
Figure 1. Example of a same- (top) and different-identity (bottom) trial from the
GFMT. Figure image from Robertson et al. (2016).
Models Face Matching Test (MFMT) (Dowsett & Burton, 2014)
The Models Face Matching Test (MFMT) is similar in design to the GFMT,
but consists of more challenging face stimuli. The stimuli in the models face
matching test are images of models, with each image captured from a different
photoshoot. These images capture large differences in appearance across multiple
images of the same identity. In the MFMT participants view face image pairs and
make same or different identity decisions for each image pair. As in the GFMT,
this test uses percentage accuracy measures of performance.
4
Eilidh Noyes and Alice J. O'Toole
The Pixelated Lookalike Test (Noyes & Jenkins, 2017)
The Pixelated Lookalike Test (PLT) is a challenging face matching task
designed to assess the effect of familiarity on face matching performance for
individuals of very similar appearance presented in poor quality images (Noyes &
Jenkins, 2017). The task emulates that of a standard face matching test, with
participants making same or different identity judgements to image pairs. Match
trials in this task consist of two different images of a celebrity, whereas mismatch
trials (different identity trials) consist of a celebrity image alongside an image of a
professional celebrity 'lookalike' for that celebrity. This means that the people in
the mismatch trials appear very similar, making this task challenging. The task is
made even more difficult by presenting the images in pixelated form (see Figure
2).
Figure 2. Example of the type image pairs used in the PLT. Image pair on the left
shows different identities (with the imposter image on the right) and the pair on
the left shows two images of the same identity. Figure from Robertson et al.
(2016).
Pixelation helps to mirror the image conditions that may be encountered in
poor quality surveillance footage or digital
imaging devices. Following
completion of the face matching task, participants are measured on their
familiarity with each of the celebrities in the task. This familiarity measure allows
for performance accuracy to be broken down according to personal levels of
familiarity with each face. Therefore, accuracy for each individual is known for
faces that are completely unfamiliar, somewhat familiar, and extremely familiar.
A percentage of correct responses is thus obtained for each participant at each
familiarity band.
What is a Super-Recogniser?
5
The Navon Test (Navon, 1977)
The Navon test is used to provide a measure of processing bias for global
compared to local image features (Navon, 1977). In the test, participants are
presented with stimuli taking the appearance of an alphabetical letter, but the
shape of this letter is formed by the configuration of many small letters. For
example, participants may see a large 'H' on the screen but a close look at the 'H'
may reveal that it is built up from many small 'T's. Participants are asked to
report the large letter (in half of the trials) and the small letter (in the other half of
the trials). Calculation of response error allows detection of bias to global or local
features.
Composite Face Test (Young, Hellawell & Hay, 1987)
The composite face test is a test of holistic processing (Young, Hellawell, &
Hay, 1987). Participants view composite and non-composite face images, in
random order, and are asked to make identity judgments for either the top or
bottom portion of the face (block dependant). A composite image consists of the
top half of one face combined with the bottom half of another face, aligned into
the configuration of a normal face. A non-composite image again consists of the
top half of one face combined with the bottom half of another face, but misaligned
horizontally. Naming accuracy is recorded and compared for the top section of the
face presented in composite and non-composite form, and the bottom section of
the face in composite and non-composite presentation format. Mean reaction time
for accurate identification is generally lower for the non-composite image
condition. This is explained by holistic processing-two identity halves presented
together are 'combined' and perceived as a new identity.
Famous Face Recognition Test (Lander et al. 2001)
In the Famous Face Recognition Test participants are presented with 30
celebrity face images, as well as 10 unfamiliar faces (half of all images were male,
the other female). Each of these images are presented one at a time and are viewed
in either moving or static form (half of the images viewed in each of these
conditions). The participants' task is to identify each face, by naming or providing
a unique identifying fact about the face. If the face was an unknown (unfamiliar)
face, the correct answer is to state that the face is unknown. After the task
participant familiarity with the familiar faces is verified by providing them with a
list of names of the faces that they had seen and asking whether they should have
recognised the face of that celebrity. Percentage accuracy of response is
calculated.
6
Eilidh Noyes and Alice J. O'Toole
REFERENCES
Burton, a M., White, D., & McNeill, A. (2010). The Glasgow Face Matching Test.
Behavior Research Methods, 42, 286–91.
Dowsett, A. J., & Burton, a. M. (2014). Unfamiliar face matching: Pairs out-perform
individuals and provide a route to training. British Journal of Psychology, 106,
433–445.
Duchaine, B., & Nakayama, K. (2006). The Cambridge Face Memory Test: Results
for neurologically intact individuals and an investigation of its validity using
inverted face stimuli and prosopagnosic participants. Neuropsychologia, 44,
576–585.
Noyes, E., & Jenkins, R. (2017). Improving Performance on a Difficult Face
Matching Task. Perception, 46, 226.
Robertson, D. J., Noyes, E., Dowsett, A., Jenkins, R., & Burton, A. M. (2016). Face
recognition by Metropolitan Police Super-recognisers. Plos One, 11, e0150036.
Russell, R., Duchaine, B., & Nakayama, K. (2009). Super-recognizers: people with
extraordinary face recognition ability. Psychonomic Bulletin & Review, 16,
252–257.
Young, A., Hellawell, D., & Hay, D. (1987). Configural information in face
perception. Perception, 16, 747–759.
|
1910.05271 | 1 | 1910 | 2019-10-08T19:33:49 | A Test for Shared Patterns in Cross-modal Brain Activation Analysis | [
"q-bio.NC",
"cs.LG",
"stat.ML"
] | Determining the extent to which different cognitive modalities (understood here as the set of cognitive processes underlying the elaboration of a stimulus by the brain) rely on overlapping neural representations is a fundamental issue in cognitive neuroscience. In the last decade, the identification of shared activity patterns has been mostly framed as a supervised learning problem. For instance, a classifier is trained to discriminate categories (e.g. faces vs. houses) in modality I (e.g. perception) and tested on the same categories in modality II (e.g. imagery). This type of analysis is often referred to as cross-modal decoding. In this paper we take a different approach and instead formulate the problem of assessing shared patterns across modalities within the framework of statistical hypothesis testing. We propose both an appropriate test statistic and a scheme based on permutation testing to compute the significance of this test while making only minimal distributional assumption. We denote this test cross-modal permutation test (CMPT). We also provide empirical evidence on synthetic datasets that our approach has greater statistical power than the cross-modal decoding method while maintaining low Type I errors (rejecting a true null hypothesis). We compare both approaches on an fMRI dataset with three different cognitive modalities (perception, imagery, visual search). Finally, we show how CMPT can be combined with Searchlight analysis to explore spatial distribution of shared activity patterns. | q-bio.NC | q-bio |
A Test for Shared Patterns
in Cross-modal Brain Activation Analysis
Elena Kalinina1,2,*, †, ‡, Fabian Pedregosa3, †, **, Vittorio Iacovella2, Emanuele Olivetti1,2,
and Paolo Avesani1,2
1NeuroInformatics Laboratory (NILab), Bruno Kessler Foundation, Trento, Italy
2Centro Interdipartimentale Mente e Cervello (CIMeC), University of Trento, Italy
3Department of Electrical Engineering, UC Berkeley, California
*e [email protected]
†These authors contributed equally to the paper.
‡Work done while PhD student at NeuroInformatics Laboratory (NILab) and Centro Interdipartimentale Mente e
Cervello (CIMeC).
** Work done while post-doctoral researcher at the Department of Electrical Engineering, UC Berkeley.
ABSTRACT
Determining the extent to which different cognitive modalities (understood here as the set of cognitive processes underlying
the elaboration of a stimulus by the brain) rely on overlapping neural representations is a fundamental issue in cognitive
neuroscience. In the last decade, the identification of shared activity patterns has been mostly framed as a supervised learning
problem. For instance, a classifier is trained to discriminate categories (e.g. faces vs. houses) in modality I (e.g. perception)
and tested on the same categories in modality II (e.g. imagery). This type of analysis is often referred to as cross-modal
decoding. In this paper we take a different approach and instead formulate the problem of assessing shared patterns across
modalities within the framework of statistical hypothesis testing. We propose both an appropriate test statistic and a scheme
based on permutation testing to compute the significance of this test while making only minimal distributional assumption.
We denote this test cross-modal permutation test (CMPT). We also provide empirical evidence on synthetic datasets that our
approach has greater statistical power than the cross-modal decoding method while maintaining low Type I errors (rejecting a
true null hypothesis). We compare both approaches on an fMRI dataset with three different cognitive modalities (perception,
imagery, visual search). Finally, we show how CMPT can be combined with Searchlight analysis to explore spatial distribution
of shared activity patterns.
Introduction
Functional MRI recordings enable the investigation of activation patterns that characterize the working brain. The main goal is
to detect whether the neural pattern of a region of interest correlates with a cognitive task, like, for example, object category
identification. Such investigations are usually focused on a specific cognitive modality, i.e.: visual perception, real auditory,
visual imagery, auditory imagery, etc. A qualitative discrimination task can be designed to extract relevant information from
patterns activated by two (or more) stimuli categories (like body and car) in one of these cognitive modalities.
Identification of activation patterns that are shared across modalities has been the subject of numerous neurocognitive
studies, with such modalities as mental calculations1, sensory/motor stimulation2 or words and picture-viewing3, to name a few.
The most common approach here is to cast the problem of identifying common activation patterns as a supervised learning, or
brain decoding, problem4, 5. Nastase and colleagues9 point out that successful classification in this setting allows to conclude
that neural patterns elicited by relevant cognitive factors in one modality generalize accross to the patterns in the other modality.
In other words, a low misclassification error on a modality which is different from the one used to train the classifier
provides empirical evidence that a given region of interest is involved in the cognitive task encoded in both modalities. In
the literature, such approach is referred to as cross-modal decoding analysis (CMDA). Statistical significance of its result can
be assessed using a t-test on the accuracy obtained by the classifier1, 6, or by a permutation test based on computing the null
distribution2, 7, 8 of such statistic. However, CMDA suffers from a number of practical issues. As we are going to discuss in
section Methods, the accuracy of a decoding model is often low in a cross-modal setting, which most probably implies an
exaggerated amount of Type II errors (failure to reject the null hypothesis).
Neuroscientific investigations into the activity patterns in the fMRI data studies can be formulated as "confirmatory" or
"exploratory" analysis. Confirmatory analysis is centered on a pre-established region of interest (ROI). The CMDA method
presented above is often used for the confirmatory approach, when it is run on the data coming from a predefined ROI (or a set
of predefined ROIs). Exploratory analysis aims at localization of areas containing information about the presented stimuli.
Here, CMDA is employed in conjunction with Searchlight technique to explore the spatial structure of cross-modal activations9.
The outcome of the Searchlight procedure are maps, where each voxel is assigned some quantitative measure of information
that the voxel contains about the stimulus. The procedure of obtaining the maps consists in applying decoding sphere by sphere
on time series extracted from the sphere voxels, and most commonly used information measure to produce Searchlight maps is
classification accuracy. Maps are first calculated individually for each subject and then pooled together to run group analysis,
where the significance of the obtained values is typically established running t-tests voxelwise with respect to chance level.
Classification accuracy and t-tests have been subject to numerous criticisms with regard to their role in Searchlight analysis23.
Besides, use of Searchlight for cross-modal analysis faces interpretation challenges introduced by the asymmetries both in
accuracies and p-values when training and testing on data coming from two different cognitive modalities9.
In this work we develop a permutation test for the investigation of shared patterns across modalities that we denote
cross-modal permutation test (CMPT). This test builds on a long tradition of randomization inference in the statistics literature,
which can be traced back to the first half of the 20th century10 -- 12. Permutation tests have recently seen renewed interest in
neuroimaging13 -- 16 thanks to their minimal distributional assumptions and the availability of cheap computational resources. We
provide empirical evidence on synthetic datasets that this method reduces Type II errors (failure to reject a false null hypothesis)
while maintaining Type I errors comparable (incorrect rejection of the null hypothesis) with respect to CMDA. Our results
highlight particular advantages of CMPT in the small sample/high dimensional regime, a setting of practical importance in
neuroimaging studies. Next, we compare CMPT and CMDA on an fMRI study of three cognitive modalities: visual attention,
imagery and perception. We conduct confirmatory analysis comparing the performance of CMPT and CMDA when identifying
the presence of shared patterns within a functionally defined Region of Interest (ROI). Finally, we present the results of an
exploratory analysis with Searchlight making use of the proposed CMPT test for the information based mapping to explore the
presence of common patterns between different modalities at the whole brain level. The use of CMPT allows to overcome major
methodological drawbacks that had been pointed out for Searchlight in the literature23.
Methods
Cross-modal permutation test (CMPT)
In this section we describe a statistical test for cross-modal activation pattern analysis that we denote CMPT. We formulate the
problem of assessing cross-modal activation as a hypothesis testing problem and propose an inference procedure for this test
based on a permutation schema.
Setting. We assume that the experimental task consists of two modalities (e.g., auditory and visual perception) and each
image in the dataset containing an activation pattern has an associated condition A or B (e.g., two stimuli categories like human
body and car). In total, we observe n activation patterns for each modality, corresponding to the number of conditions in the
experiment, where each activation image is a mean image (averaged by the number of trials in the experiment) representative
of one condition, and the goal is to decide whether there is a common condition effect across the different modalities. Let
us formalize this in the language of statistical hypothesis testing. Consider the set of pairs Z(X,Y ) = {(X1,Y1), . . . , (Xn,Yn)}
sampled iid from some unknown probability distribution P, where X = {X1, . . . ,Xn} (resp. Y = {Y1, . . . ,Yn}) are the activation
patterns corresponding to the first (resp. second) modality, and where the experimental paradigm is designed such that the Xi
and Yi are associated with the same condition (but different modality). Since the image pairs belong to the same condition, as
long as there is a condition effect shared across modality, the sequences X and Y cannot be independent. We hence formulate
the null hypothesis (which we want to reject) that both sequences are independent and so their joint probability distribution P
factorizes over their marginal:
H0 : P = PX × PY , with PX, PY the marginal distribution of P.
(1)
Test statistic. Given the set of image pairs X = {X1, . . . ,Xn} and Y = {Y1, . . . ,Yn} described in the previous paragraph, let
A (resp. B) be the set of indices for the A (resp. B) category. We define XA = 1A ∑a∈A Xa (resp. XB = 1B ∑b∈B Xb) as
the average of activation patterns in X with index in A (resp. B). YA (resp. YB) are defined in similar way as the average
of activation patterns in Y with index in A (resp. B). Note that the index set is computed from images of X (and not Y ) on
both cases. This asymmetry will be useful when designing the permutation scheme. Consider also that we have access to a
similarity measure between images that we denote by ρ. For simplicity we will initially suppose that this measure is the Pearson
correlation coefficient, although we will see later that this can be generalized to any similarity measure between images.
We now have all necessary ingredients to present the test statistic that we propose to distinguish the null hypothesis from
2/17
the alternative. This test statistic has values in [−1,1] and is defined as
(cid:16)
(cid:124)
T (X,Y ) =
1
4
ρ(XA ,YA ) + ρ(XB,YB)
− (ρ(XA ,YB) + ρ(XB,YA ))
within-condition similarity
between-condition similarity
(cid:123)(cid:122)
(cid:125)
(cid:124)
(cid:123)(cid:122)
(cid:125)
(cid:17)
.
(2)
At first, its form might seem strange. Let us give two intuitions on the form of this test statistic:
1. As a difference of similarities. The test statistic can be split as a difference of two terms. The first term is the sum of
similarities for images from the same condition (and different modalities), while the second term is a sum of similarities
for images of different conditions (and different modalities). Hence, large values of the test statistic are achieved whenever
the within-condition similarity is larger than the between-condition similarity, bringing evidence for the existence of a
condition-specific activation across modalities.
2. As a singularity test. If we compute all pairwise similarities between the images XA , XB, YA and YB, we obtain 4 scalars
that can be arranged in a 2-by-2 matrix as follows:
(cid:20)ρ(XA ,YA ) ρ(XA ,YB)
(cid:21)
ρ(XB,YA ) ρ(XB,YB)
.
(3)
Under the null hypothesis, the samples X and Y are independent and so YA ≈ YB (recall that the indexing was derived
from images in X). Whenever YA = YB the matrix above becomes colinear. A standard way to test for colinearity
is through its determinant. Computing the determinant of the above equation we obtain our test statistic (modulo the
normalizing factor 1
4).
Statistical inference. We will estimate the distribution of this test statistic under the null hypothesis from the sample by
repeatedly computing the test statistic over a permuted version of the initial sample, a technique often known as permutation or
randomization test. For this to be valid, it is necessary to identify the quantities that we wish to permute and verify that under
the null hypothesis, all permutations yield the same sample distribution17.
Consider the sequence Xπ which results from a random reordering of the activation images in X and the sequence of pairs
Z(Xπ ,Y ). Under the null hypothesis, since the probability distribution factorizes over its marginal, the permuted sequence is
distributed as P(cid:48)
X is the distribution of Xπ. Now, by the iid assumption made previously (which is commonplace
in the context of permutation testing), this distribution is invariant to permutations, and so P(cid:48)
X = PX and the condition is verified.
After computing the permuted test statistic for a large number of random permutations (typically around 10000), the significance
of this test, i.e., the probability of observing a test statistic equal or as large as the one obtained, can be computed as
X × PY , where P(cid:48)
p =
number of times{T (Xπ ,Y ) ≥ T (X,Y )}
number of permutations
.
(4)
Extensions. This test extends naturally to the setting of group analysis. In this case, the test statistic (2) can be taken as the
sum over all subjects of the subject-specific test statistic. Ideally, the same permutation should be used across subjects to obtain
each value of the permuted test statistic18. It is theoretically possible to perform a two-tailed test using this test statistic. A large
negative value of the test statistic would also bring evidence to reject the null hypothesis of independence. However, since the
neuroscientific interpretation of such negative values is not useful for our practical purpose, we will only use the one tailed test
in this paper.
For simplicity, we have considered ρ the Pearson correlation coefficient as similarity measure, but the method remains valid
using any other similarity measures. The Pearson correlation, being a measure of the linear correlation, works best when the
effect is (close to) linear, but other more complex similarities can be used such as a (negative) Malahanobis19 or Wasserstein20
distance.
Relationship with cross-modal decoding analysis (CMDA). CMDA can be regarded within the same hypothesis testing
framework outlined before, but with a different test statistic. In CMDA, the test statistic is the accuracy of a classifier on images
from one modality when it was trained on images from the other modality.
Since both CMDA and CMPT follow the same permutation test approach to computing significance, both rely implicitly
on a label exchangeability assumption behind the data-generating process. As we have seen in the previous subsection, a
sufficient condition for this is to assume that the data we observe is sampled iid. Note that this iid assumption is on the pairs
from different modalities (Xi,Yi) and also on the experimental paradigm but not on the decoding train/test split, which divides
the data by modality and is obviously not iid. This is a much weaker assumption than the distributional assumptions made by
3/17
traditional parametric methods, it is important to keep in mind that permutation tests are not fully assumption-free methods and
at the bare minimum require exchangeability of the observations.
A practical difference between both approaches is that the cross-modal permutation test is symmetric with respect to
modalities while brain decoding is not. That is, CMPT would yield the same p-value regardless of the order in which the
different modalities are labeled. This is not true for CMDA, where two possible tests can be performed (train on A and test on B
or train on B and test on A), and both can (and typically do) yield different p-values.
Datasets
Synthetic datasets
We construct a synthetic dataset according to a model in which the signal is a superposition of a modality-specific effect
(MX ,XY ), a condition-specific effect (Ci) and a Gaussian noise (εi):
Xi = α ·Ci + β · MX + εi,Yi = α ·Ci + β · MY + εi
where α,β are scalars that regulate the amount of modality-specific and condition-specific signal in the image, respectively. We
then generated a total of 20 different images according to this model, considering two different modality-specific signals and
two different condition-specific signals, all of them randomly generated from a Gaussian distribution. We generate 3 versions
of this dataset, one with 10 voxels, one with 100 and another one with 1000 voxels.
fMRI dataset
We performed empirical analysis of the data coming from a neurocognitive study of visual attention. fMRI data were collected
to investigate object categorization during preparatory activity in a visual search experiment, designed in a similar manner to
the one illustrated in21.
Participants. 24 participants (8 male, mean age 27.1, st.dev. 4.3 years) were recruited and accessed the research facility.
All participantssubjects, before starting the experiment, signed a form confirming their informed consent to participate in
the experimental study. After the experiment, they received monetary compensation. Each participant was instructed in
advance about undergoing 2 experimental sessions (S1, S2) on two different days. The data on all three modalities in question
(perception, imagery and visual search) were acquired in the same session, S1. Out of all 24, 22 completed a significant part
(6/22) or the whole (16/22) of S1. During both S1 and S2, participants were also given other tasks, which we do not report here.
Out of 16 participants that underwent the whole S1 only nine participants completed 4 runs of both perception/imagery task
and visual search task (8 functional runs in total). Other participants failed to reach 4 runs at least in one of the task types. So,
for the analysis we are using the data of the 9 participants that have the total of 8 functional runs each. The tasks are explained
in the next section.
Stimuli. Two distinctive stimulus categories were presented to the participants throughout the tasks: people (whole body
image) and cars. Participants were instructed to deal with these categories in three different ways: in perception modality, they
had to attend to 8 presentations of 16 seconds long blocks of different instances of the same category (people, here depicted by
whole body figures with no face, or cars), interspersed with 16 seconds long fixation periods. Participants were equipped with a
two-button box. They were requested to perform a one-back task - i.e., to press a specific button whenever they detected the
same image repeated twice in a row. In imagery modality participants were instructed to close their eyes and mentally visualize
instances of the category, indicated by the letter cue shown at the beginning of the trial. They had to press the response button
whenever they achieved a mental image that was sufficiently detailed, and then they had to switch to mental visualization of
another instance of the same category. At the end of the 16 seconds block, an auditory cue told the participants to open their
eyes and go on with the experiment. Perception and Imagery blocks were randomly presented within the same functional run.
In visual search modality, participants were briefly (450 ms) shown images representing natural scenes (e.g.: crowded places,
urban landscapes, etc ...). They were instructed through a visual cue (letter) to look for instances of one of the two categories
within the scene. After scene presentation, they had 1.6 s to attend to the presentation of a mask and to give a positive or
negative response by pressing a button. Visual search preparatory periods occurring between the presentation of the cue and
the presentation of the scene had different lengths: 2, 4, 6, 8 or 10 seconds. Participants did not know in advance neither the
lengths of preparatory period, nor their order of presentation throughout the task, which was random. Participants also had to
perform a block-designed task, where we alternated presentation of images of intact and scrambled everyday objects, in order
to functionally define an object selective region of interest (ROI) localized in the temporal - occipital cortex (Fig. 1). So, for
perception and imagery the overall number of trials was 32 per modality (8 trials per 4 runs). Each visual search run consisted
of 40 trials, where for each particular type of delay duration there were 8 trials. The total number of visual search trials is 160
(40 per 4 runs), while for each duration there are 32 trials (8 per 4 runs) in the dataset. All experimental procedures had been
4/17
Figure 1
approved by the Ethical Committee of the University of Trento and were carried out in accordance with applicable guidelines
and regulations on safety and ethics.
Images were acquired with a 4T Bruker (https://www.bruker.com/) scanner. For each participant, we
Data acquisition.
started both experimental sessions by acquiring a structural scan using a 3D T1-weighted Magnetization Prepared RApid
Gradient Echo (MPRAGE) sequence (TR/TE = 2700/4.18 ms, flip angle = 7◦, voxel size = 1 mm isotropic, matrix = 256× 224,
176 sagittal slices). Perception / Imagery and Visual Search tasks were performed while acquiring, respectively, 177 and 195
functional scans with the following parameters: (TR/TE = 2000 / 33 ms, flip angle = 73◦, voxel size = 3x3x3 mm, 1 mm slice
spacing, matrix = 64x64, 34 axial slices covering the entire brain), during session 1. Same acquisition parameters were used for
165 scans acquired during Functional Localizer task in session 2.
Preprocessing. For data preprocessing FSL tools were used along with in-house built Python code. In all functional runs
5 initial volumes were discarded as dummy volumes. The skull was removed from both functional and structural images to
extract the brain. Functional images were subsequently corrected for slice timing and motion artifacts. Transformation of the
functional images to standard space was carried out in the following sequence. First structural scans were coregistered to the
mean functional scan of each experimental run. Structural-in-functional-space images were then coregistered to standard (MNI)
space to finally compute affine parameters. To extract task-related effects from functional localizer data, beta maps for both
localizer conditions (Intact vs. Scrambled objects) were computed with linear regression and next fed into contrast analysis
(Intact vs. Scrambled). This analysis resulted in a ROI located in bilateral temporal occipital cortex. We selected one cluster
including 625 voxels from each hemisphere, ending up with a bilateral ROI with 1250 voxels overall. We applied the ROI mask
to functional data coming from Perception / Imagery and Visual Search tasks and we obtained matrices containing time-series
of the ROI voxels. For CMPT analysis, 1250x16 matrices of Perception / Imagery data and 1250x40 matrices of Visual Search
data were considered per run.
Results
For the rest of the paper we will refer to the rejection of the null hypothesis with the traditional significance of 0.05 without
explicitly mentioning this number.
Experiments on synthetic data
In Figure 2 we plot the resulting p-value after performing both CMPT and CMDA on the synthetic dataset described in section
Methods, for varying magnitudes of the condition-specific effect (α) and different image sizes. In the case of decoding, this
p-value was computed as described in27. For α = 0, the dataset has no condition-specific signal and so the test is not expected
to produce a statistically significant result. Indeed, the p-value of CMPT is around 0.5. Note that because of the discreteness of
the test statistic (test set accuracy), the average p-value need not converge towards 0.5 as α goes to zero in the case for CMDA.
As the magnitude of the effect (α) increases, the method that yields a lower p-value has greater statistical power, because it is
able to reject the null hypothesis with a greater probability. We can see in the figure, that in general CMDA p-values are higher,
which translates into a lower probability of rejecting the null hypothesis under this approach and hence higher Type II error.
In Figure 3 (top row) we can see that in the absence of signal (α = 0), the distribution of p-values generated with CMPT(for
6000 repetitions) is relatively flat, showing that the false positive rate (Type I error) for a significance level of β is at the
5/17
Figure 2
Figure 3
6/17
0.00.20.40.60.81.0Magnitude of condition effect ()0.00.10.20.30.40.50.6p-value10 voxels image0.00.20.40.60.81.0Magnitude of condition effect ()0.00.10.20.30.40.50.6100 voxels image0.00.20.40.60.81.0Magnitude of condition effect ()0.00.10.20.30.40.50.61000 voxels imageCMPTDecoding0.05 significance level expected value of β . In the bottom row of that figure, we can see the same experiment for CMDA. In this case because of the
discreteness of the test statistic, the distribution is not completely flat.
From the simulation results (Figure 2) we see that the average p-values yielded by CMPT are always below those of CMDA.
This implies that smaller effects can be detected, and hence, that CMPT has a higher sensitivity than CMDA. Furthermore, this
effect is replicated across images with different number of voxels, highlighting the benefits of CMPT in the high-dimensional
setting, which is of great practical importance in neuroimaging.
Comparison of CMDA and CMPT on fMRI data
In this section we assess the agreement or disagreement in detecting shared activation patterns between CMDA and CMPT. The
similarity of activation maps for the the discrimination of body vs. car categories was computed for the following pairs of
modalities: perception and imagery, imagery and visual search. The visual search modality was investigated more in detail by
first considering all durations of preparation periods put together and then analysing separately different delays (2, 4, 6, 8 and
10s). This analysis was meant to emphasize the issue of small sample size typical for neuroscientific data. The comparison
between CMDA and CMPT took into account additional elements such as the choice of ROI and the type of encoding of
activation maps.
The ROI chosen for the analysis was the Object Selective Cortex (OSC) map shown in Fig. 1. We analysed separately the
performance of methods for the left part of the ROI, right part of the ROI and the whole ROI. Encoding of the activation maps was
of two types: raw BOLD and beta maps. For the raw BOLD encoding, the volumes were selected that corresponded to the peak
of the hemodynamic response function (HRF, as rendered by SPM software - www.fil.ion.ucl.ac.uk/spm/doc/)
convolved with the boxcar function that represented the experimental manipulation. One volume was selected per trial, and for
CMPT the volumes were averaged to produce a single representative volume per subject per condition per modality. For beta
encoding, beta maps were calculated trialwise using linear regression; for CMPT the maps were averaged over trials, too, to
produce a single beta map per subject per condition per modality.
CMDA was performed by training a logistic regression classifier with (cid:96)2 regularization on the trials of one of the modalities.
The regularization parameter was selected according to a nested cross-validation scheme (leave-one-run out). The accuracy was
then estimated on a test set from another modality. The training and test process was replicated for each subject. Then, the
p-value was computed for the group using the permutation scheme described in2. The resulting p-values are reported in Table
1.
CMPT group analysis was carried out as described in section Cross-modal permutation test (CMPT). The similarity distance
between activation maps that we used is the Pearson correlation measure, both for raw BOLD volume and beta maps encoding.
The significance of the proposed test statistics was computed by a permutation scheme with 10.000 iterations to estimate the
null distribution. The resulting p-values are reported in Table 1.
In Table 1 we report only one result for the comparison between CMDA and CMPT related to raw BOLD volume encoding:
the cross-modal analysis between Perception and Imagery. In this case none of the methods detect a meaningful shared
activation pattern. Beta maps encoding on the other hand seems to be a more efficient representation. Chen and colleagues22
demonstrate that using beta values is a way to get rid of intrinsic variabilities of BOLD signal throughout the brain and,
specifically, within a single area. In our case, this means that beta values are more representative of the effect size than raw
BOLD signal changes during task-on periods. For this reason in the presentation of results we only focus on results obtained
with data encoded with beta maps.
The results in Table 1 confirm our expectations about the presence of common patterns between modalities in Object
Selective Cortex. At the same time, CMPT appears to have higher statistical power and sensitivity in revealing these patterns.
The results reported in Table 1 illustrate two main scenarios: both CMDA and CMPT show significant p-values or only CMPT.
In light of the simulation results we may argue, that since CMDA has a higher false error rate (Figure 2), in case of such
disagreements the CMPT result is more reliable. This argument is further supported by the additional empirical evidence that
the false positive rate or Type I error is similar for the two tests, limiting the risk of the disagreement being biased by a more
optimistic rejection of the null hypothesis.
Cross-modal analysis results for perception vs. imagery with beta maps encoding are in agreement between CMDA and
CMPT when the two hemispheres are considered individually, namely the left and right OSC respectively. When the analysis is
extended to the joint ROI, the number of trials remains constant, while the number of voxels double. In this case the classifier is
affected by the higher dimensionality of data, and CMDA does not succeed in rejecting the null hypothesis.
The empirical results also support the claim that CMPT is more robust not only in high-dimensional but also in small sample
setting. Simulations show that the Type II error of CMPT is below that of CMDA in the small sample regime (Fig. 2). We may
find analogous behaviour for the cross-modal analysis of visual search vs. imagery. If we consider the cumulative trials of
visual search, irrespective of the delays, CMDA rejects the null hypothesis. When we restrict the cross-modal analysis to single
delays of preparation period for visual search, the number of trials drop from 160 to 32. In this case CMDA fails to reject the
7/17
Figure 4
8/17
null hypothesis while CMPT does not.
CMPT results are in line with the view that we should expect the presence of shared activity patterns between perceived and
imagined object categories. CMPT analysis also confirms that the presence of these patterns can be expected in high-level visual
areas processing information about object categories. We are going to further elaborate on this point in the discussion section.
On the other hand, CMDA results appear to be affected by the data sample size relative to the high dimensionality of data.
Exploratory data analysis with CMPT
We ran Searchlight analysis of the whole brain, using CMPT. First, we wanted to show if and how inserting CMPT as the
elementary unit within the Searchlight framework could identify voxels that store information related to common activation
patterns for two different cognitive modalities. Next, we intended to compare the spatial profiles of the exploratory analysis
with the ROI individuated for the confirmatory analysis.
To construct group level maps, we referred to the procedure we illustrated in section Cross-modal Permutation Test (CMPT)
at a ROI level. Here we consider the spheres centered on each voxel as ROIs: we first compute the "true" statistic and then
we proceeded by using permutations. We started by computing single-participants' CMPT - Searchlight maps, where each
voxel was considered as the center of a sphere (r= 8), and calculated the T-statistic. Then, we summed up single-participants'
T-values and ended up with the true group-level statistic. Next, we created N=10000 permutations of the session labels, and
we subsequently constructed 10000 averaged beta maps for each of the conditions based on the permuted labels - that is, two
maps per subject per modality per permutation. In this way, we made sure that the data coming from different participants
were tested against the same permutations in a uniform way. We then applied CMPT procedure on these permuted maps by
first computing an individual T-statistic and then by summing up the group values - that is, we constructed an ad-hoc null
distribution. Finally, we simply counted how many times the "true" group statistic was higher than the permuted group statistic,
and we transformed the count in a fraction of the total number of permutations, obtaining a p-value. This value was assigned to
the voxel in the center of the sphere. The procedure was repeated for each voxel within the gray matter mask. We ended up
with a CMPT Searchlight map of p-values coming from a combination of permutation-like tests.
In Figures 4 and 5 we present results coming from the exploratory analysis. Fig. 4 showcases the overlap between the
whole OSC ROI and informative voxels identified by Searchlight in the occipital-temporal cortex for the cross-modal pair of
perception vs. imagery only. In Fig. 5, we put together fragments of maps for the pairs of cognitive processes where our
confirmatory analysis yielded significant results, namely perception vs. imagery, visual search (delay 8s) vs. imagery, visual
search (all delays) vs. imagery (see Table 1 ). For merely illustrative purposes, the maps were thresholded at the conventional
significance level of 0.05 (as we are not aiming at significant cluster identification, no correction for multiple comparisons was
carried out). In the left column of Fig. 5 we demonstrate the overlaps between the ROI identified in the course of the group
analysis (Fig. 1) and the portions of the map that signal the presence of information about common patterns between two
cognitive processes for a single slice (z=-16). The fact that the ROI identified contains a high portion of informative voxels is
further illustrated by the histograms in the right column of the same figure. These are histograms of the p-values of the voxels
within the ROI. We can see that all three histograms have a skewed shape, signalling the presence of a rather large number of
voxels with p-values under 0.05 in the ROI.
For comparison, we also ran CMDA Searchlight with the same sphere size (r=8) for the same modality pairs: perception vs.
imagery, visual search (all delays) vs. imagery, visual search (delay 8) vs. imagery. The analysis was performed with Matlab
8.5.0, MathWorks, NatickMA, USA using in-house code and Libsvm library (https://www.csie.ntu.edu.tw/ cjlin/libsvm/).
Classifier used for producing the maps was an SVM classifier with a linear kernel as implemented in the Libsvm library. For
each subject, two Searchlight maps were obtained for each modality pair, one where the classifier was trained on the Imagery
data and tested on the other modality data, and one where the assignment of train - test data was reversed. Then, these two maps
were averaged as suggested in9 yielding a single map per subject per modality pair. For the group analysis, a one-sample t-test
against chance level (50 %) was performed using SPM software. The resulting group maps were thresholded at the significance
level of 0.05 and cluster size of 10 voxels. Then we compared the group maps to the OSC ROI selected for the confirmatory
analysis. The results are presented in Table 2. It shows percentages of voxels within the OSC ROI that were identified by
the CMDA Searchlight as informative about shared patterns between two modalities. The numbers concerning the size of
intersection between the Searchlight map and the ROI are given both as an absolute number of voxels within the ROI and in
terms of percentages.
The problem of asymmetry between classifier results when swapping train and test modalities in a cross-modal setting is
well attested for the pair of perception vs. imagery. The accuracies obtained with the classifier trained on the imagery data have
been shown to be consistently higher than after training on perception data (28, 33). To minimize the impact of this asymmetry in
cross-modal investigations, it was suggested to average the maps resulting from different train-test combinations for a pair
of modalities9. However, the authors of the paper showed that the divergence in accuracy numbers obtained with different
train-test combinations in their data was insignificant.
9/17
Figure 5
10/17
CMDA Searchlight results in our dataset seem to be rather seriously affected by the issues stemming from the accuracy
asymmetries. First, for the pair of perception vs. imagery, CMDA Searchlight trained on imagery data identifies a high number
of voxels within the OSC mask, both in right and left OSC. If trained on perception data, the Searchlight finds a much lower
number of voxels within the same area, all of them in the right OSC. In the averaged mask, the number of the voxels that
survive is nearly 10 times lower than that identified by the Searchlight trained on imagery data (49 against 432). This same kind
of asymmetry is even more prominent for the pair of imagery vs. visual search: Searchlight, trained on imagery data, identifies
voxels within the OSC ROI, while it does not identify any in the same area if trained with visual search data (neither all delays,
nor delay 8). This result makes us pose a question about the extent to which the voxels idenitified belong actually to really
shared patterns between modalities, or we should rather talk about voxels in one modality that are informative about the patterns
in the other modality. Our overall conclusion about CMDA Searcchlight is that its use might be questionable in cases when
notable asymmetry is expected, as is the case with perception vs. imagery. For some modality pairs asymmetry does not seem
to be a big issue, as is the case with the data used in9, and the use of CMDA Searchlight could be more justified for these data.
Discussion
The patterns of brain activity that are shared between the cognitive processes of perception and imagery have been the subject
of quite numerous studies. The question investigated was if we can arrive at abstract, top-down object representations28, 29
containing distinguishing features30, 31 that will have common neural substrate both for viewed and imagined object categories30.
To test for the presence of shared patterns, many studies used cross-modal decoding - namely, multivariate pattern analysis
with SVM classifiers28, 32, 33. Significant cross-modal classification accuracies were taken as the evidence in favour of the
presence of shared activity patterns. In some studies, correlation-based analysis was also performed to visualize and estimate
similarity between these patterns in terms of distance 28, 32. What emerged from these studies was the view that, indeed, visual
imagery activates the same areas that contain information about visually perceived stimuli 30, 34, and shared patterns for stimulus
categories in these two processes can be established28, 32, 33, 35 -- 37. The areas where these common representations were found
include the ventral temporal pathway, lateral occipital cortex 28, 32, 33, 36 and extrastriate cortex29, 32, 35, 36. The question of shared
patterns in early visual areas, such as V1, remains controversial 28, 32. Horikawa31 showed that it depended on the feature type:
lower visual features had similar representations for perception and imagery in lower visual areas, while the same was true for
higher visual features in higher visual areas. Cichy33 arrived at a similar conclusion about the subdivision of features: although
they did not find significant accuracies for decoding object categories in lateral early visual cortex, they could identify shared
representations of object locations in these areas.
Top-down attention patterns mediate attention biases during perception and affect behavioural performance in attention
related tasks. In case of visual attention, these patterns can be revealed in visual search experiments via activity in the
category-related object selective areas during preparatory delays38. Several studies attempted at demonstrating the high-level
nature of the preparatory patterns through cross-modal analysis, mostly with visual perception as the other modality 21, 38, 39. As
object representations in the brain obtained during imagery tasks are thought to be closer to high-level top-down representations
of objects in visual cortex 30, 35, the hypothesis naturally suggests itself that we can expect these patterns to show up also during
visual search preparatory periods. We tried to shed light on this hypothesis using both CMDA and CMPT on visual imagery and
visual search data. Besides, we ran cross-modal analysis separately for preparatory periods of varying length (between 2 and
10 seconds) to get insights into preparatory dynamics. We were expecting that only certain dealys would result significant,
conforming different hypotheses about this dynamics. For instance, if only shorter delays (2-4 s) had resulted significant, that
could be evidence in favour of transitory and cue-related nature of the preparatory activity in the Object Selective Cortex. If, on
the other hand, we had seen significant results in the longer delays, that could reveal the fact that it takes time for the activity to
build up. First, we see that both methods confirm expectations about imagery patterns being more high level than perception.
None of the methods yielded significant results in the pairs of visual search vs. perception. As for the presence of the shared
patterns between visual imagery and visual search, we are faced again with limitations of CMDA as a method: its results can be
significant and it can reveal the presence of shared patterns between preparatory periods and imagery, but this type of analysis
needs a lot of data. On the other hand, CMPT can reveal shared patterns even with fewer data as is the case with 8 seconds
delay. Further study is needed to uncover the temporal dynamics of the preparatory top-down patterns. We hypothesize that
delays shorter that 8 s do not allow the preparatory activity to build up, while in case of 10 s the delay it is too long, and the
subject might be loosing concentration after a certain period of time.
We placed CMPT side by side with other standard data analysis techniques in order to examine whether this approach
could be as informative as others. We have shown that in confirmatory, top-down contexts CMPT can yield better results than
CMDA. However, it is necessary to mention one limitation of the method. One of the overarching questions in the study of
visual imagery is identifying neural representations of categorical features in the form of brain activation maps30, 35. CMPT
method cannot provide insights into the location of the discriminative patterns at a ROI level. Despite being a more robust test
for cross-modal analysis, CMPT is not appropriate to investigate the shape of shared pattern within a given ROI. In this case,
11/17
CMPT doesn't support a sensitivity analysis at the voxel level needed to compute granular brain maps of activations that are
common between modalities. On the other hand, CMDA (at least when linear classifiers are used) contains a vector of weights
that can give some clues about the relevance of the input features. However, CMPT combined with Searchlight technique can be
a helpful method to locate brain regions that contain information about common patterns between modalities.
We took advantage of one strong point of the Searchlight analysis, its "modular" nature: Searchlight might be thought of as a
generic framework of data examination that can subsume various analysis techniques as elementary units. Searchlight is widely
used in neuroimaging, but it suffers from a number of issues. Conducting Searchlight analysis has several major advantages:
first, it can be run on the whole brain, no prior ROI selection is required. Next, it avoids the "curse of dimensionality" of full
brain classification, by reducing the number of features used at each point by the classifier. Finally, it has proven to be quite
successful in identifying subject specific activation patterns23. The maps produced with Searchlight are of the same nature as
the maps obtained with the univariate GLM approach, but they are based on a more fine-grained pattern identification from
multiple voxels and better reflect the spatial properties of the BOLD signal (that is, adjacent voxels have similar activation
patterns). However, major criticisms of the Searchlight approach regard the use of classification accuracy as the information
measure and the t-test as the method to obtain group significances. As is pointed out in23, SVM classifiers can correctly classify
even with a few number of highly informative voxels and when weakly informative voxels are numerous enough. Both of these
behaviours can cause distortions in a map: in the first case, all searchlights overlapping with one of a few informative voxels
will be significant. In this way, the number of informative voxels is overestimated. In the second case, the cause of distortions is
"discontinuous information detection": groups of weakly informative voxels will be missed out if their size is below a certain
threshold, but can be judged significant if you just add a single voxel. That leads to underestimation of the number of significant
clusters just because the number of weakly significant voxels does not reach a certain mass. Efficiency of using classifiers with
Searchlight depends strongly on the classifier parameters and sphere size23, 24. In24, the point is raised against interpretability
of classifier accuracy with neuroscientific data: unlike distance measures, its value depends on the properties of the dataset
(amount of training data and what kind of data is used as test data) and not only on the presence of a particular effect in the data.
Besides, the authors point out that capturing interactions of several factors in a factorial experimental design cannot be cast as a
classification task. So, addressing these methodological issues for Searchlight can significantly improve this valuable tool and
make its result more scientifically rigorous.
Classification accuracy is not the only way to represent information content. In the original paper by Kriegeskorte 25 the
metrics used was Mahalanobis distance between the distributions corresponding to stimulus categories. In 24 the authors build
on the probabilistic model of the data proposing a cross-validated multivariate ANOVA (MANOVA) as the informational
content measure. In19 three various measures - classification accuracy, Euclidean/Mahalanobis distance, and Pearson correlation
distance - are compared for reliability in the context of Searchlight analysis. In this paper, it was shown that "continuous
crossvalidated distance estimators" such as Euclidean/Mahalanobis distance or Pearson correlation should be preferred for
Searchlight because they are more interpretable from the neuroscientific viewpoint.
Another bunch of critical remarks concerns the use of t-tests for assessing significance at the group level. Certain properties
of neuroscientific data make the use of t-tests questionable for this purpose, "particularly, the low number of observations and
the non-gaussianity of the probability distribution of accuracy. As a consequence, several assumptions of the t-statistic are
not met, rendering the procedure invalid from a theoretical point of view" 26. However, it is not the only option here. In26
a non-parametric test for group significance and cluster inference was proposed based on permutations and bootstrapping
procedure. Nastase and colleagues9 also opt for permutation tests in Searchlight context.
Methodologically, using CMPT in conjunction with the Searchlight technique for cross-modal pattern analysis has several
advantages over the common Searchlight procedure because it does not rely neither on classification accuracy nor on the t-tests
and hence avoids the common methodological pitfalls. At the same time, we are following the suggestions in the literature that
are considered more appropriate for the Searchlight. First, the test statistic proposed in equation 2 that is used as the measure
of information contained at each voxel is based on Pearson correlation and is interpretable in terms of similarity. Second, group
significance is tested non-parametrically with permutation tests that do not make assumptions about the shape of the data
distribution. We found that CMPT integrated into Searchlight has proven effective also to explore and, potentially, confirm what
we observed using top-down, ROI-based analysis, which suggests both robustness and efficiency of the CMPT Searchlight in
fMRI data analysis.
However, it is important to note that confirmatory and exploratory analyses report different p-values. While it is possible
to qualitatively compare the outcomes of these two analyses, plainly putting their p-values side by side might be misleading.
CMPT-ROI p-values come from an extended, functionally well-defined area including 625 or 1250 voxels. CMPT-SL analysis
spans over the whole brain sphere by sphere, extracting results from spheres including about 200 voxels each. This means that
p-values coming from confirmatory and exploratory analysis should not be compared on a purely quantitative level.
The question of shared patterns between various cognitive modalities is relevant not only for object categorization in visual
processing. It is fundamental in the study of interactions between top-down and bottom-up processing streams in the human
12/17
brain in general. Further directions of study could include using the CMPT method and the CMPT Searchlight technique with a
wider number of other cognitive modalities, such as auditory or linguistic40 -- 42. Besides, we could investigate other areas that
can share representations with imagery - for instance, working memory areas35. Finally, the CMPT method could be tried with
other types of neuroimaging data - as, for example, EEG motor imagery data for Brain-Computer interfaces 43 or MEG data
44, 45.
Acknowledgements
The research was partially funded by the Autonomous Province of Trento, Call "Grandi Progetti 2012", project "Characterizing
and improving brain mechanisms of attention - ATTEND" and by Centro Internazionale per la Ricerca Matematica (CIRM).
The authors gratefully acknowledge the contribution of Marius Peelen to the disucssions of experiment design and data analysis
and to Valentina Borghesani for valuable feedback on the manuscript.
Author contributions statement
E.K., V.I. and P.A participated in designing the experiment. E.K. and V.I. implemented the design, conducted the experiment,
acquired the data and preprocessed the data. All authors analysed the data. E.K., F.P. and E.O. implemented data analysis
pipelines. F.P. designed, implemented and conducted experiments on simulated data. E.K., F.P., V.I. and P.A. wrote the
manuscript.
Additional information
The authors declare no competing interests.
Data and code availability.
Processed data (ROI beta-maps, BOLD volumes and Searchlight maps) are available from a public repository on Github
https://github.com/elena-kalinina/Code_CMPT_paper. The code used to generate and analyze synthetic
data is available from the same repository along with data analysis code and Searchlight implementation. Sharing the whole
dataset involves the consent of third-parties who participated in the experiment design. If the consensus is reached, the data
could be made publicly available.
References
1. Knops, A., Thirion, B., Hubbard, E. M., Michel, V. & Dehaene, S. Recruitment of an area involved in eye movements
during mental arithmetic. Sci. 324, 1583 -- 1585 (2009).
2. Etzel, J. A., Gazzola, V. & Keysers, C. Testing simulation theory with cross-modal multivariate classification of fMRI
data. PLoS ONE 3(11), URL http://dx.doi.org/10.1371/journal.pone.0003690. DOI 10.1371/jour-
nal.pone.0003690. (2008).
3. Shinkareva, S. V., Malave, V. L., Mason, R. A., Mitchell, T. M. & Just, M. A. Commonality of neural representations of
words and pictures. NeuroImage 54, 2418 -- 2425 (2011).
4. Haynes, J.-D. & Rees, G. Decoding mental states from brain activity in humans. Nat. Rev. Neurosci. 7, 523 -- 534 (2006).
5. Kaplan, J. T., Man, K. & Greening, S. G. Multivariate cross-classification: applying machine learning techniques to charac-
terize abstraction in neural representations. Front. Hum. Neurosci. 9, 151 URL http://journal.frontiersin.
org/article/10.3389/fnhum.2015.00151. DOI 10.3389/fnhum.2015.00151. (2015).
6. Majerus, S. et al. Cross-modal decoding of neural patterns associated with working memory: evidence for attention-based
accounts of working memory. Cereb. cortex 26, 166 -- 179 (2016).
7. Vetter, P., Smith, F. W. & Muckli, L. Decoding sound and imagery content in early visual cortex. Curr. Biol. 24, 1256 -- 1262
(2014).
8. Kaiser, D., Azzalini, D. C. & Peelen, M. V. Shape-independent object category responses revealed by MEG and fMRI
decoding. J. Neurophysiol. 115(4), 2246 -- 50 (2016).
9. Nastase, S. A., Halchenko, Y. O., Davis, B. & Hasson, U. Cross-modal searchlight classification: methodological challenges
and recommended solutions. In Pattern Recognition in NeuroImaging (PRNI), 2016 International Workshop on, 1 -- 4. IEEE.
URL https://ieeexplore.ieee.org/document/7552355. DOI 10.1109/PRNI.2016.7552355. (2016).
13/17
10. Fisher, R. A. The Design of Experiments. Oliver Boyd, Edinb. (1935).
11. Pitman, E. J. Significance tests which may be applied to samples from any populations. Suppl. to J. Royal Stat. Soc. 4,
119 -- 130 (1937).
12. Lehmann, E. L. & Stein, C. On the theory of some non-parametric hypotheses. The Annals Math. Stat. 20(1), 28 -- 45
(1949).
13. Nichols, T. E. & Holmes, A. P. Nonparametric permutation tests for functional neuroimaging: a primer with examples.
Hum. brain mapping 15, 1 -- 25 (2002).
14. Eklund, A., Nichols, T. E. & Knutsson, H. Cluster failure: why fmri inferences for spatial extent have inflated false-positive
rates. Proc. Natl. Acad. Sci. 113(28), 7900 -- 7905 (2016).
15. Woolrich, M. W., Beckmann, C. F., Nichols, T. E. & Smith, S. M. Statistical analysis of fMRI data. In fMRI Techniques
and Protocols 41, 183 -- 239 Springer-Verlag, New York. (2009).
16. Winkler, A. M., Ridgway, G. R., Douaud, G., Nichols, T. E. & Smith, S. M. Faster permutation inference in brain imaging.
NeuroImage 141, 502 -- 516 (2016).
17. Lehmann, E. L. & Romano, J. P. Testing Statistical Hypotheses. Springer-Verlag New York. (2005).
18. Etzel, J. A. MVPA permutation schemes: permutation testing for the group level. In Pattern Recognition in NeuroImaging
(PRNI), 2015 International Workshop on, 65 -- 68. IEEE. URL https://ieeexplore.ieee.org/document/
7270849. DOI 10.1109/prni.2015.29. (2015)
19. Walther, A. et al. Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage 137, 188 -- 200 (2016).
20. Gramfort, A., Peyr´e, G. & Cuturi, M. Fast optimal transport averaging of neuroimaging data. Preprint at URL http:
//arxiv.org/abs/1503.08596. 1503.08596. (2015).
21. Peelen, M. V., Fei-Fei, L. & Kastner, S. Neural mechanisms of rapid natural scene categorization in human visual cortex.
Nat. 460, 94 -- 97 (2009).
22. Chen, X., Pereira, F., Lee, W., Strother, S. & Mitchell, T. Exploring predictive and reproducible modeling with the
single-subject FIAC dataset. Hum. brain mapping 27, 452 -- 461 (2006).
23. Etzel, J. A., Zacks, J. M. & Braver, T. S. Searchlight analysis: promise, pitfalls, and potential. NeuroImage 78, 261 -- 269
(2013).
24. Allefeld, C. & Haynes, J.-D. Searchlight-based multi-voxel pattern analysis of fMRI by cross-validated MANOVA.
NeuroImage 89, 345 -- 357 (2014).
25. Kriegeskorte, N., Goebel, R. & Bandettini, P. Information-based functional brain mapping. Proc. Natl. Acad. Sci. United
States Am. 103, 3863 -- 3868 (2006).
26. Stelzer, J., Chen, Y. & Turner, R. Statistical inference and multiple testing correction in classification-based multi-voxel
pattern analysis (MVPA): Random permutations and cluster size control. NeuroImage 65, 69 -- 82 (2013).
27. Ojala, Markus, Garriga, Gemma C, Permutation tests for studying classifier performance. J. Mach. Learn. Res. 11,
1833 -- 1863 (2010)
28. Reddy, L., Tsuchiya, N. & Serre, T. Reading the mind's eye: decoding category information during mental imagery.
NeuroImage 50, 818 -- 825 (2010).
29. Ishai, A. Seeing faces and objects with the "mind's eye". Arch. italiennes de biologie 148, 1 -- 9 (2010).
30. Roldan, S. M. Object recognition in mental representations: directions for exploring diagnostic features through visual
mental imagery. Front. Psychol. 8 URL http://dx.doi.org/10.3389/fpsyg.2017.00833. DOI 10.3389/fp-
syg.2017.00833. (2017).
31. Horikawa, T. & Kamitani, Y. Generic decoding of seen and imagined objects using hierarchical visual features. Nat.
Commun. 8 URL http://dx.doi.org/10.1038/ncomms15037. DOI 10.1038/ncomms15037. (2017).
32. Lee, S.-H., Kravitz, D. J. & Baker, C. I. Disentangling visual imagery and perception of real-world objects. NeuroImage
59, 4064 -- 4073 (2012).
33. Cichy, R. M., Heinzle, J. & Haynes, J.-D. Imagery and perception share cortical representations of content and location.
Cereb. Cortex 22, 372 -- 380 (2012).
34. Farah, M. J. The neural basis of mental imagery. Trends neurosciences 12, 395 -- 399 (1989).
14/17
35. Pearson, J., Naselaris, T., Holmes, E. A. & Kosslyn, S. M. Mental imagery: functional mechanisms and clinical applications.
Trends cognitive sciences 19, 590 -- 602 (2015).
36. Stokes, M., Thompson, R., Cusack, R. & Duncan, J. Top-down activation of shape-specific population codes in visual
cortex during mental imagery. The J. neuroscience : official journal Soc. for Neurosci. 29, 1565 -- 1572 (2009).
37. Anderson, A. J. J., Bruni, E., Lopopolo, A., Poesio, M. & Baroni, M. Reading visually embodied meaning from the brain:
Visually grounded computational models decode visual-object mental imagery induced by written text. NeuroImage 120,
309 -- 322 (2015).
38. Stokes, M., Thompson, R., Nobre, A. C. & Duncan, J. Shape-specific preparatory activity mediates attention to targets in
human visual cortex. Proc. Natl. Acad. Sci. 106, 19569 -- 19574 (2009).
39. Peelen, M. V. & Kastner, S. A neural basis for real-world visual search in human occipitotemporal cortex. Proc. Natl.
Acad. Sci. United States Am. 108, 12125 -- 12130 (2011).
40. Simanova, I., Hagoort, P., Oostenveld, R. & van Gerven, M. A. Modality-independent decoding of semantic information
from the human brain. Cereb. cortex 24, 426 -- 434 (2014).
41. Simanova, I., Francken, J. C., de Lange, F. P. & Bekkering, H. Linguistic priors shape categorical perception. Lang. Cogn.
Neurosci. 31, 159 -- 165 (2016).
42. Borghesani, V., Pedregosa, F., Eger, E., Buiatti, M. & Piazza, M. A perceptual-to-conceptual gradient of word coding
along the ventral path. In Pattern Recognition in Neuroimaging, 2014 International Workshop on, 1 -- 4. IEEE. URL
https://ieeexplore.ieee.org/document/6858512. DOI 10.1109/prni.2014.6858512. (2014)
43. Choi, K. Electroencephalography (EEG)-based neurofeedback training for brain-computer interface (BCI). Exp. brain
research 231, 351 -- 365 (2013).
44. Dikker, S. & Pylkkanen, L. Predicting language: MEG evidence for lexical preactivation. Brain language 127, 55 -- 64
(2013).
45. Hirschfeld, G., Zwitserlood, P. & Dobel, C. Effects of language comprehension on visual processing - MEG dissociates
early perceptual and late N400 effects. Brain language 116, 91 -- 96 (2011).
Modalities
P vs I
P vs I
P vs I
P vs I
V vs I (delay 2 s)
V vs I (delay 4 s)
V vs I (delay 6 s)
V vs I (delay 8 s)
V vs I (delay 10 s)
V vs I (all delays)
ROI
L+R
L+R
L
R
L+R
L+R
L+R
L+R
L+R
L+R
Encoding CMDA CMPT
0.75
0.001
0.001
0.008
0.589
0.141
0.303
0.004
0.461
0.001
0.12
0.08
0.001
0.005
0.272
0.387
0.165
0.659
0.249
0.012
Bold
Beta
Beta
Beta
Beta
Beta
Beta
Beta
Beta
Beta
Table 1. Comparison of cross-modal decoding analysis (CMDA) and cross-modal permutation test (CMPT). The values
refer to p-value for the discrimination task of body vs. car in each pair of the cognitive modalities listed in the extreme left
column, where P stands for perception, I for imagery (I) and V for visual search.
15/17
Modalities
P vs I
P vs I
P vs I
V vs I
V vs I
V vs I
P vs I
P vs I
P vs I
V vs I
V vs I
V vs I
P vs I
P vs I
P vs I
V vs I
V vs I
V vs I
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
V vs I (delay 8)
mean
mean
mean
single P I
single I P
single V I
single I V
single I V
single V I
ROI Map Type
L
L
L
L
L
L
L
L
L
R
R
R
R
R
R
R
R
L
single V I
single I V
single I V
single V I
single P I
single I P
mean
mean
mean
mean
mean
single P I
single I P
L+R
L+R
L+R
L+R
L+R single V I
L+R single I V
L+R
L+R single I V
L+R single V I
mean
Intersection N Vox
Intersection %
4
0
239
5
0
12
0
29
0
45
13
193
0
0
1
0
10
0
49
13
432
5
0
13
0
39
0
0.64
0
38.24
0.8
0
1.92
0
4.64
0
7.2
2.08
30.88
0
0
0.16
0
1.6
0
3.92
1.04
34.56
0.4
0
1.04
3.12
0
0
.
Table 2. Percentage of voxels within the OSC ROI identified by the CMDA Searchlight as informative about shared patterns
between two modalities. The cognitive modalities are listed in the extreme left column. P stands for perception, I for imagery
and V for visual search. In the second column, it is shown which part of the OSC mask was used for calculations: left (L),
right (R) or the whole mask (L+R). Number of voxels in L+R ROI is equal to 1250; each half of it (both L and R) has 650
voxels. The third column shows which type of map was used for calculations: an averaged map (denoted as mean), or a single
training - test pair map (denoted as single). In the names of the single maps, first comes the training modality, and second the
test modality. In the last two columns you can find the numbers concerning the size of intersection between the map in column
3 and the ROI in column 2. In column 4, the intersection size is given as an absolute number of voxels within the ROI. In the
last, extreme right column, it is given in terms of percentages.
Figure 1. Object Selective Cortex (OSC) group map in temporal-occipital cortex delineated based on the functional localizer
(intact vs. scrambled objects). Its right (R) hemisphere part is shown in red, left (L) in blue. X, Y and Z locate the coordinates
of the slices.
Figure 2. Statistical power as a function of the effect size. We compare the obtained p-value (y) for different magnitudes of
the condition-specific effect (x) and different image sizes (10, 100 and 1000 voxels). Shaded areas correspond to the standard
deviation. Across these different scenarios, CMPT yields a higher significance than the decoding-based approach.
Figure 3. The distribution of p-values generated with CMPT (top) and CMDA (bottom) in the absense of condition-specific
effect after 6000 repetitions of the experiment. p-values are on the x axis, while the frequencies are shown on the y axis. In case
of CMPT (top) the distribution converges to the uniform. The false positive rate (Type I error) for a significance level of β is at
the expected value of β . For CMDA (bottom), the distribution is not fully flat due to the discreteness of the scoring rule.
Figure 4. Overlap between the whole OSC ROI and informative voxels identified by Searchlight in the occipital-temporal
cortex for the cross-modal pair of perception vs. imagery. The ROI (the same as in Figure fig:roi) is shown in blue. P-values
in the Searchlight map are colorcoded. The maps were thresholded at the conventional significance level of 0.05, and the
color code shows values between 0.95 (corresponding to the p-value of 0.05) in red and 1 (corresponding to the p-value of
0.00) in yellow. The leftmost column of the figure shows sagitarial (top) and coronal (bottom) slices (at x= 42 and y = -74
correspondingly) where the rectangle shaded in blue delimits the axial slices further presented in columns 2 - 4. These are six
16/17
slices of the axial plane taken at steps of 8 between coordinates z = -24 and z = 16. We can see that in all these six slices, there
are overlaps between the OSC ROI and areas where Searchlight identified voxels informative about shared activity patterns
between perception and imagery.
Figure 5. CMPT - Searchlight analysis results. The rows represent results coming from the pairs of cognitive processes
where our confirmatory analysis yielded significant p-values (Table 1): perception vs. imagery (top), visual search (delay of 8 s)
vs. imagery (middle), visual search (all delays) vs. imagery (bottom). In each row, the left column shows the overlap between
the Object Selective Cortex ROI previously defined for confirmatory analysis (Fig. 1) and the Searchlight maps obtained for
this pair. In the right column, there is a histogram of p-values within the OSC ROI coming from the Searchlight map. The left
column showcases axial slices of Searchlight maps for the corresponding pairs, focusing on posterior region. Slices are taken at
z= -16, maps were thresholded at the significance level of 0.05, and the color code shows values between 0.95 (corresponding to
the p-value of 0.05) in red and 1 (corresponding to the p-value of 0.00) in yellow. The OSC ROI is presented in green. Overlaps
between the green area and colorcoded maps can be observed for all three pairs. The right column further illustrates the fact
that the confirmatory OSC ROI identified contains a high portion of informative voxels. In each histogram, the x axis represents
the p-values, while their frequencies (in % from the overall number of voxels in the area) are ordered along the y-axis. All three
histograms are skewed to the right, signalling the presence of a rather large number of voxels with p-values under 0.05 in the
ROI.
17/17
|
1307.5728 | 2 | 1307 | 2013-11-07T10:39:01 | How adaptation currents change threshold, gain and variability of neuronal spiking | [
"q-bio.NC"
] | Many types of neurons exhibit spike rate adaptation, mediated by intrinsic slow $\mathrm{K}^+$-currents, which effectively inhibit neuronal responses. How these adaptation currents change the relationship between in-vivo like fluctuating synaptic input, spike rate output and the spike train statistics, however, is not well understood. In this computational study we show that an adaptation current which primarily depends on the subthreshold membrane voltage changes the neuronal input-output relationship (I-O curve) subtractively, thereby increasing the response threshold. A spike-dependent adaptation current alters the I-O curve divisively, thus reducing the response gain. Both types of adaptation currents naturally increase the mean inter-spike interval (ISI), but they can affect ISI variability in opposite ways. A subthreshold current always causes an increase of variability while a spike-triggered current decreases high variability caused by fluctuation-dominated inputs and increases low variability when the average input is large. The effects on I-O curves match those caused by synaptic inhibition in networks with asynchronous irregular activity, for which we find subtractive and divisive changes caused by external and recurrent inhibition, respectively. Synaptic inhibition, however, always increases the ISI variability. We analytically derive expressions for the I-O curve and ISI variability, which demonstrate the robustness of our results. Furthermore, we show how the biophysical parameters of slow $\mathrm{K}^+$-conductances contribute to the two different types of adaptation currents and find that $\mathrm{Ca}^{2+}$-activated $\mathrm{K}^+$-currents are effectively captured by a simple spike-dependent description, while muscarine-sensitive or $\mathrm{Na}^+$-activated $\mathrm{K}^+$-currents show a dominant subthreshold component. | q-bio.NC | q-bio | How adaptation currents change
threshold, gain and variability of
neuronal spiking
Josef Ladenbauer1,2∗, Moritz Augustin1,2, Klaus
Obermayer1,2
1 Neural Information Processing Group, Technische
Universitat Berlin, Berlin, Germany
2 Bernstein Center for Computational Neuroscience Berlin,
Berlin, Germany
∗Correspondence:
Josef Ladenbauer
Technische Universitat Berlin
Department of Software Engineering and Theoretical
Computer Science
Neural Information Processing
Marchstr. 23, MAR 5049
10587 Berlin, Germany
E-mail: [email protected]
Abstract
Many types of neurons exhibit spike rate adaptation, me-
diated by intrinsic slow K+-currents, which effectively in-
hibit neuronal responses. How these adaptation currents
change the relationship between in-vivo like fluctuating
synaptic input, spike rate output and the spike train statis-
tics, however, is not well understood.
In this computa-
tional study we show that an adaptation current which
primarily depends on the subthreshold membrane volt-
age changes the neuronal input-output relationship (I-
O curve) subtractively, thereby increasing the response
threshold, and decreases its slope (response gain) for low
spike rates. A spike-dependent adaptation current alters
the I-O curve divisively, thus reducing the response gain.
Both types of adaptation currents naturally increase the
mean inter-spike interval (ISI), but they can affect ISI vari-
ability in opposite ways. A subthreshold current always
causes an increase of variability while a spike-triggered
current decreases high variability caused by fluctuation-
dominated inputs and increases low variability when the
average input is large. The effects on I-O curves match
those caused by synaptic inhibition in networks with asyn-
chronous irregular activity, for which we find subtractive
and divisive changes caused by external and recurrent inhi-
bition, respectively. Synaptic inhibition, however, always
increases the ISI variability. We analytically derive expres-
sions for the I-O curve and ISI variability, which demon-
strate the robustness of our results. Furthermore, we show
how the biophysical parameters of slow K+-conductances
contribute to the two different types of adaptation currents
and find that Ca2+-activated K+-currents are effectively
captured by a simple spike-dependent description, while
muscarine-sensitive or Na+-activated K+-currents show a
dominant subthreshold component.
Introduction
Adaptation is a widespread phenomenon in nervous sys-
tems, providing flexibility to function under varying ex-
ternal conditions. At the single neuron level this can be
observed as spike rate adaptation, a gradual decrease in
spiking activity following a sudden increase in stimulus
intensity. This type of intrinsic inhibition, in contrast to
the one caused by synaptic interaction, is typically medi-
ated by slowly decaying somatic K+-currents which accu-
mulate when the membrane voltage increases. A number
of slow K+-currents with different activation characteris-
tics have been identified. Muscarine-sensitive (Brown and
Adams 1980; Adams et al. 1982) or Na+-dependent K+-
channels activate at subthreshold voltage values (Schwindt
et al. 1989; Kim and McCormick 1998), whereas Ca2+-
dependent K+-channels activate at higher, suprathreshold
values (Brown and Griffith 1983; Madison and Nicoll 1984;
Schwindt et al. 1992).
Such adaptation currents,
for example, mediate fre-
quency selectivity of neurons (Fuhrmann et al. 2002;
Benda et al. 2005; Ellis et al. 2007), where the preferred
frequency depends on the current activation type (Dee-
myad et al. 2012). They promote network synchroniza-
tion (Sanchez-Vives and McCormick 2000; Augustin et al.
2013; Ladenbauer et al. 2013) and are likely involved in
the attentional modulation of neuronal response proper-
ties by acetylcholine (Herrero et al. 2008; Soma et al. 2012;
McCormick 1992).
It has been hypothesized that these
complex effects are produced by changing the relationship
between synaptic input and spike rate output (I-O curve)
(Deemyad et al. 2012; Benda and Herz 2003; Soma et al.
2012; Reynolds and Heeger 2009). For example, changing
the I-O curve of a neuron subtractively sharpens stimu-
lus selectivity, whereas a divisive change downscales the
neuronal response but preserves selectivity (see (Wilson
et al. 2012) in the context of synaptic inhibition). It was
also suggested, that adaptation currents affect the neu-
ral code via their effect on the inter-spike interval (ISI)
statistics (Prescott and Sejnowski 2008). So far, effects of
adaptation currents on I-O curves have been studied con-
sidering constant current inputs disregarding input fluc-
tuations (Prescott and Sejnowski 2008; Deemyad et al.
2012) and it has remained unclear how different types of
adaptation currents affect ISI variability. Therefore, in
this contribution we systematically examine how voltage-
dependent subthreshold and spike-dependent adaptation
currents change neuronal I-O curves as well as the ISI dis-
tribution for typical in-vivo like input statistics, and how
the biophysical parameters of slow K+-conductances con-
tribute to the two types of adaptation currents.
We address these questions by studying spike rates and
ISI distributions of model neurons with subthreshold and
spike-triggered adaptation currents, subject to fluctuat-
ing in-vivo like inputs, and we compare the results to
those induced by synaptic inhibition. Specifically, we use
the adaptive exponential integrate-and-fire (aEIF) neuron
model (Brette and Gerstner 2005), which has been shown
to perform well in predicting the subthreshold properties
(Badel et al. 2008) and spiking activity (Jolivet et al.
2008; Pospischil et al. 2011) of cortical neurons. To an-
alytically demonstrate the changes of I-O curves and ISI
variability we derive explicit expressions for these proper-
ties based on the simpler perfect integrate-and-fire neuron
model (see, e.g., (Gerstein and Mandelbrot 1964)) with
adaptation (aPIF). Finally, using a detailed conductance-
based neuron model we quantify the subthreshold and
spike-triggered components of various slow K+-currents
and compare the effects of specific K+-channels on the
3
1
0
2
v
o
N
7
]
.
C
N
o
i
b
-
q
[
2
v
8
2
7
5
.
7
0
3
1
:
v
i
X
r
a
1
I-O curve and ISI variability.
Materials and Methods
aEIF neuron with noisy input current
We consider an aEIF model neuron receiving synaptic in-
put currents. The subthreshold dynamics of the mem-
brane voltage V is given by
C
dV
dt
= Iion(V ) + Isyn(t),
(1)
where the capacitive current through the membrane with
capacitance C equals the sum of ionic currents Iion and
the synaptic current Isyn. Three ionic currents are taken
into account,
Iion(V ) := −gL(V −EL)+gL ∆T exp
−w. (2)
(cid:18) V − VT
(cid:19)
∆T
The first term on the right-hand side describes the leak
current with conductance gL and reversal potential EL.
The exponential term with threshold slope factor ∆T and
effective threshold voltage VT approximates the fast Na+-
current at spike initiation, assuming instantaneous activa-
tion of Na+-channels (Fourcaud-Trocm´e et al. 2003). w is
the adaptation current which reflects a slow K+-current.
It evolves according to
τw
dw
dt
= a(V − Ew) − w,
(3)
with adaptation time constant τw. Its strength depends
on the subthreshold membrane voltage via conductance a.
Ew denotes its reversal potential. When V increases be-
yond VT, a spike is generated due to the exponential term
in eq. (2). The downswing of the spike is not explicitly
modelled, instead, when V reaches a value Vs ≥ VT, the
membrane voltage is reset to a lower value Vr. At the same
time, the adaptation current w is incremented by a value
of b, implementing the mechanism of spike-triggered adap-
tation. Immediately after the reset, V and w are clamped
for a refractory period Tref , and subsequently governed
again by eqs. (1) -- (3).
The aEIF model
range
of neuronal
and
Brette
al.
2008). We selected the following parameter val-
C = 1 µF/cm2,
ues
gL = 0.05 mS/cm2,
∆T = 1.5 mV,
Ew = −80 mV,
VT = −50 mV,
Vs = −40 mV, Vr = −70 mV and Tref = 1.5 ms (Badel
et al. 2008; Destexhe 2009; Wang et al. 2003). The adap-
tation parameters a and b were varied within reasonable
ranges, a ∈ [0, 0.06] mS/cm2, b ∈ [0, 0.3] µA/cm2.
neurons:
EL = −65 mV,
τw = 200 ms,
subthreshold dynamics
patterns
can reproduce a wide
(Touboul
(Naud
et
to model
cortical
2008)
and
spike
The synaptic input consists of a mean µ(t) and a fluc-
tuating part given by a Gaussian white noise process η(t)
with δ-autocorrelation and standard deviation σ(t),
Isyn(t) = C [µ(t) + σ(t)η(t)] .
(4)
Eq. (4) describes the total synaptic current received by
KE excitatory and KI inhibitory neurons, which produce
instantaneous postsynaptic potentials (PSPs) JE > 0 and
JI < 0, respectively. For synaptic events (i.e. presynaptic
2
spike times) generated by independent Poisson processes
with rates rE (t) and rI(t), the infinitesimal moments µ(t)
and σ(t) are expressed as
µ(t) = JE KE rE (t) + JIKIrI(t),
σ(t)2 = J 2E KE rE (t) + J 2IKIrI(t),
(5)
(6)
assuming large numbers KE , KI and small magnitudes of
JE , JI (Tuckwell 1988; Renart et al. 2004; Destexhe and
Rudolph-Lilith 2012). This diffusion approximation well
describes the activity in many cortical areas (Shadlen and
Newsome 1998; Destexhe et al. 2003; Compte et al. 2003;
Maimon and Assad 2009). The parameter values were
JE = 0.15 mV, JI = −0.45 mV, KE = 2000, KI = 500
and rE , rI were varied in [0, 50] Hz.
In addition, we
directly varied µ and σ over a wide range of biologically
plausible values.
Membrane voltage distribution and spike rate
In the following we describe how we obtain the distribu-
tion of the membrane voltage p(V, t) and the instantaneous
spike rate r(t) of a single neuron at time t for a large num-
ber N of independent trials. Note that by trial we refer to
a solution trajectory of the system of stochastic differential
equations eq. (1) -- (4) for a realization of η(t).
¯w(t)
1/N(cid:80)N
First, to reduce computational demands and enable
further analysis, we replace the adaptation current w
in eqs. (2) -- (3) by its average over trials,
:=
i=1 wi(t), where i is the trial index (Gigante et al.
2007a). Neglecting the variance of w across trials is valid
under the assumption that the dynamics of the adapta-
tion current is substantially slower than that of the mem-
brane voltage, which is supported by empirical observa-
tions (Brown and Adams 1980; Sanchez-Vives and Mc-
Cormick 2000; Sanchez-Vives et al. 2000; Stocker 2004).
The instantaneous spike rate at time t can be estimated by
the average number of spikes in a small interval [t, t + ∆t],
(cid:90) t+∆t
N(cid:88)
(cid:88)
i=1
t
k
r∆t(t) :=
1
N ∆t
δ(s − tk
i )ds,
(7)
where δ is the delta function and tk
i denotes the k-th spike
time in trial i. In the limit N → ∞, ∆t → 0, the prob-
ability density p(V, t) obeys the Fokker-Planck equation
(Risken 1996; Tuckwell 1988; Renart et al. 2004),
p(V, t) +
q(V, t) = 0,
(8)
with probability flux q(V, t) given by
∂
∂t
C
(cid:18) Iion(V ; ¯w)
∂
∂V
(cid:19)
q(V, t) :=
+ µ(t)
p(V, t) − σ(t)2
2
∂
∂V
p(V, t).
(9)
Iion(V ; ¯w) denotes the sum of ionic currents (cf. eq. (2))
where w is replaced by the average adaptation current ¯w
which evolves according to
τw
d ¯w
dt
= a((cid:104)V (cid:105)p(V,t) − Ew) − ¯w + τw b r(t).
(10)
(cid:104)·(cid:105)p indicates the average with respect to the probability
density p (Brunel et al. 2003; Gigante et al. 2007b). To
account for the reset of the membrane voltage, the prob-
ability flux at Vs is re-injected at Vr after the refractory
period has passed, i.e.,
q(V, t) − lim
V (cid:37)Vr
lim
V (cid:38)Vr
q(V, t) = q(Vs, t − Tref ).
(11)
The boundary conditions for this system are reflecting for
V → −∞ and absorbing for V = Vs,
V →−∞ q(V, t) = 0,
lim
p(Vs, t) = 0,
(12)
and the (instantaneous) spike rate is obtained by the prob-
ability flux at Vs,
r(t) = q(Vs, t).
(13)
Note that p(V, t) only reflects the proportion of trials
where the neuron is not refractory at time t, given by
P (t) =(cid:82) Vs−∞ p(v, t)dv (< 1 for Tref > 0 and r(t) > 0). The
eq. (10) is calculated as (cid:104)V (cid:105)p(V,t) =(cid:82) Vs−∞ vp(v, t)dv/P (t).
total probability density that the membrane voltage is V
at time t is given by p(V, t) + pref (V, t), with refractory
density pref (V, t) = (1 − P (t)) δ(V − Vr). Since p(V, t)
does not integrate to unity in general, the average in
The dynamics of the average adaptation current ¯w(t)
reflecting the non-refractory proportion of trials is well
captured by eq. (10) as long as Tref is small compared to
τw.
In this (physiologically plausible) case ¯w(t) can be
considered equal to the average adaptation current over
the refractory proportion of trials.
Steady-state spike rate
We consider the membrane voltage distribution of an aEIF
neuron with noisy synaptic input, described by the equa-
tions (8) -- (13), has reached its steady-state p∞. p∞ obeys
∂p∞(V )/∂t = 0 or equivalently,
q∞(V ) = 0,
(14)
∂
∂V
(cid:18) Iion(V ; ¯w)
C
(cid:19)
with steady-state probability flux q∞ given by
q∞(V ) =
+ µ
p∞(V )− σ2
2
∂
∂V
p∞(V ), (15)
subject to the reset condition,
q∞(V ) − lim
V (cid:37)Vr
lim
V (cid:38)Vr
and the boundary conditions,
q∞(V ) = q∞(Vs),
(16)
V →−∞ q∞(V ) = 0
lim
p∞(Vs) = 0
(17)
The steady-state spike rate is given by r∞ = q∞(Vs)
and the steady-state mean adaptation current reads
¯w∞ = a((cid:104)V (cid:105)∞ − Ew) + τwbr∞. We multiply both sides of
eq. (14) by V and integrate over the interval (−∞, Vs], as-
suming that p∞(V ) tends sufficiently quickly toward zero
for V → −∞ (Brunel 2000; Brunel et al. 2003), to obtain
an equation which relates the steady-state spike rate and
mean membrane voltage,
(cid:104)(cid:104)V (cid:105)∞−EL +∆T
(cid:68)
exp
(cid:16) V −VT
(cid:17)(cid:69)
∆T
(cid:105)
µa − gL
r∞ =
∞
/C
,
∆V + τwb/C
(18)
where µa := µ − a((cid:104)V (cid:105)∞−Ew)/C, ∆V := Vs − Vr (here
and in the following) and (cid:104)·(cid:105)∞ denotes the average with
3
A + a
(cid:115)(cid:18)
(cid:19)2
,
respect to the density p∞(V ). The spike rate r∞ is given
by eq. (18) only for nonnegative values of the numerator
(i.e., µa − gL[. . . ]/C ≥ 0); otherwise, r∞ is defined to be
zero. For simplicity, the refractory period Tref is omitted
here. Note, that the steady-state spike rate for Tref (cid:54)= 0
can be calculated as r∞/(1 + r∞Tref ). We cannot express
p∞(V ) explicitly and thus the expressions for the aver-
ages with respect to p∞(V ) in eq. (18) are not known.
However, in the case gL = 0, which simplifies the aEIF
model to the aPIF model, an explicit expression for (cid:104)V (cid:105)∞
can be derived. We multiply eq. (14) by V 2 and integrate
over [−∞, Vs] on both sides (assuming again that p∞(V )
quickly tends to zero for V → −∞) to obtain
(cid:104)V (cid:105)∞ =
1
2a
Vs + Vr
2
−
A − a
Vs + Vr
2
+ B
(19)
where A = µC + aEw and B = 2aσ2C[1 + τwb/(C∆V )].
I-O curve
The I-O curve is specified by the spike rate as a function of
input strength. Here we consider two types of I-O curves:
a time-varying (adapting) I-O curve and the steady-state
I-O curve. In particular, we obtain the adapting I-O curve
as the instantaneous spike rate response to a sustained
input step (with a small baseline input) as a function
of step size. This curve changes (adapts) over time and
it eventually converges to the steady-state I-O curve.
As arguments of these (adapting and steady-state) I-O
functions we consider presynaptic spike rates (Figs. 2C,
4B and eq. (38)), input mean and standard deviation1
(Figs. 2D, 4B and eq. (36)) and input mean for fixed
values of input standard deviation (Fig. 8A).
ISI distribution
We calculate the ISI distribution for an aEIF neuron
which has reached a steady-state spike rate r∞ :=
limt→∞ r(t) by solving the so-called first passage time
problem (Risken 1996; Tuckwell 1988). Consider an ini-
tial condition where the neuron has just emitted a spike
and the refractory period has passed. That is, the mem-
brane voltage is at the reset value Vr and the adap-
tation current, which we have replaced by its trial av-
erage (see above), takes the value ¯w0, where ¯w0 will
be determined self-consistently (see below).
In each
of N (simultaneous) trials, we follow the dynamics of
the neuron given by dVi/dt = [Iion(Vi; ¯w) + Isyn(t)]/C,
its mem-
brane voltage crosses the value Vs and record that spike
time Ti. The set of times Ti + Tref then gives the ISI
distribution. Finally, we determine ¯w0 by imposing that
the mean ISI matches with the known steady-state spike
i=1 Ti + Tref = r−1∞ . According to this
calculation scheme, the ISI distribution can be obtained
in the limit N → ∞ by solving the Fokker-Planck system
eqs. (8) -- (9) with mean adaptation current governed by
d ¯w/dt = [a(1/N(cid:80)N
rate, i.e., 1/N(cid:80)N
i=1 Vi − Ew) − ¯w]/τw, until
τw
d ¯w
dt
= a((cid:104)V (cid:105)p(V,t) − Ew) − ¯w,
(20)
1Note that because of two arguments we obtain a surface instead
of a curve in this case.
subject to the boundary conditions (12) and initial con-
ditions p(V, 0) = δ(V − Vr), ¯w(0) = ¯w0. Note that the
re-injection condition eq. (11) is omitted (see also the dif-
ference between eqs. (10) and (20)) because here each trial
i ends once Vi(t) crosses the value Vs. The ISI distribu-
tion is given by the probability flux at Vs (Tuckwell 1988;
Ostojic 2011), taking into account the refractory period,
pISI(T ) =
q(Vs, T − Tref )
0
for T ≥ Tref
for T < Tref .
(21)
(cid:40)
Finally, ¯w0 is determined self-consistently by requiring
(cid:104)T(cid:105)pISI = r−1∞ . The coefficient of variation (CV) of ISIs
is then calculated as
(cid:113)(cid:104)T 2(cid:105)pISI − (cid:104)T(cid:105)2
pISI
CV :=
(cid:104)T(cid:105)pISI
.
(22)
An ISI CV value of 0 indicates regular, clock-like spiking,
whereas for spike times generated by a Poisson process
the ISI CV assumes a value of 1. For a demonstration
of the ISI calculation scheme described above see Fig. 1.
The results based on the Fokker-Planck equation and
numerical simulations of the aEIF model with fluctuating
input are presented for an increased subthreshold and
spike-triggered adaptation current in separation.
ISI CV for the aPIF model
To calculate the ISI CV we need the first two ISI moments,
cf. eq. (22). The mean ISI for the aPIF neuron model is
simply calculated by the inverse of the steady-state spike
rate, cf. eq. (18), derived in the previous section,
µ. Combining eqs. (22),(23) and (25) the ISI CV reads
(cid:112)σ2∆V /µa − τ 2
CV =
wb2/C 2 − 2τwb∆V /C
.
(26)
∆V + τwb/C
Neuronal network
To investigate the effects of recurrent (inhibitory) synap-
tic inputs on the neuronal response properties (spike rates
and ISIs), we consider a network instead of a single neuron,
consisting of NE excitatory and NI inhibitory aEIF neu-
rons (with separate parameter sets). The two populations
are recurrently coupled in the following way (see Fig. 4A).
Each excitatory neuron receives inputs from K extEE exter-
nal excitatory neurons which produce instantaneous PSPs
of magnitude J extEE with Poisson rate rextEE (t). Analogously,
each inhibitory neuron receives inputs from K extIE exter-
nal excitatory neurons producing instantaneous PSPs of
magnitude J extIE with Poisson rate rextIE (t).
In addition,
each excitatory neuron receives inputs from K recEI randomly
selected inhibitory neurons of the network with synap-
tic strength (i.e., instantaneous PSP magnitude) J recEI and
each inhibitory neuron receives inputs from K recIE randomly
selected excitatory neurons of the network with synaptic
strength J recIE . This network setup was chosen to exam-
ine the effects caused by recurrent inhibition and compare
them to the effects produced by external inhibition for
single neurons described above. To reduce the parameter
space, recurrent connections within the two populations
in the network were therefore omitted. The total synaptic
current for each neuron of the network can be described
using eq. (4), where the parameters µ(t) and σ(t) for ex-
citatory neurons are given by
(cid:104)T(cid:105)pISI = r−1∞ =
∆V + τwb/C
µa
,
(23)
µ(t) = J extEE K extEE rextEE (t) + J recEI K recEI rpopI
σ(t)2 = (J extEE )2K extEE rextEE (t) + (J recEI )2 K recEI rpopI
(t),
(t)
(27)
(28)
where we consider µa > 0 (here and in the following). We
approximate the second ISI moment by solving the first
passage time problem for the Langevin equation
exp(−t/τw) + ση(t),
(24)
dV
dt
= µa − ¯w0
C
with initial membrane voltage Vr and boundary voltage
Vs. That is, we replace (cid:104)V (cid:105)p(V,t) by its steady-state value
(cid:104)V (cid:105)∞ in eq. (20), which is justified by large τw (as already
assumed). The first passage time density (which is equiv-
alent to pISI) and the associated first two moments for this
type of Langevin equation can be calculated as power se-
ries in the limit of small ¯w0 (Urdapilleta 2011). ¯w0 is then
determined self-consistently by imposing eq. (23). Here
we approximate the second ISI moment by using only the
most dominant term of the power series, which yields (the
zeroth order approximation) (Urdapilleta 2011),
(cid:104)T 2(cid:105)pISI =
σ2∆V + µa∆V 2
µ3
a
.
(25)
Including terms of higher order leads to a complicated
expression for (cid:104)T 2(cid:105)pISI which has to be evaluated numer-
ically. We additionally considered the first order term
(not shown) and compared the results of both approxima-
tions (see Results). Effectively, the approximation above,
eq. (25), is valid for small levels of spike-triggered adapta-
tion current and mean input, since ¯w0 increases with b and
4
and for inhibitory neurons,
µ(t) = J extIE K extIE rextIE (t) + J recIE K recIE rpopE
σ(t)2 = (J extIE )2K extIE rextIE (t) + (J recIE )2 K recIE rpopE
(t),
(t)
(29)
(30)
(t) and rpopI
(Brunel 2000; Augustin et al. 2013). rpopE
(t)
are the spike rates of the excitatory and inhibitory
neurons of the network, respectively. Here we consider
large populations of neurons instead of a large number
of trials.
In fact, averaging over a large number of
trials in this setting is equivalent to averaging over large
populations due to the random and sparse connectivity.
In the limit NE , NI → ∞ we obtain a system two coupled
Fokker-Planck equations, one for the excitatory popula-
tion, described by eqs. (8) -- (13),(27),(28), and one for the
inhibitory population, given by eqs. (8) -- (13),(29),(30).
Note that r(t) in eqs. (10) and (13) is replaced by the
spike rates of the excitatory and inhibitory popula-
tions, rpopE
(t), respectively. We solve this
system to obtain the steady-state spike rate for each
population, rpopE,∞ and rpopI,∞. Once these quantities are
known, we calculate the ISI distribution, cf. eq. (21), for
the excitatory population (i.e.
for any neuron of that
population) as described above, using eqs. (27) -- (28) for
the (steady-state) moments of the synaptic current. The
neuron model parameter values were as above for the
single neuron, with a = 0.015 mS/cm2, b = 0.1 µA/cm2
for excitatory neurons and a = b = 0 for inhibitory
(t) and rpopI
neurons, since adaptation was found to be weak in fast-
spiking interneurons compared to pyramidal neurons (La
Camera et al. 2006). The network parameter values were
J extEE = J extIE = 0.15 mV, K extEE = K extIE = 800, constant
rextEE ∈ [0, 80] Hz, J recEI ∈ [−0.75, −0.45] mV, K recEI = 100,
constant rextIE ∈ [6, 14] Hz, J recIE ∈ [0.05, 0.2] mV and
K recIE = 400.
Numerical solution
We treated the Fokker-Planck equations for the aPIF
model analytically. In case of the aEIF model, we solved
these equations forward in time using a first-order finite
volume method on a non-uniform grid with 512 grid
[−200 mV, Vs] and the implicit
points in the interval
Euler integration method with a time step of 0.1 ms for
the temporal domain. For more details on the numerical
solution, we refer to (Augustin et al. 2013).
Detailed conductance-based neuron model
For validation purposes we used a biophysical Hodgkin-
Huxley-type neuron model with different types of slow K+-
currents. The membrane voltage V of this neuron model
obeys the current balance equation
C
dV
dt
= I − IL − INa − IK − ICa − IKs,
(31)
where C = 1 µF/cm2 is the membrane capacitance and
I denotes the injected current. The ionic currents consist
of a leak current, IL = gL(V − EL), a spike-generating
Na+-current, INa = gNa(V )(V − ENa), a delayed recti-
fier K+-current, IK = gK(V )(V − EK), a high-threshold
Ca2+-current, ICa = gCa(V )(V − ECa), and a slow K+-
current IKs. gx denote the conductances of the respec-
tive ion channels and Ex are the reversal potentials. We
separately considered three types of slow K+-current: a
Ca2+-activated current (IKs ≡ IKCa) which is associated
with the slow after-hyperpolarization following a burst of
spikes (Brown and Griffith 1983), a Na+-activated cur-
rent (IKs ≡ IKNa) (Schwindt et al. 1989), and the voltage-
dependent muscarine-sensitive (M-type) current (IKs ≡
IM) (Brown and Adams 1980). The leak current depends
linearly on the membrane potential. All other ionic cur-
rents depend on V in a non-linear way as described by
the Hodgkin-Huxley formalism. We adopted the somatic
model from (Wang et al. 2003) and included the M-current
with dynamics described (for the soma) by (Mainen and
Sejnowski 1996). The conductances underlying the cur-
rents INa, IK, ICa and IM are given by gNa = ¯gNam3∞h,
gK = ¯gKn4, gCa = ¯gCas2∞ and gM = ¯gMu, respectively,
with steady-state gating variables m∞ = αm/(αm + βm),
αm = −0.4(V + 33)/(exp(−(V + 33)/10) − 1), βm =
16 exp(−(V + 58)/12) and s∞ = 1/[1 + exp(−(V + 20)/9)].
The dynamic gating variables x ∈ h, n, u are governed by
dx
dt
= αx(1 − x) − βxx,
(32)
where αh = 0.28 exp(−(V + 50)/10), βh = 4/[1 +
exp(−(V + 20)/10)], αn = −0.04(V + 34)/[exp(−(V +
34)/10) − 1], βn = 0.5 exp(−(V + 44)/25), αu = 3.209 ·
10−4(V + 30)/[1 − exp(−(V + 30)/9)] and βu = −3.209 ·
10−4(V + 30)/[1− exp((V + 30)/9)]. The channel opening
and closing rates αx and βx are specified in ms−1 and the
membrane voltage V in the equations above is replaced by
d[Ca]
its value in mV. The conductance for the Ca2+-activated
slow K+-current IKCa is given by gKCa = ¯gKCa[Ca]/([Ca]+
κ), where the intracellular Ca2+-concentration [Ca] satis-
fies
= −αCaICa − [Ca]
τCa
(33)
with αCa = 6.67 · 10−4 µM cm2/(µA ms), τCa = 240 ms
and κ = 0.03 mM. The conductance for the Na+ --
activated slow K+-current IKNa is described by gKNa =
¯gKNa0.37/(1 + (/[Na])3.5) where = 38.7 mM and the
intracellular Na+-concentration [Na] is governed by
dt
(cid:18) [Na]3
[Na]3 + ϑ3 − γ
(cid:19)
(34)
d[Na]
dt
= −αNa − 3ϕ
with αNa = 0.3 µM cm2/(µA ms), ϕ = 0.6 µM/ms, ϑ =
15 mM and γ = 0.132. We varied the peak conduc-
tances of the three slow K+-currents IKCa, IKNa, IM in
the ranges ¯gKCa ∈ [2, 8] mS/cm2, ¯gKNa ∈ [2, 8] mS/cm2
(Wang et al. 2003) and ¯gM ∈ [0.1, 0.4] mS/cm2 (Mainen
and Sejnowski 1996). The remaining parameter values
were C = 1 µF/cm2, gL = 0.1 mS/cm2, EL = −65 mV,
ENa = 55 mV, EK = −80 mV, ECa = 120 mV (Wang
et al. 2003).
The differences of the slow K+-currents (IKCa, IKNa
and IM) is effectively expressed by their steady-state volt-
age dependence and time constants. Therefore, we fur-
ther considered a range of biologically plausible steady-
state conductance-voltage relationships and timescales us-
ing the generic description of a slow K+-current, IKs =
¯gKs ω(V )(V − EK), with peak conductance ¯gKs and gating
variable ω(V ) given by
= ω∞(V ) − ω,
dω
dt
τω
(35)
where ω∞(V ) = 1/[1 + exp(−(V − α)/β)]. The shape
of the steady-state curve ω∞(V ) was changed by the pa-
rameters α ∈ [−40, −10] mV (half-activation voltage),
β ∈ [6, 12] mV (inverse steepness) and the time con-
stant τω was varied in [100, 300] ms. The model equations
were solved using a second order Runge-Kutta integration
method with a time step of 10 µs.
To examine the effects of slow K+-currents on the I-O
curve and ISI variability for noisy input, we additionally
considered the synaptic current described by eq. (4) for
i.e., we used I ≡ Isyn in
the detailed neuron model,
eq. (31).
Subthreshold and spike-triggered components of
biophysical slow K+-currents
To assess how the relative levels of subthreshold adapta-
tion conductance (parameter a) and spike-triggered adap-
tation current increments (parameter b) in the aEIF model
reflect different types of slow K+-currents, we quantified
their subthreshold and spike-triggered components using
the detailed conductance-based neuron model. First, we
fit the steady-state adaptation current w∞ = a(V − Ew)
from the aEIF model to the respective K+-current IKs
of the Hodgkin-Huxley-type model in steady-state over a
range of subthreshold values for the membrane voltage,
V ∈ [−70, −60] mV. Thereby we obtained an estimate for
a. In the second step, we measured the absolute and rel-
ative change of IKs elicited by one spike. This was done
5
s
s
s
) − IKs(tpre
by injecting a slowly increasing current ramp into the de-
tailed model neuron and measuring IKs just before and
after the first spike that occurred. Specifically, the ab-
solute change of current caused by a spike was given by
∆IKs := IKs(tpost
), where the time points tpre
and tpost
were defined by the times at which the mem-
brane potential crosses a value close to threshold (we chose
−50 mV) during the upswing and downswing of the spike,
respectively. ∆IKs provides an estimate for b. The relative
change of K+-current was ∆I rel
). Here
we only fitted the parameters a and b of the aEIF model.
For an alternative fitting procedure which comprises all
model parameters, we refer to (Brette and Gerstner 2005).
Ks := ∆IKs/IKs(tpre
s
s
Results
Spike rate adaptation, gain and threshold modu-
lation in single neurons
We first examine the responses of single aEIF neurons
with and without an adaptation current, receiving inputs
from stochastically spiking presynaptic excitatory and in-
hibitory neurons. The compound effect of the individual
synaptic inputs is represented by an ongoing fluctuating
input current whose mean and standard deviation depend
on the synaptic strengths and spike rates of the presy-
naptic cells (cf. eqs. (4) -- (6) in Materials and Methods
and Fig. 2A). The neurons naturally respond to a sudden
increase in spike rate of the presynaptic neurons (an in-
put step) with an abrupt increase in spike rate and mean
membrane voltage, see Fig. 2B. Without an adaptation
current, both quantities remain unchanged after that in-
crease. In case of a purely subthreshold adaptation cur-
rent (a > 0, b = 0 in the aEIF model) which is present
already in absence of spiking, the rapid increase of mean
membrane voltage causes the mean adaptation current to
build up slowly, which in turn leads to a gradual decrease
in spike rate and mean membrane voltage. Note that the
mean membrane voltage is decreased in the absence of
spiking (before the increase of input) compared to the neu-
ron without adaptation. In case of a purely spike-triggered
adaptation current (a = 0, b > 0 in the aEIF model), the
sudden increase in spike rate leads to an increase of mean
adaptation current, which again causes the spike rate and
mean membrane voltage to decrease gradually.
The adapting I-O curve of neurons with and without
an adaptation current, that is, the time-varying spike rate
response to a step in presynaptic spike rates as a function
of the step size, is shown in Fig. 2C. Interestingly, the two
types of adaptation current affect the spike rate response
in different ways. A subthreshold adaptation current shifts
the I-O curve subtractively and thus increases the thresh-
old for spiking. In addition, it decreases the response gain
for low (output) spike rates. If the adaptation current is
driven by spikes on the other hand, the I-O curve changes
divisively, that is, the response gain is reduced over the
whole range of spike rate values but the response thresh-
old remains unchanged. It can be recognized that for a
given type of adaptation current the adapting I-O curve
evaluated shortly after the input steps and the steady-
state I-O curve are changed qualitatively in the same way.
Thus, for the following parameter exploration and analyt-
ical derivation we focus on (changes of) the steady-state
I-O relationship.
We next explore the effects of an adaptation current
on the steady-state spike rate for a wide range of input
statistics, that is, different values of the mean µ and the
standard deviation σ of the fluctuating total synaptic in-
put, see Fig. 2D. If excitatory and inhibitory inputs are
approximately balanced, the standard deviation σ of the
compound input is large compared to its mean µ. The
spike rate increases with an increase of either µ or σ, or
both. A subthreshold adaptation current increases the
threshold for spiking in terms of µ as well as σ. A spike-
triggered adaptation current however does not change the
threshold for spiking but reduces the gain of the spike
rate as a function of µ or σ. Thus, the differential effects
of both types of adaptation current are robust across dif-
ferent input configurations. Note that the I-O curve as a
function of mean input µ changes additively for increased
levels of standard deviation σ while its slope (i.e., gain)
decreases, particularly for small values of µ. This can be
recognized by the contour lines in Fig. 2D and is most
prominent for increased subthreshold adaptation. Con-
sequently, this type of adaptation current increases the
sensitivity of the steady-state spike rate to noise intensity
for low spike rates.
In order to analytically demonstrate the differential ef-
fects of subthreshold and spike-triggered adaptation cur-
rents on the (steady-state) I-O curve, we consider the aPIF
neuron model, which is obtained by neglecting the leak
conductance (gL = 0) in the aEIF model. This allows
to derive an explicit expression for the steady-state spike
rate,
µ − a((cid:104)V (cid:105)∞−Ew)/C
r∞ =
,
∆V + τwb/C
(36)
where the mean membrane voltage (cid:104)V (cid:105)∞ with respect to
the steady-state distribution p∞(V ) is given by eq. (19)
and ∆V := Vs − Vr is the difference between spike and
reset voltage; r∞ = 0 for µ < a((cid:104)V (cid:105)∞ − Ew)/C (see
Materials and Methods). Equation (36) mathematically
demonstrates the subtractive component of the effect a
subthreshold adaptation current (a > 0) produces when
the mean membrane voltage is larger than the reversal
potential Ew of the (K+) adaptation current. Taking the
derivative of eq. (36) with respect to µ further reveals
that an increase of a reduces the gain when the input
fluctuations (σ) are large compared to the mean (µ).
A spike-triggered adaptation current (b > 0) however
produces a purely divisive effect which can be pronounced
even for small current increments b if the adaptation
timescale τw is large.
Differential effects of adaptation currents on spik-
ing variability
We next investigate how adaptation currents affect ISIs
for different input statistics. For that reason we calculate
the distribution of times at which the membrane voltage of
an aEIF neuron crosses the threshold Vs for the first time,
which is equivalent to the distribution of ISIs (see Mate-
rials and Methods). These ISI distributions are shown in
Fig. 3A for neurons with different levels of subthreshold or
spike-triggered adaptation and a given input. An increase
of either type of adaptation current (via parameters a and
b) naturally increases the mean ISI. Interestingly, while
subthreshold adaptation leads to ISI distributions with
6
long tails, spike-triggered adaptation causes ISI distribu-
tions with bulky shapes. These differential effects on the
shape of the ISI distribution lead to opposite changes of
the coefficient of variation (CV, cf. eq. (22)) which quan-
tifies the variability of ISIs. An increase of subthreshold
adaptation curent produces an increase of CV, whereas
an increase of spike-triggered adaptation current leads to
a decreased ISI variability. How these effects on the CV of
ISIs depend on the statistics (µ and σ) of the fluctuating
input is shown in Fig. 3B,C. With or without an adap-
tation current, if the mean µ is large, that is, far above
threshold, and the standard deviation σ is comparatively
small, the neuronal dynamics is close to deterministic and
the firing is almost periodic, hence the CV is small. In con-
trast, if µ is close to the threshold and σ is large (enough),
the ISI distribution will be broad as indicated by the large
CV. A subthreshold adaptation current either leads to an
increased CV or leaves the ISI variability unchanged. In
case of a spike-triggered adaptation current the effect on
the CV depends on the input statistics. This type of adap-
tation current causes a decrease of the high ISI variability
in the region (of the µ, σ-plane) where the mean input µ is
small, and an increase of the low ISI variability for larger
values of µ.
We analytically derived an approximation of the ISI CV
for the aPIF model, which emphasizes the opposite effects
of the two types of adaptation current. It is obtained as
CV =
∆V + τwb/C
(37)
(same as eq. (26)), where µa := µ − a[(cid:104)V (cid:105)∞−Ew]/C
is the effective mean input which is again assumed to
be positive and takes into account the counteracting
subthreshold adaptation current.
The steady-state
mean membrane voltage (cid:104)V (cid:105)∞ is given by eq. (19) (see
Materials and Methods). Equation (37) mathematically
demonstrates that an increase of subthreshold adaptation
curent (a > 0) causes an increase of CV as long as (cid:104)V (cid:105)∞
is larger than Ew, that is, the mean membrane voltage
is not too hyperpolarized. An increase of spike-triggered
adaptation current (b > 0) on the other hand leads to a
reduction of ISI variability. Note that this approximation
is only valid for small values of mean input (µ) and adap-
tation current increment (b). It does not account for the
increase of CV caused by spike-triggered adaptation for
large levels µ, cf. Fig. 3C. Both (input dependent) effects
of spike-triggered adaptation on the ISI variability can be
captured by a refined approximation of the CV compared
to eq. (37) (not shown, see Materials and Methods for an
outline), which requires numerical evaluation.
(cid:112)σ2∆V /µa − τ 2
wb2/C 2 − 2τwb∆V /C
model parameters within each population and sparse ran-
dom connectivity (see Materials and Methods). Fig. 4B
shows how the steady-state I-O curve of excitatory neu-
rons, i.e., the spike rate rpopE,∞ as a function of the external
(input) spike rate rextEE , is changed by external excitation
to the inhibitory neurons (via rextIE ) and by the strengths
of the recurrent excitatory and inhibitory synapses (J recIE
and J recEI ), respectively. An increase of external excitation
to the inhibitory population (via rextIE ) changes the I-O
curve subtractively, thus increasing the response thresh-
old, while an increase of recurrent excitation to the in-
hibitory neurons (via J recIE ) has a purely divisive effect,
that is, the gain is reduced. On the other hand, an in-
crease of recurrent inhibition to the excitatory neurons
(via J recEI ) affects the I-O curve in both ways.
We demonstrate these effects analytically for a network
of perfect integrate-and-fire (PIF) model neurons (instead
of aEIF neurons). That is, we disregard the adaptation
current here for simplicity (a = b = 0), since it does not
change the results qualitatively. An explicit expression
for the steady-state spike rate of the excitatory neurons,
rpopE,∞, can be derived using eq. (36) for all the neurons
in the network with mean input µ given by eq. (27) for
excitatory neurons and by eq. (29) for inhibitory neurons.
We solve for rpopE,∞ self-consistently to obtain,
rpopE,∞ =
J extEE K extEE rextEE ∆V + J recEI K recEI J extIE K extIE rextIE
∆V 2 − J recIE K recIE J recEI K recEI
.
(38)
The equation above states that rpopE,∞ is directly pro-
portional to the strength of external excitation to the
excitatory population, negatively proportional to the
strength of external excitation to the inhibitory pop-
ulation (since J recEI < 0) and inversely proportional to
the strength of recurrent excitation, where all propor-
tionalities include an offset.
Eq. (38) clearly shows
that the effect of external excitation to the inhibitory
population is purely subtractive (since J recEI < 0), the
effect of recurrent excitation (to the inhibitory popu-
lation) is purely divisive, and the effect of recurrent
inhibition (to the excitatory population) includes both
components. For comparison, consider a single (non-
adapting) PIF neuron receiving (external) excitatory
and inhibitory input. Using eq. (36) with mean input
µ given by eq. (5), the steady-state spike rate of this
neuron reads r∞ = (JE KE rE + JIKIrI)/∆V . Thus, an
increase of external inhibition affects the I-O curve of
an excitatory neuron in the same way (subtractively)
as an increase of external excitation to the inhibitory
population within a recurrent network as described above.
Differential effects of synaptic inhibition on I-O
curves
Here we examine how synaptic input received from a pop-
ulation of inhibitory neurons affect gain and threshold of
spiking. We consider that the neuron we monitor belongs
to a population of excitatory neurons which are recurrently
coupled to neurons from an inhibitory population, as de-
picted in Fig. 4A: Each neuron of the network receives
excitatory synaptic input from external neurons and ad-
ditional synaptic input from a number of neurons of the
other population. The specific choice of the monitored
excitatory neuron does not matter because of identical
Effects of synaptic inhibition on spiking variability
We next investigate how inhibitory synaptic input changes
the ISI variability of the neurons (from the excitatory pop-
ulation) in the network described above. An increase of
external excitation to the inhibitory neurons (via rextIE ),
and the strengths of the recurrent synapses (J recIE and J recEI )
individually, leads to an increase of the mean ISI and an
increased tail of the ISI distribution, as shown in Fig. 5A.
Furthermore, an increase of rextIE or the magnitude of J recIE
or J recEI , each causes the coefficient of variation of ISIs
(CVpopE
) to increase, see Fig. 5B. Thus, an increase of in-
hibition always leads to an increase of spiking variability.
7
.
An increase of external excitation to the excitatory neu-
rons (via rextEE ), on the other hand, leads to a decrease of
CVpopE
To demonstrate these effects analytically we derived
CVpopE
for a network of PIF model neurons using eqs. (26) --
(28), where we obtained the steady-state spike rate of
the inhibitory neurons, rpopI,∞, analogously to rpopE,∞ (as de-
scribed above). Below, we express CVpopE
as a function of
either rextIE , J recIE or J recEI , and lump together all other fixed
parameters in a number of constants,
(c1rextIE + c2)/(c3 − c4rextIE )
c5J recIE + c6
(c7(J recEI )2 − c8J recEI )/(c9 + c10J recEI ).
CVpopE =
(39)
The constants c1, . . . , c10 in eq. (39) are non-negative func-
tions of the fixed parameters. Clearly, an increase of rextIE
or the magnitudes of J recIE and J recEI each produce an in-
crease of CVpopE
(since J recEI < 0). Considering a single
PIF neuron receiving (external) excitatory and inhibitory
input for comparison, we use eq. (37) with mean µ and
standard deviation σ of the input given by eqs. (5) and
(6), respectively, to express the CV as
(cid:115)
CV =
J 2E KE rE + J 2IKIrI
∆V (JE KE rE + JIKIrI)
.
(40)
The effect of
Note that eq. (40) is only valid for positive mean input
(JE KE rE + JIKIrI > 0). Again, ISI variability increases
with inhibition.
inhibition on spiking
variability can be understood intuitively as follows.
Inhibitory synaptic input reduces the mean total synaptic
input µ and increases its standard deviation σ for the
target neuron (population), which in turn causes an
increase of ISI variability.
Subthreshold and spike-triggered components of
slow K+-currents
Here we examine how the two types of an adaptation cur-
rent in the aEIF model reflect different slow K+-currents
in a detailed conductance-based neuron model. First,
we consider three prominent slow K+-currents: a Ca2+-
activated after-hyperpolarization current (IKCa), a Na+-
activated current (IKNa) and the voltage-dependent M-
current (IM). Fig. 6A shows how the conductances asso-
ciated with these K+-currents depend on the membrane
voltage in the steady state, compared to the steady-state
spike-generating Na+-conductance. The threshold mem-
brane voltage at which a spike is elicited in response to a
slowly increasing input current is primarily determined by
the conductance-voltage relationship for Na+. The thresh-
old value lies in the interval where this curve has a positive
slope (the precise value depends on the peak conductances
of all currents and on the input). The curve gNa,∞(V )
thus indicates the subthreshold and suprathreshold mem-
brane voltage ranges. In the subthreshold voltage range
the conductance gKCa,∞ is almost zero, while the conduc-
tances gKNa,∞ and gM,∞ reach significant values close to
the voltage threshold. Thus, the curves in Fig. 6A indicate
that IKCa is activated by spikes, while IM and particularly
IKNa can be increased in the absence of spiking.
The results of the fitting procedure in Fig. 6B,C show
the absolute and relative amounts of current triggered by a
8
spike versus its subthreshold level quantified by the voltage
independent conductance a. IKCa has a dominant spike-
triggered component as expected, while IKNa shows a very
small increment caused by a spike compared to the sub-
threshold component. IM, on the other hand, shows signif-
icant levels of both components. Note, however, that the
amount of IM elicited by a spike is smaller compared to
the level of IM that can be caused by subthreshold mem-
brane depolarization without spiking (since ∆I rel
Ks < 1 for
IKs ≡ IM, see Fig. 6C).
relationship for K+,
We further considered a range of biologically plausible
slow K+-currents. That is, we varied the steady-state
gKs,∞(V ),
conductance-voltage
within a realistic range, as shown in Fig. 7A, and
quantified the subthreshold and spike-triggered com-
ponents for each of these K+-currents, see Fig. 7B,C.
The value of
subthreshold conductance a naturally
increases with the fraction of K+-conductance present
at subthreshold voltage values. For the quantification
of spike-triggered current increments we also considered
different K+ time constants τω. The absolute value of
current increment ∆IKs decreases with increasing τω
and changes only slightly with changes of the shape
of the conductance-voltage curve gKs,∞(V ) (via the
parameters α, β). However, the current increment caused
by a spike relative to the amount of current already
present in the absence of spiking (∆I rel
Ks) is strongly
determined by gKs(V ). ∆I rel
increases with an in-
Ks
crease of half-activation voltage (parameter α), steepness
(via parameter β) and with decreasing time constant (τω).
Effects of slow K+-currents on I-O curve and ISI
variability
Here we examine how the different types of slow K+-
current affect the I-O curve and spiking variability of
uncoupled conductance-based model neurons subject to
noisy inputs and compare the effects to those caused by
subthreshold and spike-triggered adaptation in aEIF neu-
rons. Without a slow K+-current, the spike rate I-O curve
does not change over time, see Fig. 8A. An increase of IKCa
has a purely divisive effect on the I-O curve while an in-
crease of IM changes this curve in a mostly subtractive and
slightly divisive way. For both types of slow K+-current
the adapting spike rates reach their steady-state values in
less than 500 ms. These effects are consistent with our
results based on the aEIF model, given that IKCa pre-
dominantly depends on spikes and IM includes both, sub-
threhold as well as spike-triggered, components (Fig. 6B).
In case of increased IKNa, on the other hand, the steady-
state I-O curve is significantly altered in both ways (sub-
tractively and divisively), and the spike rates adapts very
slowly, that is, steady-state rates are reached after several
seconds. At first sight, this seems contradictory to the ef-
fect predicted above for subthreshold adaptation, consid-
ering that the amount of IKNa triggered by a spike is small
compared to its subthreshold level. Since the timescale of
IKNa is very large (Fig. 8A and (Wang et al. 2003)) even a
small spike-triggered component leads to a significant divi-
sive change of the steady-state I-O curve, cf. eq. (36). This
divisive effect is caused by K+-current building up slowly
because of small current increments triggered repeatedly
by repetitive spiking and very slow decay between spikes
due to the large timescale of the current.
Considering ISI variability, an increase of IKCa reduces
the CV for small values of mean input µ and increases
the CV for larger values of µ, see Fig. 8B. An increase of
each of the other slow K+-currents, IKNa and IM, leads
to an increase of ISI CV in general. These effects are
consistent with those caused by subthreshold and spike-
triggered adaptation currents in the aEIF model, consid-
ering the subthreshold and spike-triggered components of
IKCa, IKNa and IM, respectively (Fig. 6). Thus, the re-
sults from the detailed conductance-based neuron model
are in agreement with the results based on the adaptive
integrate-and-fire models presented above.
Discussion
In this study, we have systematically examined how adap-
tation currents and synaptic inhibition modulate the
threshold and gain of spiking as well as ISI variability in
response to fluctuating inputs resulting from stochastic
synaptic events. Based on a simple neuron model with
subthreshold and spike-triggered adaptation components
we used analytical and numerical tools to describe spike
rates and ISIs for a wide range of input statistics. We
then measured subthreshold and spike-triggered compo-
nents of different types of slow K+-currents using detailed
conductance-based model neurons and we validated our
(analytical) results from the simple neuron model by nu-
merical simulations of the detailed model.
We have shown that a purely subthreshold voltage-
dependent adaptation current increases the threshold for
spiking and reduces the gain at low spike rates in the pres-
ence of input fluctuations. This type of current produces
a long-tailed ISI distribution and thus leads to an increase
of variability for a broad range of input statistics. A spike-
triggered adaptation current, on the other hand, causes a
divisive change of the I-O curve, thereby reducing the re-
sponse gain but leaving the response threshold unaffected,
irrespective of the input noise intensity. This type of cur-
rent decreases the ISI CV for fluctuation-dominated inputs
but increases the CV when the mean input is strong, i.e.,
it reduces the sensitivity of spiking variability to the mean
input. For comparison, an increase of external inhibition
leads to a subtractive shift of the I-O curve while an in-
crease of recurrent inhibition changes it divisively. The ISI
variability, however, is increased by both types of synaptic
inhibition.
We have further demonstrated that the Ca2+-activated
after-hyperpolarization K+-current is effectively captured
by a simple description based on spike-triggered incre-
ments, while the muscarine-sensitive and Na+-activated
K+-currents, respectively, have dominant subthreshold
components. Despite its small spike-triggered component,
the Na+-dependent K+-current also substantially affects
the neuronal gain, due to its large timescale.
Methodological aspects
Our approach involves the diffusion approximation and
Fokker-Planck equation, both of which have been widely
applied to analyze the spike rates of scalar IF type neurons
in a noisy setting, see e.g. (Amit and Brunel 1997; Brunel
2000; Fourcaud-Trocm´e et al. 2003; Burkitt 2006; Roxin
et al. 2011). Our assumption of separated timescales be-
tween slow adaptation and fast membrane voltage dynam-
ics has also been frequently used in such a setting (Brunel
et al. 2003; La Camera et al. 2004; Gigante et al. 2007b;
Richardson 2009; Augustin et al. 2013). While most of
these previous studies concentrated on spike rate dynam-
ics, here we focused on asynchronous (non-oscillatory)
activity. To examine ISI distributions we extended the
method described previously for scalar IF models, which
is based on the first passage time problem (Tuckwell 1988;
Ostojic 2011), to the aEIF model, accounting for the dy-
namics of the adaptation current between spikes. Further-
more, we analytically derived an expression for the steady-
state spike rate (i.e., steady-state I-O relationship) based
on (Brunel et al. 2003) and an approximation of the ISI CV
using recent results from (Urdapilleta 2011) for the per-
fect IF model with two types of adaptation currents (aPIF
model). The I-O functions we calculated can be used to re-
late (adaptive) spiking neuron models to linear-nonlinear
cascade models, which describe the instantaneous spike
rate of a neuron by applying to the stimulus signal suc-
cessively a linear temporal filter and a static nonlinear
function (Ostojic and Brunel 2011). Such cascade models
have proven valuable for studying how sensory inputs are
mapped to neuronal activity (see, e.g., (Schwartz et al.
2006; Pillow et al. 2008)).
It is worth noting that our approach further allows to
easily calculate the power spectrum P and (normalized)
autocorrelation function A of the neuronal spike train once
the ISI distribution has been obtained, via the relation
P(ω) = A(ω) = r∞Re
,
(41)
(cid:18) 1 + pISI(ω)
1 − pISI(ω)
(cid:19)
to memoryless
(so-called renewal)
where A and pISI denote the Fourier transforms of the
autocorrelation function and ISI distribution,
respec-
tively, see (Gerstner and Kistler 2002). Eq. (41) strictly
applies
stochastic
processes and an adaptation mechanism usually leads
to a violation of this requirement for a model neuron
subject to fluctuating input. Here we have derived a
renewal process (Vi(t), ¯w(t)) from the original non-renewal
process (Vi(t), wi(t)) by averaging the adaptation current
and self-consistently determining its reset value (see
section ISI distribution in Materials and Methods). An
alternative approach that allows for the application of the
above relationship eq. (41) to adapting model neurons
has recently been described in (Naud and Gerstner 2012).
Modulation of spike rate threshold and gain
Purely subtractive and divisive changes of the I-O curve by
subthreshold and spike-triggered adaptation, respectively,
have previously been shown for model neurons considering
constant current inputs but neglecting input fluctuations
(Prescott and Sejnowski 2008; Ladenbauer et al. 2012).
These theoretical results describe the effects shown in re-
cent in-vitro experiments which involved blocking the low-
threshold M current and a Ca2+-activated K+-current sep-
arately (Deemyad et al. 2012) (Fig. 3); see also (Alaburda
et al. 2002) (Fig. 3), (Smith et al. 2002) and (Miles et al.
2005) (Fig. 1) for experimental evidence of either effect.
Here we have shown that a subthreshold adaptation cur-
rent also causes a reduction of response gain (in addition
to an increase of response threshold) when the fluctuations
of the input are strong compared to its mean. On the other
9
hand, a spike-triggered adaptation current decreases the
response gain over the whole input range, irrespective of
the level of input fluctuations. These results apply for
adapting as well as the adapted (steady) states2. When
considering the onset I-O curve, i.e., the immediate re-
sponse to a sudden increase of input, an increased level of
spike-triggered adaptation current due to pre-adaptation
has been shown to produce a rather subtractive change
(Benda et al. 2010). This, however, does not contradict
our results. On the contrary, either type of adaptation
current (subthreshold or spike-triggered) naturally leads
to a subtractive change of the onset I-O curve for neurons
which are pre-adapted to an increased input (not shown).
Notably, when considering conductance based noisy
synaptic inputs, an increase in balanced synaptic back-
ground activity can also reduce the spike rate gain (Chance
et al. 2002; Burkitt et al. 2003) and external inhibition can
reduce the gain and increase the response treshold at the
same time (Mitchell and Silver 2003). This means, the
response gain can change due to external inputs that are
independent of the activity of the target neuron, which
can be understood as follows. An increase of noisy (ex-
citatory or inhibitory) synaptic conductance leads to an
increase of total membrane conductance, which causes a
purely subtractive change of the I-O curve, and an increase
in synaptic current noise, which causes an additive change
of the I-O curve and decreases its slope (particularly for
small input strengths) (Chance et al. 2002) (Fig. 3). Both
effects combined lead to the observed change of response
gain. The two separate components are included in our re-
sults. An increase of membrane conductance (represented
by gL in the aEIF model) subtracts from the spike rate
response, see eq. (18), and the abovementioned effects of
an increase of noise intensity σ have been described in the
section Results (see Fig. 2D).
Modulation of response gain is an important phe-
nomenon, particularly in sensory neurons, because neu-
ronal sensitivity to changes in the input is amplified or
downscaled without changing input selectivity. A spike-
dependent adaptation current thus represents a cellular
mechanism by which this is achieved. For example, neu-
ronal response gain increases during selective attention
(McAdams and Maunsell 1999).
It has been shown in-
vivo that the neuromodulator acetylcholine (ACh) con-
tributes substantially to attentional upregulation of spike
rates (Herrero et al. 2008). Cholinergic changes of neu-
ronal excitability and response gain (Soma et al. 2012) in
turn are likely produced via downregulation of slow K+-
currents (Madison et al. 1987; McCormick 1992; Sripati
and Johnson 2006). Together with our results, these ob-
servations suggest that excitability and response gain of
cortical neurons are controlled by neuromodulatory sub-
stances through (de)activation of subthreshold and spike-
triggered K+-currents, respectively.
We have further shown that external inhibitory synaptic
inputs change the I-O curve subtractively, which is consis-
tent with the results of a previous numerical study using
a conductance based neuron model without consideration
of noise (Capaday 2002). Recurrent synaptic (feedback)
2Note that in case of a very large adaptation timescale a (small)
spike-triggered adaptation current has a negligible effect on the
adapting I-O curve, evaluated shortly after the input steps, but a
significant effect on the steady-state I-O curve (see Fig. 8A).
inhibition, which is a function of the neuron's spike rate,
on the other hand, reduces the response gain. This is in
agreement with the results obtained by (Sutherland et al.
2009) based on IF type neurons subject to noisy inputs.
Recent in-vivo recordings from mouse visual cortex have
shown that distinct types of inhibitory neurons produce
these differential effects (i.e., subtractive and divisive
changes of I-O curves) at their target neurons (Wilson
et al. 2012). Functional connectivity analysis suggests
that the inhibitory neurons which changed the I-O curve
of their target neurons subtractively were less likely
connected recurrently to the recorded targets than the
inhibitory neurons which changed their targets' responses
divisively (Wilson et al. 2012). Applying our results based
on the simple network model the observed differential
effects caused by the two types of inhibitory cells can
thus be explained by their patterns of connectivity with
the target cells.
Effects on ISI variability
We have shown that a spike-triggered adaptation current
reduces high ISI variability at low spike rates (when input
fluctuations are strong compared to the mean) and in-
creases low ISI variability at high spike rates (caused by a
large mean input). This result is in agreement with a pre-
vious numerical simulation study (Liu and Wang 2001) but
seems to disagree with other theoretical work (Wang 1998;
Prescott and Sejnowski 2008; Schwalger et al. 2010) at first
sight. Wang (1998) and later Prescott & Sejnowski (2008)
showed that spike-driven adaptation reduces the ISI CV
at low spike rates but they did not find an increase of ISI
CV at higher spike rates in their simulation studies. The
reason for this is that the ISI CVs of adapting and non-
adapting neurons were compared at equal spike rates (i.e.,
at equal mean ISIs) but different input statistics. That is,
the input to the adapting neurons was adjusted to com-
pensate for the change of spike rate (or mean ISI) caused
by the adaptation currents. Increasing the mean input to
the adapting neurons to achieve equal mean ISIs, however,
decreases its ISI CV (cf. eq. (37)). Here we compare the
ISI statistics across different neurons for equal inputs. On
the other hand, Schwalger et al. (2010) analyzed the ISI
statistics of perfect IF model neurons with spike-triggered
adaptation and found that this type of adaptation always
leads to an increase of ISI CV in response to a noisy in-
put current. Their approach is similar to the one presented
here but differs in that the dynamics of the adaptation cur-
rent was neglected in (Schwalger et al. 2010), see Fig. 1B
(bottom panel) for a visualization of that difference. As-
suming a stationary adaptation current leads to a reduced
effective mean input to the neuron, leaving the input vari-
ance unchanged, which always causes increased ISI vari-
ability, cf. eq. (37). Together with theoretical work show-
ing that a spike-dependent adaptation current causes neg-
ative serial ISI correlations (Prescott and Sejnowski 2008;
Farkhooi et al. 2011) our results suggest that spike rate
coding is improved by such a current for low-frequency in-
puts (Prescott and Sejnowski 2008; Farkhooi et al. 2011).
In contrast, an adaptation current which is predomi-
nantly driven by the subthreshold membrane voltage usu-
ally leads to an increase of ISI CV, as we have demon-
strated. This seems to be not consistent with a previ-
ous study (Prescott and Sejnowski 2008) where subthresh-
10
old adaptation was found to produce a small decrease of
ISI variability. The apparent discrepancy is caused by
differences in the presentation of the data: Prescott &
Sejnowski (2008) compared the ISI CVs for equal spike
rates as explained above. That is, the mean input was
adjusted to obtain equal mean ISIs for adapting and
non-adapting neurons but the input variance remained
unchanged. However, increasing the mean input (µ in
eq. (37)) to the adapting neuron counteracts the effect of
subthreshold adaptation on the effective mean input (µa in
eq. (37)). Consequently, one cannot observe an increased
ISI CV in neurons with subthreshold adaptation currents
when the mean input to these neurons is increased. Note
that our results do not contradict those in (Prescott and
Sejnowski 2008), but instead reveal that an increase of a
subthreshold adaptation current always causes an increase
of ISI CV for given input statistics and an increase of a
spike-dependent adaptation current leads to an increase of
ISI CV if the mean input is large.
Finally, we have shown that an increase in synaptic in-
hibition increases the ISI variability, regardless of whether
this inhibition originates from an external population of
neurons or from recurrently coupled ones. An intuitive ex-
planation for this effect is that increased inhibitory input
reduces the mean input but increases the input variance,
see eqs. (5) -- (6). The reason why recurrent synaptic inhi-
bition and spike-triggered adaptation change the ISI vari-
ability in opposite ways in a fluctuation-dominated input
regime could be the different timescales. Synaptic inhibi-
tion usually acts on a much faster timescale than adapta-
tion currents whose time constants range from about one
hundred milliseconds to seconds. Thus, recurrent synaptic
inhibition in contrast to spike-triggered adaptation cannot
provide a memory trace of past spiking activity (over a
duration of several ISIs) that could shape the ISI distri-
bution. Notably, our results on ISIs in a network setting
strictly apply to networks in asynchronous states. Recur-
rent synaptic inhibition, however, can also mediate oscil-
latory activity (Brunel 2000; Brunel et al. 2003; Isaacson
and Scanziani 2011; Augustin et al. 2013) where the vari-
ability of ISIs might be affected differently.
Acknowledgements
Augustin M, Ladenbauer J, Obermayer K. How
adaptation shapes spike rate oscillations in recurrent
neuronal networks. Front Comput Neurosci 7: 1 -- 11,
2013.
Badel L, Lefort S, Brette R, Petersen CCH, Ger-
stner W, Richardson MJE. Dynamic I-V curves
are reliable predictors of naturalistic pyramidal-neuron
voltage traces. J Neurophysiol 99: 656 -- 666, 2008.
Benda J, Herz AVM. A universal model for spike-
frequency adaptation. Neural Comput 15: 2523 -- 2564,
2003.
Benda J, Longtin A, Maler L. Spike-frequency adap-
tation separates transient communication signals from
background oscillations. J Neurosci 25: 2312 -- 2321,
2005.
Benda J, Maler L, Longtin A. Linear versus nonlinear
signal transmission in neuron models with adaptation
currents or dynamic thresholds. J Neurophysiol 104:
2806 -- 2820, 2010.
Brette R, Gerstner W. Adaptive exponential integrate-
and-fire model as an effective description of neuronal
activity. J Neurophysiol 94: 3637 -- 3642, 2005.
Brown DA, Adams PR. Muscarinic suppression of a
novel voltage-sensitive K+ current in a vertebrate neu-
rone. Nature 283: 673 -- 676, 1980.
Brown DA, Griffith WH. Calcium-activated outward
current in voltage-clamped hippocampal neurones of the
guinea-pig. J Physiol 337: 287 -- 301, 1983.
Brunel N. Dynamics of sparsely connected networks of
excitatory and inhibitory spiking neurons. J Comput
Neurosci 8: 183 -- 208, 2000.
Brunel N, Hakim V, Richardson M. Firing-rate res-
onance in a generalized integrate-and-fire neuron with
subthreshold resonance. Phys Rev E 67: 051916, 2003.
Burkitt AN. A review of the integrate-and-fire neuron
model: I. Homogeneous synaptic input. Biol Cybern 95:
1 -- 19, 2006.
This work was supported by DFG in the framework
of collaborative research center SFB910. We thank
Maziar Hashemi-Nezhad for helpful comments on the
manuscript.
Burkitt AN, Meffin H, Grayden DB.
Study of
neuronal gain in a conductance-based leaky integrate-
and-fire neuron model with balanced excitatory and in-
hibitory synaptic input. Biol Cybern 89: 119 -- 125, 2003.
References
Adams PR, Brown DA, Constanti A. Pharmacologi-
cal inhibition of the M-current. J Physiol 332: 223 -- 262,
1982.
Alaburda A, Perrier JF, Hounsgaard J. An M-like
outward current regulates the excitability of spinal mo-
toneurones in the adult turtle. J Physiol 540: 875 -- 881,
2002.
Amit DJ, Brunel N. Model of global spontaneous ac-
tivity and local structured activity during delay periods
in the cerebral cortex. Cereb Cortex 7: 237 -- 252, 1997.
Capaday C. A re-examination of the possibility of con-
trolling the firing rate gain of neurons by balancing ex-
citatory and inhibitory conductances. Exp Brain Res
143: 67 -- 77, 2002.
Chance FS, Abbott LF, Reyes AD. Gain modulation
from background synaptic input. Neuron 35: 773 -- 782,
2002.
Compte A, Constantinidis C, Tegner
J,
Raghavachari S, Chafee MV, Goldman-Rakic
PS, Wang XJ.
Temporally irregular mnemonic
persistent activity in prefrontal neurons of monkeys
during a delayed response task. J Neurophysiol 90:
3441 -- 3454, 2003.
11
Deemyad T, Kroeger J, Chacron MJ. Sub- and
suprathreshold adaptation currents have opposite ef-
fects on frequency tuning. J Physiol 590: 4839 -- 4858,
2012.
La Camera G, Rauch A, Luscher HR, Senn W, Fusi
S. Minimal models of adapted neuronal response to in
vivo-like input currents. Neural Comput 16: 2101 -- 2124,
2004.
Destexhe A. Self-sustained asynchronous irregular states
and up-down states in thalamic, cortical and thalam-
ocortical networks of nonlinear integrate-and-fire neu-
rons. J Comput Neurosci 27: 493 -- 506, 2009.
La Camera G, Rauch A, Thurbon D, Luscher HR,
Senn W, Fusi S. Multiple time scales of temporal
response in pyramidal and fast spiking cortical neurons.
J Neurophysiol 96: 3448 -- 3464, 2006.
Destexhe A, Rudolph M, Par´e D.
The high-
conductance state of neocortical neurons in vivo. Nat
Rev Neurosci 4: 739 -- 751, 2003.
Destexhe A, Rudolph-Lilith M. Neuronal Noise. New
York, USA: Springer, 2012.
Ellis LD, Mehaffey WH, Harvey-Girard E, Turner
RW, Maler L, Dunn RJ. SK channels provide a novel
mechanism for the control of frequency tuning in elec-
trosensory neurons. J Neurosci 27: 9491 -- 9502, 2007.
Farkhooi F, Muller E, Nawrot M. Adaptation reduces
variability of the neuronal population code. Phys Rev
E 83: 050905, 2011.
Fourcaud-Trocm´e N, Hansel D, van Vreeswijk C,
Brunel N. How spike generation mechanisms deter-
mine the neuronal response to fluctuating inputs. J
Neurosci 23: 11628 -- 11640, 2003.
Fuhrmann G, Markram H, Tsodyks M. Spike fre-
quency adaptation and neocortical rhythms. J Neuro-
physiol 88: 761 -- 770, 2002.
Ladenbauer J, Augustin M, Shiau L, Obermayer
K. Impact of adaptation currents on synchronization of
coupled exponential integrate-and-fire neurons. PLoS
Comput Biol 8: e1002478, 2012.
Ladenbauer J, Lehnert J, Rankoohi H, Dahms T,
Scholl E, Obermayer K. Adaptation controls syn-
chrony and cluster states of coupled threshold-model
neurons. Phys Rev E 88: 042713, 2013.
Liu YH, Wang XJ. Spike-frequency adaptation of a gen-
eralized leaky integrate-and-fire model neuron. J Com-
put Neurosci 10: 25 -- 45, 2001.
Madison DV, Lancaster B, Nicoll RA. Voltage clamp
analysis of cholinergic action in the hippocampus. J
Neurosci 7: 733 -- 741, 1987.
Madison DV, Nicoll RA. Control of the repetitive dis-
charge of rat CA1 pyramidal neurones in vitro. J Physiol
354: 319 -- 331, 1984.
Maimon G, Assad JA. Beyond poisson:
increased
spike-time regularity across primate parietal cortex.
Neuron 62: 426 -- 440, 2009.
Gerstein GL, Mandelbrot B. Random walk models
for the spike activity of a single neuron. Biophys J 4:
41 -- 68, 1964.
Mainen ZF, Sejnowski TJ. Influence of dendritic struc-
ture on firing pattern in model neocortical neurons. Na-
ture 382: 363 -- 366, 1996.
Gerstner W, Kistler WM. Spiking Neuron Models.
Cambridge, UK: Cambridge University Press, 2002.
Gigante G, Del Giudice P, Mattia M. Frequency-
dependent response properties of adapting spiking neu-
rons. Math Biosci 207: 336 -- 351, 2007a.
Gigante G, Mattia M, Del Giudice P. Diverse
population-bursting modes of adapting spiking neurons.
Phys Rev Lett 98: 148101, 2007b.
Herrero JL, Roberts MJ, Delicato LS, Gieselmann
MA, Dayan P, Thiele A. Acetylcholine contributes
through muscarinic receptors to attentional modulation
in V1. Nature 454: 1110 -- 1114, 2008.
Isaacson JS, Scanziani M. How inhibition shapes cor-
tical activity. Neuron 72: 231 -- 243, 2011.
Jolivet R, Schurmann F, Berger TK, Naud R, Ger-
stner W, Roth A. The quantitative single-neuron
modeling competition. Biol Cybern 99: 417 -- 426, 2008.
Kim U, McCormick DA. Functional and ionic proper-
ties of a slow afterhyperpolarization in ferret perigenic-
ulate neurons in vitro. J Neurophysiol 80: 1222 -- 1235,
1998.
McAdams CJ, Maunsell JHR. Effects of attention
on orientation-tuning functions of single neurons in
macaque cortical area V4. J Neurosci 19: 431 -- 441,
1999.
McCormick DA. Neurotransmitter actions in the tha-
lamus and cerebral cortex and their role in neuromod-
ulation of thalamocortical activity. Progr Neurobiol 39:
337 -- 388, 1992.
Miles GB, Dai Y, Brownstone RM. Mechanisms un-
derlying the early phase of spike frequency adaptation
in mouse spinal motoneurones. J Physiol 566: 519 -- 532,
2005.
Mitchell SJ, Silver RA. Shunting inhibition modulates
neuronal gain during synaptic excitation. Neuron 38:
433 -- 445, 2003.
Naud R, Gerstner W. Coding and decoding with
adapting neurons: a population approach to the peri-
stimulus time histogram.
PLoS Comput Biol 8:
e1002711, 2012.
Naud R, Marcille N, Clopath C, Gerstner W. Firing
patterns in the adaptive exponential integrate-and-fire
model. Biol Cybern 99: 335 -- 347, 2008.
12
Ostojic S.
Inter-spike interval distributions of spiking
neurons driven by fluctuating inputs. J Neurophysiol
106: 361 -- 373, 2011.
Schwindt PC, Spain WJ, Crill WE. Calcium-
dependent potassium currents in neurons from cat sen-
sorimotor cortex. J Neurophysiol 67: 216 -- 226, 1992.
Shadlen MN, Newsome WT. The variable discharge
of cortical neurons: implications for connectivity, com-
putation, and information coding. J Neurosci 18: 3870 --
3896, 1998.
Smith MR, Nelson AB, Du Lac S. Regulation of
firing response gain by calcium-dependent mechanisms
in vestibular nucleus neurons. J Neurophysiol 87: 2031 --
2042, 2002.
Soma S, Shimegi S, Osaki H, Sato H. Cholinergic
modulation of response gain in the primary visual cortex
of the macaque. J Neurophysiol 107: 283 -- 291, 2012.
Sripati AP, Johnson KO. Dynamic gain changes during
attentional modulation. Neural Comput 18: 1847 -- 1867,
2006.
Stocker M. Ca(2+)-activated K+ channels: molecular
determinants and function of the SK family. Nat Rev
Neurosci 5: 758 -- 770, 2004.
Sutherland C, Doiron B, Longtin A. Feedback-
induced gain control in stochastic spiking networks. Biol
Cybern 100: 475 -- 489, 2009.
Touboul J, Brette R. Dynamics and bifurcations of
the adaptive exponential integrate-and-fire model. Biol
Cybern 99: 319 -- 334, 2008.
Tuckwell HC. Introduction to Theoretical Neurobiology,
volume 2. Cambridge, UK: Cambridge University Press,
1988.
Urdapilleta E. Survival probability and first-passage-
time statistics of a Wiener process driven by an expo-
nential time-dependent drift. Phys Rev E 83: 021102,
2011.
Wang XJ. Calcium coding and adaptive temporal com-
putation in cortical pyramidal neurons. J Neurophysiol
79: 1549 -- 1566, 1998.
Wang XJ, Liu Y, Sanchez-Vives MV, McCormick
DA. Adaptation and temporal decorrelation by single
neurons in the primary visual cortex. J Neurophysiol
89: 3279 -- 3293, 2003.
Wilson NR, Runyan CA, Wang FL, Sur M. Division
and subtraction by distinct cortical inhibitory networks
in vivo. Nature 488: 343 -- 348, 2012.
Ostojic S, Brunel N.
From spiking neuron mod-
els to linear-nonlinear models. PLoS Comput Biol 7:
e1001056, 2011.
Pillow JW, Shlens J, Paninski L, Sher A, Litke
AM, Chichilnisky EJ, Simoncelli EP.
Spatio-
temporal correlations and visual signalling in a complete
neuronal population. Nature 454: 995 -- 999, 2008.
Pospischil M, Piwkowska Z, Bal T, Destexhe A.
Comparison of different neuron models to conductance-
based post-stimulus time histograms obtained in corti-
cal pyramidal cells using dynamic-clamp in vitro. Biol
Cybern 105: 167 -- 180, 2011.
Prescott SA, Sejnowski TJ. Spike-rate coding and
spike-time coding are affected oppositely by different
adaptation mechanisms. J Neurosci 28: 13649 -- 13661,
2008.
Renart A, Brunel N, Wang XJ. Mean-field theory
of irregularly spiking neuronal populations and work-
ing memory in recurrent cortical networks. In: Com-
putational Neuroscience - A Comprehensive Approach,
edited by J Feng, New York, USA: CRC Press, 425 -- 484,
2004.
Reynolds JH, Heeger DJ. The normalization model of
attention. Neuron 61: 168 -- 185, 2009.
Richardson M. Dynamics of populations and networks
of neurons with voltage-activated and calcium-activated
currents. Phys Rev E 80: 1 -- 16, 2009.
Risken H. The Fokker-Planck Equation. New York, USA:
Springer, 1996.
Roxin A, Brunel N, Hansel D, Mongillo G, van
Vreeswijk C. On the distribution of firing rates in
networks of cortical neurons. J Neurosci 31: 16217 --
16226, 2011.
Sanchez-Vives MV, McCormick DA. Cellular and
network mechanisms of rhythmic recurrent activity in
neocortex. Nat Neurosci 3: 1027 -- 1034, 2000.
Sanchez-Vives MV, Nowak LG, McCormick DA.
Membrane mechanisms underlying contrast adaptation
in cat area 17 in vivo. J Neurosci 20: 4267 -- 4285, 2000.
Schwalger T, Fisch K, Benda J, Lindner B. How
noisy adaptation of neurons shapes interspike interval
histograms and correlations. PLoS Comput Biol 6:
e1001026, 2010.
Schwartz O, Pillow JW, Rust NC, Simoncelli EP.
Spike-triggered neural characterization. J Vis 6: 484 --
507, 2006.
Schwindt PC, Spain WJ, Crill WE. Long-lasting
reduction of excitability by a sodium-dependent potas-
sium current in cat neocortical neurons. J Neurophysiol
61: 233 -- 244, 1989.
13
Figure 1: Steady-state spike rates and ISI distributions of single neurons. A, from top to bottom: Spike
times, instantaneous spike rate (r∆t) histogram, membrane voltage (Vi), membrane voltage histogram and adaptation
current (wi) of an (adapted) aEIF neuron with a = 0.06 mS/cm2, b = 0 (left) and a = 0, b = 0.18 µA/cm2 (right)
√
ms for N = 5000 trials. Spike times and
driven by a fluctuating input current with µ = 2.5 mV/ms, σ = 2 mV/
adaptation current are shown for a subset of 10 trials, the membrane voltage is shown for one trial. Results from
numerical simulations are shown in grey. Results obtained using the Fokker-Planck equation are indicated by orange
lines and include the instantaneous spike rate (r), the membrane potential distribution (p) and the mean adaptation
current ( ¯w). r, p and ¯w were calculated from the eqs. (13), (8) and (10), respectively. These quantities have reached
their steady state here. The time bin for r∆t was ∆t = 2 ms, for the other parameter values see Materials and Methods.
B, top panel: ISI histogram corresponding to the N trials in A and ISI distribution (pISI, orange line) calculated via
the first passage time problem (eq. (21)). B, center and bottom panels: Membrane voltage and adaptation current
trajectories from one trial in A, but rearranged such that just after each spike the time is set to zero. Histograms for
the adaptation current just after the spike times are included. The time-varying mean adaptation current from the
first passage time problem (eq. (20)) and the steady-state mean adaptation current from A (eq. (10)) are indicated by
solid and dashed orange lines, respectively. All histograms (in A and B) represent the data from all N trials.
14
Figure 2: Spike rate adaptation, gain and threshold modulation in single neurons. A: Cartoon of a
single neuron visualizing the input parameters and output quantities. B: Instantaneous spike rate r (top panel),
mean membrane voltage (cid:104)V (cid:105)p (center panel, squares) and mean adaptation current ¯w (center panel, solid lines) of
an aEIF neuron without adaptation, a = b = 0 (left), and with either a purely subthreshold adaptation current,
a = 0.06 mS/cm2, b = 0 (center) or a spike-triggered adaptation current, a = 0, b = 0.3 µA/cm2 (right), in response to
a sudden increase in synaptic drive (bottom panel). C: I-O curves of the neurons in B, i.e., spike rate r as a function
of presynaptic spike rates rE , rI. Here, rE = rI, but excitation is stronger than inhibition, due to the coupling
parameter values (see Materials and Methods). The I-O curves represent the spike rate response of the neurons to a
sudden increase of rE and rI, measured in steps of 50 ms after that increase (light to dark colors). Dots indicate the
evolution of the spike rate corresponding to the input in B. D: Steady-state spike rate r∞ as a function of the mean
µ and standard deviation σ of the fluctuating input. Note that µ and σ are determined by the number of presynaptic
neurons, their (Poisson) spike rates and synaptic strengths, cf. eqs. (5) -- (6). The dashed lines in D indicate the values
of µ and σ which correspond to the presynaptic spike rates in C, circles mark the values of the moments corresponding
to the increased input in B.
15
Figure 3: Changes of spiking variability in single neurons. A: ISI distribution (pISI) of a single aEIF neuron
in response to a fluctuating input with mean µ = 0.75 mV/ms and standard deviation σ = 3.25 mV/
ms, for
a = 0, 0.03, 0.06 mS/cm2, b = 0 (top) and a = 0, b = 0, 0.15, 0.3 µA/cm2 (bottom). B: ISI coefficient of variation (CV)
as a function of µ and σ, for a neuron without adaptation, a = b = 0 (left), and with either a subthreshold adaptation
current, a = 0.06 mS/cm2, b = 0 (center) or a spike-triggered adaptation current, a = 0, b = 0.3 µA/cm2 (right).
Circles indicate the values of µ and σ used in A. C: Change of ISI CV caused by a subthreshold (left) or spike-triggered
(right) adaptation current as a function of µ and σ. The white regions in B and C indicate the parameter values for
which the ISI CV was not computed, because r∞ < 1 Hz.
√
16
Figure 4: Gain and threshold modulation caused by network interaction. A: Cartoon of the network visual-
izing the coupling parameters. B, top panel: Steady-state spike rate of excitatory aEIF neurons, rpopE,∞ (solid lines) and
inhibitory aEIF neurons, rpopI,∞ (dashed lines), as a function of rextEE , for rextIE = 6, 10, 14 Hz (left), J recIE = 0.05, 0.1, 0.2 mV
(center), J recEI = −0.45,−0.6,−0.75 mV (right). Inset cartoons visualize the varied parameters as specified on the top
left. If not indicated otherwise, J recEI = −0.6 mV, rextIE = 10 Hz and J recIE = 0.1 mV. For the other parameter values see
Materials and Methods. B, bottom panel: Steady-state spike rate rpopE,∞ as a function of the input parameters µ and
σ for the excitatory neurons. Solid lines and dots in the top panel correspond to those of equal color in the bottom
panel.
17
Figure 5: Changes of spiking variability caused by network interaction. A: ISI distributions (pISI) of excitatory
aEIF neurons for rextEE = 50 Hz. J recEI = −0.6 mV, rextIE = 10 Hz and J recIE = 0.1 mV if not indicated otherwise. B: ISI
CV for excitatory neurons (CVpopE
) as a function of rextEE . Color code as in A. Dots indicate the input and ISI CV
values for the ISI distributions in A. Insets: ISI CV as a function of the input parameters µ and σ for the excitatory
neurons. Lines and dots (insets) correspond to those of equal color in B.
Figure 6: Subthreshold and spike-triggered components of IKCa, IKNa and IM. A: Conductances for the slow
K+-currents INa, IKCa, IKNa and IM in steady state as a function of the membrane voltage, normalized to a peak value
of 1 mS/cm2. B and C: Subthreshold conductance a and spike-triggered absolute increment ∆IKs (B) and relative
increment ∆I rel
Ks (C) obtained from the fitting procedure (see Materials and Methods) for the conductance-based model
neurons with ¯gKCa ∈ [2, 8] mS/cm2 and ¯gKNa = ¯gM = 0 (dots), ¯gKNa ∈ [2, 8] mS/cm2 and ¯gKCa = ¯gM = 0 (squares),
¯gM ∈ [0.1, 0.4] mS/cm2 and ¯gKCa = ¯gKNa = 0 (diamonds). Darker symbols indicate larger conductance values.
18
Figure 7: Subthreshold and spike-triggered components of a range of slow K+-currents A: Steady-state
K+-conductance gKs,∞(V ) = ¯gKsω∞(V ) as a function of the membrane voltage, for the generic Hodgkin-Huxley-type
description of a slow K+-current (see Materials and Methods), with half-activation voltage α = −40 mV (left curves),
α = −10 mV (right curves), inverse steepness β = 6, 9, 12 mV and peak conductance ¯gKs = 1 mS/cm2. The dashed
curve indicates the Na+-conductance gNa,∞(V ) of the conductance-based model, normalized to a maximum value of
1 mS/cm2. B: Subthreshold conductance a obtained from the fitting procedure for different values of the parameters
α and β. C: Absolute and relative spike-triggered increments ∆IKs (top panel) and ∆I rel
Ks (bottom panel), respectively,
as a function of α, for τω = 100 ms (left) and τω = 300 ms (right).
19
Figure 8: Effects of IKCa, IKNa and IM on I-O curve and ISI variability. A: Spike rate of a conductance-based
model neuron without slow K+-currents, ¯gKCa = ¯gKNa = ¯gM = 0 (black), and with either type of slow K+-current
included, ¯gKCa = 8 mS/cm2 (red), ¯gKNa = 8 mS/cm2 (blue), ¯gM = 0.4 mS/cm2 (green), in response to a sudden
√
increase of mean input µ, measured in four subsequent time intervals of 250 ms after that increase (light to dark
colors). The baseline mean input was µ = 0.05 mV/ms and the input standard deviation was σ = 0.5 mV/
ms.
Average values over 50 independent trials are shown. The adapting I-O curve of the neuron with increased IKNa
(¯gKNa = 8 mS/cm2) converges very slowly to the steady-state curve (dashed blue) measured 20 s after the increase in
√
µ. B: ISI CV of the neurons in A as a function of mean input µ for low (left), medium (center), and high (right) noise
ms), respectively. The ISIs were collected over an interval of 10 s after the steady-state
intensity (σ = 1, 1.5, 2 mV/
spike rates were reached, in 50 independent trials.
20
|
1811.08498 | 1 | 1811 | 2018-11-20T21:44:16 | Global Sensitivity Analysis of High Dimensional Neuroscience Models: An Example of Neurovascular Coupling | [
"q-bio.NC"
] | The complexity and size of state-of-the-art cell models have significantly increased in part due to the requirement that these models possess complex cellular functions which are thought--but not necessarily proven--to be important. Modern cell models often involve hundreds of parameters; the values of these parameters come, more often than not, from animal experiments whose relationship to the human physiology is weak with very little information on the errors in these measurements. The concomitant uncertainties in parameter values result in uncertainties in the model outputs or Quantities of Interest (QoIs). Global Sensitivity Analysis (GSA) aims at apportioning to individual parameters (or sets of parameters) their relative contribution to output uncertainty thereby introducing a measure of influence or importance of said parameters. New GSA approaches are required to deal with increased model size and complexity; a three stage methodology consisting of screening (dimension reduction), surrogate modeling, and computing Sobol' indices, is presented. The methodology is used to analyze a physiologically validated numerical model of neurovascular coupling which possess 160 uncertain parameters. The sensitivity analysis investigates three quantities of interest (QoIs), the average value of $K^+$ in the extracellular space, the average volumetric flow rate through the perfusing vessel, and the minimum value of the actin/myosin complex in the smooth muscle cell. GSA provides a measure of the influence of each parameter, for each of the three QoIs, giving insight into areas of possible physiological dysfunction and areas of further investigation. | q-bio.NC | q-bio | Global Sensitivity Analysis of High Dimensional Neuroscience Models: An
Example of Neurovascular Coupling
J.L. Hart · P.A. Gremaud · T. David
8
1
0
2
v
o
N
0
2
]
.
C
N
o
i
b
-
q
[
1
v
8
9
4
8
0
.
1
1
8
1
:
v
i
X
r
a
Abstract The complexity and size of state-of-the-art cell models have significantly increased in part
due to the requirement that these models possess complex cellular functions which are thought -- but not
necessarily proven -- to be important. Modern cell models often involve hundreds of parameters; the values
of these parameters come, more often than not, from animal experiments whose relationship to the human
physiology is weak with very little information on the errors in these measurements. The concomitant
uncertainties in parameter values result in uncertainties in the model outputs or Quantities of Interest
(QoIs). Global Sensitivity Analysis (GSA) aims at apportioning to individual parameters (or sets of
parameters) their relative contribution to output uncertainty thereby introducing a measure of influence
or importance of said parameters. New GSA approaches are required to deal with increased model size and
complexity; a three stage methodology consisting of screening (dimension reduction), surrogate modeling,
and computing Sobol' indices, is presented. The methodology is used to analyze a physiologically validated
numerical model of neurovascular coupling which possess 160 uncertain parameters. The sensitivity
analysis investigates three quantities of interest (QoIs), the average value of K + in the extracellular
space, the average volumetric flow rate through the perfusing vessel, and the minimum value of the
J.L.H. was supported in part by the National Science Foundation (NSF) award NSF DMS-1522765.
P.A.G. was supported in part by National Science Foundation (NSF) awards NSF DMS-1522765 and DMS-1745654.
J.L. Hart
Department of Mathematics, North Carolina State University, Raleigh, NC, USA, E-mail: [email protected],
P.A. Gremaud
Department of Mathematics, North Carolina State University, Raleigh, NC, USA, E-mail: [email protected]
T. David
Department of Mechanical Engineering, University of Canterbury, New Zealand, E-mail: [email protected]
actin/myosin complex in the smooth muscle cell. GSA provides a measure of the influence of each
parameter, for each of the three QoIs, giving insight into areas of possible physiological dysfunction and
areas of further investigation.
Keywords: neurovascular coupling, global sensitivity analysis, model parameters
1 Introduction
Over the past 20 years, the use of computational models to describe physiological phenomena has grown
spectacularly. This growth has provided significant advantages to the experimental community in that
it can furnish results of in silico experiments which are either ethically or physically impossible in the
laboratory. However, these models have also increased in complexity with a concomitant increase in the
associated number of parameters (for a variety of reasons, most notably the requirement that the models
should possess complex cellular functions which are thought but not necessarily proven to be important).
These parameters, in defining the relevant phenomenon, more often than not come from a plethora
of animal experiments whose relationship to the human physiology is weak. In addition, experimental
results in the public domain provide very little information on the errors in these measurements.
Even simple physiological models tend to be highly non-linear; their range of applicability and reliabil-
ity can only be assessed through careful analysis. For large complex systems, the sensitivity of quantities
of interest to model parameters is a priori unclear. Herein lies one of the difficulties of modeling: what
effect do uncertainties in parameters determined from experiment have on the output of a non-linear
numerical model? Because of both model complexity and high dimensionality, sensitivity analysis is com-
putationally demanding and may require several ad hoc steps -- such as screening (reducing the parameter
dimension). Importantly, in analysing sensitivities, we can learn significant facts about the physiology of
the system which would have stayed hidden under the premise of simply producing results. From a purely
physiological perspective, an understanding of the dominant cellular mechanisms resulting in cerebral
tissue perfusion after neuronal stimulation would be of particular and important interest.
In the above context, we investigate the sensitivity of the neurovascular coupling (NVC) response
(see Section 2), which has a large parameter dimension where most (if not all) of the parameter values
2
come from non-human experiment with an inherent (unknown) error. We denote by y = (y1, . . . , yN )
the state variables of the model and by θ = (θ1, . . . , θP ) the uncertain parameters of the model. The
evolution of the state variables is governed by a system of ordinary differential equations (ODEs)
dy
dt
= f (y, θ),
(1.1)
where f is a known function of its arguments. Equation (1.1) is completed with a set of initial conditions
y(0) = y0; here, we simply take y0 as the equilibrium solution at the parameters' nominal values, i.e.,
f (y0, ¯θ) = 0, where ¯θ = (¯θ1, . . . , ¯θP ) denotes the nominal values of the parameters. The Supplementary
Material contains code and the nominal values of parameters, further information can be found in [6].
Based on physiological considerations, we examine three quantities of interest (QoI), see Section 2.
Let q be one of our three QoIs; while determined from the state variables, i.e., from y, q is ultimately a
function of the parameters alone (and possibly time), i.e.
q = g(θ).
(1.2)
The overall goal of our numerical study is to determine which of the uncertain parameters θ1, . . . , θP are
the most/least influential for each QoI. There is a substantial amount of research currently being done
in applied mathematics and statistics in the corresponding field of Global Sensitivity Analysis (GSA).
How to meaningfully define "influential" or "non-influential" and how to develop methods applicable to
high-dimensional problems are two significant challenges of the field [1, 14, 15, 17, 32, 35].
The present NVC model contains N = 67 state variables and P = 160 uncertain parameters. The
complexity of the model and large parameter space dimension preclude a direct application of GSA tools.
Part of our contribution in this paper is to show how multiple GSA tools may be combined to analyse
such problems.
2 Physiological Model: Neurovascular Coupling
The NVC response, i.e., the ability to locally adjust vascular resistance as a function of neuronal activ-
ity, is believed to be mediated by a number of different signaling mechanisms. A mechanism based on a
metabolic negative feedback theory was first proposed in [34]. According to this theory, neural activity
leads to a drop in oxygen or glucose levels and increases in CO2, adenosine, and lactate levels. All of these
3
signals could dilate arterioles and hence were believed to be part of the neurovascular response. How-
ever, recent experiments illustrated that the NVC response is partially independent of these metabolic
signals [22, 23, 25, 31, 33]. An alternative to this theory was proposed where the neuron releases signal-
ing molecules to directly or indirectly affect the blood flow. Many mechanisms such as the potassium
(K+) signaling mechanism [11], the nitric oxide (NO) signaling mechanism, or the arachidonic acid to
epoxyeicosatrienoic acid (EET) pathway are found to contribute to the neurovascular response [3].
The K+ signaling mechanism of NVC seems to be supported by significant evidence, although new
evidence shows that the endfoot astrocytic calcium (Ca2+) could play a significant role. The K+ signaling
hypothesis mainly utilises the astrocyte, positioned to enable the communication between the neurons and
the local perfusing blood vessels. The astrocyte and the endothelial cells (ECs) surrounding the perfusing
vessel lumen exhibit a striking similarity in ion channel expression and thus can enable control of the
smooth muscle cell (SMC) from both the neuronal and blood vessel components [24]. Whenever there is
neuronal activation K+ ions are released into the extracellular space (ECS) and synaptic cleft (SC). The
astrocyte is depolarised by taking up K+ released by the neuron and releases it into the perivascular
space (PVS) via the endfeet through the BK channels [12]. This increase in ECS K+ concentration
(3− 10 mM) near the arteriole hyperpolarises the SMC through the inward rectifying K+ (KIR) channel,
effectively closing the voltage-gated Ca2+ channel, reducing smooth muscle cytosolic Ca2+ and thereby
causing dilation. Higher K+ concentrations in the PVS cause contraction due to the reverse flux of the
KIR channel [10].
In maintaining a relatively homeostatic condition the neuron uses a considerable amount of energy (
as noted below) on ensuring a specific concentration of ions (and ion gradients) in the extracellular space.
This concentration of K+and Na+can be envisaged as one of the inputs to the astrocyte in determining
the nutrient flux to the cerebral tissue. Parameters which model/affect the energy usage are therefore
important in simulating neurovascular coupling. In order not to increase the complexity of the numerical
model further than is necessary an approximation needs to be made on how to model the energy usage
and the associated parameters. Estimates of the relative demands of the cerebral processes that require
energy were given based on different experimental data by Ames [2]. Vegetative processes that maintain
this homeostasis including protein synthesis accounted for 10 − 15% of the total energy consumption.
4
The costliest function seems to be in restoring the ionic gradients during neural activation. The sodium
potassium (Na+/K+) exchange pump is estimated to consume 40 − 50%, while the Ca2+ influx from
organelles and extracellular fluid consumes 3 − 7%. Processing of neurotransmitters such as uptake or
synthesis consumes 10 − 20%, while the intracellular signaling systems which includes activation and
inactivation of proteins consumes 20 − 30%. The rest of the energy is estimated to be consumed by the
axonal and dendritic transport in both directions. Given the distribution of energy amongst the cellular
pathways it is assumed for this model that variation in oxygen concentration in the cerebral tissue affects
only the Na+/K+ exchange pump.
Previous work [30] has provided the construction of an experimentally validated numerical (in silico)
model based on experimental data to simulate the functional magnetic resonance imaging (fMRI) blood-
oxygen-level dependent (BOLD) signal associated with NVC along with the associated metabolic and
blood volume responses. An existing neuron model [28, 29] has been extended to include an additional
transient sodium (Na+) ion channel (NaT) expressed in the neuron, and integrated into a complex NVC
model [7, 8, 19]. This present model is based on the hypothesis that the K+ signaling mechanism of
NVC is the primary contributor to the vascular response and the Na+/K+ exchange pump in the neuron
is the primary consumer of oxygen during neural activation. The model contains 317 parameters, most
of which come from non-human experiments. Based on the work by [8] and [20], we have chosen a
subset of parameters defining basic pathways, such as the nitric oxide and potassium pathways, that are
considered important for the normal function of neurovascular coupling. We model the uncertainty of
the chosen parameters by representing them as random variables. The remaining parameters are fixed to
nominal values; they include leak terms, characteristic oxygen and other species concentrations, buffer
concentrations, volume surface ratios etc. . . By permitting variability in only the parameters that support
these pathways, the dimension of the parameter space is reduced from 317 to 160, which greatly facilitates
our analysis. The algorithms defined below can be used to investigate other complex models including
that of neurovascular coupling. However, for this initial work, we constrain ourselves to the above subset.
Figure 1 shows the components and main pathways of the neurovascular coupling model (version 2.0).
The numerical model outlined in sketch form in Figure 1 is fully defined in the Supplementary Material
and has been developed over a number of years [7, 8, 10].
5
Fig. 1: Graphic Sketch of NVU version 2.0, showing the basic components of neuron, astrocyte, smooth
muscle cell, endothelial cell, lumen and extracellular space. ion channels, pumps and pathways from
neuron to endothelial cell are also shown in addition to the basis for evaluation of the fMRI blood
oxygenation level signal (BOLD) protocol.
In order to induce a variation in radius of a vessel which perfuses the associated cerebral tissue
following a neuronal stimulation, an input current is used. The numerical experiments solve equation (1.1)
in the presence of short duration electrical (current) stimuli as displayed in Figure 2. For the presented
cases, two input profiles are utilised. A rectangular pulse of width 10 seconds and a experimental pulse
sequence (stimulating the whisker pad of a rat) used in the work of [40] which has the same magnitude
as the rectangular pulse but a duration of sixteen seconds followed by second pulse (which is not used in
this analysis). Further information about the experiment and the results can be found in [40]. For this
second case, the stimulus was a current injection at the rat whisker pad. This induced an increase in
6
ERERPumpPumpIPR3LeakSRPumpIP R3LeakCICRCICR0CationExtrusionStretchKResidualCa2+cplVcplIP3cplPLCLeakwall shear stresseNOSNOPVSSMCLUNOECStretchNaCaClk1k-1k3k2k4EEEb6c5cNONONONOBKVOCCNaKExtrusionIP R3KIRTRPV4BKA + MA + MpAMpAMKKKKKKK1425637RadiusEETACSCNONa3HCONBCKCC1NKCC1NaKKClρO2CBFCBVCMRO2HbRBOLDO2O2ECSNMDAPostsynaptic terminalnNOSNODendriteSoma/axonNa leakK leakNaPKAKDRNMDAKDRNaPKAK leakNa leakNaT[KB][B]+[K ]+BufferingInput current I(t)NaK pumpO2O2O2NaK pumpO2O2O2O2Ca2+ionK+ionIP3Na+ionagonistClion'voltage'degradationionchannelionpumpreceptorHCO3ion--GluEETstretchcGMPCaMdiffusionNOO2Nitric oxideOxygenO2O2Glu neuronal activity in the somato-sensory cortex which subsequently, through the neurovascular pathway,
produced a change in the radius R(t) allowing increased nutrients to perfuse into the cerebral tissue; this
is the essence of neuro-vascular coupling.
Fig. 2: Left: rectangular pulse input stimulus; right: stimulus used in lab experiments.
We wish to analyse parameter sensitivity of quantities of interest (QoIs) that are key to our un-
derstanding and quantification of neurovascular coupling. From previous work, it is known that the
K+concentration in the extracellular space (ECS) is crucial to maintaining a homeostatic state (large
concentrations of K+support the propagation of spreading depression waves) hence we choose to look at
the average value of K+in the ECS as our first QoI q1. Secondly, neurovascular coupling is the main phe-
nomenon providing oxygen and nutrients to the neuronal tissue. This is mediated by the local arteriole
dilating and (under the assumption of constant pressure) increasing the flow of blood into the tissue. We
therefore define the time averaged cerebral blood flow determined over the course of neuronal stimulation
as our second QoI, q2. The contraction/dilation of the arteriolar vessel depends on the smooth muscle cell
(SMC) concentration of Ca2+and the phosphorylation of the actin/myocin complex. Our third QoI q3
is the minimum combined concentration of the actin myosin complex, both phosphorylated and unphos-
phorylated. This allows us to analyse the functioning of the main components of the neurovascular unit
(NVU) defined as the linked components of neuron, synaptic ceft, astrocyte, perivascular space (PVS),
SMC, endothelial cell (EC), lumen (LU, the domain in which blood flows) and the ECS. Assuming the
stimulation occurs for time t between t1 and t2, t2 > t1, the QoIs are thus
7
95100105110115120Time (sec)00.0050.010.0150.020.0250.03Input Stimulus95100105110115120Time (sec)00.0050.010.0150.020.0250.03Input Stimulus -- average ECS potassium
q1 =
1
t2 − t1
[K +]ECS(s) ds,
(cid:90) t2
t1
(cid:90) t2
-- average volumetric flow rate in the cerebral tissue
(cid:18) R(s)
(cid:19)4
q2 =
1
t2 − t1
t1
R0
ds,
(2.1)
(2.2)
-- minimum combined concentration of the actin myosin complex, both phosphorylated and unphos-
phorylated
q3 = min
t,t1≤t≤t2
[AM (t) + AMp(t)],
(2.3)
which we will refer to as [AM + AMp]min throughout the article.
Other examples of sensitivity analysis studies relevant to the general field of biomedicine include
[16, 37, 38, 39].
3 Methodology
Our analysis may be described in a four step process:
(i) define a probability distribution for θ which models its uncertainty,
(ii) draw samples from this distribution,
(iii) evaluate the QoI for each of these parameter samples,
(iv) use these QoI evaluations to infer its sensitivity to each parameter.
Subsection 3.1 describes our approach for step (i). Step (ii) is easily executed using standard methods. The
computational bottleneck for our analysis is the QoI evaluations (which require ODE solves) in step (iii).
While relatively large and stiff, the ODE system (1.1) can be solved with standard tools and methods,
here through the MATLAB routine ode15s with relative and absolute tolerances of 10−4. The evaluation
of the above QoIs themselves from the ODE solutions is straightforward and can be done at low cost. Step
(iv), inferring the sensitivities, may be done in a plurality of ways. Because of the parameter dimension
and computational cost of model evaluations, we perform step (iv) with a multi-phased procedure,
(I) screening,
8
(II) surrogate modeling,
(III) and computing Sobol' indices,
described in Subsections 3.2, 3.3, and 3.4 respectively. We summarize our method in Subsection 3.5.
3.1 Parameter Distribution Fitting
Describing the uncertainty attached to the parameter vector θ is an important and delicate modeling
assumption. Here, we give each θi, i = 1, . . . , 160, a nominal value ¯θi and assume each parameter to be
independent and uniformly distributed over the interval [0.9 ¯θi, 1.1 ¯θi], i.e., within ±10% of the nominal
value. Larger intervals were considered for the parameters. For the NVU model under consideration,
increasing the width of the interval resulted in many samples for which the solution exhibited atypical
or non-physical behavior, or in some cases ode15s was unable to solve the system. Our choice of ±10%
uncertainty is considered a reasonable compromise between accounting for uncertainty and ensuring
computational feasibility.
Our initial assumption of parameter independence is incorrect. In fact, ode15s is unable to solve
the system for most samples drawn under this assumption. The parameter dependencies, which are
unknown a priori, are discovered through a computational procedure akin to Approximate Bayesian
Computation [27]. Our approach, described below, is computationally intractable when applied to the
NVU model under consideration. We simplify the model by removing the stimulus, i.e. modeling the
steady state behavior of the system. A collection of S samples, θk, k = 1, 2, . . . , S, are drawn from our
initial distribution (assuming independence) and the ODE system is solved for each parameter sample.
The collection of solutions are post-processed to partition the samples {θk}S
k=1 into two subsets {θak}Sa
and {θrk}Sr
k=1, where ak, k = 1, 2, . . . , Sa denotes the samples where the steady state solution exhibited
k=1
physiologically normal behavior and rk, k = 1, 2, . . . , Sr denotes the samples where it does not. We reject
the samples {θrk}Sr
k=1 and fit a distribution to the accepted samples {θak}Sa
k=1 using standard statistical
methods. This new distribution is sampled S times, the ODE system is solved for each sample, and the
results are post-processed into accepted and rejected samples again. We continue this process iteratively
until satisfactory convergence of the fitted distribution.
9
160(cid:88)
3.2 Screening
Having determined a parameter distribution, we evaluate the QoI q = g(θ) at M different samples,
denote them θk, k = 1, 2, . . . , M . We fit a linear model to the QoI under study
g(θk) = g(θk
1 , . . . , θk
P ) ≈ β0 +
βjθk
j ,
k = 1, . . . , M.
(3.1)
j=1
This approach yields a crude (but highly efficient here) sensitivity analysis of the model with respect to
the θj's, j = 1, . . . , 160. We assign a preliminary importance measure to each θj by computing for each
of them the relative size of their coefficient in the above linear approximation, i.e.
Lj =
160(cid:80)
βj
β(cid:96)
,
(cid:96)=1
j = 1, . . . , 160.
To obtain a model with a more manageable size, we reduce the parameter space to only the θj's for
which Lj > 0.01. We denote these r parameters {θji}r
i=1. The rest of the parameters are regarded
as non-influential and treated as latent variables, even though they are uncertain, their specific values
(within the given range) have little bearing of the considered QoI. In other words, we consider the
approximation
g(θ1, . . . , θ160) ≈ h(θj1, . . . , θjr ),
(3.2)
where h is obtained from g by treating the non-influential parameters as latent. In Section 4, this reduction
yields around 15-20 parameters instead of the original 160. We use θ to denote the reduced parameter
vector.
3.3 Surrogate model
For any of the three considered QoIs, our information on the function h defined (3.2) consists of the set
of sampled values {h(θk
)}, k = 1, . . . , M . To facilitate the use of standard GSA tools, which
, . . . , θk
jr
j1
may require derivatives or variance estimations, it is both convenient and computationally advantageous
to construct an approximating function, i.e., a surrogate model. We use a sparse Polynomial Chaos (PC)
surrogate. This amounts to introducing a polynomial approximation of h of the type
h( θ) ≈ H( θ) ≡(cid:88)
cαψα( θ)
(3.3)
α
10
where the ψα's are multivariate polynomials which are orthogonal with respect to the probability distri-
bution function (PDF) p θ of θ, i.e.
(cid:90)
ψα(x)ψβ(x) p θ(x) dx = δα,β
(3.4)
where α and β are multi-indices and δα,β is the generalized Kronecker symbol. The coefficients are
computed through least-squares minimization, see the Appendix for additional discussion. All surrogate
models are validated using 10-fold cross validation.
Polynomial Chaos is by now a well documented method. We use the UQLab implementation for the
results below and refer the reader to its manual [26] for more details.
3.4 Sobol' indices
We use variance based GSA to assess the relative importance of the input parameters of H in (3.3). In
their simplest form, the total Sobol' indices [36] apportion to uncertain parameters, or sets thereof, their
relative contribution to the variance of the output. Indeed, thanks to the law of total variance, we can
decompose the variance of H( θ) as
var(H( θ)) = var(E[H( θ) θ∼i]) + E[var(H( θ) θ∼i)],
(3.5)
where θ∼i denotes all the parameters in θ except θi. If we assume now that all the input parameters of
H are known with certainty, i.e., if θ∼i is known, then the remaining variance of H( θ) is simply given
by
var(H( θ)) − var(E[H( θ) θ∼i]) = E[var(H( θ) θ∼i)].
This latter expression is thus a natural way of measuring how influential θi is; the corresponding total
Sobol' index STi is but a normalized version of this
STi =
E[var(H( θ) θ∼i)]
var(H( θ))
.
(3.6)
From this definition, one easily observes that STi ∈ [0, 1]; large values indicate that θi is important,
STi ≈ 0 implies that θi is not important. The relevance of this basic definition can be extended to time
dependent QoIs [1], stochastic QoIs [15] or correlated parameters [14].
11
3.5 Summary of the Method
Algorithm 1 provides a summary of the method.
Algorithm 1 overall numerical approach
1: while parameter distribution has not converged do
2:
3:
4:
5:
6:
for k = 1 : S do
(cid:46) sampling
sample parameter distribution −→ θk; solve (1.1) (without stimulus) −→ yk
end for
partition the parameter samples into accepted and rejected samples
fit a new distribution to the accepted samples
(cid:46) see § 3.1
7: end while
8: for k = 1 : M do
9:
sample parameter distribution −→ θk; solve (1.1) (with stimulus) −→ yk
(cid:46) sampling (final distribution)
10: end for
11: for each QoI q do
12:
13:
14:
15:
solve the least-squares problem (3.1)
identify the influential parameter vector θ = (θj1 , . . . , θjr )
fit the polynomial chaos surrogate H (3.3)
compute total Sobol' indices (3.6) of the surrogate model H
16: end for
(cid:46) linear model
(cid:46) screening
(cid:46) surrogate model
(cid:46) Sobol' indices
4 Numerical results
4.1 Parameter distribution
Applying the methodology described in Section 3.1, S = 919 parameter samples are drawn and (1.1)
is solved for each sample with no stimulus applied. Of these 919 runs, 670 of them fail due to ode15s
terminating prematurely because its time steps became too small. The remaining 249 yields 139 solutions
for which the radius reaches a stable steady state (at least for the duration of the time integration) and
110 solutions for which the radius reaches a steady state but subsequently becomes transient. The left
panel of Figure 3 displays four representative solutions for the radius; the red curves remain in steady
state for the duration of our time integration whereas the black curves revert into a transit regime after
12
some time in or near steady state. The right panel of Figure 3 displays the samples for two parameters
(defined in the buffer equation (4.1), see the Supplementary Material for details) which are found to be
highly influential in determining the behavior of the solution. A blue * indicates a sample where the solver
terminated prematurely, a yellow ♦ indicates a sample where an unstable steady state was observed, a
red ◦ indicates a sample where the solution remained in steady state. A strong correlation determining
the behavior of the solution is observed. The two parameters in question, denoted by (θ120, θ121), occur
in the buffer equation defined by
ff =
1+exp(
θ118
−[K+ ]e+θ120
θ121
)
[K +]e + B
(cid:10)
[K +]b
fr = θ118
d[K +]b
dt
= ff [K +]e(θ119 − [K +]b) − fr[K +]b
The parameters θ120 and θ121 have values 5.5 and 1.09, representing a shift and scaling in the forward
rate constant, respectively.
(4.1)
Fig. 3: Left: examples of stable (red) and unstable (black) steady state solutions. Right: samples of the
buffer parameters (θ120, θ121) using uniform independent sampling. A blue * indicates the sample yielded
a premature termination of the solver, a yellow ♦ indicates the sample yielded an unstable steady state,
a red ◦ indicates the sample yielded a stable steady state.
Observing this correlation, we use the accepted samples to fit (θ120, θ121) with a bivariate Frank
copula with beta marginals. The experiment is repeated by sampling the two correlated parameters from
13
02004006008001000Time (sec)1.822.22.42.62.83Radius (m)10-555.25.45.65.86p111.051.11.151.2p2 this bivariate distribution and all other parameter from their original uniform distributions. After four
iterations refining the joint distribution of (θ120, θ121), we were able to generate 902 out of 960 samples
which yielded solutions with stable steady states (51 solutions had unstable steady states and 7 had
premature solver terminations). This fitted distribution is used for all subsequent analysis.
Samples are drawn and the model, with a stimulus applied (in two separate cases, the 10 second
rectangular pulse and the 16 second experimental pulse), is run for each sample. This results in solutions
exhibiting three different physiological regimes; they are displayed in Figure 4 where the radius is plotted
as a function of time. The leftmost panel corresponds to the typical case when the radius increases
in response to the stimulus and then decreases when it is removed; the center panel corresponds to
an atypical case where the radius has an initial decrease in response to the stimulus; the right panel
corresponds to another atypical case where the radius reaches another steady state and does not decrease
after the stimulus is removed.
Fig. 4: Radii corresponding to samples (using the rectangular pulse stimulus). Left: curves an an increase
in response to the stimulus; center: curves with a decrease in response to the stimulus; right: curves
which settle in a different steady state.
This article focuses on the non-pathological case, corresponding to the left panel of Figure 4, so we
remove samples where the radius does not increase in response to the stimulus and decrease when it is
removed. However, we recognise that a decrease in radius upon stimulation does not necessarily mean an
incorrect result. Indeed these cases are of particular importance (due to the possibility of the existence
of cortical spreading depression [21]) and a topic of future research. This processing yields 660 samples
for analysis when the rectangular pulse stimulus is applied and 438 samples when the experimental pulse
stimulus is applied. The results presented below use these samples.
14
Exploration of the 660 retained samples indicate that "pathological" cases have higher probability
when the parameter which shifts the activation variable for the K+flux through the soma KDR channel
is reduced; however, this parameter does not characterize the solution regime by itself; it is likely that the
solution regime is characterized by a combination of several parameters. Further sampling and exploration
is required to better understand the structure in parameter space which determine the solution regime.
The 160 parameters are indexed (purely for coding reasons) and are not to be taken as a ranked
order. The tables of ranked parameters, given in the subsections below to summarize the most influential
parameters, provide the parameters in the first column, their location in the Supplementary Material in
the second column, and their total Sobol' indices for the experimental and rectangular pulses in the third
and fourth columns, respectively. Each table contains the five most influential parameters for a given
QoI, as measured by the total Sobol' indices for the experimental pulse. There are slight differences in
the ordering of the less important parameters for the rectangular pulse and experimental pulse cases; the
tables below report the ranking from the experimental pulse case. Both rectangular and experimental
pulse cases agree on the ranking of the most influential parameters.
4.2 Average ECS Potassium
Figure 5 and 6 display results for the average of the ECS potassium, defined in equation (2.1), for the
rectangular pulse stimulus and experimental pulse stimulus, respectively. In the top left panels, predic-
tions of the linear surrogate are plotted against the QoI values. The sensitivities Lj, j = 1, 2, . . . , 160,
are displayed in the top right panels. Predictions of the PC surrogate are plotted against the QoI values
in the bottom left panels. The total Sobol' indices of the PC surrogate are given in the bottom right
panels. Table 1 reports the five most important parameters and their respective total Sobol' indices.
The first and third parameters in Table 1 are scaling and shift parameters (θ62 and θ63) for the
activation gating variable, m4 in the dendrite NaP channel respectively, whose ODE is defined as
dm4
dt
= m4α(1 − m4) − m4βm4,
m4α =
m4β =
1
6(1 + exp(−(θ62vd + θ63)))
exp(−(θ62vd + θ63))
6(1 + exp(−(θ62vd + θ63)))
,
.
15
(4.2)
Fig. 5: Average ECS potassium QoI with a rectangular pulse stimulus. From left to right and top to
bottom: linear surrogate predictions, linear surrogate importance measure, PC surrogate predictions,
total Sobol' indices for PC surrogate.
where vd is the dendrite membrane potential.
The nominal values of θ62 and θ63 are 0.143 and 5.67, respectively. These effectively define the charac-
teristic time scale and forcing function in the rate equation for the open probability of the persistent
sodium channel. The second most important parameter determines the strength of the conductance in
the K+leak ion channel. The fourth and fifth parameters, θ75 and θ76, are the shift and scale of the
neuron membrane potential in the ODE for the activation variable for the K flux through dendritic KDR
channel, defined as
(cid:18)
m6α = θ74
vd + θ75
1 − exp(−(θ76vd + θ76θ75))
(cid:19)
,
(4.3)
where the nominal values for θ74, θ75, and θ76 are 0.016, 34.9, and 0.2, respectively.
16
55.566.577.58Model Value55.566.577.58Surrogate PredictionLinear SurrogateLinear Surrogate120406080100120140160Parameter Index00.020.040.060.080.1Importance Measure55.566.577.58Model Value55.566.577.58Surrogate PredictionPC SurrogatePC Surrogate61626374757677787994116119120141Parameter Index00.10.20.3Total Sobol' Index Fig. 6: Average ECS potassium QoI with experimental pulse stimulus. From left to right and top to
bottom: linear surrogate predictions, linear surrogate importance measure, PC surrogate predictions,
total Sobol' indices for PC surrogate.
The results for this specific QoI are similar for both the rectangular pulse and experimental pulse
stimulus. In both cases, the QoI is approximated with reasonable accuracy by a linear surrogate and with
higher accuracy by the PC surrogate. The most important parameters are shared in both cases. Notice
that parameter θ120, defined in (4.1), appears to be important in the linear surrogate but unimportant
in the PC surrogate. This is because it is strongly correlated with parameter θ121 (also defined in defined
in (4.1)), see Figure 3, and as a result the coefficient in the linear surrogate may be very large because
its effect is offset by the effect of θ121. To decorrelate inputs, the PC surrogate is build with only θ120
instead of both θ120 and θ121. It subsequently has minimal importance.
17
55.566.577.58Model Value55.566.577.58Surrogate PredictionLinear SurrogateLinear Surrogate120406080100120140160Parameter Index00.020.040.060.080.1Importance Measure55.566.577.58Model Value55.566.577.58Surrogate PredictionPC SurrogatePC Surrogate61626374757678849498116117119120122141Parameter Index00.10.20.3Total Sobol' Index Parameter
Identification in Supplementary Material Total Sobol' Index (exp.) Total Sobol' Index (rect.)
θ62
θ61
θ63
θ75
θ76
Nominal value 0.143 in equation (61)
gK,leakd in equation (30)
Nominal value 5.67 in equation (61)
Nominal value 34.9 in equation (54)
Nominal value 0.2 in equation (54)
0.3732
0.2409
0.2297
0.1112
0.0247
0.3738
0.2074
0.2449
0.1124
0.0198
Table 1: Five most influential parameters for the average ECS potassium QoI when the experimental
pulse stimulus is applied. The leftmost column is the parameter, the left-center column identifies the
parameter in the Supplementary Material, the right-center column is the total Sobol' index computed
for the parameter using the experimental pulse stimulus, and the right column is the total Sobol' index
computed for the parameter using the rectangular pulse stimulus.
We also analysed another QoI for the ECS potassium, namely, its maximum over the interval of
stimulation. The results are not reported because of their similarity to the average ECS potassium QoI
results given above.
The remaining two QoIs also present very similar results for both the rectangular pulse and ex-
perimental pulse stimulus. In the interest of conciseness, we only present figures corresponding to the
experimental pulse stimulus for these two QoIs; Tables 2 and 3 give results for both stimuli.
4.3 Average Volumetric Flow Rate
Figure 7 displays results for the average volumetric flow rate in the cerebral tissue defined by (2.2) in
the same manner as Figures 5 and 6. Table 2 reports the five most important parameters and their total
Sobol' indices. Unsurprisingly, the parameter list contains values found in the SMC/EC compartment of
the full model. However, the topmost parameter, θ141, is associated with the conductance of the inwardly
rectifying SMC KIR channel, gKIR, defined as a function of both membrane potential vSM C and the
18
K+concentration in the perivascular space [K +]P V S, given by
gKIR = exp(cid:0)θ142vSM C + θ140[K +]P V S − θ141
(cid:1)
(4.4)
as shown in [7] fitting to the data of [11]. θ141 shifts the conductance to the right for constant [K +]P V S
concentration in the perivascular space. The second parameter in Table 2, θ158 (found in the wallme-
chanics section of the model) determines the strength of influence of cytosolic [Ca2+] in determining
the reaction rate of phosphorylation of myosin [13]. Although not especially important, the third listed
parameter θ139 shifts the Nernst potential for the KIR channel to the right in the equation
vKIR = θ138[K +]P V S − θ139.
(4.5)
We again observe reasonably accurate fits by a linear surrogate and improved accuracy by a PC sur-
rogate. The surrogate predictions and important parameters for the rectangular pulse and experimental
pulse closely agree. As in Subsection 4.2, parameter θ120 appears important in the linear surrogate but
unimportant in the PC surrogate.
Parameter
Identification in Supplementary Material Total Sobol' Index (exp.) Total Sobol' Index (rect.)
θ141
θ158
θ139
θ136
θ142
z4 in equation (149)
ncross in equation (214)
z2 in equation (148)
GKi in equation (149)
z5 in equation (149)
0.4420
0.3544
0.0436
0.0362
0.0305
0.4561
0.3311
0.0529
0.0305
0.0351
Table 2: Five most influential parameters for the average volumetric flow rate QoI when the experimental
pulse stimulus is applied. The leftmost column is the parameter, the left-center column identifies the
parameter in the Supplementary Material, the right-center column is the total Sobol' index computed
for the parameter using the experimental pulse stimulus, and the right column is the total Sobol' index
computed for the parameter using the rectangular pulse stimulus.
19
Fig. 7: Average volumetric flow rate QoI with experimental pulse stimulus. From left to right and top
to bottom: linear surrogate predictions, linear surrogate importance measure, PC surrogate predictions,
total Sobol' indices for PC surrogate.
4.4 [AM + AMp]min
Figure 8 displays results for the minimum of the combined concentration of the actin myosin complex
with the experimental pulse stimulus. Table 3 reports the five most important parameters and their
total Sobol' indices. By comparing Table 3 and Table 2, we see that the [AM + AMp]min and average
volumetric flow rate QoIs share four of their five most important parameters. Although at first analysis
this should not be surprising it does suggest that the reaction rates of the actin myosin model for vessel
contraction/dilation are relatively insensitive to finding the minimum of the contraction force and that
the KIR ion channel is a vital component of the model.
20
1234Model Value11.522.533.54Surrogate PredictionLinear SurrogateLinear Surrogate120406080100120140160Parameter Index00.050.1Importance Measure1234Model Value11.522.533.54Surrogate PredictionPC SurrogatePC Surrogate14376275120135136137138139140141142154156157158160Parameter Index00.10.20.30.4Total Sobol' Index Fig. 8: [AM + AMp]min QoI with experimental pulse stimulus. From left to right and top to bottom:
linear surrogate predictions, linear surrogate importance measure, PC surrogate predictions, total Sobol'
indices for PC surrogate.
The linear surrogate does not perform as well for this QoI; however, the PC surrogate is far more
accurate; this highlights the nonlinearity of this particular QoI.
As in the previous results, parameter θ120 appears important in the linear surrogate and less important
in the PC surrogate, albeit, it is more important for this QoI than the previous ones.
5 Discussion
We note at the outset that the sensitivity analysis investigates the numerical model rather than the
physiological one. However, since the numerical model has been validated with experimental data [7, 30],
the results indicate parameters (and therefore areas) of importance for physiologists and modellers. The
21
00.10.20.30.40.5Model Value00.10.20.30.40.5Surrogate PredictionLinear SurrogateLinear Surrogate120406080100120140160Parameter Index00.050.10.15Importance Measure00.10.20.30.40.5Model Value00.10.20.30.40.5Surrogate PredictionPC SurrogatePC Surrogate16626375120135136137138139140141142154156157158160Parameter Index00.10.20.30.40.50.6Total Sobol' Index Parameter
Identification in Supplementary Material Total Sobol' Index (exp.) Total Sobol' Index (rect.)
θ141
θ158
θ139
θ142
z4 in equation (149)
ncross in equation (214)
z2 in equation (148)
z5 in equation (149)
θ120
Nominal value 5.5 in equation. (10)
0.6242
0.1239
0.0918
0.0644
0.0571
0.6203
0.1488
0.0954
0.0629
0.0240
Table 3: Five most influential parameters for the [AM + AMp]min QoI when the experimental pulse
stimulus is applied. The leftmost column is the parameter, the left-center column identifies the parameter
in the Supplementary Material, the right-center column is the total Sobol' index computed for the
parameter using the experimental pulse stimulus, and the right column is the total Sobol' index computed
for the parameter using the rectangular pulse stimulus.
integration times t1, t2 are the start and end of neuronal stimulation. Decreasing or increasing these
time does not significantly alter the ranking of the parameters. We discuss below results for each QoI
individually.
5.1 Average ECS Potassium
It is interesting to note for this particular quantity of interest that the first and third most important
parameters are associated not with a K+channel but the persistent Na+channel, NaP. The neuron model
used in this analysis was developed from the work of Kager et al [18] and Chang et al [5]. The differential
equation governing m4, the activation variable for the NaP channel, is given by equation (4.2).
The inactivation of the NaP channel is very small compared to activation and therefore these param-
eters make little impact on the extracellular K+. Scatter plots for the average of K+in the ECS against
the parameters θ61 and θ62 are shown in Figure 9. For each parameter, there is a linear trend where
increasing the parameter yields either an increase or decrease in the K+ECS. The characteristic time for
the activation variable for the NaP channel is defined as τ =
1
m4α+m4β
which, by using equation (4.2),
is constant (6 ms). The scatter plots also indicate this definition. These results show that a variation of
22
Fig. 9: Left: scatter plot of θ62 and the average ECS potassium (2.1); right: scatter plot of θ63 and the
average ECS potassium (2.1).
±10% in either θ62 or θ63 can either reduce or increase the extracellular K+by approximately 15%. We
should note that these results do not take into account the spatial buffering carried out by the astrocytic
syncytium which may have a significantly greater effect on the extracellular K+[4, 21].
The main driver for K+in the extracellular space is the K+Na+ATP-ase pump which has the general
form given by (in the dendrite)
(cid:16)
(N a+
d )3
d + N a+
N a+
d,baseline
(cid:17)3
(cid:16)
(K +
e )2
K +
e + K +
e,baseline
(cid:17)2 .
(5.1)
The ATP-ase pumps out Na+and in K+in the ratio of 3 to 2, hence the Na+in the dentrite has a large
effect on the extracellular K+as seen in equation (5.1) and strengthens the result that the NaP channel
has the most effect on the K+. One could have expected the NaT (transient Na+) channel to be prominent
however, it has a fast inactivation variable and, although it produces a larger flux, does so over a shorter
time.
5.2 Average Volumetric Flow Rate
In the presented model, the main pathway for neurovascular coupling is that of the K+pathway. Here
the astrocyte takes up K+ from the synaptic cleft and provides an efflux into the perivascular space
(PVS) via the BK channel. The smooth muscle cell detects this increase in PVS K+and through the in-
wardly rectifying channel, KIR, hypopolarises the SMC, shutting off the voltage mediated Ca2+channel.
23
0.130.140.150.16c155.566.577.58ECS Kmean+55.566.5c255.566.577.58ECS Kmean+ Ca2+is therefore reduced and the SMC dilates. As stated in the Section 4, the most important param-
eter is the shift parameter θ141 in the KIR conductance (4.4) determining the magnitude of ion flux
per unit change in the membrane potential away from the equilibrium (Nernst potential). Compared to
the other parameters defining the KIR conductance, θ141 is large with a nominal value of 12.6 whereas
θ140 = 4.2 × 10−4µM−1 and θ142 = −7.4 × 10−2mV −1. Hence, for constant membrane potential and
K+, variations in θ141 produce exponentially large variations in the KIR channel conductance allowing
substantial efflux of Ca2+from the SMC and a dilation of the vessel(cid:0) dR
dt > 0(cid:1).
The second most important parameter is the power index for the cytosolic SMC Ca2+which mediates
the four state latch model of Hai and Murphy [13] that is used in this model, in particular the rate of
phosphorylation of myosin and the actin-myosin complex. Variations in Ca2+as predominantly dictated
by the KIR channel will therefore have a direct effect on the dilation/contraction properties of the SMC
and hence the perfusing vessel radius. The remaining three parameters have significantly lower total
Sobol' indices and therefore only make small contributions to the QoI.
5.3 [AM + AMp]min
We see similar parameters appearing as most important for this QoI and the average volumetric flow
rate QoI, which is to be expected given the strong relationship between the radius dilation/contraction
phenomenon and the total quantity of actin/myosin complex and its phosphorylated compliment. In fact,
this QoI serves as a test of the statistical mechanism in that the only two non-repeating parameters (in
the list of five most important) between the volumetric flow rate QoI and the actin/myson complex QoI
are θ120 and θ136, which, as noted above, are significantly less import than the leading parameters.
6 Conclusion
A three stage methodology is presented for global sensitivity analysis of numerical cell models with a
large number of parameters. To the authors knowledge, this is the first type of analysis which investigates
neuroscience models of such size. The analysis investigated three quantities of interest pertaining to a
24
numerical model of neurovascular coupling. The results indicated several prevalent features of the model.
A significant influence of the persistent N a+ channel activation variable on the average extracellular
space K +, a two parameter set (the inwardly rectifying K + channel shift parameter and the index for
the cytosolic Ca2+) which characterizes most of the variability in the average volumetric flow rate, and
strong similarities between the most influential parameters for the average volumetric flow rate and
minimum value of the combined actin/myocin complex.
In addition to the results reported in this article, four other QoI's were considered. Two of them, the
maximum and average potassium concentration in the Astrocyte, were omitted because the parameter
to QoI mapping is nearly constant and hence global sensitivity analysis is not necessary. Specifically,
the mean of the QoI is approximately 37 times larger than its standard deviation for both cases. The
other two unreported QoIs correspond to lag times. The first being the duration of time between the
application of the stimulus and the minimum of the phosphorylated actin myosin complex, and the
second being the duration of time between the application of the stimulus and the maximum of the
radius. In both cases, the QoI exhibited highly nonlinear behavior which we were unable to approximate
with linear or PC surrogates trained on the existing data. In fact, fitting such nonlinearities would likely
require more samples than is computationally feasible for this model. The linear surrogate had 59%
and 47% relative L2 errors for these two QoIs, respectively. Our global sensitivity analysis methodology
was unsuccessful because the linear surrogate was an unreliable tool for screening. Defining the QoI
as the maximum/minimum value instead of the time lag makes the analysis more tractable. These
maximum/minimum value QoIs were also considered and yielded similar results to the QoIs reported in
the article.
For a given model and collection of samples, the methodology presented in this article may be appli-
cable for some QoIs and intractable for others. The success of our method depends upon the surrogate
models being sufficiently accurate. A general principle is that QoIs defined as averages will be more
amenable for analysis than, for instance, minimum values or lag times. A practical benefit of our method
is that any QoI may be considered without requiring additional model evaluations. The sampling and
ODE solves are executed once, followed by computing the QoIs and performing global sensitivity analysis,
which may be easily repeated for many different QoIs.
25
Appendix
Determining the coefficients cα in the Polynomial Chaos surrogate (3.3) is challenging. Ideally, one would
solve the least squares problem
(cid:32)
g(θk) −(cid:88)
α
M(cid:88)
k=1
min
(cid:33)2
cαψα( θ)
(A.1)
to determine the coefficients. This approach is not currently feasible for the problems considered in
this article. If there are, for instance, 18 input variables (θji 's) then a 3rd degree polynomial has 1330
unknown coefficients and a 4th degree polynomial has 7315 unknown coefficients. With less than 1000
sample points, as in our case, (A.1) will admit infinitely many solutions which interpolate the data but
will yield poor approximations of the QoI. Rather, we seek an approximate solution of (A.1) for which
most of the coefficients are exactly 0. This may be achieved by introducing a penalty term and solving
M(cid:88)
(cid:32)
g(θk) −(cid:88)
min
(cid:33)2
(cid:88)
cαψα( θ)
+ λ
cα
(A.2)
k=1
α
α
instead of (A.1). Adding the sum of absolute values of the coefficients encourages a sparse solution,
i.e. one with many 0 coefficients. However, it comes at the cost of making the objective function non-
differentiable and hence (A.2) requires a more sophisticated optimization approach in comparison to
(A.1). A plurality of well documented methods exist for solving (A.2). In this article we use Least Angle
Regression (LAR) [9] with its implementation in [26], and a maximum polynomial degree of 5.
Because the basis function of the Polynomial Chaos surrogate are orthogonal with respect to the
PDF p θ, the variance and conditional expectation in (3.6) may be computed analytically as a function
of the coefficients. Hence the total Sobol' indices of the Polynomial Chaos surrogate are given in closed
form as a function of the coefficients.
References
1. Alexanderian A, Gremaud P, Smith R (2018) Variance-based sensitivity analysis for time-dependent
processes, under revision, https://arxiv.org/abs/1711.08030
2. Ames A (2000) CNS energy metabolism as related to function. Brain Research Reviews 34:42 -- 68
26
3. Attwell D, Buchan AM, Charpak S, Lauritzen M, MacVicar Ba, Newman Ea (2010) Glial and
neuronal control of brain blood flow. Nature 468(7321):232 -- 243
4. Bellot-Saez A, K´ekesi O, Morley JW, Buskila Y (2017) Astrocytic modulation of neuronal excitability
through K + spatial buffering. Neuroscience and Biobehavioral Reviews 77:87 -- 97
5. Chang JC, Brennan KC, He D, Huang H, Miura RM, Wilson PL, Wylie JJ (2013) A mathematical
model of the metabolic and perfusion effects on cortical spreading depression. PloS One 8(8):e70469
6. Dormanns K, Brown RG, David T (2015) Neurovascular coupling: a parallel implementation. Fron-
tiers in Computational Neuroscience 9(September):1 -- 17
7. Dormanns K, van Disseldorp EMJ, Brown RG, David T (2015) Neurovascular coupling and the
influence of luminal agonists via the endothelium. Journal of Theoretical Biology 364:49 -- 70
8. Dormanns K, Brown R, David T (2016) The role of nitric oxide in neurovascular coupling. Journal
of Theoretical Biology 394:1 -- 17
9. Efron B, Hastie T, Johnstone I, Tibshirani R (2004) Least angle regression. Annals of Statistics
32(11):407 -- 499
10. Farr H, David T (2011) Models of neurovascular coupling via potassium and EET signalling. Journal
of Theoretical Biology 286(1):13 -- 23
11. Filosa JA, Bonev AD, Straub SV, Meredith AL, Wilkerson MK, Aldrich RW, Nelson MT (2006)
Local potassium signaling couples neuronal activity to vasodilation in the brain. Nature Neuroscience
9(11):1397 -- 1403
12. Filosa JA, Blanco VMVM, Filosa JA Blanco VM, Filosa JA, Blanco VMVM (2007) Neurovascular
coupling in the mammalian brain. Experimental Physiology 92(4):641 -- 646
13. Hai CM, Murphy RA (1988) Cross-bridge phosphorylation and regulation of latch state in smooth
muscle. American Journal of Physiology - Cell Physiology 254(1):C99 -- C106
14. Hart J, Gremaud P (2018) An approximation theoretic perspective of Sobol' indices with dependent
variables, to appear in Int. J. for Uncertainty Quant.
15. Hart J, Alexandrian A, Gremaud P (2017) Efficient computation of Sobol' indices for stochastic
models. SIAM J Sci Comput 39:A1514 -- A1530
27
16. Hsieh NH, Reisfeld B, Bois FY, Chiu WA (2018) Applying a global sensitivity analysis workflow to
improve the computational efficiencies in physiologically-based pharmacokinetic modeling. Frontiers
in Pharmacology 9(588)
17. Iooss B, Saltelli A (2017) Introduction: Sensitivity analysis. In: Ghanem R, Higdon D, Owhadi H
(eds) Handbook of Uncertainty Quantification, Springer, pp 1103 -- 1122
18. Kager H, Wadman WJ, Somjen GG (2000) Simulated seizures and spreading depression in a neuron
model incorporating interstitial space and ion concentrations. Journal of Neurophysiology 84(1):495 --
512
19. Kenny A, Plank MJ, David T (2017) The role of astrocytic calcium and TRPV4 channels in neu-
rovascular coupling. Journal of Computational Neuroscience doi:10.1007/s10827 -- 017 -- 0671 -- 7, DOI
10.1007/s10827-017-0671-7
20. Kenny A, Plank MJ, David T (2017) The role of astrocytic calcium and TRPV4 channels in neu-
rovascular coupling. Journal of Computational Neuroscience (in press) 41(1):97 -- 114
21. Kenny A, Plank M, David T (2018) Macro scale modelling of cortical spreading depression and the
role of astrocytic gap junctions. Journal of Theoretical Biology 458:78 -- 91
22. Leithner C, Royl G, Offenhauser N, Fuchtemeier M, Kohl-Bareis M, Villringer A, Dirnagl U, Lindauer
U (2010) Pharmacological uncoupling of activation induced increases in CBF and CMRO2. Journal
of Cerebral Blood Flow and Metabolism: Official Journal of the International Society of Cerebral
Blood Flow and Metabolism 30(2):311 -- 322
23. Lindauer U, Leithner C, Kaasch H, Rohrer B, Foddis M, Fuchtemeier M, Offenhauser N, Steinbrink J,
Royl G, Kohl-Bareis M, Dirnagl U (2010) Neurovascular coupling in rat brain operates independent
of hemoglobin deoxygenation. Journal of Cerebral Blood Flow & Metabolism 30(4):757 -- 768
24. Longden TA, Hill-Eubanks DC, Nelson MT (2015) Ion channel networks in the control of cerebral
blood flow. Journal of Cerebral Blood Flow & Metabolism 36:492 -- 512
25. Makani S, Chesler M (2010) Rapid rise of extracellular pH evoked by neural activity is generated by
the plasma membrane calcium ATPase. Journal of Neurophysiology 103(2):667 -- 676
26. Marelli S, Sudret B (2014) UQLab: A framework for uncertainty quantification in Matlab. 2nd Int
Conf on Vulnerability, Risk Analysis and Management
28
27. Marin JM, Pudlo P, Robert CP, Ryder RJ (2012) Approximate Bayesian computational methods.
Statistics and Computing 22(6):1167 -- 1180
28. Mathias EJ, Plank MJ, David T (2017) A model of neurovascular coupling and the BOLD response:
PART I. Computer Methods in Biomechanics and Biomedical Engineering 20(5):508 -- 518
29. Mathias EJ, Plank MJ, David T (2017) A model of neurovascular coupling and the BOLD response:
PART II. Computer Methods in Biomechanics and Biomedical Engineering 20(5):519 -- 529
30. Mathias EJ, Kenny A, Plank MJ, David T (2018) NeuroImage Integrated models of neurovascular
coupling and BOLD signals: Responses for varying neural activations. NeuroImage 174(February):69 --
86
31. Mintun MA, Lundstrom BN, Snyder AZ, Vlassenko AG, Shulman GL, Raichle ME (2001) Blood
flow and oxygen delivery to human brain during functional activity: Theoretical modeling and ex-
perimental data. Proc Natl Acad Sci USA 98(12):6859 -- 6864
32. Owen A (2014) Sobol' indices and Shapley values. SIAM/ASA J Uncertainty Quant 2:245 -- 251
33. Powers WJ, Hirsch IB, Cryer PE (1996) Effect of stepped hypoglycemia on regional cerebral blood
flow response to physiological brain activation. Am J Physiol 270(2 Pt 2):H554 -- -9
34. Roy CSCS, Sherrington CSS (1890) On the regulation of the blood-supply of the brain. The Journal
of Physiology 11(1-2):85
35. Saltelli A, Ratto M, Andres T, Campolongo F, Cariboni J, Gatelli D, Saisana M, Tarantola S (2008)
Global sensitivity analysis: The primer. Wiley
36. Saltelli A, Annoni P, Azzini I, Campolongo F, Ratto M, Tarantola S (2010) Variance based sensitivity
analysis of model output. Design and estimator for the total sensitivity index. Computer Physics
Communications 181:259 -- 270
37. Sobie EA (2009) Parameter sensitivity analysis in electrophysiological models using multivariable
regression. Biophysical Journal 96:1264 -- 1274
38. Tennøe S, Halnes G, Einevoll GT (2018) Uncertainpy: A Python toolbox for uncertainty quantifica-
tion and sensitivity analysis in computational neuroscience, http://dx.doi.org/10.1101/274779
39. Witthoft A, Filosa JA, Karniadakis GE (2013) Potassium buffering in the neurovascular unit: models
and sensitivity analysis. Biophysical Journal 105(9):2046 -- 2054
29
40. Zheng Y, Pan Y, Harris S, Billings S, Coca D, Berwick J, Jones M, Kennerley A, Johnston D, Martin
C, Devonshire IM, Mayhew J (2010) A dynamic model of neurovascular coupling: Implications for
blood vessel dilation and constriction. NeuroImage 52(3):1135 -- 1147
30
|
1907.00263 | 1 | 1907 | 2019-06-29T19:28:25 | A Power Efficient Artificial Neuron Using Superconducting Nanowires | [
"q-bio.NC",
"cond-mat.supr-con",
"cs.NE"
] | With the rising societal demand for more information-processing capacity with lower power consumption, alternative architectures inspired by the parallelism and robustness of the human brain have recently emerged as possible solutions. In particular, spiking neural networks (SNNs) offer a bio-realistic approach, relying on pulses analogous to action potentials as units of information. While software encoded networks provide flexibility and precision, they are often computationally expensive. As a result, hardware SNNs based on the spiking dynamics of a device or circuit represent an increasingly appealing direction. Here, we propose to use superconducting nanowires as a platform for the development of an artificial neuron. Building on an architecture first proposed for Josephson junctions, we rely on the intrinsic nonlinearity of two coupled nanowires to generate spiking behavior, and use electrothermal circuit simulations to demonstrate that the nanowire neuron reproduces multiple characteristics of biological neurons. Furthermore, by harnessing the nonlinearity of the superconducting nanowire's inductance, we develop a design for a variable inductive synapse capable of both excitatory and inhibitory control. We demonstrate that this synapse design supports direct fanout, a feature that has been difficult to achieve in other superconducting architectures, and that the nanowire neuron's nominal energy performance is competitive with that of current technologies. | q-bio.NC | q-bio | A Power Efficient Artificial Neuron Using Superconducting
Nanowires
Emily Toomey1, Ken Segall2, Karl K. Berggren*1
1Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science,
Cambridge, MA 02139, USA
2Colgate University, Department of Physics, Hamilton, NY 13346, USA
*contact: [email protected]
ABSTRACT
With the rising societal demand for more information-processing capacity with lower power
consumption, alternative architectures inspired by the parallelism and robustness of the human brain have
recently emerged as possible solutions. In particular, spiking neural networks (SNNs) offer a bio-realistic
approach, relying on pulses analogous to action potentials as units of information. While software
encoded networks provide flexibility and precision, they are often computationally expensive. As a result,
hardware SNNs based on the spiking dynamics of a device or circuit represent an increasingly appealing
direction. Here, we propose to use superconducting nanowires as a platform for the development of an
artificial neuron. Building on an architecture first proposed for Josephson junctions, we rely on the
intrinsic nonlinearity of two coupled nanowires to generate spiking behavior, and use electrothermal
circuit simulations to demonstrate that the nanowire neuron reproduces multiple characteristics of
biological neurons. Furthermore, by harnessing the nonlinearity of the superconducting nanowire's
inductance, we develop a design for a variable inductive synapse capable of both excitatory and inhibitory
control. We demonstrate that this synapse design supports direct fanout, a feature that has been difficult to
achieve in other superconducting architectures, and that the nanowire neuron's nominal energy
performance is competitive with that of current technologies.
Keywords: artificial neuron, superconductor, nanowire, spiking neural network, artificial synapse
1. INTRODUCTION
The human brain has long been a subject of fascination due to the wide variety of complex operations
made possible by groups of a single component -- the neuron. Now, as computation needs are rapidly
approaching the limits of traditional von Neumann architectures, the neuron's unique features have led it
to become a source of inspiration for new directions for the advancement of computing. Unlike
conventional computing schemes, the human brain benefits from characteristics such as extensive
parallelism and robustness to errors, allowing it to operate efficiently despite slow speeds on the order of
a few Hz. These appealing qualities have spurred the development of technologies that use the brain as a
platform for information processing, ranging from small scale modeling of single neuron dynamics to
large-scale parallel computing.
At the heart of this concept are Spiking Neural Networks (SNNs), which seek to mimic the
spiking dynamics of the brain in order to encode information, with the additional benefit of possibly
offering new insight into the brain's functionality. In this scheme, spikes serve as the tokens of
information, while neurons act analogously to logic gates, producing a single output in response to a
combination of multiple inputs [1]. Past approaches to SNNs vary both in degree of bio-realism and in
how the spikes are implemented. While software approaches that hard-code spiking dynamics offer
flexibility and precision, they are computationally expensive. As a result, other SNNs have used a
hardware implementation, relying on devices with intrinsic dynamics that replicate neuron behavior as a
means of reducing computation costs. Hardware approaches have been explored in a wide variety of
platforms, including CMOS [2], magnetic materials [3], memristors [4] [5], and superconducting
Josephson junctions (JJs) [6] [7]. CMOS circuits provide large-scale integration, but are heavy on power
dissipation and need many components to achieve biological realism. Memristors and magnetic materials
have device characteristics which emulate neural dynamics, but can be slow and are also not energy
efficient. Josephson junctions are a fast and energy-efficient technology, but need more components in
large networks due to poor fan-out/fan-in properties; furthermore, their weak action potentials are unable
to be viewed directly, making diagnostics difficult.
These developing technologies highlight several characteristics that are critical for an artificial
neuron: (1) inherent device dynamics that are capable of producing spiking behavior; (2) tunable synapses
between neurons to allow for the expansion into larger networks with adjustable connectivity; and (3) low
power dissipation, both in the dynamic firing state of the neuron and in the static state. As has been
emphasized in prior literature[2], optimizing static power dissipation is particularly vital to the goal of
creating an energy efficient network.
Here we propose an artificial neuron based on superconducting nanowires operating at cryogenic
temperatures. Superconductors are prime candidates for generating low-power spiking behavior due to
their inherent nonlinearity and negligible static power dissipation. Building on an architecture first
implemented in Josephson junctions [6], we rely on the coupling between two shunted nanowires to act
analogously to a two-channel neuron for an action potential. We start by describing the nonlinear
dynamics of superconducting nanowires, and then present the architecture of the nanowire-based neuron.
Using electrothermal circuit simulations, we demonstrate that the device is able to replicate several
behaviors of a single neuron. Finally, we present a synapse design and discuss the advantages of the
nanowire neuron, including fanout and an energy figure of merit two orders of magnitude better than that
of competing technologies, which are critical in moving towards the parallelism of the human brain.
2. THE NANOWIRE NEURON MODEL
Superconducting nanowires possess an inherent nonlinearity that serves as the building block of our
artificial neuron model. We begin by briefly describing this nonlinearity in a single nanowire, and then
present the nanowire neuron circuit and its basic operation principles.
2.1 Relaxation Oscillations
The intrinsic nonlinearity of superconducting nanowires makes them ideal candidates for the hardware
generation of spiking behavior. When a bias current flowing through a superconducting nanowire exceeds
a threshold known as the critical current (Ic), superconductivity breaks down and the nanowire becomes
resistive, producing a voltage. The nanowire only switches back to the superconducting state once the
bias current is reduced below a level called the retrapping current (Ir), and the resistive portion (the
"hotspot") cools down. When the nanowire is placed in parallel with a shunt resistor, this switching
process participates in electrothermal feedback with the shunt, producing relaxation oscillations[8].
Relaxation oscillations can be viewed in the context of a simplified action potential. Like the Na+
influx and K+ outflux currents of a neuron, the influx and outflux currents from the nanowire to the shunt
resistor are governed by different timescales, τ1 and τ2. As shown in Figure 1a, the rising edge of the
output voltage is defined by τ1=L/(Rs+Rhs), where L is the inductance of the nanowire, Rs is the shunt
resistance, and Rhs is the resistance of the nanowire hotspot, usually on the order of ~ 1-10 kΩ.
Conversely, the outflux current occurs when the nanowire is no longer resistive and the bias current is
redirected from the shunt; this reduced resistance results in a slower time constant τ2= L/Rs which defines
the falling edge of the output voltage. For typical nanowires devices, τ1 ~ 100 ps and τ2 ~ 1 ns. The two
currents are "gated," as shown by the insets in Fig. 1a, by the state of the nanowire -- when the state is
resistive (producing a voltage), the influx current flows into the shunt, and when it is superconducting, the
outflux current flows back to the nanowire.
The inductance of superconducting nanowires is dominated by an intrinsic material property
known as kinetic inductance [9], and is defined per unit length of the structure. As a result, it is possible
to tune these time constants by changing the length of the nanowire, with a longer wire leading to a higher
inductance and thus a longer timescale. Figure 1b shows a scanning electron micrograph of a typical
superconducting nanowire with a meandering geometry designed for maximizing the total device
inductance. An example of experimentally observed relaxation oscillations for such a device is displayed
in Figure 1c.
Figure 1: Relaxation oscillations in superconducting nanowires, which serve as the foundation of the
nanowire neuron's spiking behavior. (a) Simplified model of a relaxation oscillation. The output voltage is
defined by two time constants. The rising edge occurs when the superconducting nanowire has switched into the
resistive state (Rhs > 0), and the bias current is redirected to the shunt resistor. The falling edge takes place after the
nanowire regains superconductivity (Rhs = 0), and the bias current is redirected from the shunt resistor. (b) Example
of a superconducting nanowire in a long, meandered design for obtaining a high kinetic inductance. (c)
Experimentally measured relaxation oscillations in a long superconducting nanowire shunted by 50 Ω.
2.2 The Neuron Model
Although a shunted nanowire on its own produces oscillations analogous to action potentials, as the bias
current increases, the output signal eventually accumulates a voltage offset. This effect deviates from true
neuron behavior, as the cell must maintain a constant resting potential (approximately -70 mV). To
overcome this difference, we have implemented a neuron architecture based on one that was first
proposed for Josephson junctions [6], as shown in Figure 2. The circuit consists of two shunted
nanowires -- the main oscillator and the control oscillator -- linked together in a superconducting loop. A
bias current Ibias is applied to both oscillators such that they are each biased right below their critical
currents, but in opposite directions. To trigger an action potential, a small input current pulse Iin (Fig. 2a)
is applied and sums with the bias current to exceed Ic of the main oscillator, causing it to switch (Fig. 2d).
The control does not fire since the input opposes the direction of its bias.
Once the main oscillator switches, current is added to the superconducting loop in the
counterclockwise direction (Fig. 2b), which sums with the bias current to fire the control oscillator (Fig.
2c). The control oscillator removes counterclockwise current from the loop, allowing the main oscillator
to fire again. Without the presence of the control oscillator, the main oscillator would only be able to fire
once, since the counterclockwise current added to the loop would reduce the total current through the
nanowire of the main oscillator below its Ic. The voltage from the main oscillator node (Fig. 2e) serves as
the spiking output that is carried down to the next neuron via the synapse. Unlike the output from the
single shunted nanowire, the output of the two-nanowire circuit does not accumulate a bias offset, making
it a suitable spiking signal.
In the context of the two-channel neuron model, the main oscillator acts analogously to the Na+
influx current by adding flux to the superconducting loop in the form of a circulating current. The control
oscillator acts analogously to the K+ outflux current by reducing the circulating current, resetting the
neuron and allowing the main oscillator to fire again. As described in Ref. [8], the rate at which each
oscillator fires depends on the magnitude of the bias current, paralleling the voltage-dependent rate
constants of ion gates in the Hodgin-Huxley model [10].
Figure 2: Circuit simulations of the two-nanowire soma, where the two oscillators act analogously to the two
ion channels in the simplified neuron model. (a) Input pulse, Iin = 4 µA. (b) Current through the loop inductor. (c)
Current through the control nanowire. The control nanowire reduces the amount of counter-clockwise current
circulating in the loop, allowing the main nanowire to fire again. (d) Current through the main nanowire. (e) Output
voltage pulse that is sent to the synapse. For these simulations, the critical current of the control nanowire is Ic,control
= 30 µA, the critical current of the main nanowire is Ic,main = 30 µA, and Ibias = 58.6 µA.
3. SINGLE NEURON CHARACTERISTICS
Neurons display a wide variety of traits unique to certain populations, allowing them to collectively
interact to achieve varied and complex tasks. While no single neuron possesses all possible traits, the
basic functionality of an artificial neuron can be evaluated by demonstrating some common bio-realistic
characteristics. Here we present multiple neuron behaviors that can be achieved with the nanowire
neuron, using electrothermal circuit simulations conducted in LTSpice. The simulations implement
material-specific characteristics and nanowire hotspot dynamics as described in previous literature [11]
[12], and have been shown to reliably reproduce experimental data pertaining to nanowire relaxation
oscillations [8].
3.1 Threshold Response
A general characteristic of biological neurons is their inability to fire unless the input signal exceeds a
certain threshold. Figure 3a shows the threshold voltage response of the nanowire neuron when the bias
current is held constant and the input current is varied. As evident in the plot, the neuron does not begin
firing until the input current passes a threshold, defined by when the sum of the bias and input current
through the main nanowire exceed its Ic. Above the threshold, the peak voltage of the spike output is
essentially constant. However, as emphasized by Izhikevich [13], biological neurons have a threshold that
may be varied by previous activity, such as an inhibiting input that reduces it. Figure 3b and c illustrate
this process ("threshold variability") in the nanowire neuron; an initial subthreshold input pulse (Fig. 3b)
fails to elicit a spike, while a later input pulse of the same magnitude triggers a spike (Fig. 3c) after a
smaller negative (inhibitory) pulse reduces the firing threshold. It should be noted that when the preceding
pulse was of the opposite polarity (excitatory), no spike was triggered. This behavior is consistent with
the expectations of Ref. [13].
Figure 3: Firing threshold of the two-nanowire neuron. (a) Peak output voltage as a function of input current
under a constant bias (Ibias = 58.6 µA). The plot illustrates that the neuron does not spike until the input current
exceeds 4 µA, at which point it fires with output voltages of the same amplitude. Inset shows the time domain
voltage output of the neuron for different input currents. (b) Input to the neuron, leading to a reduction in firing
threshold by a preceding inhibitory pulse. (c) Spiking output of the neuron in response to the inputs of (b),
demonstrating that the nanowire neuron's firing threshold is variable. For this simulation, Ibias = 57.62 µA, the
excitatory inputs Iin = 4.6 µA, and the inhibitory input Iin - = -4.3 µA.
3.2 Refractory Period
In addition to exhibiting a firing threshold, the nanowire neuron displays a refractory period, which we
define to be the minimum time between two input pulses such that both pulses elicit a spike. Figure 4
illustrates this response. When two pulses are separated enough in time so that the main oscillator is
biased close to its critical current when the second input pulse arrives, then the second pulse will cause a
spike (Fig. 4a). However, if the second pulse arrives before the bias current has fully returned to the main
oscillator, then the sum of the second input pulse and the bias will not be sufficient to switch the nanowire
and trigger the neuron (Fig. 4b). As a result, the refractory period is limited by the time it takes to fully
bias the main oscillator again.
Figure 4: Refractory period of the two-nanowire neuron. (a) Response when there is sufficient time between two
inputs to each elicit a separate spike. Parameters: Ibias = 58 µA, Iin = 6 µA, Δt = 4 ns. The pink dashed lines indicate
the beginning of the rising edge of each pulse. (b) Response when there is insufficient time between two input
pulses, causing the neuron to fire only once. Parameters are the same as in (a), except Δt = 2 ns. For both cases,
panel (i) displays the current through the nanowire of the main oscillator, while panel (ii) displays the output voltage
of the neuron.
3.3 Near-Coincidence Detection
The neuron may also be under-biased such that two pulses in rapid succession are required to elicit a
spike. Figure 5 displays the response of the nanowire neuron to two input pulses with different time
delays Δt between them. In this case, in order for the second pulse to fire the neuron, the delay must be
less than 3 ns, as shown in Fig. 5b. This behavior is similar to the response of integrating neurons to high
frequency inputs [13], and demonstrates that the nanowire neuron may be used for near-coincidence
detection of pulses, depending on its biasing conditions.
Figure 5 : Near-coincidence detection of pulses. Output voltage of the neuron when the time between successive
input pulses Δt is a) 5 ns b) 3 ns. The pink dashed lines indicate the rising edge of each pulse. Two pulses must be in
rapid succession in order to fire the under-biased neuron, demonstrating that it may be used for near-coincidence
detection of input pulses. Parameters: Iin = 4.6 µA, Ibias = 57.62 µA.
3.4 Class I Behavior
Biological neurons differ in their response to varying signal strengths. Whereas Class I neurons have a
spiking frequency that increases with increasing input strength, Class II neurons maintain a constant firing
rate[13] [6]. Figure 6 illustrates the spiking behavior of the nanowire neuron at different levels of bias
current. Fig. 6a shows the time-domain voltage output of the neuron as the bias current is increased, and
suggests an increase in spiking frequency. This response is confirmed by observing the voltage output's
frequency spectrum displayed in Fig. 6b, which shows a shift in the spiking frequency to higher levels
with increasing bias. Consequently, the nanowire neuron has Class 1 behavior. The modulation of spiking
frequency by bias current demonstrates that the frequency of the nanowire neuron output may be used to
glean information about its input conditions.
Figure 6: Effect of bias current on spiking frequency. (a) Time domain simulations of the two-nanowire neuron
with different bias currents. Iin = 6 µA for all simulations. Traces have been shifted from one another in the y-axis
for clarity. (b) Fourier transform of the voltage output for each biasing condition. The shift in peak frequency with
bias current indicates that the circuit acts like a Class I neuron.
3.5 Parabolic Bursting
Some neurons, such as thalamic and dopaminergic neurons [14], display a unique mode of
behavior called bursting, in which the cell alternates between the resting and firing states. The transition
between states may be dictated by slow changes in low levels of intracellular calcium ions, which
influence the conductance of K+ [15]. As a result, a small, slowly varying signal controls the rapid
dynamics of the action potential, leading to alternating periods of resting and firing. This process is
considered to be an important aspect of electrical activity in the brain.
Due to the significance of bursting in biological neurons, past work has sought to replicate similar
behavior in platforms such as digital silicon models [16] and Josephson junction models [17] by injecting
a low-frequency ac signal into the system. Figure 7 shows the result when a similar technique is applied
to the nanowire neuron. In this case, a weak ac signal (f = 50 Hz, Iac = 4 µA) shown by the dashed red line
in Fig. 7a is coupled into the bias port of the neuron, causing the neuron to alternate between the resting
and firing states, as reflected in the resulting output voltage signal. A close examination of the timing
between adjacent spikes (see Fig. 7b) shows that the spiking frequency increases and then decreases over
the firing period, a phenomenon known as parabolic bursting[18]. This behavior was first observed in
neuron R15 of the abdominal ganglion of Aplysia [19] [20], and has since been demonstrated in many
other cells. The ability of the nanowire neuron to replicate similar dynamics may be therefore be useful
for performing a wider range of functions.
Figure 7: Parabolic bursting in the two-nanowire neuron. (a) Output voltage of the two-nanowire neuron when
the bias is coupled to a weak sinusoidal drive (f = 50 MHz, Iac = 4 µA). The red dashed curve indicates the
sinusoidal drive, shifted in the y-axis for clarity. (b) The inverse of the time between adjacent peaks shows that the
time difference follows a parabolic form. Parameters: Ibias = 59 µA, Iin = 6 µA.
3.6 Axon: Transmission Line Characteristics
After an action potential occurs in a biological neuron, the output signal propagates down the axon as if
sent through a delay line [1]. This delay is valuable in that it preserves time domain information,
potentially facilitating behaviors that rely on the recognition of specific spatio-temporal patterns [13].
Such pattern recognition is often not possible in SNNs with traditional wiring, since signals travel too
rapidly for timing information to be maintained [1]. This is not the case in superconducting transmission
lines. Recent work has shown that the high kinetic inductance of superconducting transmission lines, like
those made out of niobium nitride, results in propagation speeds of ~2% c, where c is the speed of light in
vacuum [21]. As illustrated in Figure 8, a simulated nanowire neuron output sent through a
superconducting transmission line model [22] is delayed by ~100-500 ps, close to the full width of an
action potential. In mammalian brains, the full width of the action potential is typically a few
milliseconds [23], and so by analogy the axonal delay shown in Figure 8 would also be on the millisecond
timescale. This delay would be on the order of typical delays in cortical systems, for example the cortico-
cortical delay in mammals [24]; if longer delays are needed, the transmission line can simply be made
longer.
Figure 8: A superconducting transmission line as an axon. Simulations of a superconducting transmission line
show that the spikes can be delayed on the order of ~0.5 ns, depending on the length of the structure. This could
enable the storage of timing information in addition to frequency information. Transmission line parameters:
nanowire inductance Ln = 0.3 nH/µm, nanowire capacitance Cn = 0.1 fF/µm, propagation speed v = 1.9%c,
transmission line length l = 2.5 mm. Shorter transmission lines on the order of 800 um still had delays of ~140 ps.
4. THE SYNAPSE
The collective dynamics of a neural network depend on the ability of a neuron to influence the behavior
of another downstream neuron via a synapse. Here we introduce an inductive synapse that can be
integrated with the nanowire neuron to facilitate downstream control. We start by demonstrating
excitatory and inhibitory control, and then present a scheme for tuning the synaptic strength.
4.1 The Inductive Synapse
Figure 9 illustrates the circuit schematic of an inductive synapse that may be implemented in the nanowire
neuron. Similar to the slow release of neurotransmitters in response to an action potential, the inductive
synapse relies on the slow charging of a large inductor in response to the nanowire neuron's more rapid
voltage spikes. The energy stored in the large synapse inductor is then discharged as current into the input
port of the target neuron, modulating its behavior.
Figure 9: Circuit schematic of the nanowire neuron with an inductive synapse. The output voltage of the neuron
charges the large synapse inductor Lsyn. The inductor discharges current into the input port of the downstream target
neuron, modulating its behavior. A series resistor Rseries is in place to reduce backaction into the main neuron.
The effect of the inductive synapse can be either excitatory or inhibitory based on the sign of the
bias current applied to the upstream (main) neuron. These two cases are shown in Figure 10, which
displays the voltage outputs of the main and target neurons, as well as the current through the synapse
inductor. In the excitatory case (Fig. 10a), the upstream neuron is positively biased such that the inductive
synapse discharges a positive current, causing an under-biased target neuron to fire. In the inhibitory case
(Fig. 10b), the upstream neuron is negatively biased, causing a negative current to be sent from the
synapse to the target, turning off a target that was biased in the firing state. Consequently, the inductive
synapse allows for both types of control between neurons.
Figure 10: Downstream control of the inductive synapse. (a) Excitatory control. Parameters: Ibias,main = 59 µA,
Rseries = 14 Ω, Rsyn,1 = 40 Ω, Lsyn = 0.265 µH, Rsyn,2 = 40 Ω, Ibias,target = 57.17 µA, Iin = 4.6 µA. (b) Inhibitory control.
Parameters Ibias,main = -58.6 µA, Rseries = 24 Ω, Rsyn,1 = 45 Ω, Lsyn = 0.23 µH, Rsyn,2 = 40 Ω, Ibias,target = 57.68 µA, Iin =
4.6 µA. For both cases: Panel (i) displays the output voltage of the main neuron, with the red dashed lines indicating
the rising and falling edge of the input signal; Panel (ii) displays the current through the synapse inductor; Panel (iii)
displays the output voltage of the downstream target neuron.
4.2 Variable Strength Synapse
Although the inductive synapse of Figure 10 can be engineered for both excitatory and inhibitory control
of a downstream neuron, a fundamental property of artificial neural networks is the ability to modulate
that control by adjusting synaptic strength. One possible scheme for implementing such variability in the
inductive synapse is to incorporate superconducting nanowires as a different circuit element: a tunable
inductor. A nanowire's kinetic inductance increases with increasing bias current, reaching an
enhancement of 10-20% near Ic [25]. This modulation has been incorporated into the circuit model of the
superconducting nanowire used in these simulations [12]. By placing a high inductance nanowire with an
ideal current source in parallel with the synapse inductor, the overall parallel inductance of the synapse
can be modulated, which in turn changes the amount of current sent to the target neuron. Figure 11 shows
the simulated results for the case of an inhibitory synapse. When a higher modulating current Imod is
applied to the nanowire inductor, the overall parallel inductance increases, reducing the amount of current
sent to the target. It is important to note that the polarity of the modulating current is not important, since
the change in kinetic inductance depends only on the magnitude of the modulating current in relation to
its Ic. For instance, Figure 11b illustrates that the modulation in synaptic current for Imod = 5 µA and Imod =
-5 µA is roughly the same. As a result, it is clear that the modulation is due to the change in kinetic
inductance, and not simply an injection of current by Imod in the opposing direction.
Figure 11: Modulating the inductive synapse. (a) Circuit schematic of the inductive synapse with a high-
inductance nanowire and ideal current source placed in parallel. L1, L2 << Lnanowire, Lsyn. Rseries, out has been added to
further prevent backaction from the target neuron. (b) Simulation of the current through Rseries,out in an inhibitory
synapse as a function of different modulation currents. Spikes represent backaction from the firing of the target
neuron. (Inset) Enlarged view of the boxed area. Higher modulation currents increase the overall synapse
inductance, reducing the amount of current sent to the target. The +/- 5 µA modulation currents have nearly the
same effect, showing that modulation is not polarity-dependent. Parameters: Rseries,in = 25 Ω, Rsyn,1 = 39 Ω, Rsyn,2 = 40
Ω, Rseries,out = 0.1 Ω, Lsyn = 0.45 µH, Lnanowire = 0.275 µH, Ic, nanowire = 6 µA, L1 = L2 = 50 pH, , Ibias,target = 57.65 µA,
Ibias,main = -59.5 µA
In conjunction with the plasticity of synapses, the high degree of parallelism in the brain creates a
densely connected, adaptable network able to optimize and adjust for different conditions. An example of
fanout in the nanowire neuron is shown in Figure 12. As shown in Fig. 12a, a single neuron is connected
to four target neurons through four separate tunable synapses. When each of the four modulating currents
is set to Imod = 0, the firing of all four targets is inhibited by the main neuron, as illustrated in Fig. 12b. To
weaken the connection with one of the targets (target #4), Imod,4 is set to 8 µA, turning off the inhibiting
action on target #4 but allowing it to remain on the three other target neurons, as shown in Fig. 12c.
Comparison of the synaptic currents for each of the four targets shows that the synaptic current for target
#4 is reduced as a result of the modulated inductance. This modulation illustrates that the nanowire
neuron is able to be used in a parallel network where the strength of individual synaptic connections can
be adapted.
Fan-out may be one area where nanowires will be an improvement over Josephson junctions [6],
another superconducting technology with promise for artificial neurons. Both nanowires and Josephson
junctions have quantized flux outputs (flux here being defined as the time-integral of the voltage).
However, Josephson devices have outputs of only a single flux-quantum, while nanowires have outputs of
50-100 flux quanta (i.e. 70 flux quanta for the device in Figure 2). In pushing the fan-out and fan-in to
larger systems, one expects to be eventually limited by parasitic inductances and thermal noises. In such
a case, the more substantial signal of the nanowire neuron would permit a larger fan-out and fan-in,
leading to a higher degree of parallelism. In addition to the fan-out, the nanowire voltage signals are long
enough and large enough to be digitized directly on an oscilloscope, in contrast to their Josephson
junction counterparts, allowing for more direct analysis and readout. Finally, there are a suite of three-
terminal, power-control devices made from nanowires [26], which could be fabricated alongside the
neurons and used to boost signal and match impedances.
Figure 12: Fanout of nanowire neuron with a tunable inductive synapse. (a) Simplified circuit model of the
fanout circuit. (b) Simulation when Imod = 0 µA for all four target neurons. (i) Output of the main upstream neuron.
(ii) Current through each of the four series resistors Rseries out to each of the target neurons. (iii) Output voltage of
the four target neurons, shifted on the y-axis for clarity. (c) Simulation when Imod = 0 µA for targets 1-3 and Imod = 8
µA for target #4. (i) Output of the main upstream neuron. (ii) Current through each of the four series resistors
Rseries out to each of the target neurons. The current through Rseries,out for target #4 is less than that of the other
targets, showing that it has been modulated. (iii) Output voltage of the four target neurons, shifted on the y-axis for
clarity. The first three targets have been inhibited, while the inhibitory action on target #4 has been turned off.
Parameters: Rseries,in= 7 Ω, Rsyn,1 = Rsyn,2 = 300 Ω, Lsyn = 0.4 µH, Rseries,out = 1 Ω, Lnanowire = 0.275 µH, Ic, nanowire = 6 µA,
Ibias,main = -59.5 µA, Ibias,target = 57.65 µA.
4.3 Benchmarking Synaptic Energy and Speed
The energy dissipation of both the neuron and the synapse can be calculated using LTSpice by taking the
time integral of the current-voltage product of each circuit element. Performing this analysis, we find that
the nanowire neuron has an energy of about 0.05 fJ for each action potential, while the synapse is about
an order of magnitude less at around 0.005 fJ. In large systems, it will be the synapses that will dominate;
typically if there are O(N) neurons there will be O(N2) synapses. Hence, even though the neuron
dissipates more energy, it will be the synapses that will dominate the power consumption of a large
network.
In a spiking neural network, the energy dissipated increases with the speed of the system; spiking
twice as often dissipates twice the energy, assuming the energy per spike stays constant. The appropriate
figure of merit to compare different technologies is then to take the ratio of speed and power. IBM [27]
introduced the figure of merit of synaptic operations per second per watt (SOPS/W). We include a
constant factor of about 400 W/W for the nanowire neuron to account for the cryogenic cooling costs.
Table 1 compares both the energy per spike and the SOPS/W for the nanowire neuron, the human brain,
and two CMOS technologies. We acknowledge that the estimate for the nanowire neuron is a projection
from a calculation of a single component, whereas the other entries in the table have actually been
measured on large systems. However, given that superconducting platforms have no dissipation in their
interconnects, we believe that it is reasonable to project in such a way. In doing so, it is apparent that the
nanowire neuron can be a highly competitive technology from a power and speed prospective.
Human brain NeuroGrid TrueNorth Nanowire neuron
energy/ switch 10 fJ
100 pJ
25 pJ
~0.05 fJ
SOPS/Watt
1e14
1e10
4e10
5e14
Table 1: Energy and figure of merit values for different artificial neural networks in comparison to those of
the human brain. Energy values for the human brain, NeuroGrid, and TrueNorth are taken from Ref. [28].
4.4 Discussion on Possible Synapse Alternatives
Although the fanout achieved here is far from the level of parallelism in the human brain (where each
neuron connects to thousands of neighbors), it suggests that nanowires may serve a unique purpose in the
development of future superconducting neural networks, which have thus far struggled to support fanout.
Given that nanowires can interface with both CMOS and Josephson junction circuits [29], it may be
possible for nanowire neurons to serve as intermediary devices in a network with both platforms. Indeed,
a recently proposed neural network with hybrid technologies employed superconducting nanowires as
photon detectors, relying instead on optical signals for facilitating high fanout [30] [31]. Although our
work uses nanowires solely as electrical components, they can easily be biased to act as photodetectors
[32] [33] as well, illustrating that the two different architectures would be compatible for integration.
While our work takes more of an electrical approach, the two architectures illustrate the diverse ways in
which superconducting nanowires can be used in neural networks, suggesting that a combination of the
two technologies may be possible.
Furthermore, while the inductive synapse proposed in this work relies on an external modulating
bias current, it may be possible to use a superconducting memory cell to store the value of this
modulating current, or replace it directly in the circuit. Memory cells based on superconducting nanowire
loops have been shown to reliably store states in the form of a trapped circulating current [34] [35] [36],
with recent work demonstrating that one can program and store different amounts of current in the loop in
discrete quantities [37]. Incorporating these programmable memory cells into the modulating elements of
the inductive synapse could offer multi-level storage of synaptic strengths, allowing for more complex
applications. It may also be possible for nanowire neurons to connect through alternative synapse designs,
such as one based on inductive coupling. Further analysis is needed to verify the possibility of fanout in
such a scheme.
5. CONCLUSION
By taking advantage of intrinsic nonlinearities in superconducting nanowires, we have developed a
platform for a low-power artificial neuron. In this platform, the coupling of two nanowire-based
oscillators acts analogously to the two ion channels in a simplified neuron model, producing an output
voltage spike that serves as the information-carrying token in these circuits. Using electrothermal circuit
simulations, we have shown that the nanowire neuron is able to reproduce universal biological neuron
characteristics, such as a firing threshold, as well as unique characteristics distinct to certain neural
classes, such as parabolic bursting. Furthermore, we suggested that a nanowire transmission line with a
propagation speed of ~ 2% c may be used as an axon delay line, potentially allowing spatio-temporal
information to be accessed. These collective behaviors may enable the nanowire neuron to be used in a
rich variety of operations.
In addition to harnessing the nonlinearity of the nanowire's switching dynamics, we relied on the
nonlinearity of the nanowire's kinetic inductance in order to develop a variable inductive synapse. We
demonstrated that the total parallel inductance of the synapse can be tuned with a modulating bias current,
thereby changing the strength of the signal sent to a downstream target neuron. This scheme proved to be
capable of fanout, as demonstrated by a single neuron inhibiting the firing of four target neurons. An
energy analysis of this circuit in comparison to other spiking networks illustrated that the nanowire
neuron has competitive performance in the dynamic firing state and a figure of merit two orders of
magnitude better than certain alternative platforms, while the static state benefits from the lack of power
dissipation by the superconducting elements. Although experimental realization of these results is a
subject of future work, the analysis performed here suggests that the nanowire neuron is a promising
candidate for the advancement of low-power artificial neural networks.
Looking forward, networks of superconducting nanowires could be the basis for powerful new
computer hardware. Analog blocks of highly connected neurons could be digitally linked to achieve
networks with either scale-free or small-worlds connectivity. With superconducting interconnects, entire
chips could be wired together with no cost in heat dissipation. The result would be a large-scale
neuromorphic processor which could be trained as a spiking neural network to perform tasks like pattern
recognition or used to simulate the spiking dynamics of a large, biologically-realistic network. The
combination of speed, low power dissipation, and biological realism with only a few components
suggests that nanowires could outperform or complement other existing and developing neuromorphic
hardware technologies.
ACKNOWLEDGEMENTS
The authors thank Murat Onen and Brenden Butters for scientific discussions, Akshay Agarwal and Dr.
Donnie Keithley for edits, and the entire Quantum Nanostructures and Nanofabrication group. They
would also like to acknowledge the generous support of the Bose Foundation. Emily Toomey was
supported by the National Science Foundation Graduate Research Fellowship Program (NSF GRFP)
under Grant No. 1122374.
AUTHOR CONTRIBUTIONS
E.T. performed the simulations with input from K.S. and K.K.B. All authors contributed to the design of
the circuits and the writing of the paper.
SUPPLEMENTARY MATERIAL
An example of a capacitive synapse, as well as a demonstration of the tunable inductive synapse in an
excitatory case, may be found in the Supplementary Material.
DATA AVAILABILITY
The nanowire LTSpice model used for this study can be found on GitHub at
https://github.com/karlberggren/snspd-spice. All circuit schematic and data files, including the Matlab
code used to generate the figures, are available from the authors upon request.
Works Cited
[1] S. Furber and S. Temple, "Neural systems engineering," J. R. Soc. Interface, vol. 4, no. 13, pp. 193 --
[2]
206, Apr. 2007.
I. Sourikopoulos et al., "A 4-fJ/Spike Artificial Neuron in 65 nm CMOS Technology," Front.
Neurosci., vol. 11, 2017.
[3] S. Li, W. Kang, Y. Huang, X. Zhang, Y. Zhou, and W. Zhao, "Magnetic skyrmion-based artificial
neuron device," Nanotechnology, vol. 28, no. 31, p. 31LT01, Jul. 2017.
[4] D. Querlioz, O. Bichler, and C. Gamrat, "Simulation of a memristor-based spiking neural network
immune to device variations," in The 2011 International Joint Conference on Neural Networks,
2011, pp. 1775 -- 1781.
[5] A. Thomas, "Memristor-based neural networks," J. Phys. Appl. Phys., vol. 46, no. 9, p. 093001,
Feb. 2013.
[6] P. Crotty, D. Schult, and K. Segall, "Josephson junction simulation of neurons," Phys. Rev. E, vol.
82, no. 1, p. 011914, Jul. 2010.
[7] M. L. Schneider et al., "Ultralow power artificial synapses using nanotextured magnetic Josephson
junctions," Sci. Adv., vol. 4, no. 1, p. e1701329, Jan. 2018.
[8] E. Toomey, Q.-Y. Zhao, A. N. McCaughan, and K. K. Berggren, "Frequency Pulling and Mixing of
Relaxation Oscillations in Superconducting Nanowires," Phys. Rev. Appl., vol. 9, no. 6, p. 064021,
Jun. 2018.
[9] A. J. Kerman et al., "Kinetic-inductance-limited reset time of superconducting nanowire photon
counters," Appl. Phys. Lett., vol. 88, no. 11, p. 111116, Mar. 2006.
[10] G. B. Ermentrout and D. H. Terman, "The Hodgkin -- Huxley Equations," in Mathematical
Foundations of Neuroscience, vol. 35, New York, NY: Springer New York, 2010, pp. 1 -- 28.
[11] A. J. Kerman, J. K. W. Yang, R. J. Molnar, E. A. Dauler, and K. K. Berggren, "Electrothermal
feedback in superconducting nanowire single-photon detectors," Phys. Rev. B, vol. 79, no. 10, p.
100509, Mar. 2009.
[12] K. K. Berggren et al., "A Superconducting Nanowire can be Modeled by Using SPICE,"
Supercond. Sci. Technol., 2018.
[13] E. M. Izhikevich, "Which model to use for cortical spiking neurons?," IEEE Trans. Neural Netw.,
vol. 15, no. 5, pp. 1063 -- 1070, Sep. 2004.
[14] R. Bertram, M. J. Butte, T. Kiemel, and A. Sherman, "Topological and phenomenological
classification of bursting oscillations," Bull. Math. Biol., vol. 57, no. 3, pp. 413 -- 439, Jan. 1995.
[15] J. Rinzel, "Bursting oscillations in an excitable membrane model," in Ordinary and Partial
Differential Equations, 1985, pp. 304 -- 316.
[16] T. Nanami, K. Aihara, and T. Kohno, "Elliptic and parabolic bursting in a digital silicon neuron
model," presented at the 2016 International Symposium on Nonlinear Theory and Its Applications,
Yugawara, Japan, 2016, p. 4.
[17] S. K. Dana, D. C. Sengupta, and C.- Hu, "Spiking and Bursting in Josephson Junction," IEEE
Trans. Circuits Syst. II Express Briefs, vol. 53, no. 10, pp. 1031 -- 1034, Oct. 2006.
[18] C. Soto-Treviño, N. Kopell, and D. Watson, "Parabolic bursting revisited," J. Math. Biol., vol. 35,
no. 1, pp. 114 -- 128, Nov. 1996.
[19] F. Strumwasser, "The demonstration and manipulation of a circadian rhythm in a single neuron," in
Circadian clocks, Amsterdam: North-Holland Publishing Co., 1965, pp. 442 -- 462.
[20] B. O. Alving, "Spontaneous Activity in Isolated Somata of Aplysia Pacemaker Neurons," J. Gen.
Physiol., vol. 51, no. 1, p. 29, Jan. 1968.
[21] Q.-Y. Zhao et al., "Single-photon imager based on a superconducting nanowire delay line," Nat.
Photonics, vol. 11, no. 4, pp. 247 -- 251, Apr. 2017.
[22] Q.-Y. Zhao, D. F. Santavicca, D. Zhu, B. Noble, and K. K. Berggren, "A distributed electrical
model for superconducting nanowire single photon detectors," Appl. Phys. Lett., vol. 113, no. 8, p.
082601, Aug. 2018.
[23] B. P. Bean, "The action potential in mammalian central neurons," Nat. Rev. Neurosci., vol. 8, p.
451, Jun. 2007.
[24] S. Ferraina, M. Paré, and R. H. Wurtz, "Comparison of Cortico-Cortical and Cortico-Collicular
Signals for the Generation of Saccadic Eye Movements," J. Neurophysiol., vol. 87, no. 2, pp. 845 --
858, Feb. 2002.
[25] A. J. Annunziata et al., "Tunable superconducting nanoinductors," Nanotechnology, vol. 21, no. 44,
p. 445202, Oct. 2010.
[26] A. N. McCaughan and K. K. Berggren, "A Superconducting-Nanowire Three-Terminal
Electrothermal Device," Nano Lett., vol. 14, no. 10, pp. 5748 -- 5753, Oct. 2014.
[27] P. A. Merolla et al., "A million spiking-neuron integrated circuit with a scalable communication
network and interface," Science, vol. 345, no. 6197, pp. 668 -- 673, Aug. 2014.
[28] S. Furber, "Large-scale neuromorphic computing systems," J. Neural Eng., vol. 13, no. 5, p.
051001, Aug. 2016.
[29] Q.-Y. Zhao, A. N. McCaughan, A. E. Dane, K. K. Berggren, and T. Ortlepp, "A nanocryotron
comparator can connect single-flux-quantum circuits to conventional electronics," Supercond. Sci.
Technol., vol. 30, no. 4, p. 044002, 2017.
[30] J. M. Shainline et al., "Circuit designs for superconducting optoelectronic loop neurons," J. Appl.
Phys., vol. 124, no. 15, p. 152130, Oct. 2018.
[31] J. M. Shainline, "Fluxonic Processing of Photonic Synapse Events," ArXiv190402807 Cs, Apr.
2019.
[32] G. N. Gol'tsman et al., "Picosecond superconducting single-photon optical detector," Appl. Phys.
Lett., vol. 79, no. 6, pp. 705 -- 707, Aug. 2001.
[33] F. Marsili et al., "Single-Photon Detectors Based on Ultranarrow Superconducting Nanowires,"
Nano Lett., vol. 11, no. 5, pp. 2048 -- 2053, May 2011.
[34] A. Murphy, D. V. Averin, and A. Bezryadin, "Nanoscale superconducting memory based on the
kinetic inductance of asymmetric nanowire loops," New J. Phys., vol. 19, no. 6, p. 063015, 2017.
[35] Q.-Y. Zhao et al., "A compact superconducting nanowire memory element operated by nanowire
cryotrons," Supercond. Sci. Technol., vol. 31, no. 3, p. 035009, 2018.
[36] A. N. McCaughan, E. A. Toomey, M. Schneider, K. K. Berggren, and S. W. Nam, "A kinetic-
inductance-based superconducting memory element with shunting and sub-nanosecond write
times," Supercond. Sci. Technol., 2018.
[37] E. Toomey, M. Onen, M. Colangelo, B. A. Butters, A. N. McCaughan, and K. K. Berggren,
"Bridging the Gap Between Nanowires and Josephson Junctions: A Superconducting Device Based
on Controlled Fluxon Transfer," Phys. Rev. Appl., vol. 11, no. 3, p. 034006, Mar. 2019.
|
1508.00429 | 1 | 1508 | 2015-08-03T14:22:52 | A three-threshold learning rule approaches the maximal capacity of recurrent neural networks | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Understanding the theoretical foundations of how memories are encoded and retrieved in neural populations is a central challenge in neuroscience. A popular theoretical scenario for modeling memory function is the attractor neural network scenario, whose prototype is the Hopfield model. The model has a poor storage capacity, compared with the capacity achieved with perceptron learning algorithms. Here, by transforming the perceptron learning rule, we present an online learning rule for a recurrent neural network that achieves near-maximal storage capacity without an explicit supervisory error signal, relying only upon locally accessible information. The fully-connected network consists of excitatory binary neurons with plastic recurrent connections and non-plastic inhibitory feedback stabilizing the network dynamics; the memory patterns are presented online as strong afferent currents, producing a bimodal distribution for the neuron synaptic inputs. Synapses corresponding to active inputs are modified as a function of the value of the local fields with respect to three thresholds. Above the highest threshold, and below the lowest threshold, no plasticity occurs. In between these two thresholds, potentiation/depression occurs when the local field is above/below an intermediate threshold. We simulated and analyzed a network of binary neurons implementing this rule and measured its storage capacity for different sizes of the basins of attraction. The storage capacity obtained through numerical simulations is shown to be close to the value predicted by analytical calculations. We also measured the dependence of capacity on the strength of external inputs. Finally, we quantified the statistics of the resulting synaptic connectivity matrix, and found that both the fraction of zero weight synapses and the degree of symmetry of the weight matrix increase with the number of stored patterns. | q-bio.NC | q-bio |
A three-threshold learning rule approaches the maximal capacity of
recurrent neural networks
Alireza Alemi1,2,*, Carlo Baldassi1,2, Nicolas Brunel3, Riccardo Zecchina1,2
1 Human Genetics Foundation (HuGeF), Turin, Italy
2 DISAT, Politecnico di Torino, Turin, Italy
3 Departments of Statistics and Neurobiology, University of Chicago, USA
* [email protected]
Abstract
Understanding the theoretical foundations of how memories are encoded and retrieved in neural
populations is a central challenge in neuroscience. A popular theoretical scenario for modeling
memory function is the attractor neural network scenario, whose prototype is the Hopfield model.
The model simplicity and the locality of the synaptic update rules come at the cost of a poor
storage capacity, compared with the capacity achieved with perceptron learning algorithms. Here,
by transforming the perceptron learning rule, we present an online learning rule for a recurrent
neural network that achieves near-maximal storage capacity without an explicit supervisory error
signal, relying only upon locally accessible information. The fully-connected network consists of
excitatory binary neurons with plastic recurrent connections and non-plastic inhibitory feedback
stabilizing the network dynamics; the memory patterns to be memorized are presented online
as strong afferent currents, producing a bimodal distribution for the neuron synaptic inputs.
Synapses corresponding to active inputs are modified as a function of the value of the local fields
with respect to three thresholds. Above the highest threshold, and below the lowest threshold,
no plasticity occurs. In between these two thresholds, potentiation/depression occurs when the
local field is above/below an intermediate threshold. We simulated and analyzed a network of
binary neurons implementing this rule and measured its storage capacity for different sizes of the
basins of attraction. The storage capacity obtained through numerical simulations is shown to
be close to the value predicted by analytical calculations. We also measured the dependence of
capacity on the strength of external inputs. Finally, we quantified the statistics of the resulting
synaptic connectivity matrix, and found that both the fraction of zero weight synapses and the
degree of symmetry of the weight matrix increase with the number of stored patterns.
Author Summary
Recurrent neural networks have been shown to be able to store memory patterns as fixed point
attractors of the dynamics of the network. The prototypical learning rule for storing memories
in attractor neural networks is Hebbian learning, which can store up to 0.138N uncorrelated
patterns in a recurrent network of N neurons. This is very far from the maximal capacity 2N ,
which can be achieved by supervised rules, e.g. by the perceptron learning rule. However, these
rules are problematic for neurons in the neocortex or the hippocampus, since they rely on the
1
computation of a supervisory error signal for each neuron of the network. We show here that the
total synaptic input received by a neuron during the presentation of a sufficiently strong stimulus
contains implicit information about the error, which can be extracted by setting three thresholds
on the total input, defining depression and potentiation regions. The resulting learning rule
implements basic biological constraints, and our simulations show that a network implementing
it gets very close to the maximal capacity, both in the dense and sparse regimes, across all values
of storage robustness. The rule predicts that when the total synaptic inputs goes beyond a
threshold, no potentiation should occur.
Introduction
One of the fundamental challenges in neuroscience is to understand how we store and retrieve
memories for a long period of time. Such long-term memory is fundamental for a variety of
our cognitive functions. A popular theoretical framework for storing and retrieving memories in
recurrent neural networks is the attractor network model framework [1–3]. Attractors, i.e. stable
states of the dynamics of a recurrent network, are set by modification of synaptic efficacies in
a recurrent network. Synaptic plasticity rules specify how the efficacy of a synapse is affected
by pre- and post-synaptic neural activity. In particular, Hebbian synaptic plasticity rules lead
to long-term potentiation (LTP) for correlated pre- and post-synaptic activities, and long-term
depression (LTD) for anticorrelated activities. These learning rules build excitatory feedback
loops in the synaptic connectivity, resulting in the emergence of attractors that are correlated
with the patterns of activity that were imposed on the network through external inputs. Once
a set of patterns become attractors of a network (in other words when the network “learns”
the patterns), upon a brief initial activation of a subpopulation of neurons, the network state
evolves towards the learned stable state (the network “retrieves” a past stored memory), and
remains in that state after removal of the external inputs (and hence maintains the information
in short-term memory). The set of initial network states leading to a memorized state is called
the basin of attraction, whose size determines how robust a memory is. The attractor neural
network scenario was originally explored in networks of binary neurons [1, 2], and then extended
from the 90s to networks of spiking neurons [4–7].
Experimental evidence in different areas of the brain, including inferotemporal cortex [8–11]
and prefrontal cortex [12–14], has provided support for the attractor neural network framework,
using electrophysiological recordings in awake monkeys performing delayed response tasks. In
such experiments, the monkey has to maintain information in short-term (working) memory in
a ‘delay period’ to be able to perform the task. Consistent with the attractor network scenario,
some neurons exhibit selective persistent activity during the delay period. This persistent activity
of ensembles of cortical neurons has thus been hypothesized to form the basis of the working
memory of stimuli shown in these tasks.
One of the most studied properties of attractor neural network as a model of memory is
its storage capacity, i.e. how many random patterns can be learned in a recurrent network of
N neurons in the large N limit. Storage capacity depends both on the network architecture
and on the synaptic learning rule.
In
particular, the Hopfield network [1] that uses a Hebbian learning rule has a storage capacity
of 0.138N in the limit of N → ∞ [15]. Later studies showed how the capacity depends on
the connection probability in a randomly connected network [16, 17] and on the coding level
(fraction of active neurons in a pattern) [18, 19]. A natural question is, what is the maximal
capacity of a given network architecture, over all possible learning rules? This question was
answered by Elizabeth Gardner, who showed that the capacity of fully connected networks of
In many models, the storage capacity scales with N .
2
binary neurons with dense patterns scales as 2N [20], a storage capacity which is much larger
than the one of the Hopfield model. The next question is what learning rules are able to saturate
the Gardner bound? A simple learning rule that is guaranteed to achieve this bound is the
perceptron learning rule (PLR) [21] applied to each neuron independently. However, unlike the
rule used in the Hopfield model, the perceptron learning rule is a supervised rule that needs an
explicit “error signal” in order to achieve the Gardner bound. While such an error signal might
be available in the cerebellum [22–24], it is unclear how error signals targeting individual neurons
might be implemented in cortical excitatory synapses. Therefore, it remains unclear whether and
how networks with realistic learning rules might approach the Gardner bound.
The goal of the present paper is to propose a learning rule whose capacity approaches the
maximal capacity of recurrent neural networks by transforming the original perceptron learning
rule such that the new rule does not explicitly use an error signal. The perceptron learning
rule modifies the synaptic weights by comparing the desired output with the actual output to
obtain an error signal, subsequently changing the weights in the opposite direction of the error
signal. We argue that the total synaptic inputs (‘local fields’) received by a neuron during the
presentation of a stimulus contain some information about the current error (i.e. whether the
neuron will end up in the right state after the stimulus is removed). We use this insight to build
a field dependent learning rule that contains three thresholds separating no plasticity, LTP and
LTD regions. This rule implements basic biological constraints: (a) it uses only information
local to the synapse; (b) the new patterns can be learned incrementally, i.e. it is an online rule;
(c) it does not need an explicit error signal; (d) synapses obey Dale’s principle, i.e. excitatory
synapses are not allowed to have negative weights. We studied the capacity and the size of
the basins of attraction for a binary recurrent neural network in which excitatory synapses are
endowed with this rule, while a global inhibition term controls the global activity level. We
investigated how the strength of external fields and the presence of correlations in the inputs
affect the memory capacity. Finally, we investigated the statistical properties of the connectivity
matrix (distribution of synaptic weights, degree of symmetry).
Results
The network
We simulated a network of N binary (McCulloch-Pitts) neurons, fully-connected with excitatory
synapses (Fig 1A). All the neurons feed a population of inhibitory neurons which is modeled as
a single aggregated inhibitory unit. This state-dependent global inhibition projects back onto
all the neurons, stabilizing the network and controlling its activity level. At each time step, the
activity (or the state) of neuron i (i = 1 . . . N ) is described by a binary variable si ∈ {0, 1}. The
state is a step function of the local field vi of the neuron:
si = Θ (vi − θ) ,
(1)
where Θ is the Heaviside function (Θ (x) = 1 if x > 0 and 0 otherwise) and θ is a neuronal
threshold. The local field vi represents the overall input received by the neuron from its excitatory
and inhibitory connections (Fig 1B). The excitatory connections are of two kinds: recurrent
connections from within the excitatory population, and external inputs.
The recurrent excitatory connections are mediated by synaptic weights, denoted by a matrix
W whose elements wij (the weight of the synapse from neuron j to i) are continuous non-negative
variables (wij ∈ [0,∞); wii = 0). In the following, and in all our simulations, we assume that the
weights are initialized randomly before the training takes place (see Materials and Methods).
3
Figure 1: A sketch of the network and the neuron model. A. Structure of the network.
The fully-connected network consists of N binary (si ∈ {0, 1}) neurons and an aggregated in-
hibitory unit. The global inhibition is a function of the state of the network and the external
fields, i.e. I(~x, ~s). A memory pattern ~ξ is encoded as strong external fields, i.e. ~x = X ~ξ and
presented to the network during the learning phase. B. Each neuron receives excitatory recurrent
inputs (thin black arrows) from the other neurons, a global inhibitory input (red connections),
and a strong binary external field (xi ∈ {0, X}; thick black arrows). All these inputs are summed
to obtain the total field, which is then compared to a neuronal threshold θ; the output of the
neuron is a step function of the result.
4
Therefore, in the absence of external inputs, the local field of each neuron i is given by:
N
vi =
Xj=1
where I0 (~s) represents the inhibitory input.
For the sake of simplicity, we simulated a synchronous update process, in which the activity
of each neuron si is computed from the local field vi at the previous time step, and all updates
happen in parallel.
wijsj − I0 (~s) ,
(2)
The network was designed so that, in absence of external input and prior to the training
process, it should spontaneously stabilize itself to some fixed overall average activity level f
(fraction of active neurons, or sparseness), regardless of the initial conditions. In particular, we
aimed at avoiding trivial attractors (the all-off and all-on states). To this end, we model the
inhibitory feedback (in absence of external inputs) as a linear function of the overall excitatory
activity:
I0(~s) = H0 + λ N
Xi=1
si − f N! .
(3)
the excitatory network has the desired activity level f , i.e. when PN
The parameters H0 and λ can be understood as follows: H0 is the average inhibitory activity when
i=1 si = f N ; λ measures
the strength of the inhibitory feedback onto the excitatory network. This expression can be
interpreted as a first-order approximation of the inhibitory activity as a function of the excitatory
activity around some reference value f N , which is reasonable under the assumption that the
deviations from f N are small enough.
Indeed, by properly setting these two parameters in
relation to the other network parameters (such as θ and the average connection strength) it is
possible to achieve the desired goal of a self-stabilizing network.
i ∈ {0, 1}), and assume that each entry ξµ
In the training process, the network is presented a set of p patterns in the form of strong
external inputs, representing the memories which need to be stored. We denote the patterns
as {~ξµ} (where µ = 1...p and ξµ
is drawn randomly
and independently. For simplicity, the coding level f for the patterns was set equal to the
ξµ
i = 1 with probability f , 0 otherwise. During
spontaneous activity level of the network, i.e.
the presentation of a pattern µ, each neuron i receives an external binary input xi = Xξµ
i ,
where X denotes the strength of the external inputs, which we parameterized as X = γ√N . In
addition, the external input also affects the inhibitory part of the network, eliciting a response
which indirectly downregulates the excitatory neurons. We model this effect as an additional
term H1 in the expression for the inhibitory term (Eq. 3), which therefore becomes:
i
I(~x, ~s) = H0 + H1PN
i=1 xi
f N X
N
+ λ(
Xi=1
si − f N ),
The general expression for the local field vi then reads:
vi =
N
Xj=1
wij sj + xi − I(~x, ~s).
(4)
(5)
In the absence of external fields, xi = 0 for all i, and thus Eqs. 4 and 5 reduce to the previous
expressions Eqs. 3 and 2.
The goal of the learning process is to find values of wij ’s such that the patterns {~ξµ} become
attractors of the network dynamics. Qualitatively, this means that, if the training process is
5
i=1 ξµ
i.e. whenever the Hamming distance d = PN
successful, then whenever the network state gets sufficiently close to one of the stored patterns,
i − si between the current network state and
a pattern µ is sufficiently small, the network dynamics in the absence of external inputs should
drive the network state towards a fixed point equal to the pattern itself (or very close to it).
The general underlying idea is that, after a pattern is successfully learned, some brief external
input which initializes the network close to the learned state would be sufficient for the network
to recognize and retrieve the pattern. The maximum value of d for which this property holds is
then called the basin of attraction size (or just basin size hereafter for simplicity); indeed, there
is generally a trade-off between the number of patterns which can be stored according to this
criterion and the size of their basin of attraction.
More precisely, the requirement that a pattern ~ξµ is a fixed point of the network dynamics in
the absence of external fields can be reduced to a condition for each neuron i (cfr. Eqs. 4 and 5):
N
∀i : Θ
Xj=1
wij ξµ
j − I(cid:16)~0, ~ξµ(cid:17) − θ
i .
= ξµ
(6)
This condition only guarantees that, if the network is initialized into a state ~s = ~ξµ, then it will
not spontaneously change its state, i.e. it implements a zero-size basin of attraction. A simple
way to enlarge the basin size is to make the requirement in Eq. 6 more stringent, by enforcing a
more stringent constraint for local fields:
∀i :
PN
PN
j=1 wij ξµ
j=1 wij ξµ
j − I(cid:16)~0, ~ξµ(cid:17) > θ + f√N ǫ
j − I(cid:16)~0, ~ξµ(cid:17) < θ − f√N ǫ
if ξµ
if ξµ
i = 1
i = 0,
(7)
where ǫ ≥ 0 is a robustness parameter. When ǫ = 0, we recover the previous zero-basin-size
scenario; increasing ǫ we make the neurons’ response more robust towards noise in their inputs,
and thus we enlarge the basin of attraction of the stored patterns (but then fewer patterns can
be stored, as noted above).
The three-threshold learning rule (3TLR)
In the training phase, the network is presented with patterns as strong external fields xi. Patterns
are presented sequentially in random order. For each pattern µ, we simulated the following
scheme:
Step 1: The pattern is presented (i.e. the external inputs xi are set to Xξµ
i ). A single step
of synchronous updating is performed (Eqs. 1, 4 and 5). If the external inputs are strong enough,
i.e. γ is large enough, this updating sets the network in a state corresponding to the presented
pattern.
Step 2: Learning occurs. Each neuron i may update its synaptic weights depending on 1)
their current value wt
ij , 2) the state of the pre-synaptic neurons, and 3) the value of the local
field vi. Therefore, all the information required is locally accessible, and no explicit error signals
are used. The new synaptic weights wt+1
are set to:
ij
wt
ij − ηsj,
wt
ij + ηsj,
wt
ij ,
if θ0 < vi < θ
if θ < vi < θ1
otherwise,
wt+1
ij =
6
(8)
where η is the learning rate, and θ0 and θ1 are two auxiliary learning thresholds set as
θ0 = θ − (γ + ǫ) f√N
θ1 = θ + (γ + ǫ) f√N .
(9)
(10)
We refer to this update scheme as “three-threshold learning rule” (3TRL). After some number
of presentations, we checked whether the patterns are learned by presenting a noisy version of
these patterns, and checking whether the patterns (or network states which are very close to the
patterns) are fixed points of the network dynamics.
When N ≫ 1, γ is large enough, and H1 = f X, the update rule described by Eq. 8 is
essentially equivalent to the perceptron learning rule for the task described in Eq. 7. This can
be shown as follows (see also Fig 2 for a graphical representation of the case f = 0.5 and ǫ = 0):
when a stimulus is presented, the population of neurons is divided in two groups, one for which
xi = 0 and one for which xi = X. The net effect of the stimulus presentation on the local
field has to take into account the indirect effect through the inhibitory part of the network (see
Eq. 4), and thus is equal to −f X for the xi = 0 population and to (1 − f ) X for the xi = X
population. Before learning, the distribution of the local fields across the excitatory population,
in the limit N → ∞, is a Gaussian whose standard deviation is proportional to √N , due to the
central limit theorem; moreover, the parameter H0 is set so that the average activity level of the
network is f , which means that the center of the Gaussian will be within a distance of order √N
from the neuronal threshold θ (this also applies if we use different values for the spontaneous
activity level and the pattern activity level). Therefore, if X = γ√N is large enough, the state
of the network during stimulus presentation will be effectively clamped to the desired output,
i.e. si = ξµ
for all i. This fact has two consequences: 1) the local field potential can be used to
i
detect the desired output by just comparing it to the threshold, and 2) each neuron i will receive,
. Furthermore, due to the choice
of the secondary thresholds θ0 and θ1 in Eqs. 9 and 10, the difference between the local field
and θ0 (or θ1) during stimulus presentation for the xi = 0 population (or xi = X, respectively)
as its recurrent inputs {sj}j6=i, the rest of the pattern (cid:8)ξµ
is equal to the difference between the local field and θ − f√N ǫ (or θ + f√N ǫ, respectively) in
j(cid:9)j6=i
the absence of external stimuli, provided the recurrent inputs are the same. Therefore, the value
of the local field vi during stimulus presentation in relation to the three thresholds θ, θ0 and
θ1 is sufficient to determine whether an error is made with respect to the constraints of Eq. 7,
and which kind of error is made. Following these observations, it is straightforward to map the
standard perceptron learning rule on the 4 different cases which may occur (see Fig 2), resulting
in Eq. 8.
In Fig 3 we demonstrate the effect of the learning rule on the distribution of the local field
potentials as measured from a simulation (with f = 0.5 and ǫ = 1.2): the initial distribution
of the local fields of the neurons, before the learning process takes place and in the absence of
external fields, is well described by a Gaussian distribution centered on the neuronal threshold
θ (see Fig 3A) with a standard deviation which scales as √N . During a pattern presentation,
the resulting distribution becomes a bimodal one; before learning takes place, the distribution is
given by the sum of two Gaussians of equal width, centered around θ0 + f√N ǫ and θ1 − f√N ǫ
(Fig 3B). The left Gaussian corresponds to the cases where xi = 0 and the right one to the cases
where xi = X. Having applied the learning rule, we observe that the depression region (i.e. the
interval (θ0, θ)) and the potentiation region (i.e. (θ, θ1)) gets depleted (Fig 3C). In the testing
phase, when the external inputs are absent, the left and right parts of the distribution come
closer, such that the distance between the two peaks is equal to at least 2ǫf√N (Fig 3D). This
margin between the local fields of the ON and OFF neurons makes the attractors more robust.
7
Original perceptron rule
3TLR
d
e
r
i
s
e
d
)
d
(
t
u
p
t
u
o
l
a
u
t
c
a
)
y
(
t
u
p
t
u
o
0
0
1
1
0
1
0
1
l
i
a
n
g
s
r
o
r
r
e
)
y
-
d
(
0
-1
+1
0
Δw
Ø
D
P
Ø
d = 0
y = 0
do
nothing
y = 1
depress
active w
v
d = 1
y = 0
potentiate
active w
y = 1
do
nothing
v
xi = 0
xi = X
do
nothing
depress
active w
potentiate
active w
do
nothing
v
Figure 2: The three-threshold learning rule (3TLR), and its relationship with the stan-
dard perceptron learning rule (PLR). The perceptron learning rule modifies the synaptic
weights by comparing the desired output with the actual output to obtain an error signal, subse-
quently changing the weights in the opposite direction of the error signal (see the table in the left
panel). For a pattern which is uncorrelated with the current synaptic weights, the distribution
is Gaussian (in the limit of large N ), due to the central limit theorem. H0 is set such that, on
average, a fraction f of the local fields are above the neuronal threshold θ; in the case of f = 0.5,
this means that the Gaussian is centered on θ (left panel). In our model (Fig 1B), the desired
output is given as a strong external input, whose distribution across the population is bimodal
(with two delta functions on xi = 0 and xi = X); therefore, the distribution of the local fields
during stimulus presentation becomes bimodal as well (right panel). The left and right bumps
of this distribution correspond to cases where the desired outputs are zero and one, respectively.
Note that, since the external input also elicits an inhibitory response, the neurons in the net-
work which are not directly affected by the external input (i.e. those with desired output equal
to zero) are effectively hyperpolarized. If X is sufficiently large, the two distributions do not
overlap, and the four cases of the PLR can be mapped to the four regions determined from the
three thresholds, indicated by vertical dashed lines (see text).
8
WITHOUT EXTERNAL INPUT
WITH EXTERNAL INPUT
ON neurons
OFF neurons
B
C
Total input
A
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
D
I
G
N
N
R
A
E
L
E
R
O
F
E
B
I
G
N
N
R
A
E
L
R
E
T
F
A
Figure 3: Distribution of local fields before and after learning for f = 0.5 and non-zero
robustness. A. Before learning begins, the distribution of local field of neurons is a Gaussian
distribution (due to central limit theorem) centered around neuronal threshold θ both for neurons
with the desired output zero (OFF neurons) and with the desired output one (ON neurons). The
goal is to have the local field distribution of ON neurons (red curve) to be above the threshold θ,
and that of OFF neurons to be below θ. B. Once any of the to-be-stored patterns are presented
as strong external fields, right before the learning process starts, the local field distribution of the
OFF neuron shifts toward the left-side centered around θ0+f ǫ√N , whereas the distribution of the
ON neurons moves toward the right-side, centered around θ1 − f ǫ√N , with a negligible overlap
between the two curves if the external field is strong enough. Thanks to the strong external
fields and global inhibition, the local fields of the ON and OFF neurons are well separated. C.
Due to the learning process, the local fields within the depression region [i.e. (θ0, θ)] get pushed
to the left-side, below θ0, whereas those within the potentiation region get pushed further to
the right-side, above θ1. If the learning process is successful, it will result in a region (θ0, θ1)
which no longer contain local fields, with two sharp peaks on θ0 and θ1. D. After successful
learning, once the external fields are removed, the blue and red curves come closer, with a gap
equal to 2f ǫ√N . The larger the robustness parameter ǫ, the more the gap between the left- and
right-side of the distribution. Notice that now the red curve is fully above θ which means those
neurons remain stably ON, while the the blue curve is fully below θ, which means those neurons
are stably OFF. Therefore the corresponding pattern is successfully stored by the network.
Storage capacity
Since our proposed learning rule is able to mimic (or approximate, depending on the parameters)
the perceptron learning rule, which is known to be able to solve the task posed by Eq. 7 whenever
9
a solution exists, we expect that a network implementing such rule can get close to maximal
capacity in terms of the number of memories which it can store at a given robustness level.
The storage capacity, denoted by α = p/N , is measured as a ratio of the maximum number of
patterns p which can successfully be stored to the number of neurons N , in the limit of large N .
As mentioned above, it is a function of the basin size.
We used the following definition for the basin size: a set of p patterns is said to be successfully
stored at a size b if, for each pattern, the retrieval rate when starting from a state in which a
fraction b of the pattern was randomized is at least 90%. The retrieval rate is measured by the
probability that the network dynamics is able to bring the network state to an attractor within
1% distance from the pattern, in at most 30 steps. The distance between the state of the network
and a pattern µ is measured by the normalized Hamming distance 1
i . Therefore,
at coding level f = 0.5, reaching a basin size b means that the network can successfully recover
patterns starting from a state at distance b/2.
i=1 si − ξµ
N PN
Fig 4A shows the maximal capacity as a function of the basin size for a simulated network
of N = 1001 neurons. We simulated many pairs of (α, ǫ) with different random seeds, obtaining
a probability of success for each pair. The red line shows the points for which the probability of
successful storage is 0.5, and the error bars span 0.95 to 0.05 success probability. The capacity
was optimized over the robustness parameter ǫ. The maximal capacity (the Gardner bound) in
the limit of N → ∞ at the zero basin size is αc = 2 for our model (see Materials and Methods for
the calculation), as for a network with unconstrained synaptic weights [20]. In Fig 4A, we also
compare our network with the Hopfield model. Our network stores close to the maximal capacity
at zero basin size, at least eleven times more than the Hopfield model. Across the range of basin
sizes, 3TLR achieves more than twice the capacity that can be achieved with the Hopfield model.
The enlargement of the basin of attraction was achieved by increasing the robustness pa-
rameter ǫ. We computed the maximal theoretical capacity as a function of ǫ at N → ∞ (see
Materials and Methods) and compared it to our simulations, and to the maximal theoretical
capacity of the Hopfield network. The results are shown in Fig 4B. For any given value of ǫ,
the cyan curve shows the maximum α for which the success ratio with our network was at least
0.5 across different runs. The difference between the theory and the experiments in our model
can be ascribed to several factors: the finite size of the network; the choice of the finite learning
rate η, and the fact that we imposed a hard limit on the number of pattern presentations (see
number of iterations in Table 1), while the perceptron rule for excitatory synaptic connectivity
is only guaranteed to be optimal in the limit of η → 0, with a number of presentations inversely
proportional to η [25]. Note that the correspondence between the PLR and the 3TLR is only
perfect in the large γ limit, and is only approximate otherwise, as can be shown by comparing
explicitly the synaptic matrices obtained by both algorithms on the same set of patterns (see
Materials and Methods.)
A crucial ingredient of the 3TLR is having a strong external input which effectively acts as a
supervisory signal. How strong do the external fields need to be? How much does the capacity
depend on this strength? To answer these questions, we measured the maximum number of
stored patterns as a function of the parameter γ which determines the strength of external fields
as X = γ√N . This parameter, in fact, determines how far the two Gaussian distributions of the
local field are; as shown in Fig 2, the distance between the two peaks of the distribution is X. For
large enough γ, the overlap of these two distributions is negligible and the capacity is maximal;
but as we lower γ, the overlap increases, causing the learning rule to make mistakes, i.e. when
it should potentiate, it depresses the synapses and vice versa. In our simulations with N = 1001
neurons in the dense regime f = 0.5 at a fixed epsilon ǫ = 0.3, we varied γ and computed the
maximum α that can be achieved with a fixed number of iterations (1000). The capacity indeed
gradually decreases as γ decreases, until it reaches a threshold, below which there is a sharp
10
A
B
2 Gardner bound
Our model
Hopfield model
)
c
(cid:4)
(
y
t
i
c
a
p
a
C
1.5
1
0.5
0
0
2
)
c
(cid:4)
(
y
t
i
c
a
p
a
C
1.5
1
0.5
0
0
f = 0.5
(cid:1) = 6
0.2
0.4
0.6
Basin size
0.8
1
Gardner (bounded synapses)
Gardner (unbounded synapses)
Simulation (3TLR)
f = 0.5
(cid:1) = 6
1
Robustness((cid:3))
2
3
Figure 4: Critical capacity as a function of the basin size and the robustness param-
eter. A. The red plot shows the critical capacity as a function of the size of the basins of
attraction (N = 1001 neurons in the dense regime f = 0.5) when the strength of the external
field is large (γ = 6) such that the ON and OFF neuronal populations are well separated. The
points indicate 0.5 probability of successful storage at a given basin size, optimized over the
robustness parameter ǫ . The error bars show the [0.95, 0.05] probability interval for successful
storage. The blue plot shows the performance of the Hopfield model with N = 1001 neurons.
The maximal capacity at zero basin size (the Gardner bound) is equal to 2. B. To compare the
result of simulation of our model with the analytical results, we plotted the critical capacity as
a function of the robustness parameter ǫ. The dark red curve is the critical capacity versus ǫ
for our model obtained form analytical calculations (see Materials and Methods), the cyan line
shows the result of simulations of our model, and the dark blue shows the Gardner bound for
a network with no constraints on synaptic weights. The difference between the two theoretical
curves is due to the constraints on the weights in our network.
drop of capacity (see Fig 5). With the above values for the parameters, this transition occurs at
γ ≈ 2.4.
11
1
0.5
(cid:1)
m
u
m
i
x
a
M
0
0
f = 0.5
(cid:4) = 0.3
2
(cid:3)
4
6
Figure 5: Dependence of the critical capacity on the strength of the external input.
We varied the strength of the external field (γ) in order to quantify its effect on the learning
process. The critical capacity is plotted as a function of γ at a fixed robustness ǫ = 0.3 in the
dense regime f = 0.5. The simulations show that there is a very sharp drop in the maximum α
when γ goes below ≈ 2.4.
3
2.5
2
1.5
1
0.5
)
c
(cid:1)
(
y
t
i
c
a
p
a
C
0
0
Theory
Simulation
f = 0.2
(cid:4) = 12
1
Robustness ((cid:2))
2
3
Figure 6: Capacity as a function of the robustness parameter ǫ at sparseness f = 0.2.
The theoretical calculations is compared with the simulations for f = 0.2. Note that the capacity
in the sparse regime is higher than in the dense regime.
The 3TLR can also be adapted to work in a sparser regime, at a coding level lower than 0.5.
However, the average activity level of the network is determined by H0, and their relationship
also involves the variance of the distribution of the synaptic weights when f 6= 0.5 (see Materials
and Methods). During the learning process, the variance of the weights changes, which implies
that the parameter H0 must adapt correspondingly.
In our simulations, this adaptation was
performed after each complete presentation of the whole pattern set. In practice, this additional
12
self-stabilizing mechanism could still be performed in an unsupervised fashion along with (or
in alternation with) the learning process. Using this adjustment, we simulated the network at
f = 0.2 and compared the results with the theoretical calculations. As shown in Fig 6, we can
achieve at least 70% of the critical capacity across different values of the robustness parameter
ǫ.
f = 0.2
(cid:2) = 12
(cid:3) = 3
)
c
(cid:4)
(
1
0.8
0.6
0.4
0.2
0
0
0.25
0.5
0.75
)
Figure 7: Capacity as a function of correlations in the input patterns, for f = 0.2 at
ǫ = 3.0. Patterns are organized in categories, with a correlation c with the prototype of the
corresponding category (see text).
We also investigated numerically the effect of correlations in the input patterns. The PLR
is able to learn correlated patterns as long as a solution to the learning problem exists. As the
3TLR approximates the PLR, we expect the 3TLR to be able to learn correlated patterns as
well. As a simple model of correlation, we tested patterns organized in L categories [26,27]. Each
category was defined by a randomly generated prototype. Prototypes were uncorrelated from
category to category. For each category, we then generated p/L patterns independently with a
specified correlation coefficient c with the corresponding prototype. We show in Fig 7 the results
of simulations with L = 5, f = 0.2 and ǫ = 3. The figure shows that the learning rule reaches a
capacity that is essentially independent of c, in the range 0 ≤ c ≤ 0.75.
Statistical properties of the connectivity matrix
We next investigated the statistical properties of the connectivity matrix after the learning pro-
cess. Previous studies have shown that the distribution of synaptic weights in perceptrons with
excitatory synapses becomes at maximal capacity a delta function at zero weight, plus a truncated
Gaussian for strictly positive weights [25, 28–30]. Our model differs from this setting because of
the global inhibitory feedback. Despite this difference, the distribution of weights in our network
bear similarities with the results obtained in these previous studies: the distribution exhibits a
peak at zero weight (‘silent’, or ‘potential’ synapses), while the distribution of strictly positive
weights resembles a truncated Gaussian. Finally, the fraction of silent synapses increases with
the robustness parameter (see Fig 8).
13
C
2
1.5
1
0.5
y
t
i
s
n
e
D
0
0
(cid:1) = 0
P(w<0.05)=0.01
1
3
Weight strength
2
(cid:1) = 1.2
P(w<0.05)=0.250
(cid:1) = 0.66
P(w<0.05)=0.245
2
1.5
1
0.5
2
1.5
1
0.5
0
0
4
2
4
6
0
0
2
4
6
Figure 8: Synaptic weight distributions. Comparing the distributions of the synaptic weights
at critical capacity for three different values of robustness obtained from simulation. The distri-
bution of weights approaches a Dirac-delta distribution at zero plus a truncated Gaussian. As
the patterns become more robust, the center of the partial Gaussian shifts towards the left, and
the number of silent synapses increases.
We have also computed the degree of symmetry of the weight matrix. The symmetry degree
is computed as the Pearson correlation coefficient between the reciprocal weights in pairs of
neurons. We observe a general trend towards an increasingly symmetric weight matrix as more
patterns are stored, for all values of the robustness parameter ǫ (see Fig 9).
1
0.8
0.6
0.4
0.2
(cid:2) = 0
(cid:2) = 0.6
(cid:2) = 1.2
)
j
i
i
w
j,
w
(
r
r
o
C
0
0
0.5
1
M
1.5
(cid:1))
2
Figure 9: The degree of symmetry of the weight matrix. The Pearson correlation co-
efficient between wij and wji is computed at different values of α for three values of ǫ. As α
increases the weight matrix tends to be more symmetric, but gets saturated for high α. For the
same values of α, as the robustness increases, the correlation also increases, so the weight matrix
becomes more symmetric. Error bars (across 10 runs) are smaller than the symbols.
14
Discussion
We presented a biologically-plausible learning rule that is characterized by three thresholds, and
is able to store memory patterns close to the maximal storage capacity in a recurrent neural
networks without the need of an explicit “error signal”. We demonstrated how the learning rule
can be considered a transformed version of the PLR in the limit of a strong external field. Our
network implements the separation between excitatory and inhibitory neurons, with learning
occurring only at excitatory-to-excitatory synapses. We simulated a recurrent network with
N = 1001 binary neurons, reaching to αc = 1.6 at zero basin size. We then used a robustness
parameter ǫ to enlarge the basin size. The simulations showed that we are close to the theoretical
capacity across the whole investigated range of values of ǫ. We expect that as N increases and
the learning rate gets smaller, this difference would go to zero.
Two crucial ingredients of the 3TLR are necessary: (1) strong external inputs, (2) three
learning thresholds which are set according to the statistics of inputs to the neuron. The learning
rule only uses information that is local to a synapse and corresponding neurons. Like classic
Hebbian learning rules, our 3TLR works in an online fashion. In addition, it can also perform as
a ‘palimpsest’ [31–33]: in case the total number of patterns exceeds the maximal capacity (at a
certain basin size) the network begins to forget patterns that are not being presented anymore.
Comparison with other learning rules
The 3TLR can be framed in the setting of the classic Bienenstock-Cooper-Munro (BCM) the-
ory [34, 35], with additional requirements to adapt it to the attractor network scenario. The
original BCM theory uses firing-rate units, and prescribes that synaptic modifications should be
proportional to (1) the synaptic input, and (2) a function φ (v) of the total input v (or, equiv-
alently, of the total output). The function φ (v) is subject to two conditions: (1) φ (v) ≥ 0 (or
≤ 0) when v > θ (or < θ, respectively); (2) φ (0) = 0. The parameter θ is also assumed to
change, but on a longer time scale (such that the changes reflect the statistics of the inputs); this
(metaplastic) adaptation has the goal of avoiding the trivial situations in which all inputs elicit
indistinguishable responses. This (loosely specified) framework ensures that, under reasonable
conditions, the resulting units become highly selective to a subset of the inputs, and has been
mainly used to model the developmental stages of primary sensory cortex. The arising selectiv-
ity is spontaneous and completely unsupervised: in absence of further specifications, the units
become selective to a random subset of the inputs (e.g. depending on random initial conditions).
Our model is defined on simpler (binary) units; however, if we define φ (v) = Θ (v − θ) Θ (θ1 − v)−
Θ (θ − v) Θ (v − θ0), then φ behaves according to the prescriptions of the BCM theory. Further-
more, we have essentially assumed the same slow metaplastic adaptation mechanism of BCM,
even though we have assigned this role explicitly to the inhibitory part of the network (see Ma-
terials and Methods). On the other hand, our model has additional requirements: (1) φ (v) = 0
when v < θ0 or v > θ1, (2) plasticity occurs during presentation of external inputs, which in
turn are strong enough to drive the network towards a desired state. The second requirement
ensures that the network units become selective to a specific subset of the inputs, as opposed
to a random subset as in the original BCM theory, and thus that they are able to collectively
behave as an attractor network. The first requirement ensures that each unit operates close to
critical capacity. Indeed, these additional requirements involve extra parameters with respect to
the BCM theory, and we implicitly assume these parameters to also slowly adapt according to
the statistics of the inputs during network formation and development.
A variant of the BCM theory, known as ABS rule [36, 37] introduced a lower threshold for
LTD, analogous to our θ0, motivated by experimental evidence; however, a high threshold for LTP,
15
analogous to our θ1, was not used there, or — to our knowledge — in any other BCM variant. The
idea of stopping plasticity above some value of the ‘local field’ has been introduced previously to
stabilize the learning process in feed-forward networks with discrete synapses [38–40]. Our study
goes beyond these previous works in generalizing such a high threshold to recurrent networks,
and showing that the resulting networks achieve close to maximal capacity.
Comparison with data and experimental predictions
In vitro experiments have characterized how synaptic plasticity depends on voltage [41] and fir-
ing rate [42], both variables that are expected to have a monotonic relationship with the total
excitatory synaptic inputs received by a neuron. In both cases, a low value of the controlling
variable leads to no changes; intermediate values lead to depression; and high values to potenti-
ation. These three regimes are consistent with the three regions for v < θ1 in Fig 2. The 3TLR
predicts that a fourth region should occur at sufficiently high values of the voltage and/or firing
rates. Most of the studies investigating the dependence of plasticity on firing rate or voltage have
not reported a decrease in plasticity at high values of the controlling variables, but these studies
might have not increased sufficiently such variables. To our knowledge, a single study has found
that at high rates, the plasticity vs rate curve is a decreasing function of the input rate [43].
Another test of the model consists in comparing the statistics of the synaptic connectivity with
experimental data. As it has been argued in several recent studies [25,28,30,44,45], networks with
plastic excitatory synapses are generically sparse close to maximal capacity, with a connection
probability that decreases with the robustness of information storage, consistent with short range
cortical connectivity [46,47]. Our network is no exception, though the fraction of silent synapses
that we observe is significantly lower than in models that lack inhibition. Furthermore, network
that are close to maximal capacity tends to have a connectivity matrix that has a significant
degree of symmetry, as illustrated by the over-representation of bidirectionally connected pairs
of neurons, and the tendency of bidirectionally connected pairs to form stronger synapses than
unidirectionally connected pairs as observed in cortex [47,48], except in barrel cortex [49]. Again,
the 3TLR we have proposed here reproduces this feature (Fig 9), consistent with the fact that
the rule approaches the optimal capacity.
Future directions
Our network uses the simplest possible single neuron model [50]. One obvious direction for future
work would be to implement the learning rule in a network of more realistic neuron models such as
firing rate models or spiking neuron models. Another potential direction would be to understand
the biophysical mechanisms leading to the high threshold in the 3TLR. In any case, we believe
the results discussed here provide a significant step in the quest for understanding how learning
rules in cortical networks can optimize information storage capacity.
Materials and Methods
Simulation
The main equations of the network, the neuron model, the learning rule, and the criteria for
stopping the learning algorithm are outlined in the Results section, Eqs. 1-7. We present here
additional details about network simulations.
16
Network setup before learning process
Before applying the learning rule, we required the network to have stable dynamics around a
desired activity level f . A network with only excitatory neurons is highly unstable and typically
converges towards the trivial all-off and all-on states; therefore, we implemented a global inhibi-
tion such that the network operates around activity level f . The basal inhibitory term (H0) and
the inhibitory reaction term (H1) are defined as:
H0 = (N − 1)(f ¯w − ψ) + H−1(f )p(N − 1)f σw
H1 = f γ√N − 1
(11)
(12)
2 erfc(cid:16) x√2(cid:17) and H−1 is the inverse of H, ψ is defined as θ = (N − 1)ψ; ¯w
where H (x) = 1
and σw are the mean and standard deviation of the synaptic weights, respectively. With these
definitions the network dynamics is stable in the sense that the activity level converges to f very
fast, regardless of the initial condition.
In Eq. 11, we see that H0 depends on the activity level f and on the standard deviation of
the weights σw. In the dense regime, f = 0.5, we have H−1(0.5) = 0, therefore the rightmost
term of Eq. 11 vanishes, which means that in this regime H0 is independent of σw. However,
in sparser regimes, the network must be endowed with a mechanism to adjust for the changes
in standard deviation, otherwise the learning process would bring the network out of the stable
state, changing the basal activity level. In contrast, the mean synaptic efficacy ¯w does not change
significantly during the learning process.
In all our simulations, the initial values for {wij} were sampled from a Gaussian distribution
with mean and standard deviation equal to one, after which negative values were set to zero.
This has the effect the ¯winit
is slightly higher than one. We also set wii = 0 for all i.
ij
Table 1 shows the values of the parameters used in the simulations, in the dense and sparse
regimes.
Table 1: Table of parameters in the simulation
ij
Parameter name
N
λ = ¯winit
f
ψ
θ
η
γ
# of interations (learning)
# of trials in test phase
Value in dense regime Value in sparse regime
1001
≈1.08
0.5
0.35
350
0.01 [0.001 when ǫ = 0]
6.0
1000 [10000 when ǫ = 0]
50
1001
≈1.08
0.2
0.35
350
0.01 [0.001 when ǫ = 0]
12.0
1000 [10000 when ǫ = 0]
50
Direct comparison between the 3TLR and the PLR
In order to determine the degree to which the 3TLR is able to mimic the PRL, and the effect
of deviations from the latter rule, we tested both rules on the same tasks. In these simulations,
every part of the simulation code was kept identical — including the pseudo-random numbers
used to choose the initial state and the arbitrary permutations for the update order of the units
— except for the learning rule. We tested the network in the dense case f = 0.5, at ǫ = 3, varying
17
1
0.8
0.6
0.4
0.2
(cid:2) = 6)
1245 6
P4571245 6
(cid:2) = 12)
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
)
6
=
(cid:2)
j#
w
i
0
0
0.05
0.1
0
0.2
0.15
(cid:1))
Figure 10: Direct comparions of the 3TLR and the PLR. Success probability for the 3TLR
at γ = 6 (blue curve, left axis) and the PLR (red curve); the results for the 3TLR at γ = 12
are identical to those of the PLR (red curve). The orange points show the absolute difference of
weights between the final values of the weights for the PLR at γ = 6 and the PLR (right axis):
the points show the median of the distribution, while the error bars span the 5th-95th percentiles,
showing that, while the distribution is concentrated at near-zero values, outliers appear at the
critical capacity of the 3TLR algorithm. (Note that the average value of the weights is in all cases
approximately 1.08; also compare the discrepancies with the overall distribution of the weights,
Fig 8.)
the storage load α, using 10 samples for each point. We compared the probability of solving the
learning task and the distribution of the discrepancies (absolute value of the differences) in the
values of the resulting synaptic weights. We tested two values of the parameter γ, 6 (as in Fig 4)
and 12. We found that at γ = 12 there was absolutely no difference between the two rules, while
at γ = 6 the 3TLR performed slightly worse, and significant deviations from the PLR started to
appear close to the maximal capacity of the 3TLR (see Fig 10).
Analytical calculation of the storage capacity at infinite N
Entropy calculation
In this section, we present the details of the calculations for the typical storage capacity of our
network in the limit of N → ∞, using the Gardner analysis [20, 28].
typically be found.
The capacity is defined as the maximum value of α = p/N such that a solution to Eq. 7 can
We can rewrite Eq. 7 as
∀i :
αN
Yµ=1
where
Θ
(2ξµ
i − 1)
wij ξµ
N
ξµ
N
Xj=1
Xj=1
j − f N
i − H0 − λ
− θ
H0 = N f ¯w − θ + H−1 (f ) σwpf N
λ = ¯w
− f ǫ√N
= 1
(13)
(14)
(15)
18
S
p
!
"
$
f
f
%
&
%
n
'
%
(
%
)
*
%
%
n
+
L
-
.
3
/
L
-
0
Eq. 13 becomes:
∀i :
αN
Yµ=1
Θ
(2ξµ
i − 1)
N
Xj=1
(wij − ¯w) ξµ
i − H−1 (f ) σwpf N
− f ǫ√N
= 1
Let us now consider a single unit i. We write σµ
i = (2ξµ
as Wij = wij
¯w − 1 ∈ [−1,∞), and also define
i − 1), and re-parametrize the weights
(16)
(17)
(18)
(19)
T = H−1 (f )pf
K =
.
ǫ
¯w
Dropping the index i and neglecting terms of order 1, we obtain:
αN
Yµ=1
Θ
σµ
Wjξµ
j − T
σw
¯w
N
Xj=1
√N
− f K√N
= 1
Our goal is to compute the quenched entropy of this problem, i.e. the scaled average of the
logarithm of the volume of W which satisfies the above equation:
S =
=
1
N hlog V i{ξµ,σµ}
N *logZ
1
Yj=1
N
(dWjΘ (Wj + 1))
αN
Yµ=1
Θ
σµ
N
Xj=1
Wj ξµ
j −
σw
¯w
T√N
− f K√N
+
(20)
{ξµ,σµ}
The computation proceeds along the lines of [20, 28], by using the so-called replica trick to
perform the average of the logarithm of V , exploiting the identity:
hlog V i = lim
n→0
hV ni − 1
n
,
(21)
performing the computation for integer values of n and using an analytical continuation to
perform the limit n → 0. we perform the calculation using the replica-symmetric (RS) Ansatz,
which is believed to give exact results in the case of perceptron models with continuous weights.
The final expression for the entropy depends on six order parameters; the first three are Q, q
and M , whose meaning is
Q =
q =
M =
(Wj)2
W a
j W b
j
Wj
1
1
N Xj
N Xj
√N Xj
1
where we used W a and W b to denote two different replicas of the system, which can simply be
interpreted as two independent solutions to the constraint equation. Q is called the self-overlap,
19
are the conjugate quantities Q, q and M . The entropy expression is:
in our case, while q is the mutual-overlap. The remaining order parameters
¯w (cid:1)2
and is equal to(cid:0) σw
where
q q
S(cid:16)Q, q, M, Q, q, M(cid:17) = −(cid:18)Q Q −
2 (cid:19) + αZA (Q, q, M ) + ZW (cid:16) Q, q, M(cid:17)
ZA (Q, q, M ) = Z Du*ln H K − σ(cid:0)M − T√Q(cid:1) + u (1 − f )√q
ZW (cid:16) Q, q, M(cid:17) = Z Du ln(cid:18)Z ∞
2 (q−2 Q)W 2+W (u√q− M)(cid:19) .
(1 − f )√Q − q
W e− 1
−1
!!+σ
(22)
(23)
(24)
2
u
x Du = 1
We used the usual notation Du ≡ du e−
2√2π
2 erfc(cid:16) x√2(cid:17). In the following, we will also use the shorthand G (x) = G(x)
H (x) = R ∞
H(x) .
We also used the notation h·iσ to denote the average over the output σ, i.e. hϕ (σ)iσ = f ϕ (1) +
(1 − f ) ϕ (−1) for any function ϕ. The value of the order parameters is found by extremizing S.
The notation and the following computations can be simplified using:
= du G (u) to denote Gaussian integrals, and defined
∆Q = Q − q
tσ (u) =
K − σ(cid:0)M − T√Q(cid:1) + u (1 − f )√q
(1 − f )√∆Q
2 ∆ QW 2+W (u√q− M)
∆ Q = q − 2 Q
ν (u, W ) = e− 1
The extremization of S then results in the system of equations:
∆ Q =
q =
α
p(Q − ∆Q) ∆QZ Du u hG (tσ (u))iσ
∆QZ Du hG (tσ (u)) tσ (u)iσ + ∆ Q
α
0 = Z Du hG (tσ (u)) σiσ
Q = Z Du R ∞
−1 dW W 2ν (u, W )
R ∞
−1 dW ν (u, W )
√q Z Du uR ∞
−1 dW W ν (u, W )
R ∞
−1 dW ν (u, W )
0 = Z Du R ∞
−1 dW W ν (u, W )
R ∞
−1 dW ν (u, W )
1
∆Q =
20
(25)
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
The integrals over dW in the last three equations can be performed explicitly, yielding:
Q =
q + M 2 + ∆ Q
∆ Q2
∆Q =
1
∆ Q
+
0 = −
M
∆ Q
Critical capacity
+
3
1
+
1
∆ Q
2 Z Du (cid:16)upq − M − ∆ Q(cid:17) G
−
q∆ Qq Z Du uG
−
u√q − M + ∆ Q
q∆ QZ Du G
−
q∆ Q
q∆ Q
u√q − M + ∆ Q
1
u√q − M + ∆ Q
q∆ Q
(35)
(36)
(37)
At critical capacity, the space of the solutions shrinks to a point, and the mutual overlap tends
to become equal to the self overlap: q → Q, i.e. ∆Q → 0. In this limit, the conjugate order
parameters diverge as:
q =
∆ Q =
M =
C
∆Q2
A
∆Q
B√C
∆Q
(38)
(39)
(40)
Using these conditions, and calling αc the critical value of α, the saddle point equations, 29
to 34, become:
1
A
Q =
A = H(cid:18)B −
A(cid:16)C − B√C(cid:17)
√C(cid:19)
√C
A (cid:18)G(cid:18)B −
C = αcQ(cid:10)(cid:0)1 + τ 2
A = αc hH (τσ)iσ
0 = hσ (G (τσ) − τσH (τσ))iσ
0 =
A
√C(cid:19) − BA(cid:19) − (1 − A)
σ(cid:1) H (τσ) − τσG (τσ)(cid:11)σ
(41)
(42)
(43)
(44)
(45)
(46)
(47)
where we defined
τσ =
σ(cid:0)M − T√Q(cid:1) − K
(1 − f )√Q
These equations can be solved numerically to find the six parameters αc, Q, A, B, C and M .
Note that in the special case K = 0 these equations have a degenerate solution with Q = 0
and the same αc as in the case of unbounded synaptic weights (e.g. αc = 2 for f = 0.5). This is
because in that case the original problem has the property that scaling all weights by a factor of
x is equivalent to scaling the boundary ¯w by a factor of x−1 (see Eq. 16); therefore, the optimal
strategy is to exploit this property by setting x → 0, i.e. effectively reducing the problem to the
unbounded case. Of course, this strategy can only be pursued up to the available precision in a
practical setting.
21
Acknowledgments
References
1. Hopfield JJ. Neural networks and physical systems with emergent collective computational
abilities. Proc Natl Acad Sci USA. 1982;79:2554–2558.
2. Amit DJ. Modeling brain function. Cambridge University Press; 1989.
3. Hertz J, Krogh A, Palmer RG.
Introduction to the Theory of Neural Computation.
Addison-Wesley, Redwood City; 1991.
4. Amit DJ, Brunel N. Model of global spontaneous activity and local structured activity
during delay periods in the cerebral cortex. Cerebral Cortex. 1997;7:237–252.
5. Brunel N, Wang XJ. Effects of neuromodulation in a cortical network model of object
working memory dominated by recurrent inhibition. J Comput Neurosci. 2001;11:63–85.
6. Mongillo G, Barak O, Tsodyks M. Synaptic Theory of Working Memory. Science.
2008;319:1543.
7. Barak O, Tsodyks M. Working models of working memory. Curr Opin Neurobiol.
2014;25:20–24.
8. Fuster JM, Jervey JP. Inferotemporal neurons distinguish and retain behaviourally relevant
features of visual stimuli. Science. 1981;212:952–955.
9. Miyashita Y. Neuronal correlate of visual associative long-term memory in the primate
temporal cortex. Nature. 1988;335:817–820.
10. Miyashita Y, Chang HS. Neuronal correlate of pictorial short-term memory in the primate
temporal cortex. Nature. 1988;331:68–70.
11. Nakamura K, Kubota K. Mnemonic firing of neurons in the monkey temporal pole during
a visual recognition memory task. J Neurophysiol. 1995;74:162–178.
12. Fuster JM, Alexander G. Neuron activity related to short-term memory.
Science.
1971;173:652–654.
13. Funahashi S, Bruce CJ, Goldman-Rakic PS. Mnemonic coding of visual space in the
monkey’s dorsolateral prefrontal cortex. J Neurophysiol. 1989;61:331–349.
14. Romo R, Brody CD, Hern´andez A, Lemus L. Neuronal correlates of parametric working
memory in the prefrontal cortex. Nature. 1999;399:470–474.
15. Amit DJ, Gutfreund H, Sompolinsky H. Storing infinite numbers of patterns in a spin-glass
model of neural networks. Phys Rev Lett. 1985;55:1530–1531.
16. Sompolinsky H. Neural networks with nonlinear synapses and a static noise. Phys Rev A.
1986;34:2571–2574.
17. Derrida B, Gardner E, Zippelius A. An exactly solvable asymmetric neural network model.
Europhys Lett. 1987;4:167–173.
22
18. Tsodyks M, Feigel’man MV. The enhanced storage capacity in neural networks with low
activity level. Europhys Lett. 1988;6:101–105.
19. Buhmann J, Divko R, Schulten K. Associative memory with high information content.
Phys Rev A. 1989;39:2689–2692.
20. Gardner EJ. The phase space of interactions in neural network models. J Phys A: Math
Gen. 1988;21:257–270.
21. Rosenblatt F. Principles of neurodynamics. Spartan Books, New York; 1962.
22. Marr D. A theory of cerebellar cortex. J Physiol. 1969;202:437–470.
23. Albus JS. A theory of cerebellar function. Mathematical Biosciences. 1971;10:26–51.
24. Ito M, Sakurai M, Tongroach P. Climbing fibre induced depression of both mossy fibre re-
sponsiveness and glutamate sensitivity of cerebellar Purkinje cells. J Physiol. 1982;324:113–
134.
25. Clopath C, Nadal JP, Brunel N. Storage of correlated patterns in standard and bistable
Purkinje cell models. PLoS Comput Biol. 2012;8:e1002448.
26. Parga N, Virasoro MA. The ultrametric organization of memories in a neural network.
J Phys France. 1986;47:1857–1864.
27. Brunel N, Carusi F, Fusi S. Slow stochastic Hebbian learning of classes in recurrent neural
networks. Network. 1998;9:123–152.
28. Brunel N, Hakim V, Isope P, Nadal JP, Barbour B. Optimal information storage and the
distribution of synaptic weights: perceptron versus Purkinje cell. Neuron. 2004;43:745–57.
29. Brunel N, van Rossum MC. Lapicque’s 1907 paper: from frogs to integrate-and-fire. Biol
Cybern. 2007;97:337–339.
30. Clopath C, Brunel N. Optimal properties of analog perceptrons with excitatory weights.
PLoS Comput Biol. 2013;9:e1002919.
31. M M´ezard, Nadal JP, Toulouse G. Solvable models of working memories. J Physique.
1986;47:1457–.
32. Parisi G. A memory which forgets. J Phys A: Math Gen. 1986;19:L617.
33. Amit DJ, Fusi S. Dynamic learning in neural networks with material synapses. Neural
Computation. 1994;6:957–982.
34. Bienenstock E, Cooper L, Munro P. Theory for the development of neuron selectivity:
orientation specificity and binocular interaction in visual cortex. J Neurosci. 1982;2:32–48.
35. Jedlicka P. Synaptic plasticity, metaplasticity and BCM theory. Bratislavsk´e lek´arske listy.
2002;103(4/5):137–143.
36. Brocher S, Artola A, Singer W. Intracellular injection of Ca2+ chelators blocks induction
of long-term depression in rat visual cortex. Proceedings of the National Academy of
Sciences. 1992;89(1):123–127.
23
37. Artola A, Singer W. Long-term depression of excitatory synaptic transmission and its
relationship to long-term potentiation. Trends in neurosciences. 1993;16(11):480–487.
38. Amit Y, Mascaro M. Attractor networks for shape recognition. Neural Comput.
2001;13:1415–1442.
39. Fusi S, Drew PJ, Abbott LF. Cascade models of synaptically stored memories. Neuron.
2005;45:599–611.
40. Brader JM, Senn W, Fusi S. Learning real-world stimuli in a neural network with spike-
driven synaptic dynamics. Neural Comput. 2007;19:2881–2912.
41. Ngezahayo A, Schachner M, Artola A. Synaptic activity modulates the induction of bidi-
rectional synaptic changes in adult mouse hippocampus. J Neurosci. 2000;20:2451–2458.
42. Kirkwood A, Rioult MC, Bear MF. Experience-dependent modification of synaptic plas-
ticity in visual cortex. Nature. 1996;381:526–528.
43. Wang H, Wagner JJ. Priming-induced shift in synaptic plasticity in the rat hippocampus.
J Neurophysiol. 1999;82:2024–2028.
44. Barbour B, Brunel N, Hakim V, Nadal JP. What can we learn from synaptic weight
distributions? Trends Neurosci. 2007;30:622–629.
45. Chapeton J, Fares T, LaSota D, Stepanyants A. Efficient associative memory stor-
age in cortical circuits of inhibitory and excitatory neurons. Proc Natl Acad Sci USA.
2012;109:E3614–3622.
46. Kalisman N, Silberberg G, Markram H. The neocortical microcircuit as a tabula rasa.
Proc Natl Acad Sci U S A. 2005;102:880–885.
47. Song S, Sjostrom PJ, Reigl M, Nelson S, Chklovskii DB. Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS Biol. 2005;3:e68.
48. Wang Y, Markram H, Goodman PH, Berger TK, Ma J, Goldman-Rakic PS. Heterogeneity
in the pyramidal network of the medial prefrontal cortex. Nat Neurosci. 2006;9:534–542.
49. Lefort S, Tomm C, Floyd Sarria JC, Petersen CC. The excitatory neuronal network of the
C2 barrel column in mouse primary somatosensory cortex. Neuron. 2009;61:301–316.
50. McCulloch WS, Pitts WA. A logical calculus of the ideas immanent in nervous activity.
Bull Math Biophys. 1943;5:115–133.
24
|
1212.0076 | 2 | 1212 | 2012-12-04T03:15:35 | Short term synaptic depression improves information transfer in perceptual multistability | [
"q-bio.NC",
"math.DS",
"nlin.PS"
] | Competitive neural networks are often used to model the dynamics of perceptual bistability. Switching between percepts can occur through fluctuations and/or a slow adaptive process. Here, we analyze switching statistics in competitive networks with short term synaptic depression and noise. We start by analyzing a ring model that yields spatially structured solutions and complement this with a study of a space-free network whose populations are coupled with mutual inhibition. Dominance times arising from depression driven switching can be approximated using a separation of timescales in the ring and space-free model. For purely noise-driven switching, we use energy arguments to justify how dominance times are exponentially related to input strength. We also show that a combination of depression and noise generates realistic distributions of dominance times. Unimodal functions of dominance times are more easily differentiated from one another using Bayesian sampling, suggesting synaptic depression induced switching transfers more information about stimuli than noise-driven switching. Finally, we analyze a competitive network model of perceptual tristability, showing depression generates a memory of previous percepts based on the ordering of percepts. | q-bio.NC | q-bio |
SHORT TERM SYNAPTIC DEPRESSION IMPROVES
INFORMATION TRANSFER IN PERCEPTUAL MULTISTABILITY
ZACHARY P. KILPATRICK
Abstract. Competitive neural networks are often used to model the dynamics of perceptual
bistability. Switching between percepts can occur through fluctuations and/or a slow adaptive pro-
cess. Here, we analyze switching statistics in competitive networks with short term synaptic depres-
sion and noise. We start by analyzing a ring model that yields spatially structured solutions and
complement this with a study of a space-free network whose populations are coupled with mutual
inhibition. Dominance times arising from depression driven switching can be approximated using
a separation of timescales in the ring and space-free model. For purely noise-driven switching, we
use energy arguments to justify how dominance times are exponentially related to input strength.
We also show that a combination of depression and noise generates realistic distributions of domi-
nance times. Unimodal functions of dominance times are more easily differentiated from one another
using Bayesian sampling, suggesting synaptic depression induced switching transfers more informa-
tion about stimuli than noise-driven switching. Finally, we analyze a competitive network model
of perceptual tristability, showing depression generates a memory of previous percepts based on the
ordering of percepts.
Key words. binocular rivalry, neural field, ring model, bump attractor, short term depression
1. Introduction. Ambiguous sensory stimuli with two interpretations can pro-
duce perceptual rivalry [5]. For example, two orthogonal gratings presented to either
eye lead to perception switching between one grating and then the next repetitively,
a common paradigm known as binocular rivalry [31]. Perceptual rivalry can also be
triggered by a single stimulus with two interpretations, like the Necker cube [39]. One
notable feature of the switching process in perceptual rivalry is its stochasticity -- a
histogram of the dominance times of each percept spreads across a broad range [17].
Senses other than vision also exhibit perceptual rivalry. When two different odorants
are presented to the two nostrils, a similar phenomenon occurs with olfaction, termed
"binaral" rivalry [55]. Similar experiences have been evoked in the auditory [13, 41]
and tactile [10] system.
Multiple principles govern the relationship between the strength of percepts and
the mean switching statistics in perceptual rivalry [32]. "Levelt's propositions" relate
stimulus contrast to the mean dominance times: (i) increasing the contrast of one
stimulus increases the proportion of time that stimulus is dominant; (ii) increasing
the contrast of one stimulus does not affect its average dominance time; (iii) increas-
ing contrast of one stimulus increases the rivalry alternation rate; and (iv) increasing
the contrast of both stimuli increases the rivalry alternation rate. There are also
relationships between the properties of the input and the stochastic variation in the
dominance times [6]. For instance, the distribution of dominance times is well fit
by a gamma distribution [17, 30, 47]. The fact that dominance times are not expo-
nentially distributed suggests some background slow adaptive process plays a role in
providing a nonzero peak in the dominance histograms [44]. Two commonly proposed
mechanisms for this adaptation are spike frequency adaptation and short term synap-
tic depression [29, 49, 43]. An even stronger case can be made of the existence of
adaptation in perceptual processing networks by examining results of experiments on
perceptual tristability [21]. Here, perception alternates between three possible choices
and subsequent switches are determined by the previous switch [38]. This memory
suggests switches in perceptual multistability are not purely noise-driven [37].
Most theoretical models of perceptual rivalry employ two pools of neurons, each
1
selective to one percept, coupled to one another by mutual inhibition [33, 29, 43, 42].
With no other mechanisms at work, such architectures lead to winner-take-all states,
where one pool of neurons inhibits the other indefinitely [48]. However, switches
between the dominance of one pool and the other can be initiated with the inclusion
of fluctuations [37] or an adaptive process [29, 43]. Combining the two effects leads to
dominance times that are distributed according to the gamma distribution [29, 44, 47].
Slow adaptation and noise thus serve as agents for the sampling of the stimulus
through network activity. A mutual inhibitory network would otherwise remain in
the winner-take-all state indefinitely.
In light of these observations, we wish to consider the role adaptive mechanisms
play in properly sampling ambiguous stimuli in the context of a mutual inhibitory
network. Purely fluctuation driven switching would provide a noisy sample of the two
percepts, but pure adaptation would provide an extremely reliable sampling of percept
contrast [44]. Thus, as the level of adaptation is increased and noise is decreased, one
would expect that the ability of mutual inhibitory networks to encode information
about ambiguous stimuli is vastly improved. A major point is that it is also vastly
improved over networks without any adaptation at all. We focus specifically on short
term synaptic depression [1, 45].
Using parametrized models, we will explore how synaptic depression improves the
ability of a network to extract stimulus contrasts. First, we will be concerned with how
much information can be determined about the contrast of each of the two percepts
of an ambiguous stimulus. In the case of a winner-take-all solution, only information
about a single percept could be known, since the pool of neurons encoding the other
percept would be quiescent. We will study this problem using an anatomically moti-
vated neural field model of an orientation column with synaptic depression [54, 25],
given by (1). We find that increasing the strength of synaptic depression from zero
leads to a bifurcation whereby rivalrous oscillations onset. When rivalrous switching
occurs through a combination of depression and noise, we show stronger depression
improves the transfer of information using simple Bayesian inference [24]. We also
analyze a competitive network model with depression and noise (6) to help study the
combined effects of noise and depression on perceptual switching. In particular, we
will show that the presence of synaptic depression increases the information relayed
by the output of the network. Finally, we will show trimodal stimuli to a neural field
model with synaptic depression can generate oscillations where each mode spends time
in dominance. To deeply analyze the relative contributions of noise and depression to
this switching process, we study a reduced model. This reveals depression generates
a history dependence in switching that would not arise in the network with purely
noise-driven switching.
2. Materials and Methods.
2.1. Ring model with synaptic depression. As a starting point, we will
consider a model for processing the orientation of visual stimuli [3, 7] which also
includes short term synaptic depression [54, 25]. Since GABAergic inhibition is much
faster than AMPA-mediated excitation [23], we make the assumption that inhibition
is slaved to excitation as in [2]. Reduction this disynaptic pathway and assuming
depression acts on on excitation [45], we then have the model [54, 26]
(cid:90) π/2
−π/2
τm
∂u(x, t)
∂t
= −u(x, t) +
w(x − y)q(y, t)f (u(y, t))dy + I(x) + ξ(x, t),
(1a)
2
τ
∂q(x, t)
∂t
= 1 − q(x, t) − βq(x, t)f (u(x, t)).
(1b)
Here u(x, t) measures the synaptic input to the neural population with stimulus pref-
erence x at time t, evolving on the timescale τm. Firing rates are given by taking the
gain function f (u) of the synaptic input, which we usually proscribe to be [51]
and often take the high gain limit γ → ∞ for analytical convenience, so [2, 26]
f (u) = H(u − κ) =
: u < κ,
: u ≥ κ.
f (u) =
1 + e−γ(u−κ)
,
1
(cid:26) 0
1
(2)
(3)
External input, representing flow from upstream in the visual system is prescribed by
the time-independent function I(x) [3, 7]. For the majority of our study of (1), we
employ the bimodal stimulus
I(x) = −I0 cos(4x) + Ia sin(2x),
(4)
representing stimuli at the two orthogonal angles −π/4 and π/4 and I0 controls the
mean of each peak and Ia controls the level of asymmetry between the peaks. Ef-
fects of noise are described by the stochastic processe (cid:104)ξ(x, t)(cid:105) with (cid:104)ξ(x, t)(cid:105) = 0 and
(cid:104)ξ(x, t)ξ(y, s)(cid:105) = C(x − y)δ(t − s). For simplicity, we assume the spatial correlations
have a cosine profile C(x) = π cos(x). Synaptic interactions are described by the inte-
gral term, so w(x − y) describes the strength (amplitude of w) and net polarity (sign
of w) of synaptic interactions from neurons with stimulus preference y to those with
preference x. Following previous studies, we presume the modulation of the synaptic
strength is given by the cosine
w(x − y) = cos(2(x − y)),
(5)
so neurons with similar orientation preference excite one another and those with
dissimilar orientation preference disynaptically inhibit one another [3, 15]. The factor
q(x, t) measures of the fraction of available presynaptic resources, which are depleted
at a rate βf [1, 45], and are recovered on a timescale specified by the time constant
τ [11].
By setting τm = 1, we can assume time evolves on units of the excitatory synaptic
time constant, which we presume to be roughly 10ms [18]. Experimental observations
have shown synaptic resources specified q are recovered on a timescale of 200-800ms
[1, 46], so we require τ is between 20 and 80, usually setting it to be τ = 50. Our
parameter β can then be varied independently to adjust the effective depletion rate
of synaptic depression.
2.2. Idealized competitive neural network. We also study space-free com-
petitive neural networks with synaptic depression [43]. In this way, we can make more
progress analyzing switching behavior. As a general model of networks connected by
mutual inhibition, we consider the system [29, 37, 43]
uR(t) = −u1(t) + f (IR − qL(t)uL(t)) + ξ1(t),
uL(t) = −u2(t) + f (IL − qR(t)uR(t)) + ξ2(t),
τ qR(t) = 1 − qR(t) − βuR(t)qR(t),
τ qL(t) = 1 − qL(t) − βuL(t)qL(t),
3
(6a)
(6b)
(6c)
(6d)
where uj represents the firing rate of the j = 1, 2 population. The strength of recurrent
synaptic excitation within a population is specified by the parameter α, whereas the
strength of cross-inhibition between populations is specified by β. Fluctuations are
introduced into population j with the independent white noise processes ξj with
(cid:104)xj(t)(cid:105) = 0 and (cid:104)ξj(t)ξj(s)(cid:105) = εδ(t − s).
2.3. Numerical simulation of stochastic differential equations. The spa-
tially extended model (1) was simulated using an Euler-Maruyama method with a
timestep dt = 10−4, using Riemann integration on the convolution term with N
= 2000 spatial grid points. The space clamped competitive network (6) was also
simulated using Euler-Maruyama with a timestep dt = 10−6. To generate histograms
of dominance times, we simulated systems for 20000s (2 × 106 time units).
2.4. Fitting dominance time distributions. To generate the theoretical curves
presented for exponentially distributed dominance times, we simply take the mean of
dominance times and use it as the scaling in the exponential (40). For those densi-
ties that we presume are gamma distributed, we solve a linear regression problem.
Specifically, we look for the constants c1, c2, and c3 of
f (T ) = ec1T c2 e−c3T
(7)
an alternate form of (42). Upon taking the logarithm of (7), we have the linear sum
ln f (T ) = c1 + c2 ln T − c3T.
(8)
Then, we select three values of the numerically generated distribution pn(T n) along
with its associated dominance times: (T n
j ).
We always choose T n
2 /2. It is
then straightforward to solve the linear system
2 , pn
2 = arg maxT pn(T ) as well as T n
3 , pn
2 /2 and T n
2 ); (T n
1 = T n
3 ) where pn
1 , pn
1 ); (T n
j = pn(T n
3 = 3T n
1
1
ln pn
2
ln pn
3
for the associated constants using the \ command in MATLAB.
c2
c3
2
ln T n
1 ln T n
1 ln T n
1
1 −T n
2 −T n
3 −T n
3
c1
=
ln pn
3. Results. We now discuss several results that reveal the importance of synap-
tic depression in transferring information about stimuli to competitive networks. This
is initially shown by analyzing the ring model with depression (1). However, to carry
out detailed analysis on stochastic switching a competitive network with depression,
we must reduce (1) as well as analyzing an analogous model without space (6).
3.1. Deterministic switching in the ring model. To start we will consider
the ring model with depression (1) in the absence of noise. We then have the deter-
ministic system [54, 26, 27]
(cid:90) π/2
ut(x, t) = −u(x, t) +
τ qt(x, t) = 1 − q(x, t) − βq(x, t)f (u(x, t)).
−π/2
w(x − y)q(y, t)f (u(y, t))dy + I(x),
(9a)
(9b)
In previous work, versions of (9) have been analyzed to explore how synaptic de-
pression can generate traveling pulses [54, 26], self-sustained oscillations [26], and
spiral waves in two-dimensions [27]. Here, we will extend previous work that explored
4
Fig. 1. Three possible active states of the noise-free stimulus driven ring model with depression
(9). (A) Winner take all (I0 = 0.6); (B) Rivalrous oscillations (I0 = 0.84); (C) Fusion (I0 = 1).
Other parameters are θ = 0.5, β = 1, and τ = 50.
input-driven oscillations in two-layer networks like (9) that possessed many statistics
matching binocular rivalry [25]. We will think of (9) as a model of monocular ri-
valry, since oscillations can be due to competition between representations in a single
orientation column [3]. Competition between ocular dominance columns [25] is not
necessary for our theory. For the purpose of exposition, we will employ specific forms
for the functions of (9): cosine weight (5); a Heaviside firing rate function (3); and a
bimodal input (4).
Winner take all state. We start by looking for winner-take-all solutions to the
deterministic system (9), such as that shown in Fig. 1(A). These states consist of a
single activity bump arising in the network, representing only one of the two percepts
contained in the bimodal stimulus (4). These are stationary in time, so ut = qt = 0,
implying that u = U (x) and q = Q(x). Also, they are single bump solutions, so there
will be a single region in space that is superthreshold (U (x) > κ). We assume the right
stimulus is represented by a bump, although we can derive analogous results when
the left stimulus is represented. We then have a steady state solution determined by
(cid:90) π/4+a
U (x) =
Q(x) = [1 + βH(U (x) − κ)]
π/4−a
−1 ,
cos(2(x − y))Q(y)dy − I0 cos(4x) + Ia sin(2x),
(10)
(11)
since U (x) > κ when x ∈ (π/4 − a, π/4 + a), so that by plugging (11) into (10) and
using the trigonometric identity cos(2(x − y)) = cos(2x) cos(2y) + sin(2x) sin(2y) we
have
U (x) = A cos(2x) + B sin(2x) − I0 cos(4x) + Ia sin(2x),
(12)
where the multiplicative constants A, B can be computed
(cid:90) π/4+a
(cid:90) π/4+a
1
A =
1 + β
π/4−a
cos(2x)dx = 0,
B =
1
1 + β
π/4−a
sin(2x)dx =
sin(2a)
1 + β
.
Therefore, by simplifying the threshold condition, U (π/4 ± a) = κ, we have
U (π/4 ± a) =
sin(4a)
2(1 + β)
+ I0 cos(4a) + Ia cos(2a) = κ.
(13)
5
The implicit equation (13) can be solved numerically using root finding algorithms.
For symmetric inputs (Ia = 0), we can solve (13) explicitly
(cid:34)
1 ±(cid:112)1 + 4(1 + β)2(I 2
2(1 + β)(I0 + κ)
(cid:35)
0 − κ2)
a =
tan−1
1
2
.
(14)
Thus, we can fully characterize winner-take-all solutions
U (x) =
sin(2a)
1 + β
sin(2x) − I0 cos(4x) + Ia sin(2x).
(15)
The advantage of having this solution is that we can relate the parameters of the
model to the existence of the winner-take-all state, where we would expect to only see
single bump solutions. To do so, we need to look at a second condition that must be
satisfied, U (x) < κ for all x /∈ (π/4 − a, π/4 + a). Since the function (15) is bimodal
across (−π/2, π/2), we check the other possible local maximum at x = −π/4 as
U (π/4) = I0 − Ia − sin(2a)
1 + β
< κ.
(16)
At the point in parameter space where the (16) is violated, a bifurcation occurs, so
the winner-take-all state ceases to exist. This surface in parameter space is given by
the equation
I0 = κ + Ia +
sin(2a)
1 + β
,
(17)
along with the explicit formula for the bump half-width (14). Beyond the bifurcation
boundary (17), one of two behaviors can occur. Either there is a symmetric two-bump
solution that exists, the fusion state [52, 4, 43], or rivalrous oscillations [32, 5].
Fusion state. It has been observed in many experiments on ambiguous stimuli
that sufficiently strong contrast rivalrous stimuli can be perceived as a single fused im-
age [4, 8]. This should not be surprising, considering stereoscopic vision and audition
behave in exactly this way [52]. However, the contrast necessary to evoke this state
with dissimilar images is much higher than with similar images [5]. In the network
(9), the fusion state (Fig. 1(C)) is represented as two disjoint bumps. Therefore
(cid:34)(cid:90) −π/4+b
(cid:90) π/4+a
(cid:35)
U (x) =
1
1 + β
−π/4−b
π/4−a
+
cos(2(x − y))dy − I0 cos(4x) + Ia sin(2x).
Computing the integral terms, we find
U (x) =
S(x, a) − S(x, b)
1 + β
− I0 cos(4x) + Ia sin(2x),
(18)
where S(x, y) = sin2(x + y) − sin2(x − y). The solution can be specified by requiring
the threshold conditions U (−π/4 ± b) = U (π/4 ± a) = κ are satisfied
cos(2a)[sin(2a) − sin(2b)]
1 + β
cos(2b)[sin(2b) − sin(2a)]
1 + β
+ I0 cos(4a) + Ia cos(2a) = κ,
+ I0 cos(4b) − Ia cos(2b) = κ,
(19)
(20)
6
which we can solve numerically to relate the asymmetry of inputs Ia to the half-
widths a, b of each bumps. In the the case of symmetric inputs, U (x) = −I0 cos(4x),
and it is straightforward to find the two bump widths explicitly. We will now study
rivalrous oscillations by simply constructing them using a fast-slow analysis. We can
also numerically identify the boundary of various behaviors of (1), as shown in Fig.
3.
Rivalrous oscillations. Oscillations can occur, where the two bump locations
trade dominance successively (Fig. 1(B)). As in Levelt's proposition (i), increasing
the contrast of a stimulus leads to that stimulus being in dominance longer. This infor-
mation is not revealed when the system is stuck in a winner-take-all state. Therefore,
the introduction of synaptic depression into the system (9) leads to an increase in
information transfer. We will also examine how well (9) recapitulates Levelt's other
propositions concerning the mean dominance of percepts.
To analyze (9) for oscillations, we assume that the timescale of synaptic depression
τ (cid:29) 1, long enough that we can decompose (9) into a fast and slow system [29, 25].
Synaptic input u then tracks the slowly varying state of the synaptic scaling term q.
We also assume that q is essentially piecewise constant in space, in the case of the
Heaviside nonlinearity (3), which yields
u(x, t) ≈
cos(2(x − y))q(y, t)H(u(y, t) − κ)dy − I0 cos(4x),
(21)
(cid:90) π/2
−π/2
and q is governed by (9b). To start, we will also assume a symmetric bimodal input.
This way, we can simply track q in the interior of one of the bumps, given qi(t) =
q(π/4, t). Assuming a switch has just occurred, where the left bump has escaped
suppression, to pin down the right, so
τ qi(t) = 1 − qi(t),
t ∈ (0, T ),
qi(0) = q0,
(22)
where T is the amount of time each percept is in dominance, and q0 is the synaptic
strength within a bump region immediately prior to its shutting off. After the right
bump escapes dominance of the left bump
τ qi(t) = 1 − qi(t) − βqi(t),
t ∈ (T, 2T ).
(23)
Solving (22) and (23) simultaneously, we have
which can be solved explicitly for the dominance time
q0 =
1
+
β
e−T /τ − (1 − q0)e−2T /τ ,
1 + β
1 + β
(cid:34)
β +(cid:112)β2 − 4(1 + β)(1 − q0)[(1 + β)q0 − 1]
T = τ ln
2(1 + β)q0 − 2
(cid:35)
,
(24)
so that we now must specify the value q0. We can examine the fast equation (21),
solving for the form of the slowly narrowing right bump during its dominance phase
u(x, t) = qi(t)
= qi(t)(cid:2)sin2(x + a(t)) − sin2(x − a(t))(cid:3) − I0 cos(4x).
cos(2(x − y))dy − I0 cos(4x)
π/4−a(t)
(25)
(cid:90) π/4+a(t)
7
We can solve for the slowly evolving width a(t) of this bump by requiring the threshold
condition u(π/4 ± a(t), t) = κ to yield
qi(t)
2
sin(4a(t)) + I0 cos(4a(t)) = κ,
and then solving using trigonometric identities
(cid:34)
qi(t) +(cid:112)qi(t)2 + 4(I 2
0 − κ2)
(cid:35)
.
(26)
2(I0 + κ)
a(t) =
tan−1
1
2
We can also identify the maximal value of qi(t) = q0 which still leads to the right bump
suppressing the left. Once qi(t) falls below q0, the other bump escapes suppression,
flipping the dominance of the current bump. This is the point at which the other
bump of (25) rises above threshold, as defined by the equation
u(−π/4, t) = I0 − q0 sin(2a0) = κ.
Combining this with the equation (26), we have an algebraic equation for q0 given
(I0 − κ)2 = q2
0
2q2
8I 2
0 + 4(I 2
0 + 8I0κ + 2q2
0 + 4(I 2
0 + 4(I 2
0 − κ2)
0 − κ2)
0 − κ2) + 2q0
0 + 2q0
(cid:112)q2
(cid:112)q2
(cid:112)(I0 − κ)(3I0 + κ)
3I0 + κ
,
(27)
which is straightforward to solve for
2I0
q0 =
and we have excluded an extraneous negative solution. Interestingly, the amplitude of
synaptic depression is excluded from (27), but we do know based on (22) and (23) that
q0 ∈ ([1 + β]−1, 1). This establishes a bounded region of parameter space in which we
can expect to find rivalrous oscillations, which we use to construct a partitioning of
parameter space in Fig. 3. We can also now approximate the dominance time using
(24) with (27), as shown in Fig. 2(D).
In the case of an asymmetric bimodal input (Ia > 0), we can also solve for explicit
approximations to the dominance times of the right TR and left TL populations.
Following the same formalism as for the symmetric input case
(cid:34)
(cid:34)
β + qd +(cid:112)(β + qd)2 − 4(1 + β)(1 − qL)[(1 + β)qR − 1]
β − qd +(cid:112)(β − qd)2 − 4(1 + β)(1 − qR)[(1 + β)qL − 1]
2(1 + β)qR − 2
(cid:35)
(cid:35)
,
,
(28)
(29)
TR = τ ln
TL = τ ln
2(1 + β)qL − 2
where qd = (1 + β)(qR − qL), in terms of the local values qL and qR of the synaptic
scaling in the right and left bump immediately prior to their suppression. Notice in
the case qL = qR, then qd = 0 and (28) and (29) both reduce to (24). We now need to
examine the fast equation (21) to identify these two values. This is done by generating
two implicit equations for the half-width aR and qR at the time of a switch
qR
2
sin(4aR) + I0 cos(4aR) + Ia cos(2aR) = κ,
I0 − Ia − qR sin(2aR) = κ,
(30)
(31)
8
Fig. 2. Dependence of rivalry on the amplitudes of the bimodal input (4). (A) Dominance
times are both T ≈ 1s when input is symmetric (I0 = 0.84, Ia = 0). (B) Dominance time of right
input (TR ≈ 0.9s) is longer than left (TL ≈ 0.6s) for asymmetric input (IR = 0.9, IL = 0.84).
Notice TR is unchanged from case (A). (C) Dominance times are both shorter of higher contrast
symmetric stimulus (I0 = 0.9, Ia = 0). (D) Increasing the strength of the symmetric (Ia = 0)
bimodal input (4) decreases the dominance time T of both populations. Our theory (black) computed
from fast-slow analysis (24) fits results of numerical simulations (blue) well. (E) For asymmetric
inputs (Ia (cid:54)= 0), we find that varying IR = I0 + Ia while keeping IL = I0 − Ia fixed changes the
dominance times of the left percept TL (blue) much more than that of the right percept TR (red).
Other parameters are κ = 0.5, β = 1, and τ = 50.
which we can solve explicitly for
aR =
cos−1
1
2
(cid:20) κ
2I0
(cid:21)
,
+
1
2
and
(cid:112)(3I0 + κ)(I0 − κ)
2I0(IL − κ)
,
qR =
(32)
(33)
where IL = I0 − Ia is the strength of input to the left side of the network. Likewise,
we can find the value of the synaptic scaling in the left bump immediately prior to
its suppression
,
(34)
(cid:112)(3I0 + κ)(I0 − κ)
2I0(IR − κ)
qL =
where IR = I0 + Ia is the strength of input to the right side of the network. Using
the expressions (33) and (34) we can now compute the dominance time formulae (28)
and (29), showing the relationship between inputs and dominance times in Fig. 2
(E). Notice that all of Levelt's propositions are essentially satisfied. Changing the
strength of the right stimulus IR has a very weak effect on the dominance time of the
9
Fig. 3. Partitions of parameter space into various stimulus induced states of (9). (A) Plotted
as a function of network threshold κ and strength I0 of the bimodal input (4) when β = 1. (B)
Plotted as a function of synaptic depression strength β and strength I0 of bimodal input (4) when
κ = 0.5. Parameter τ = 50.
right percept. However, it does increase the overall alternation rate and decrease the
proportion of time the left percept remains in dominance.
We also note there is a critical strength of synaptic depression β below which
rivalrous oscillations do not occur. When synaptic depression is sufficiently strong,
the winner-take-all state ceases to exist. Beyond this critical synaptic depression
strength, the network either supports rivalrous oscillations or a fusion state. Either
way, information is conveyed to the network that would otherwise be kept hidden. We
show this in Fig. 3(B). In this way, synaptic depression can improve the information
transfer of the network (9). In fact, we will show that it does so in a way that is much
more reliable than noise.
3.2. Purely stochastic switching in the ring model. We will now study
rivalrous switching brought about by fluctuations. In particular, we ignore depression
and examine the noisy system
(cid:35)
(cid:34)
(cid:90) π/2
−π/2
du(x, t) =
−u(x, t) +
w(x − y)f (u(y, t))dy + I(x)
dt + dξ(x, t).
(35)
where (cid:104)ξ(x, t)(cid:105) = 0 and (cid:104)ξ(x, t)ξ(y, s)(cid:105) = εC(x− y)δ(t− s) defines the spatiotemporal
correlations of the system.
To start, we can simply consider (35) in the absence of noise
ut(x, t) = −u(x, t) +
w(x − y)f (u(y, t))dy + I(x).
(36)
(cid:90) π/2
−π/2
This model has been studied extensively [2, 3], so we will not perform an in depth
analysis of stationary bump solutions. We are interested in the winner-take-all state.
For a cosine weight (5), Heaviside firing rate (3), and bimodal stimulus (4) these can
be computed as a special case of (15) and lie at either x = ±π/4, so
U±(x) = sin(2a±) sin(2x) − I0 cos(4x) ± Ia sin(2x),
(37)
10
Fig. 4. Purely noise induced switching of dominance in the depression-free ring model (35)
(A) Numerical simulations of the system for the various input strengths of the symmetric (Ia = 0)
bimodal input (4) with (A) I0 = 0.8, (B) I0 = 0.9, and (C) I0 = 1.0. Distributions of dominance
times computed numerically (blue bars) with the exponential distribution (40) with numerically com-
puted mean (cid:104)T(cid:105) (red) superimposed for (D) I0 = 0.8 has (cid:104)T(cid:105) ≈ 1.2s, (e) I0 = 0.9 has (cid:104)T(cid:105) ≈ 0.70s,
and (f ) I0 = 1.0 has (cid:104)T(cid:105) ≈ 0.45s. Other parameters are κ = 0.5 and ε = 0.04.
and we can apply the threshold condition U±(±π/4 + a±) = κ, so
sin(4a±) + I0 cos(4a±) ± Ia cos(2a±) = κ.
1
2
In the case of a symmetric input, a± = a and we can solve (38) explicitly
(cid:34)
1 +(cid:112)1 + 4(I 2
2(I0 + κ)
(cid:35)
.
0 − κ2)
a =
tan−1
1
2
(38)
(39)
Since we have no synaptic depression in the model (36), we cannot rely on deter-
ministic mechanisms to generate switches between one winner-take-all state and an-
other. Therefore, we will consider the effects of introducing a small amount of noise
(0 < ε (cid:28) 1), reflective of synaptic fluctuations. We focus on the spatial correlation
function C(x) = cos(x). Noise can generate switches in between the two dominant
states (Fig. 4). As the strength of both input contrasts are increased, switches be-
tween percepts occur more often. In fact, there is an exponential dependence of the
mean dominance time (cid:104)T(cid:105) on the strength I0 of the bimodal input (4). We will pro-
vide an argument, using energy methods, as to why this occurs in our analysis of the
simplified system (6).
We will study switching in the case that the suppressed bump comes on and
shuts off the bump that is currently on. This mechanism is known as escape [48]. As
shown in Fig. 4(A-C), dominance times decrease with input strength on average.
Here escape is a noise induced effect [37], rather than a deterministic depression-
induced effect [43, 25]. However, as opposed to depression-induced switching, there is
a substantial spread in the possible dominance times for a given set of parameters (Fig.
4(D-F)). Thus, by examining two dominance times back to back, an observer would
11
Fig. 5. Sampling of the strength of each input based on noise-induced transitions. (A) Mean
dominance time (cid:104)T(cid:105) as a function of the strength I0 of the symmetric (Ia = 0) bimodal input (4).
(B) Probability that the right input IR was higher than the left output IL, based on the sampling
n cycles (2n switches between percepts), in the case of symmetric inputs IL = IR = 0.9. Notice
it takes close to 2000 cycles before p[IR > ILT ∗(n)] ≈ 0.5. Other parameters are κ = 0.5 and
ε = 0.04.
have difficulty telling if the input strengths were roughly the same or not. Recently,
psychophysical experiments have been carried out where an observer must identify
the higher contrast of two rivalrous percepts [36], showing humans perform quite well
in at this task. Results of [36] suggest humans most likely use Bayesian inference
in discerning information about visual percepts. This is in keeping with previous
observations that humans' visual perception of objects likely carries out Bayesian
inference [24].
Notice in Fig. 5(B) that the likelihood an observer assigns to IR > IL approaches
1/2 as the number of observations n increases. We compute p[IR > ILT ∗(n)], the
probability an observer would presume IR > IR conditioned on dominance time pairs
from n cycles T ∗(n) =
, numerically here. How-
ever, as the number of cycles n → ∞, the exponential distributions approximately
defining the identical probability densities pR(TR) = pL(TL) = p(T ) will be fully
sampled. We can calculate
L , ..., T (n)
T (1)
R , T (1)
L , T (2)
R , T (2)
R , T (n)
L
(cid:110)
(cid:111)
p(IR > ILT ∗(∞)) =
(cid:90) ∞
(cid:90) x
0
0
p(x)p(y)dydx =
1
(cid:104)T(cid:105)2
e−(x+y)/(cid:104)T(cid:105)dydx =
1
2
,
(cid:90) ∞
(cid:90) x
0
0
as in Fig. 5(B). In the case where depression, in the absence of noise, drives switches,
we would expect this limit to be approached much more quickly. We will also examine
how a combination of depression and noise affects an observer's ability to discern
contrast differences.
We explore this further in the case of asymmetric inputs, showing that dominance
times are still specified by roughly exponential distributions as shown in Fig. 6. When
IR > IL, even though the means satisfy (cid:104)TR(cid:105) > (cid:104)TL(cid:105) (Fig. 6(E)), the exponential
distributions p(TR) and p(TL) have considerable variance, also given by the means.
Therefore, randomly sampled values from these distributions may satisfy TR < TL.
Were an observer to use one such sample as a means for guessing the inputs that
generated them, they would guess IR < IL, rather than the correct IR > IL. In terms
of conditional probabilities, we can expect situations where p(IR > ILT ∗(n)) < 1/2
12
Fig. 6. (A) Single realization of the stochastic neural field (35) with asymmetric (Ia > 0)
(B)
inputs IR = 0.92 and IL = 0.88, leads to longer dominance times for right percept TR.
Expected likelihood p[IR > IRT ∗(n)] that the right input IR is stronger than left IL based on n
comparisons of dominance times TR and TL sampled. Blue line is theoretical prediction (41) of the
limit as n → ∞. Numerically computed dominance time distributions (blue bars) are well fit by the
exponential distribution (40) for the (C) left ((cid:104)TL(cid:105) ≈ 0.5s) and (D) right ((cid:104)TR(cid:105) ≈ 1s) percepts.
(E)Dependence of mean dominance times (cid:104)TR(cid:105) and (cid:104)TL(cid:105) on the strength of the right input IR.
Black curves are best fits to exponential functions of IR. (F) Expected likelihood p[IR > ILT ∗(∞)]
right input IR is stronger than left IL in the limit of high sample number n → ∞, as computed
theoretically by (41). Other parameters are κ = 0.5, and ε = 0.04.
for finite n, even though IR > IL. We can quantify this effect numerically, as shown
in Fig. 6(F). In the limit n → ∞, we find uncertainty continues to creep in, since
fluctuations continually give an observer misleading information. Since the marginal
distributions are approximately exponential
pj(Tj) = e−Tj /(cid:104)Tj(cid:105)/(cid:104)Tj(cid:105)
j = L, R,
(40)
13
Fig. 7. Switching in the stochastic ring model with depression (1) with asymmetric inputs
(Ia > 0). (A) Single realization for asymmetric inputs with IR = 0.92 and IL = 0.88, which
leads to right percept dominating longer. (B) Distribution of left percept dominances times pL(TL)
over 1000s is well fit by a gamma distribution (42). (C) Distribution of right percept dominance
times pR(TR) across 1000s is well fit by a gamma distribution (42). Other parameters are κ = 0.5,
β = 0.2, τ = 50, and ε = 0.01.
we can approximate the conditional probability
p[IR > ILT ∗(∞)] =
=
(cid:90) ∞
(cid:90) x
0
0
1
(cid:104)TR(cid:105)(cid:105)TL(cid:105)
pR(x)pL(y)dydx
(cid:90) ∞
(cid:90) x
e−x/(cid:105)TR(cid:105)e−y/(cid:105)TL(cid:105)dydx
= 1 −
0
0
(cid:104)TL(cid:105)
(cid:104)TR(cid:105) + (cid:104)TL(cid:105) =
(cid:104)TR(cid:105)
(cid:104)TR(cid:105) + (cid:104)TL(cid:105) .
(41)
Observe that the approximation we make using the formula (41) accurately estimates
the limit p(IR > ILT ∗(∞)) as shown in Fig. 6(B). This is the likelihood that an
observer performing Bayesian sampling of the probability densities (40) will predict
IR > IL. Recent psychophysical experiments suggests humans would perform this
task of contrast differentiation in this way [24, 36].
We see from our analysis that when switches are generated by noise, rather than
deterministic depression, the means dominance times still obey Levelt's propositions
to some extent (Fig. 6(E)). This would allow for an accurate comparison of input
strengths IR and IL based on the means (cid:104)TR(cid:105) and (cid:104)TL(cid:105). However, when consider-
ing a more realistic observer, that could only compare successive dominance times,
accurately discerning the comparison of the input contrasts is more difficult. This be-
comes much more noticeable when the input contrasts are quite close to one another,
as in Fig. 6(F). We will explore now how introducing depression along with noise
improves discernment of the input contrasts by an observer using simple comparison
of dominance times.
3.3. Switching through combined depression and noise. We now study
the effects of combining noise and depression in the full ring model of perceptual
rivalry (1). Numerical simulations of (1) reveal that noise-induced switches occur
robustly, even in parameter regimes where the noise-free system supports no rivalrous
oscillations, as shown in Fig. 7. Rather than dominance times being distributed
exponentially, they roughly follow a gamma distribution [17, 30]
pj(Tj) =
1
σkΓ(k)
j exp [−Tj/σ] ,
T k
k > 1.
(42)
14
Fig. 8. Comparing the probability densities of dominance times in the stochastic ring model
with depression (1) for various levels of noise and depression. (A) No depression and ε = 0.04 (B)
Depression strength β = 0.2 and ε = 0.01. (C) Depression strength β = 0.4 and ε = 0.0025. (D)
Expected likelihood p[IR > IRT ∗(∞)] the right input IR is stronger than the left IL based in the
limit of an infinite number of samples of the dominance times TR and TR for the parameters in (A)
(pink); (B) (magenta); and (C) (red). Other parameters are τ = 50 and κ = 0.5.
As opposed to the exponential distribution, (42) is peaked away from zero at Tj = kσ,
which is also the mean of the distribution. Therefore, two distributions of dominance
times with different means will be more easily discerned from one another. We show
this in Fig. 8(B) by superimposing the two distributions from Fig. 7 on top of
one another. They clearly separate better than in the probability densities of purely
noise-driven switching, shown in Fig. 8(A). As the strength of synaptic depression is
increased even further, keeping the mean dominance times (cid:104)TR(cid:105) and (cid:104)TL(cid:105) the same,
probability densities separate even further (Fig. 8(C)). We summarize how this sep-
aration improves the inference of input contrast difference in Fig. 8(D). As the
strength β of depression is increased and noise is decreased, an observer's ability to
discern which input was stronger is improved. The likelihood assigned to IR being
greater than IL is a sigmoidal function of IR whose steepness increases with β. For no
noise, the likelihood function is simply a step function H(IR > IL), implying perfect
discernment.
3.4. Analyzing switching in a reduced model. We now perform similar
analysis on a reduce competitive network model (6) and extend some of the results for
the ring model. One of the advantages is that we can construct an energy function [20],
15
which provides us with intuition as to the exponential dependence of mean dominance
times on input strengths in the noise-driven case. In particular, we will analyze (6)
where the firing rate function is Heaviside (3), starting with the case of no noise
uR = −uR + H(IR − qLuL),
τ qR = 1 − qR − βuRqR,
uL = −uL + H(IL − qRuR)
τ qL = 1 − qL − βuLqL.
(43a)
(43b)
First, we note (43) has a stable winner-take-all solution in the jth population (j =
R, L) for Ij > 0 and Ik < 1/(1 + β) (k (cid:54)= j). Second, a stable fusion state exists when
both IL, IR > 1/(1 + β). Coexistent with the fusion state, there may be rivalrous
oscillations, as we found in the spatially extended system (9). To study these, we
make a similar fast-slow decomposition of the model (43), assuming τ (cid:29) 1 to find
uj's possess the quasi-steady state
uR = H(IR − qLuL),
uL = H(IL − qRuR).
(44)
so we expect uj = 0 or 1 almost everywhere. Therefore, we can estimate the domi-
nance time of each stimulus using a piecewise equation for the slow subsystem
(cid:26) 1 − qj − βqj
1 − qj
τ qj =
: uj = 1,
: uj = 0,
j = R, L.
(45)
Combining the slow subsystem (45) with the quasi-steady state, we can use self-
consistency to solve for the dominance times TR and TL of the right and left popula-
tions. We simply note that switches will occur through escape mechanism, when the
cross-inhibition between populations becomes weak enough such that the suppressed
population's (j) input becomes superthreshold, so Ij = qk. Using (45) as we did in
the spatial system, we find
(cid:34)
(cid:34)
β − Id +(cid:112)(β − Id)2 − 4(1 − IR)(1 + β)[(1 + β)IL − 1]
β + Id +(cid:112)(β + Id)2 − 4(1 − IL)(1 + β)[(1 + β)IR − 1]
2(1 + β)IL − 2
(cid:35)
(cid:35)
2(1 + β)IR − 2
TR = τ ln
TL = τ ln
,
,
(46)
(47)
where Id = (1 + β)[IR − IL]. For symmetric stimuli, IL = IR = I, both (46) and (47)
reduce to
(cid:34)
β +(cid:112)β2 − 4(1 − I)(1 + β)[(1 + β)I − 1]
(cid:35)
,
(48)
T = τ ln
2(1 + β)I − 2
using which we can solve for the critical input strength I above which only the fusion
state exists
I =
2 + β
2(1 + β)
,
(49)
in the case of symmetric inputs. We show in Fig. 9 that this asymptotic approxima-
tions (46) and (47) of the dominance times match well with the results of numerical
simulations. Levelt's propositions are recapitulated well.
16
Fig. 9. Dominance times computed adiabatically in the noise-free competitive network with
depression (43). (A) Plot of dominance times T as a function of the strength of a symmetric
input IR = IL = I to the competitive system (43) show fast-slow theory (curve) match numerical
simulations (dots) very well. (B) Dominance times TL and TR as a function of right input IR
keeping IL = 0.6 fixed as computed by theory (curves) in (46) and (47) fits numerically computed
(dots) very well. Other parameters are β = 1 and τ = 50.
Fig. 10. Noise induced transitions in depression-free two population network (50). (A) Single
realization with IR = IL = I = 0.9 for the right uR (red) and left uL (black) population activities.
(B) Mean dominance time (cid:104)T(cid:105) as a function of input strength I computed numerically (red dots)
and fit to the theoretically derived exponential function (51). (C) Mean dominance times (cid:104)TR(cid:105) and
(cid:104)TL(cid:105) as a function of the right input strength IR while IL = 0.95 is fixed. Other parameters are
ε = 0.01.
Now, we study noise-induced switching in the competitive network. We can sep-
arate timescales to study the effects of depression and noise together. To start, we
consider the limit of no depression β → 0, so that
uR = −uR + H(IR − uL) + ξR,
uL = −uL + H(IL − uR) + ξL,
(50a)
(50b)
where ξj are independent white noise processes with variance ε. We show a single
realization of the competitive network in Fig. 10(A). Most of the time, the dynamics
remains close to one of the winner-take-all attractors where uj = 1 and uk = 0
(j = R, L and k (cid:54)= j). Occasionally, noise causes large deviations where the suppressed
population's activity rises above threshold, causing the once dominant population to
then be suppressed. We plot the relationship between the strength of the inputs
and the mean dominance times in Fig. 10. Notice, in the case of symmetric inputs
17
IL = IR = I, we can fit this relationship to the exponential
(cid:104)T(cid:105) ≈ A exp[B(1 − I)].
(51)
To understand why this is so, we can study the energy function associated with the
system (50). Notice, this network is essentially the classic two neuron flip-flop Hopfield
network. As has been shown before [20], for symmetric inputs the energy function for
this network can be defined
E[uR, uL] = H(I − uR)H(I − uL) − I [H(I − uR) + H(I − uL)] ,
(52)
so we can compute the energy difference between the winner-take-all and fusion states
E[1, 0] = −I,
E[1, 1] = 1 − 2I,
(53)
respectively. By taking the difference between these two quantities, we find ∆E =
1− I, which well approximates the exponential dependence of the dominance times T ,
as shown by the fit in Fig. 10(B). This provides the intuition as to this relationship.
In the same way, we can write down the energy function in the case where the
inputs are non-symmetric IR (cid:54)= IL, which is [12]
E[uR, uL] = H(IL − uR)H(IR − uL) − ILH(IL − uR) − IRH(IR − uL),
(54)
which means the energy depth of the right and left winner-take-all states are ∆ER =
1 − IL and ∆EL = 1 − IR, respectively. Thus, as we observe in Fig. 10(C), we
expect the dominance time of each population to depend upon the strength of the
other stimulus according to
(cid:104)TR(cid:105) ≈ AR exp [BR(1 − IL)] ,
(cid:104)TL(cid:105) ≈ AL exp [BL(1 − IR)] .
(55)
Interestingly, this simple model agrees well with the qualitative predictions of Levelt
propositions (i-iv) in this high contrast input regime. Now, we will see how including
synaptic depression in the model generates distributions of dominance times that are
more similar to those observed experimentally [17, 30, 6].
Finally, we show that the network with depression and noise generates gamma
distributed dominance times, as the spatially extended system does. In addition, we
provide some analytic intuition as to how gamma distributed dominance times may
arise in the fast slow system. First, we display as single realization of the network
(6) in Fig. 11(A) along with a plot of an adiabatically computed energy function
E[uR, uL, qR, qL] for the system. To compute the energy function, we first note that
in the limit of slow depression recovery time τ (cid:29) 1, we can assume the energy of the
system will be defined simply by (52) augmented by the synaptic scalings imposed by
qR and qL [34]. In the fully general case, where inputs IR and IL may be asymmetric
we have
E[uR, uL, qR, qL] =H(IL − qRuR)H(IR − qLuL)
− IL
qR
H(IL − qRuR) − IR
qL
H(IR − qLuL).
(56)
A similar energy function was previously used in a model with spike frequency adap-
tation [37]. Here, we are able to derive the energy function from the model (6).
Therefore, the energy gap between a winner-take-all state and the fusion state will
18
Fig. 11. (A) Single realization of the network (6) with depression and noise. Activity variables
uR (black) and uL (blue) stay close to attractors at 0 and 1, aside from depression or noise induced
switching. Depression variables qR (red) and qL (green) slowly exponentially change in response to
the states of uR and uL. (B) Probability density p(T ) of dominance times T sampled over 1000s,
well fit by a gamma distribution (42). Parameters are ε = 0.036, β = 0.2, τ = 50, and I = 0.8.
Fig. 12. Distribution of dominance times for (A) right and (B) left populations fit with red
and blue gamma distributions (42) respectively, in the network (6) with depression and noise in the
case of asymmetric inputs IR = 0.82 and IL = 0.78, sampled over 1000s. The right population has
longer dominance times. Other parameters are β = 0.2, τ = 50, and ε = 0.036.
be time-dependent, varying as the synaptic scaling variables qR and qL change. The
energy difference between the right dominant state and fusion is
∆ER(t) = 1 − IL
qR(t)
,
∆EL(t) = 1 − IR
qL(t)
,
(57)
for the right and left population respectively.
Notice that dominance times of stochastic switching (Fig. 11) in (6) are dis-
tributed roughly according to a gamma distribution (42). Superimposing the prob-
ability density of right (left) dominance times on the left (right) probability density,
we see they are reasonably separated. Using the analysis we performed for the spa-
tially extended system, we could also show that depression improves discernment of
the input contrast difference. Mainly here, we wanted to provide a justification as
to the relationship between input strength and mean dominance times. Using en-
19
Fig. 13. Perceptual tristability. Examples of images with three possible interpretations.
(A) Three overlapping gratings. Redrawn with permission from [38]. (B) 'Mother, father, and
daughter.' Redrawn with permission from [16]. Staring at tristable images for long enough leads
to the perception switching between the three possible interpretations. (C) Numerical simulation of
showing the activity variables u1, u2, u3 and the second synaptic scaling variable q2 (cyan) of the
three population network (58) driven by symmetric stimulus I = 0.7. (D) Relationship between the
strength of the stimulus I and the dominance times T computed using fast-slow analysis (black) and
numerics (red dots). Other parameters are β = 1 and τ = 50.
ergy arguments, we have provided reasoning behind why Levelt's propositions are
still preserved in this model, when noise is included, even when switches are noise-
induced. Increasing one input leads to a reduction in the energy barrier between the
other population's winner-take-all state and the fusion state. This leads to the other
population's dwell time being shorter.
3.5. Switching between three percepts. Finally, we will compare the trans-
fer of information in competitive networks that process more than two inputs. Re-
cently, experiments have revealed that perceptual multistability can switch between
three or four different percepts [16, 9, 38, 22]. In particular, the work of [38] charac-
terized some of the switching statistics during the oscillations of perceptual tristabil-
ity. Fig. 13(A,B) shows examples of tristable percepts. Since dominance times are
gamma distributed and there is memory evident in the ordering of percepts, the pro-
cess is also likely governed by some slow adaptive process in addition to fluctuations.
We will pursue the study of perceptual tristability in a competitive neural network
20
model with depression and noise. In the case of three different percepts, a Heaviside
firing rate (3), and symmetric inputs I1 = I2 = I3 = I, we study the system
u1 = −u1 + H(I − q2u2 − q3u3),
u2 = −u2 + H(I − q1u1 − q3u3),
u3 = −u3 + H(I − q1u1 − q2u2),
τ qj = 1 − qj − βujqj,
j = 1, 2, 3.
(58a)
(58b)
(58c)
(58d)
We are interested in rivalrous oscillations, which do arise in this network for certain
parameter regimes (Fig. 13(C)). As per our previous analysis, we perform a fast-slow
decomposition of our system. In the case of symmetric inputs, we use our techniques
to compute the dominance time T of a population as it depends on input strength
I. Our analysis follows along similar lines to that carried out for the two population
network, where we assume τ (cid:29) 1. We find
(1 − I)(1 + β) +(cid:112)(1 + β)(1 − I)[3I(1 + β) + β − 3]
(cid:35)
(cid:34)
T = τ ln
2[(1 + β)I − 1]
,
(59)
which compares very well with numerically computed dominance times in Fig. 13(D).
While perceptual tristability has not been explored very much experimentally [16,
38, 22], observations that have been made suggest that relationships between mean
dominance time and input contrast may be similar to the two percept case [22]. In
our model, we see that as the input strength is increased, dominance times decrease.
One other important point is that percept dominance occurs in the same order every
time (Fig. 13(C)): one, two, three. There are no "switchbacks." We will show that
this can occur in the noisy regime, which degrades information transfer.
Now, we seek to understand how noise alters the switching behavior when added
to the deterministic network (58). Thus, we discuss the three population competitive
network with noisy depression
u1 = −u1 + H(I − q2u2 − q3u3),
u2 = −u2 + H(I − q1u1 − q3u3),
u3 = −u3 + H(I − q1u1 − q2u2),
τ qj = 1 − qj − βujqj + ξj,
j = 1, 2, 3,
(60a)
(60b)
(60c)
(60d)
where ξj are identical independent white noise processes with variance ε. In Fig. 14,
we show the noise in (60) degrades two pieces of information carried by dominance
switches: the switching time and the direction of switching. Notice that as the ampli-
tude of noise ε is increased, the dominance times become more spread out. Thus, there
is a less precise characterization of the input strength in the network. Concerning the
direction of switching, we see that the introduction of noise makes "switch backs"
more likely. We define a "switch back" as a series of three percepts that contains the
same percept twice (e.g. 1 → 3 → 1). This as opposed to a "switch forward," which
contains all three percepts (e.g. 1 → 3 → 2). Statistics like these were analyzed
from psychophysical experiments of perceptual tristability, using an image like Fig.
13(A) [38]. The main finding of [38] concerning this property is that switch forwards
occurred more often than chance would suggest. Therefore, they proposed that some
slow process may be providing a memory of the previous image. We suggest short
term depression as a candidate substrate for this memory. As seen in Fig. 14, the bias
21
Fig. 14. Noise degrades two sources of information provided by dominance switches. (A) In
the absence of noise, switches always move "forward," so that the previous percept perfectly predicts
the subsequent percept. Dominance times accumulate at a single value too. (B) For slightly higher
levels of noise (
ε = 0.0002), "switch backs" can occur where the subsequent percept is the same
√
as the previous percept. Also, the distribution of dominance times spreads. (C) For stronger noise
(
ε = 0.0004). Other parameters are I0 = 0.6, β = 1, and τ = 50.
√
Fig. 15. The probability of a switch being in the forward direction in simulations of (60)
as a function of the amplitude ε of noise. As ε increases, network switches behave in more of a
Markovian way, not reflecting any memory of the previous percept. Therefore, information of the
previous percept is lost as soon as a switch occurs.
22
in favor of switching forward persists even for substantial levels of noise. The idea
of short term plasticity as a substrate of working memory was also recently proposed
in [35]. Our results extends this idea, suggesting synaptic mechanisms of working
memory may be useful in visual perception tasks, such as understanding ambiguous
images. In Fig. 15, we show that the process of dominance switching becomes more
ε is increased even a modest amount. In the limit
Markovian as the level of noise
of large noise, the likelihoods of "switch forwards" and "switch backs" are the same.
√
4. Discussion. Mechanisms underlying stochastic switching in perceptual ri-
valry have been explored in a variety of psychophysical [17, 30, 6], physiological [31, 5],
and theoretical studies [33, 29, 37]. Since psychophysical data is widely accessible,
it can be valuable to use the hallmarks of its statistics as benchmarks for theoreti-
cal models. For instance, the fact that dominance time distributions are unimodal
functions peaked away from zero suggests that some adaptive process must underlie
switching in addition to noise [29, 6, 44].
In addition, [36] recently suggested the
visual system may sample the posterior distribution of interpretations of bistable im-
ages. This type of sampling can be well modeled by attractor networks analogous to
those presented here [37]. Therefore, many dominance time statistics from perceptual
rivalry experiments can be employed as points of reference for physiologically based
models of visual perception. New data now exists concerning tristable images showing
this process also likely is guided by a slow adaptive process in addition to fluctuations
[38].
We have studied various aspects of competitive neuronal network models of per-
ceptual multistability that include short term synaptic depression. First, we were able
to analyze the onset of rivalrous oscillations in a ring model with synaptic depression
[54, 25]. Stimulating the network with a bimodal input leads to winner-take-all so-
lutions, in the form of single bumps, in the absence of synaptic depression. As the
strength of synaptic depression is increased, the network undergoes a bifurcation
which leads to slow oscillations whose timescale is set by that of synaptic depression.
Each stimulus peak is represented in the network by a bump whose dominance time is
set by the height of each peak. Thus, synaptic depression reveals information about
the stimulus that would otherwise be masked by the lateral inhibitory connectivity
of the network. The inclusion of noise in the network leads to dominance times that
are exponentially (gamma) distributed in the absence (presence) of synaptic depres-
sion. Motivated by recent work exploring how visual perception may exploit Bayesian
sampling on posterior distributions [24, 19, 36], we considered the simple task of an
observer trying to infer the contrast of stimuli based on dominance times. We found
Bayesian sampling of the dominance times discerning input contrast differences better
as switches become more depression driven and less noise-driven. Thus, short term
depression improves information transfer of networks that process ambiguous images
in multiple ways.
We also used energy methods in simple space-clamped neural network models to
understand how a combination of noise and depression interact to produce switching
in competitive neural networks. Using the energy function derived by Hopfield for
analog neural networks, we justify the exponential dependence of dominance times
upon input strength in purely noise-driven switching. Studying an adiabatically de-
rived energy function for the case of slow depression, we also show how depression
works to reduce the energy barrier between winner-take-all states, leading to the slow
timescale that defines the peak in depression-noise generated switches. Finally, using
a three population space-clamped neural network, we analyzed depression and noise
23
generated switching that may underlie perceptual tristability. We found this network
also sustained some of the same relationships between input contrast and dominance
times as the two population network. Also, we found that when switches are gen-
erated by depression there is an ordering to the population dominance that is lost
when switches are noise generated. This is due to the memory generated by short
term depression [35], so the switching process is non-Markovian due to the inherent
slow timescale in the background. However, even small amounts of noise can wash this
memory away. Thus, since recent psychophysical experiments reveal a non-Markovian
property to percept ordering, this provides further support for the idea that a slow
adaptive process underlies percept switching.
Note to analytically study the relationship between dominance times and input
contrast in the noisy system, we resorted to a simple space-clamped neural network.
In future work, we plan to develop energy methods for spatially extended systems like
(35). Such methods have seen success in analyzing stochastic partial differential equa-
tion models such as Ginzburg-Landau models [14]. Energy functions have recently
been developed for neural field models, but have mostly been studied as a means of
determining global stability in deterministic systems [53, 28, 40]. We proposed that
by deriving the specific potential energy of spatially extended neural fields, it may
be possible to approximate the transition rates of solutions from the vicinity of one
attractor to another. In the system (35), there should be some separatrix between
the two winner-take-all states that must be crossed in order for a transition to occur.
The least action principle states that there is even a specific point on this separatrix
through which the dynamics most likely flows [14]. Finding this with an energy func-
tion in hand would be straightforward would allow us to relate the parameters of the
model to the distribution of dominance times. This would provide a better theoretical
framework for interpreting data concerning rivalry of spatially extended images, such
as those that produce waves [50].
REFERENCES
[1] L. F. Abbott, J. A. Varela, K. Sen, and S. B. Nelson, Synaptic depression and cortical
gain control., Science, 275 (1997), pp. 220 -- 224.
[2] S. Amari, Dynamics of pattern formation in lateral-inhibition type neural fields., Biol Cybern,
27 (1977 Aug 3), pp. 77 -- 87.
[3] R. Ben-Yishai, R. L. Bar-Or, and H. Sompolinsky, Theory of orientation tuning in visual
cortex, Proc. Natl. Acad. Sci. USA, 92 (1995), pp. 3844 -- 3848.
[4] R. Blake, A neural theory of binocular rivalry, Psychol Rev, 96 (1989), pp. 145 -- 67.
[5] R. Blake and N. Logothetis, Visual competition, Nature Reviews Neuroscience, 3 (2002),
pp. 1 -- 11.
[6] J. W. Brascamp, R. van Ee, A. J. Noest, R. H. A. H. Jacobs, and A. V. van den Berg,
The time course of binocular rivalry reveals a fundamental role of noise, J Vis, 6 (2006),
pp. 1244 -- 56.
[7] P. C. Bressloff and J. D. Cowan, An amplitude equation approach to contextual effects in
visual cortex, Neural Comput, 14 (2002), pp. 493 -- 525.
[8] A. Buckthought, J. Kim, and H. R. Wilson, Hysteresis effects in stereopsis and binocular
rivalry, Vision Res, 48 (2008), pp. 819 -- 30.
[9] G. Burton, Successor states in a four-state ambiguous figure, Psychonomic Bulletin and Re-
view, 9 (2002), pp. 292 -- 297.
[10] O. Carter, T. Konkle, Q. Wang, V. Hayward, and C. Moore, Tactile rivalry demonstrated
with an ambiguous apparent-motion quartet, Curr Biol, 18 (2008), pp. 1050 -- 4.
[11] F. S. Chance, S. B. Nelson, and L. F. Abbott, Synaptic depression and the temporal
response characteristics of v1 cells, J Neurosci, 18 (1998), pp. 4785 -- 99.
[12] T. Chen and S. I. Amari, Stability of asymmetric hopfield networks, IEEE Trans Neural Netw,
12 (2001), pp. 159 -- 63.
24
[13] D. Deutsch, An auditory illusion, Nature, 251 (1974), pp. 307 -- 309.
[14] W. E, W. Ren, and E. Vanden-Eijnden, Minimum action method for the study of rare events,
Communications on Pure and Applied Mathematics, 57 (2004), pp. 637 -- 656.
[15] D. Ferster and K. D. Miller, Neural mechanisms of orientation selectivity in the visual
cortex, Annu Rev Neurosci, 23 (2000), pp. 441 -- 71.
[16] G. H. Fisher, "mother, father, and daughter": a three-aspect ambiguous figure, Am J Psychol,
81 (1968), pp. 274 -- 7.
[17] R. Fox and J. Herrmann, Stochastic properties of binocular rivalry alternations, Attention,
Perception, & Psychophysics, 2 (1967), pp. 432 -- 436.
[18] M. Hausser and A. Roth, Estimating the time course of the excitatory synaptic conductance
in neocortical pyramidal cells using a novel voltage jump method, J Neurosci, 17 (1997),
pp. 7606 -- 25.
[19] J. Hohwy, A. Roepstorff, and K. Friston, Predictive coding explains binocular rivalry: an
epistemological review, Cognition, 108 (2008), pp. 687 -- 701.
[20] J. J. Hopfield, Neurons with graded response have collective computational properties like
those of two-state neurons, Proc Natl Acad Sci U S A, 81 (1984), pp. 3088 -- 92.
[21] J.-M. Hupe, Dynamics of m´enage `a trois in moving plaid ambiguous perception, J Vis, 10
(2010), p. 1217.
[22] J.-M. Hup´e and D. Pressnitzer, The initial phase of auditory and visual scene analysis,
Philos Trans R Soc Lond B Biol Sci, 367 (2012), pp. 942 -- 53.
[23] Y. Kawaguchi and Y. Kubota, Gabaergic cell subtypes and their synaptic connections in rat
frontal cortex, Cereb Cortex, 7 (1997), pp. 476 -- 86.
[24] D. Kersten, P. Mamassian, and A. Yuille, Object perception as bayesian inference, Annu
Rev Psychol, 55 (2004), pp. 271 -- 304.
[25] Z. P. Kilpatrick and P. C. Bressloff, Binocular rivalry in a competitive neural network
with synaptic depression, SIAM J Appl Dyn Syst, 9 (2010), pp. 1303 -- 1347.
[26] Z. P. Kilpatrick and P. C. Bressloff, Effects of synaptic depression and adaptation on
spatiotemporal dynamics of an excitatory neuronal network, Physica D: Nonlinear Phe-
nomena, 239 (2010), pp. 547 -- 560. Mathematical Neuroscience.
[27] Z. P. Kilpatrick and P. C. Bressloff, Spatially structured oscillations in a two-dimensional
excitatory neuronal network with synaptic depression, J Comput Neurosci, 28 (2010),
pp. 193 -- 209.
[28] S. Kubota and K. Aihara, Analyzing global dynamics of a neural field model, Neural Pro-
cessing Letters, 21 (2005), pp. 133 -- 141.
[29] C. R. Laing and C. C. Chow, A spiking neuron model for binocular rivalry, J Comput
Neurosci, 12 (2002), pp. 39 -- 53.
[30] S. R. Lehky, Binocular rivalry is not chaotic, Proceedings of the Royal Society of London.
Series B: Biological Sciences, 259 (1995), pp. 71 -- 76.
[31] D. A. Leopold and N. K. Logothetis, Activity changes in early visual cortex reflect monkeys'
percepts during binocular rivalry, Nature, 379 (1996), pp. 549 -- 53.
[32] W. J. M. Levelt, On binocular rivalry, Institute for Perception RVO -- TNO, Soesterberg, The
Netherlands, 1965.
[33] K. Matsuoka, The dynamic model of binocular rivalry, Biol Cybern, 49 (1984), pp. 201 -- 8.
[34] J. F. Mejias, H. J. Kappen, and J. J. Torres, Irregular dynamics in up and down cortical
states, PLoS One, 5 (2010), p. e13651.
[35] G. Mongillo, O. Barak, and M. Tsodyks, Synaptic theory of working memory, Science, 319
(2008), pp. 1543 -- 6.
[36] R. Moreno-Bote, D. C. Knill, and A. Pouget, Bayesian sampling in visual perception,
Proc Natl Acad Sci U S A, 108 (2011), pp. 12491 -- 6.
[37] R. Moreno-Bote, J. Rinzel, and N. Rubin, Noise-induced alternations in an attractor net-
work model of perceptual bistability, J Neurophysiol, 98 (2007), pp. 1125 -- 39.
[38] M. Naber, G. Gruenhage, and W. Einhauser, Tri-stable stimuli reveal interactions among
subsequent percepts: rivalry is biased by perceptual history, Vision Res, 50 (2010), pp. 818 --
28.
[39] J. Orbach, D. Ehrlich, and H. A. Heath, Reversibility of the necker cube. I. an examination
of the concept of "satiation of orientation", Percept Mot Skills, 17 (1963), pp. 439 -- 58.
[40] M. R. Owen, C. R. Laing, and S. Coombes, Bumps and rings in a two-dimensional neural
field: splitting and rotational instabilities, New Journal of Physics, 9 (2007), p. 378.
[41] D. Pressnitzer and J.-M. Hup´e, Temporal dynamics of auditory and visual bistability reveal
common principles of perceptual organization, Curr Biol, 16 (2006), pp. 1351 -- 7.
[42] J. Seely and C. C. Chow, Role of mutual inhibition in binocular rivalry, J Neurophysiol, 106
(2011), pp. 2136 -- 50.
25
[43] A. Shpiro, R. Curtu, J. Rinzel, and N. Rubin, Dynamical characteristics common to neu-
ronal competition models, J Neurophysiol, 97 (2007), pp. 462 -- 73.
[44] A. Shpiro, R. Moreno-Bote, N. Rubin, and J. Rinzel, Balance between noise and adaptation
in competition models of perceptual bistability, J Comput Neurosci, 27 (2009), pp. 37 -- 54.
[45] M. Tsodyks, K. Pawelzik, and H. Markram, Neural networks with dynamic synapses, Neural
Comput., 10 (1998), pp. 821 -- 835.
[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 (1997),
pp. 719 -- 723.
[47] R. van Ee, Stochastic variations in sensory awareness are driven by noisy neuronal adaptation:
evidence from serial correlations in perceptual bistability, J Opt Soc Am A Opt Image Sci
Vis, 26 (2009), pp. 2612 -- 22.
[48] X. Wang and J. Rinzel, Alternating and synchronous rhythms in reciprocally inhibitory model
neurons, Neural Computation, 4 (1992), pp. 84 -- 97.
[49] H. R. Wilson, Computational evidence for a rivalry hierarchy in vision, Proc Natl Acad Sci
U S A, 100 (2003), pp. 14499 -- 503.
[50] H. R. Wilson, R. Blake, and S. H. Lee, Dynamics of travelling waves in visual perception,
Nature, 412 (2001), pp. 907 -- 10.
[51] H. R. Wilson and J. D. Cowan, A mathematical theory of the functional dynamics of cortical
and thalamic nervous tissue, Biol Cybern, 13 (1973), pp. 55 -- 80.
[52] J. M. Wolfe, Stereopsis and binocular rivalry, Psychol Rev, 93 (1986), pp. 269 -- 82.
[53] S. Wu, S.-I. Amari, and H. Nakahara, Population coding and decoding in a neural field: a
computational study., Neural Comput, 14 (2002), pp. 999 -- 1026.
[54] L. C. York and M. C. W. van Rossum, Recurrent networks with short term synaptic depres-
sion, J Comput Neurosci, 27 (2009), pp. 607 -- 20.
[55] W. Zhou and D. Chen, Binaral rivalry between the nostrils and in the cortex, Curr Biol, 19
(2009), pp. 1561 -- 5.
26
|
1703.04176 | 1 | 1703 | 2017-03-12T21:04:39 | Locally embedded presages of global network bursts | [
"q-bio.NC",
"math.DS",
"nlin.CD"
] | Spontaneous, synchronous bursting of neural population is a widely observed phenomenon in nervous networks, which is considered important for functions and dysfunctions of the brain. However, how the global synchrony across a large number of neurons emerges from an initially non-bursting network state is not fully understood. In this study, we develop a new state-space reconstruction method combined with high-resolution recordings of cultured neurons. This method extracts deterministic signatures of upcoming global bursts in "local" dynamics of individual neurons during non-bursting periods. We find that local information within a single-cell time series can compare with or even outperform the global mean field activity for predicting future global bursts. Moreover, the inter-cell variability in the burst predictability is found to reflect the network structure realized in the non-bursting periods. These findings demonstrate the deterministic mechanisms underlying the locally concentrated early-warnings of the global state transition in self-organized networks. | q-bio.NC | q-bio | Locally embedded presages of global network bursts
Satohiro Tajima1,2,3,*, Takeshi Mita4, Douglas J. Bakkum5, Hirokazu Takahashi4, and Taro Toyoizumi3
1) Department of Neuroscience, University of Geneva, CMU, Rue Michel-Servet 1, Genève, 1211, Switzerland.
2) JST PRESTO, Japan Science and Technology Agency, 4-1-8 Honcho, Kawaguchi, Saitama, 332-0012, Japan.
3) RIKEN Brain Science Institute, 2-1, Hirosawa, Wako, Saitama, 351-0198, Japan.
4) Research Center for Advanced Science and Technology, the University of Tokyo, 4-6-1, Komaba, Meguro-ku, Tokyo, 153-8904,
Japan.
5) ETH Zurich, Department of Biosystems Science and Engineering, Basel, Switzerland.
*[email protected], CMU, Rue Michel-Servet 1, Genève, 1211, Switzerland.
Abstract
Spontaneous, synchronous bursting of neural population is a widely observed phenomenon in nervous networks, which is
considered important for functions and dysfunctions of the brain. However, how the global synchrony across a large number
of neurons emerges from an initially non-bursting network state is not fully understood. In this study, we develop a new
state-space reconstruction method combined with high-resolution recordings of cultured neurons. This method extracts
deterministic signatures of upcoming global bursts in "local" dynamics of individual neurons during non-bursting periods.
We find that local information within a single-cell time series can compare with or even outperform the global mean field
activity for predicting future global bursts. Moreover, the inter-cell variability in the burst predictability is found to reflect
the network structure realized in the non-bursting periods. These findings demonstrate the deterministic mechanisms
underlying the locally concentrated early-warnings of the global state transition in self-organized networks.
1
Introduction
The collective firing dynamics of neural population have been related to emergent functions and dysfunctions in the brain.
Unraveling the structure of dynamics emerging from the collective behavior in neural networks is an ultimate issue in
studies of complex systems (1). The simplest and most profound example of collective dynamics is synchronous burst: the
simultaneous occurrence of closely-spaced action potentials ("bursts") across a large number of neurons in the network.
Within neural circuits in vivo, the synchronous bursts are believed to serve important roles in information storage and
transmission (2, 3) as well as in disease states including epileptic seizures (4–6). Intriguingly, neurons cultured in vitro also
display spontaneous synchronous bursts akin to those observed in vivo (7–9), indicating that the collective bursting is a
universal behavior of neural networks in a wide variety of preparations.
Classic biophysical models of the synchronized bursts assume broadly diverging synaptic interaction that propagates
activity of a single cell to the remaining neural population (4, 10–12). This view has been partially supported by
experiments demonstrating that stimulating a single neuron can change the large-scale dynamics of neural populations (13–
15). These models and experiments suggest a close link between single neuron activity and the global state of a neural
population. On the other hand, the previous stochastic models do not specify when a global burst occurs; forecasting a
spontaneous occurrence of synchronous burst from the preceding individual neural activities in non-bursting states remains
a highly challenging issue, particularly in real nervous systems which feature the heterogeneity of local-global coupling
(e.g., non-uniform coupling strength between individual neural activity and the overall firing of the population) (16). Indeed,
recent studies demonstrate potentially diverse dynamical mechanisms accompanied by complex but reproducible patterns
of activity sequences leading to the synchronous bursts (17, 18), implying the existence of nonlinear deterministic
mechanisms around the burst initiations.
The present paper investigates the deterministic aspects of burst initiation. In particular, we question whether individual
neuron dynamic in non-bursting periods can predict an upcoming synchronous burst in cultured neural population. To cope
with the heterogeneous nature of local-global coupling, we introduce a model-free method to quantify the deterministic
relationships among individual neuron traces and the global state, based on a nonlinear state-space reconstruction technique.
Using this method, we report that dynamic of only one neuron can robustly predict the upcoming synchronous burst in the
neural population, at high signal-to-noise ratio. Surprisingly, the burst predictability with an appropriately-chosen neuron
even outperforms that with the mean field activity of neural population, demonstrating that macroscopic fluctuation of
neural population is better predicted by the microscopic dynamic of a specific single cell, rather than the macroscopic state
itself. This apparently counterintuitive property is explained based on mathematical property termed 'embedding' in
nonlinear dynamical systems and the heterogeneous input-output interactions among neurons in the network shaping non-
bursting spontaneous dynamics.
Results
Spontaneous synchronous bursts in cultured neurons
Rat cortical neurons were grown on high-dense CMOS-based multi-electrode arrays for 12-41 days-in-vitro (MEAs; Fig.
1a) (19, 20). The devices allow us to simultaneously stimulate and record from multiple neurons lying on the array surface
at a high spatiotemporal resolution (Methods). Each recording yielded simultaneous traces of 60–98 cells at a temporal
resolution of 2 kHz. Figure 1b shows the cell distribution recorded with a representative array preparation. In our
preparations of sparse culturing, the cell bodies were well isolated, and the action potential of each cell was identified
accurately based on its spatial location and waveform. Spike action-potentials were detected using standard LimAda
method (21).
We first analyzed the spontaneous activities of the recorded neural population. During the spontaneous firing, the neurons
showed intermittent synchronous bursts, which were interleaved with silent periods (Fig. 1c). To describe the macroscopic
network dynamics, we defined population activity by the mean firing rate across all recorded neurons (Fig. 1d). The
synchronous bursting period were defined by time bins in which the population activity exceeded a threshold determined
for each preparation, according to the previously established method based on inter-spike interval distributions (20) (Fig.
1d, inset). The timing of synchronous burst occurrence did not show any clear periodicity. Duration of bursts varied from
about 200 to 800 ms. The different preparations showed distinct complex patterns of collective activity dynamics when
visualized with standard dimensionality reduction techniques such as principal component analysis (Supplementary Fig.
S1).
2
Predicting bursts
To capture the potentially complex temporal structures in neural activities around bursting dynamics, we developed a
model-free method that is able to quantify the general relationships between the network and single-element dynamics in
high-dimensional nonlinear systems. Because the method features tracing the reconstructed state-space trajectories in the
retrograde direction, from now on we term the method as trace-back embedding (TBE). The TBE extends the nonlinear
forecasting techniques based on state-space reconstruction using the delay-embedding theorems in deterministic dynamical
systems (22, 23). Compared to the previous nonlinear forecasting methods (24–26), it has two additional advantages. First,
the method focuses particularly on the relationships between a global mean-field variable and local variables. Second, the
method quantifies the predictability of a global synchronous burst based on temporally distant dynamics of each local
variable. The basic protocol of TBE is summarized below.
Figure 2 illustrates the TBE protocol (further details are provided in Methods). We first reconstruct the dynamics of
population activity and each single-neuron activity respectively in delay-coordinate state-spaces (Fig. 2a, b):
bt
d = (bt, bt-τ …, bt-(d-1)τ),
xi,t
d = (xi,t, xi,t-τ …, xi,t-(d-1)τ),
(1)
(2)
d and xi,t
where bt, and xi,t are the population mean-field and neuron i's activity at time t, respectively; d denotes the embedding
dimensions (number of delay-coordinates); τ is the unit delay size. In this study, the local time series xi,t was defined by
each neuron i's spike train smoothed with a Gaussian kernel, and bt was defined by the mean of xi,t over all the
d as the "global" and "local" state
simultaneously recorded neurons (Methods). For convenience, we refer to bt
trajectories, respectively. The TBE method measures the accuracy of synchronous-burst prediction by seeking "presages"
of global bursts in each local state trajectories, based on the accuracy of forecasting using a nearest-neighbor model (Fig.
2b, see also the figure legend for a step-by-step description of the protocol). The rationale behind the TBE protocol is as
follows: if neuron i has enough information to predict the future synchronous burst occurring after a time span –Δt (–Δt>0),
neuron i's state up to time –Δt before the targeted burst should be already differentiated from the states without any future
bursting, forming a cluster of states that deviate from those predicting non-bursting periods in the reconstructed state space.
Similarly, we define the accuracy of "postdiction" of bursts (i.e., detecting a burst event based on neural activities after its
occurrence) with the same protocol as that for prediction except for using time span Δt with a positive value. We iterated
this procedure for all the neurons to characterize the burst predictabilities in individual neurons. In addition, we quantified
the "self-predictability" of the mean-field activity by replicating the above procedure based on its own (global) state
trajectory, bt, instead of a local state trajectory in individual neurons, xi,t.
Using this method, we investigated the predictability of synchronous bursts based on the individual neural activities or the
population mean-field. In particular, we addressed two key questions: (i) how accurately can each single neuron predict
the future synchronous bursting events, and (ii) if a subset of neurons predicts synchronous bursts better than others, how
are they related to the underlying neural network structure?
Single-neuron dynamics can predict global network bursts
First, we asked how accurately individual neurons predict (or postdict) the occurrence of synchronous bursts, and how their
predictions compare to that based on the population mean-field, which defines synchronous burst. We quantified the rate
of successful synchronous-burst predictions in each single neuron with keeping the false-positive rate at 5% (Methods).
Interestingly, in some cells (e.g., Fig. 3a), the rates of successful burst predictions with the TBE method could be relatively
high (>50% of trials) even when the network appeared to be silent in terms of the global mean-field firing rate (Fig. 3d, –
1000 ~ –500ms). Corresponding to this, the success rate of TBE-based burst detection using a single neuron dynamics
could be even higher than that with the global mean field dynamics (Fig. 3b). The predictability of burst occurrence varied
among neurons, suggesting the heterogeneity of burst predictability across neurons. However, the accuracy of predictions
in individual neurons were highly consistent among different sets of burst trials within each preparation (ρ=0.79~0.93,
P<2×10-23, Spearman's rank correlation between the first and the second halves of trials; Fig. 3e), demonstrating that this
cell-to-cell variability reflects a robust property of each neuron. Across all the preparations, about 1/3 of neurons were
found to predict synchronous bursts robustly better than population mean-field over some prediction span, Δt (Fig. 3g).
The fact that only a single neuron can compare or even outperform the population mean at predicting the mean-field
dynamics may appear surprising, considering that the single-cell activities fluctuate due to spiking much more than the
mean-field activity (e.g., compare the black traces in Figs. 3c and 3d). However, it is explained by the delay-embedding
theorems (22, 23) that the global state information could be reconstructed from observation of time series in a single
variable participating in the dynamical system (Supplementary Fig. S2). The present result suggests that such a
mathematical property can be demonstrated in a biological system. Importantly, to reconstruct the global state requires, we
need to observe multiple time points rather than the momentary state in the time-series of the observed variable (which is
3
why the theorem is called "delay-embedding" theorem). It suggests that the information about the global state is carried by
the temporal patterns including multiple time points, rather than momentary snapshots of activity in neurons.
To explore the importance of temporal patterns of activities in predicting bursts compared to the momentary firing rates,
we compared the current TBE-based prediction accuracy with an analogous index based on momentary firing rates, xi,t or
bt (the red traces in Figs. 3c and 3d, respectively; see also Methods). In single neurons, the TBE method tended to provide
more accurate prediction than the momentary-activity-based method (P<5×10-14 in all the individual preparations, sign test
with paired samples; Fig. 3f). This superiority of TBE to the momentary firing rate is not likely to be due to the difference
in the sampled spike numbers because it remained even if we lengthened the bin size for computing activity so as to have
the same total duration as used in TBE (P<6×10-3 in all the preparations except for Chip 2427 [which had much lower
firing rate than others], sign test with paired samples; Fig. 3f, the inset). Finally, when we used the momentary firing rates
instead of the TBE with temporal sequence, the fraction of cells whose burst detectability outperformed that of the mean
field reduced from ~1/3 to ~1/5 (Fig. 3g). Together, these results suggest that the temporal patterns in neural activity, not
only momentary activity, are crucial for early detection of synchronous bursts.
Burst predictability reflects network structures realized in non-bursting periods
Next, we examined whether the predictability of synchronous bursts corresponds to any specific network structure defined
by synaptic interactions. The structural correspondences of burst predictability were investigated in three steps (Figs. 4):
we first re-analyzed the non-bursting spontaneous activities to infer the directed causal network structure; next, the inferred
causal networks were verified by comparing them to synaptic interactions measured by an additional electrical stimulation
experiment with the same preparations; finally, we related each neuron's network connectivity with its burst prediction
accuracy.
In the causal network analysis, we applied a previously proposed method that is capable of detecting weak nonlinear
coupling (26) (Methods). Theoretically, the method detects causal interactions including indirect ones (27). Applying this
method to all the neuron pairs showing non-bursting spontaneous activities yielded a matrix for each preparation that
represents pairwise, directed interactions within each preparation (Fig. 4a).
The relevance of the inferred causal couplings to synaptic interactions among neurons was verified using independent data
in which synaptic interactions were directly measured (Fig. 4a, b). The present CMOS-based MEA system allowed us to
stimulate the neural tissue local to given electrodes at sub-millisecond accuracy, directly activating a single neuron's soma
at a high temporal precision with few artifact (19). A subset of neurons was electrically stimulated while the evoked
activities in other neurons were monitored simultaneously. According to the evoked response latencies, we identified the
short-latency (< 10 ms) and long-latency (≥ 10 ms) downstream cells for each stimulated neuron, between which the latter
was expected to include more multi-synaptic interactions and thus weaker causal effects. As expected, we confirmed that
the causality inferred based on non-bursting spontaneous data consistently reflected the identified short- and long-latency
interactions (Fig. 4b). Specifically, the inferred causality differentiated the short- and long-latency downstream cells from
the remaining population (P = 0.0001, sign test, data pooled across the all preparations). Based on these confirmations, we
used the inferred causality as proxies for the effective synaptic interactions in the network during the non-bursting periods.
Finally, we compared these causal network structures and the burst predictabilities, finding that the burst predictability in
each neuron displayed a positive correlation to the average inferred causality between it and the other neurons (Figs. 4d
and 4e). We classified putative excitatory and inhibitory neurons based on the spike waveforms (Methods). Interestingly,
the excitatory neurons predicted bursts better than the inhibitory neurons when we compared the neurons showing strong
causality (P<0.002, both for the causality for input and output connections, rank sum test), but such cell-type dependency
was unclear in the neurons having weak interactions (P>0.5; Figs. 4d). These results suggest that the stronger excitatory
network interaction with other cells underlie the higher burst predictability. The above results together led us to conclude
that the cell-to-cell variability in burst predictability is likely to reflect the heterogeneous network structures that shape
non-bursting activity dynamics.
Discussion
In this study, we have focused on the relationship between local and global neural properties by analyzing how single cell
activity predicts the population mean-field dynamics. The newly introduced TBE method revealed the heterogeneity of
neurons also in the predictability of upcoming occurrence of a synchronous burst of a neural population, which is
temporally distant in order of hundred milliseconds. The bursts were predicted accurately by dynamics of some single
neurons. Remarkably, such neuron even outperformed the population mean-field dynamics itself in terms of prediction
accuracy, particularly when we extract the information from the temporal patterns of neural activity rather than the
momentary firing rates. From causal network analysis combined with the electrical stimulation experiment, the
4
heterogeneity of burst predictability was explained by the structures of neural networks: the "hub"-like neurons having
stronger synaptic interactions with other cells can predict the upcoming global network burst better than others. In addition,
the present results based on the cultured neurons indicate the heterogeneous relationships between single neurons and
population activity are basic properties observed in the nervous network that is self-organized without external stimulus.
Classic models of the synchronized population burst assume diverging interaction that propagates activity of a cell to the
other cells (4, 10–12). A variant of this model features "avalanche" like bursting sequences, which models synchronous
burst with noise being amplified and propagated through the network (9, 18, 28–32). These models, focusing mainly on
the stochastic aspects of burst initiation and propagation, have made successes in explaining a variety of statistical
properties of bursting neural networks. On the other hand, it has been challenging to predict the occurrence of synchronous
burst, partially because they can be triggered by complex and heterogeneous mechanisms (33). The present study, focusing
on deterministic aspects of the neural population dynamics that transit from silent state to synchronous burst, demonstrates
that the burst occurrence is predictable to the extent based on preceding activity traces in single neurons. This is in line
with studies suggesting that the spatiotemporal pattern of spontaneous fluctuation in neural population activity is not purely
random but restricted to a limited number of configurations (34, 35). The existence of specific neurons predicting the
upcoming burst is also consistent with the previous idea of 'leader neurons' that fire at the beginning of bursts (8, 36).
Notably, however, the predictor of the global burst is not identical to the increased activities in those single neurons as
expected from previous studies (13, 30, 36), because the magnitude of FR alone is not sufficient to explain the highly
accurate predictability based on TBE method (Fig. 3e). Rather, the earliest presages are likely to be embedded in the
dynamical patterns of those neurons.
On the other hand, it should be noted that not everything is predictable in a deterministic manner before the burst occurrence.
A most relevant example is the trial-to-trial fluctuation in burst size (the height of peak population rate during a burst). In
contrast to the case of the burst detection, the cell-to-cell heterogeneity in the burst-size predictability (Methods) was not
always consistent (ρ = –0.289 ~ –0.077, P>0.4, Spearman's rank correlation between the odd and even trials). Neither the
global state could predict the size of upcoming burst (ρ = 0.091 ~ 0.35, P>0.1, Spearman's rank correlation between actual
and predicted burst sizes). These facts suggest that the burst size is determined in a different mechanism from that triggering
the burst: the former is more likely to follow a stochastic mechanism during the burst (as previously described in
probabilistic activity-propagation models (18, 32)) whereas, somewhat counterintuitively, the latter seems to follow more
deterministic process.
Finally, the predictability of upcoming burst occurrence based on single-cell dynamics suggests an effective methodology
to capture the early warnings of population state transition in a variety of networks, including epileptic seizures in the brain
(5, 37), rumor propagation in social networks (38–40), or epidemic spreading in human networks (41, 42). In clinical
purposes, for example, accurate prediction and characterization of pathologically synchronized neural firing is an important
issue (5, 37). Although nonlinear features of field potential sequence have been proposed as useful markers of epileptic
state (37, 43–46), the predictability of global dynamics based on single neuron activity, to the best knowledge of ours, has
been not thoroughly studied. In particular, the present study suggests that single neuron could even outperform the direct
observation of global state itself in predicting the global state dynamics. The success of the present state-space
reconstruction-based method implies a new approach to detect the early warnings of transitions in global network state
based on local rather than global features.
Methods
Cell culturing
All experimental protocol was approved by the Committee on the Ethics of Animal Experiments at the Research Center
for Advanced Science and Technology, the University of Tokyo (Permit Number: RAC130106).
A protocol for cell culturing had been reported previously (19). Briefly, E18 Wistar rat cortices were dissociated using
trypsin and mechanical trituration. 20-40k/µL neurons and glia were seeded over an area of ~12 mm2 on top of the CMOS
chip. Layers of poly (ethyleneimine) followed by laminin were used to adhere cells. Plating media consisted of Neurobasal-
B27 supplemented with 10% horse serum and 0.5 mM GlutaMAX during the first 24 hours. Growth media consisted of
DMEM supplemented with 10% horse serum, 0.5 mM GlutaMAX, and 1 mM sodium pyruvate. Experiments were
conducted inside an incubator to control of environmental conditions (36°C and 5% CO2).
CMOS-based recording and stimulation of network activity
Cultured neuron activities were simultaneously recorded with high-density microelectrode array as described before (19,
20). Cortical networks were grown over 11,011-electrode CMOS-based MEAs (47, 48), which provide enough spatial and
5
temporal resolution to detect action potentials from any neuron lying on the array: 1.8 × 2.0 mm2 area containing 8.2 × 5.8
μm2 electrodes with 17.8 μm pitch, sampled at 20 kHz. Subsets of 126 electrodes can be read-out (and stimulated) at one
time, and electrode selection can be re-configured within a few milliseconds. To identify the locations of neurons growing
over the array, a sequence of about one hundred recording configurations were scanned across the whole array while
recording spontaneous activity. Locations of active somata were identified because action potentials could be usually
detected from multiple nearby electrodes. Electrode selection was then re-configured such that a single electrode was
assigned to each identified soma, and spontaneous activities were measured simultaneously from all of these cells. The
putative neuron types (excitatory or inhibitory) were estimated based on spike shapes (Supplementary Fig. S3). We could
recode from 93, 47, 98, 92 and 53 neurons in Chip 1437, 1440, 1444, 2427 and 2440, respectively.
Microstimulation-elicited network activities were then investigated to characterize a pairwise synaptic strength between
neurons. An adequate stimulating electrode was explored such that a single target neuron was directly activated through
axonal stimulation and that the directly evoked spikes were exclusively measured from the target cell. The directly evoked
spikes could be easily distinguished from post-synaptic activations because they were very reliable (i.e., 100 spikes elicited
out of 100 stimulation trials) and exhibited a small temporal jitter (Supplementary Fig. S4). The microstimulation was a
single, positive-negative, biphasic pulse with a charge-balanced amplitude of ± 300-900 mV and a duration of 200 µs/phase,
and was delivered 100 times every 3 sec.
Burst detection
Burst in spontaneous activity was detected modifying a protocol that was previously established by authors' group (20).
The previous study proposes to determine the threshold for burst detection based on the distribution of inter-spike interval
(ISI) of consecutive spikes (20). In the bursting neurons, the ISI distribution typically has a bimodal structure whose valley
can be used as an objective criterion for burst detection. In this study, we first computed sequences of population firing
rate that were normalized such that it ranges from 0 to 1. We defined this as the sequence of global state, bt, where t is the
time. This was to use a common burst detection threshold across different preparations of neuron cultures, which generally
vary in terms of the absolute firing rate. We used 5 ms bin for the firing rate computation. We next derived the distributions
of inverse firing rate over the bins. This yielded an ISI distributions for each preparation, in which we confirmed that all
of them had bimodal shapes (Fig. 1d, inset). For each preparation, we selected the smallest ISI providing the valley of
distribution as the burst detection threshold. The burst periods were determined by the at least three consecutive sequences
of bins that have average ISIs under this threshold, identifying 26 bursts on average for each preparation. We defined the
peak burst timing by selecting the center of the bin having the smallest average ISI (i.e., the largest global state) within
each burst period.
Estimation of synaptic connectivity from electrical stimulation
Synaptic connectivity was estimated based on the evoked responses during electrical stimulation experiment. To reliably
stimulate a single neuron, we searched a stimulation site around a target neuron in each MEA that could elicit an action
potential exclusively at the target neuron. After the careful selection of stimulation sites, we could evoke the action
potentials at almost 100% probability for each single stimulation, with very little jitters in the timings of action potentials
in the stimulated neurons. We first aligned the spike raster to the timings of stimulation. We next computed the sequence
of firing rate xi,t in a way described above. The firing rates before the stimulation (-500 ms < t < -100 ms) were used to
produce the reference probability distribution of each neuron's firing rate, P(xi). Next, we computed the probability, P(xi >
xi,t), for each time bin centered at t in a short post-stimulation period (2.5 ms < t < 15 ms); the responses during 0 ms < t <
2.5 ms were omitted to eliminate the artifacts of electrical stimulation. We defined the smallest time t that satisfies P(xi >
xi,t) < 0.01 in the post-stimulation period as the latency of neuron i. The neurons that had at least one time-bin satisfying
this condition were defined as "downstream cells of the stimulated neuron; the neurons that do not have any such time bin
were defined as "non-downstream" cells.
According to the evoked response latencies, we identified the short- (< 10 ms) and long-latency (≥ 10 ms) downstream
cells for each stimulated neuron. The multi-synaptic downstream cells for a stimulated neuron comprised relatively small
fraction n (17% on average) of the entire population. This indicates that they are subsets of neurons receiving effectively
strong input from the stimulated cell via relatively a small number of path length, although the interaction with longer path-
length, which was not detected here, could include the larger fraction of the cell population.
Causal network analysis
Pairwise causal interaction among neurons was estimated based on a variant of convergent cross-mapping (CCM). CCM
was developed by Sugihara et al. (26) as an extension of nonlinear forecasting method of time sequence based on nearest-
neighbor models (24, 25, 49). The method is capable of detecting relatively weak causal interactions in deterministic
dynamical systems (which can include some stochastic components) if the system and observations are deterministic, the
6
embedding dimension is sufficiently large, and data size is sufficiently large for the given embedding dimension. A variant
of CCM that can be used for systems, in which the variables have heterogeneous time-scales (such as in neural system),
was developed by the authors (27).
Suppose that we want to know the interaction from neuron x to another neuron y. We first reconstruct the state-space
𝑑𝑚𝑎𝑥 = (𝑥𝑡, 𝑥𝑡−𝜏, … , 𝑥𝑡−(𝑑𝑚𝑎𝑥−1)𝜏), where 𝑑𝑚𝑎𝑥 represents the
trajectories of each neuron 𝑥 in the delay-coordinates, 𝒙𝑡
maximum number of dimensions (number of delay-coordinates) to be considered, 𝑡 is the time point, and 𝜏 is the unit delay
length. We used 𝜏 = 5 ms dmax = 8 in this study. In the embedding-based analyses, we convolved the firing-rate sequence
of each neuron with Gaussian kernel that has half-width-of-half-height of 25 ms, and normalized such that each neuron
trace ranges from 0 to 1. To avoid a known vulnerability of the state-space reconstruction to the time-scale heterogeneity
𝑑𝑚𝑎𝑥 to a randomized coordinate space by multiplying a square random matrix, 𝑹, from
(50), we projected delay vector 𝒙𝑡
𝑑 is constructed by selecting
the left to obtain a transformed vector: 𝒙𝑡
𝑑𝑚𝑎𝑥. The causal interaction form neuron x to another neuron y is quantified based
the first 𝑑 (≤ 𝑑𝑚𝑎𝑥) components of 𝒙𝑡
on the correlation coefficient, 𝜌𝑦𝑡,𝑦(𝒙𝑡
𝑑), between the true (𝑦𝑡) and forecast (𝑦𝑡(𝒙𝑡
𝑑𝑚𝑎𝑥. A 𝑑-dimensional delay vector 𝒙𝑡
𝑑)) signals, where
𝑑𝑚𝑎𝑥 = 𝑹 𝒙𝑡
𝑦𝑡(𝒙𝑡
𝑑) = ∑ 𝑤(𝒙𝑡′
𝑑 − 𝒙𝑡
𝑑) 𝑦𝑡′
𝑡′ s.t. 𝒙𝑡′
𝑑)
𝑑 ∈𝐵(𝒙𝑡
with the k-nearest neighbor set 𝐵(𝒙𝑡
𝑑) of 𝒙𝑡
𝑑 in the delay-coordinate space. We set 𝑘=4, weight,
𝑤(𝒙𝑡′
𝑑 − 𝒙𝑡
𝑑) =
exp(−𝒙𝑡′
𝑑 − 𝒙𝑡
𝑑)
𝑑 − 𝒙𝑡
𝑑) exp(−𝒙𝑡′
,
𝑑)
𝛴𝑡′ s.t. 𝒙
𝑑 ∈𝐵(𝒙𝑡
𝑡′
𝑑
𝑑 and 𝒙𝑡
in the data analysis. Note that 𝜌𝑦𝑡,𝑦(𝒙𝑡
𝑑 − 𝒙𝑡
𝑑 as the square distance between 𝒙𝑡′
and 𝒙𝑡′
prediction. We selected the optimal dimension d=d* so as to maximize the prediction performance, 𝜌𝑦𝑡,𝑦(𝒙𝑡
of d* was distributed from 2 to 4 in the current data.
𝑑) = 1 means perfect
𝑑). Typical value
Trace-back embedding method and burst predictability
The trace-back embedding (TBE) method is an extension of the above CCM algorithm. The method extends the original
CCM (i) by relating the global state to local components while the original protocol only quantified the relationships
between local variables, and (ii) by quantifying on the predictability of temporally distant global state based on local
variable dynamics. This is done by finding a mapping between the system's global state bt and each neuron 𝑥𝑡−𝛥𝑡 including
temporal lag 𝛥t, instead of simultaneous prediction among individual neurons. For burst detection, we set t as the time
corresponding to the peak of global state during a burst period, 𝑡 ∈ {𝑡1, … , 𝑡𝑛}, where n is the number of burst trials in the
experiment. We first reconstruct the state-space trajectory for neuron x and the normalized global state b in the
d = (bt, bt-τ …, bt-(d-1)τ). It is mathematically shown that both
(randomized) delay-coordinates, xt
of these trajectories reproduce the topological structure (i.e., preserving continuity but allowing deformations) of attractor
dynamics that they participate, with which a (near-) future state of each variable is accurately predicted from the current
state of other variables, if the dynamics (approximately) follows deterministic mechanisms (22, 23, 26, 27). Practically,
however, the accuracy of prediction is limited by various factors including data size, process/observation noises, as well as
the uncertainty due to asymmetric causal interactions.
d = (xt, xt-τ …, xt-(d-1)τ) and bt
The forecast of the global state 𝑏
𝑡 based on a single neuron x with time lag 𝛥𝑡 is given by
𝑏
𝑡 (𝒙𝑡−𝛥𝑡
𝑑
) =
∑
𝑑
𝑤(𝒙𝑡′−𝛥𝑡
𝑑
− 𝒙𝑡−𝛥𝑡
) 𝑏𝑡′
𝑑
𝑡′ s.t. 𝒙
𝑡′−𝛥𝑡
𝑑
∈𝐵(𝒙𝑡−𝛥𝑡
)
with the k-nearest neighbor set 𝐵(𝒙𝑡−𝛥𝑡
same weight function,
𝑑
) of 𝒙𝑡−𝛥𝑡
𝑑
, in the same manner s in CCM. In the data analysis we used k=4 and the
𝑑
𝑤(𝒙𝑡′−𝛥𝑡
𝑑
− 𝒙𝑡−𝛥𝑡
) =
exp(−𝒙𝑡′−𝛥𝑡
𝑑
𝑑
− 𝒙𝑡−𝛥𝑡
) exp(−𝒙𝑡′−𝛥𝑡
𝑑
𝛴𝒙
𝑑
𝑡′−𝛥𝑡
𝑑
∈𝐵(𝒙𝑡−𝛥𝑡
)
− 𝒙𝑡−𝛥𝑡
𝑑
.
)
The success rate in burst detection was defined by the ratio between the number of burst trials such that 𝑏
𝑡𝑛 > 𝜃𝑥, where
𝜃𝑥 is the threshold of burst detection for neuron x. In the present study, we defined 𝜃𝑥 by the 95 percentile in the null-
hypothesis distribution P(𝑏
)), where 𝑡Null is randomly selected time outside the bursting period. The burst-
size predictability in each cell was defined by correlation between actual and predicted burst sizes, average over range of
time span -200 ms < Δt < -50 ms.
𝑡Null (𝒙𝑡Null−𝛥𝑡
𝑑
7
Similarly, the forecast based on the history of the global state itself is given by the same protocol except for using b instead
of x:
𝑏
𝑡 (𝒃𝑡−𝛥𝑡
𝑑
) =
∑
𝑑
𝑤(𝒃𝑡′−𝛥𝑡
𝑑
− 𝒃𝑡−𝛥𝑡
) 𝑏𝑡′
.
𝑑
𝑡′ s.t. 𝒃
𝑡′−𝛥𝑡
𝑑
∈𝐵(𝒃𝑡−𝛥𝑡
)
This quantifies the self-predictability of burst based on the global state itself, through the success rate compute based on
the corresponding threshold 𝜃𝑏, which in the present study was defined by 95 percentile of null-hypothesis distribution
P(𝑏
)).
𝑡Null (𝒃𝑡Null−𝛥𝑡
𝑑
Acknowledgements
The authors are grateful to Dr. Urs Frey at RIKEN Quantitative Biology Center, Kobe, Japan, and Prof. Andreas
Hierlemann at ETH Zurich, Basel, Switzerland, for providing CMOS-based MEA with their technical supports. This work
was supported by JST PRESTO (ST), Asahi Glass Foundation, and KAKENHI (26630089), Brain/MINDS from AMED
(TT), and RIKEN Brain Science Institute (TT).
Conflict of interests
The authors declare no competing financial interests.
Author contributions
Designed research: ST, TM, HT, TT
Performed research: -
Contributed unpublished reagents/analytic tools: TM, DJB, HT
Analyzed data: ST, TT
Wrote the paper: ST, TM, HT, TT
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Chialvo DR (2010) Emergent complex neural dynamics. Nat Phys 6(10):744–750.
Lisman JE (1997) Bursts as a unit of neural information: Making unreliable synapses reliable. Trends Neurosci
20(1):38–43.
Izhikevich EM, Desai NS, Walcott EC, Hoppensteadt FC (2003) Bursts as a unit of neural information: Selective
communication via resonance. Trends Neurosci 26(3):161–167.
Traub RD, Wong RK (1982) Cellular mechanism of neuronal synchronization in epilepsy. Science (80- ) 216:745–
747.
McCormick D, Contreras D (2001) On the cellular and network bases of epileptic seizures. Annu Rev Physiol
63:815–846.
Staley KJ, Dudek FE (2006) Interictal Spikes and Epileptogenesis. Epilepsy Curr 6(6):199–202.
Feinerman O, Segal M, Moses E (2007) Identification and dynamics of spontaneous burst initiation zones in
unidimensional neuronal cultures. J Neurophysiol 97(4):2937–2948.
Eckmann JP, Jacobi S, Marom S, Moses E, Zbinden C (2008) Leader neurons in population bursts of 2D living
neural networks. New J Phys 10. doi:10.1088/1367-2630/10/1/015011.
Yang H, Shew WL, Roy R, Plenz D (2012) Maximal variability of phase synchrony in cortical networks with
neuronal avalanches. J Neurosci 32(3):1061–72.
Johnston D, Brown T (1981) Giant synaptic potential hypothesis for epileptiform activity. Science (80- ) 211:294–
297.
Bush PC, Douglas RJ (1991) Synchronization of bursting action potential discharge in a model network of
neocortical neurons. Neural Comput 3(1):19–30.
12.
Pinsky PF, Rinzel J (1994) Intrinsic and network rhythmogenesis in a reduced Traub model for CA3 neurons. J
8
Comput Neurosci 1(1–2):39–60.
13.
Li C-YT, Poo M-M, Dan Y (2009) Burst spiking of a single cortical neuron modifies global brain state. Science
(80- ) 324(5927):643–6.
14. Miles R, Wong RK (1983) Single neurones can initiate synchronized population discharge in the hippocampus.
Nature 306(5941):371–373.
15.
Takahashi H, et al. (2012) Light-addressed single-neuron stimulation in dissociated neuronal cultures with sparse
expression of ChR2. BioSystems 107(2):106–112.
16.
Okun M, et al. (2015) Diverse coupling of neurons to populations in sensory cortex. Nature 521:511–515.
17.
18.
19.
Nowotny T, Rabinovich M (2007) Dynamical origin of independent spiking and bursting activity in neural
microcircuits. Phys Rev Lett 98(12):128106:1-4.
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(22):5216–29.
Bakkum DJ, et al. (2013) Tracking axonal action potential propagation on a high-density microelectrode array
across hundreds of sites. Nat Commun 4:2181.
20.
Bakkum DJ, et al. (2013) Parameters for burst detection. Front Comput Neurosci 7(January):193:1-12.
21. Wagenaar D, Demarse TB, Potter SM (2005) MeaBench: A toolset for multi-electrode data acquisition and on-line
analysis. 2nd Int IEEE EMBS Conf Neural Eng 2005:518–521.
22.
Takens F (1981) Detecting strange attractors in fluid turbulence. Dynamical Systems and Turbulence, Lecture
Notes in Mathematics, eds Rand DA, Young L-S (Springer-Verlag, Berlin), pp 366–381.
23.
Sauer T, Yorke JA, Casdagli M (1991) Embedology. J Stat Phys 65(3/4):579–616.
24.
25.
Sugihara G, May R (1990) Nonlinear forecasting as a way of distinguishing chaos from measurement error in time
series. Nature 344(6268):734–741.
Tsonis AA, Elsner JB (1992) Nonlinear prediction as a way of distinguishing chaos from random fractal sequences.
Nature 358:217–220.
26.
Sugihara G, et al. (2012) Detecting causality in complex ecosystems. Science (80- ) 338(6106):496–500.
27.
28.
29.
Tajima S, Yanagawa T, Fujii N, Toyoizumi T (2015) Untangling brain-wide dynamics in consciousness by cross-
embedding. PLOS Comput Biol 11(11):e1004537.
Orlandi JG, Soriano J, Alvarez-Lacalle E, Teller S, Casademunt J (2013) Noise focusing and the emergence of
coherent activity in neuronal cultures. Nat Phys 9(9):582–590.
Shew WL, Yang H, Yu S, Roy R, Plenz D (2011) Information capacity and transmission are maximized in balanced
cortical networks with neuronal avalanches. J Neurosci 31(1):55–63.
30.
Hahn G, et al. (2010) Neuronal avalanches in spontaneous activity in vivo. J Neurophysiol 104(6):3312–22.
31.
Shriki O, et al. (2013) Neuronal avalanches in the resting MEG of the human brain. J Neurosci 33(16):7079–90.
32.
Beggs JM, Plenz D (2003) Neuronal Avalanches in Neocortical Circuits. J Neurosci 23(35):11167–11177.
33.
34.
35.
36.
37.
Pinto DJ, Patrick SL, Huang WC, Connors BW (2005) Initiation, propagation, and termination of epileptiform
activity in rodent neocortex in vitro involve distinct mechanisms. J Neurosci 25(36):8131–40.
Ikegaya Y, et al. (2004) Synfire chains and cortical songs: temporal modules of cortical activity. Science (80- )
304(5670):559–64.
Luczak A, Barthó P, Harris KD (2009) Spontaneous events outline the realm of possible sensory responses in
neocortical populations. Neuron 62(3):413–25.
Eytan D, Marom S (2006) Cellular/Molecular Dynamics and Effective Topology Underlying Synchronization in
Networks of Cortical Neurons. J Neurosci 26(33):8465–8476.
Elger CE, Lehnertz K (1998) Seizure prediction by non-linear time series analysis of brain electrical activity. Eur
J Neurosci 10(1997):786–789.
38.
Golder S (2010) Tweet , Tweet , Retweet : Conversational Aspects of Retweeting on Twitter.
39.
40.
Ishii A, et al. (2012) The "hit" phenomenon: a mathematical model of human dynamics interactions as a stochastic
process. New J Phys 14(6):63018.
Zhou J, Liu Z, Li B (2007) Influence of network structure on rumor propagation. Phys Lett Sect A Gen At Solid
State Phys 368(6):458–463.
41.
Gross T, D'Lima CJD, Blasius B (2006) Epidemic dynamics on an adaptive network. Phys Rev Lett 96(20):1–4.
9
42.
43.
Pastor-Satorras R, Vespignani A (2001) Epidemic spreading in scale-free networks. Phys Rev Lett 86(14):3200–
3203.
Burns SP, et al. (2014) Network dynamics of the brain and influence of the epileptic seizure onset zone. Proc Natl
Acad Sci U S A. doi:10.1073/pnas.1401752111.
44. Wang Y, Goodfellow M, Taylor PN, Baier G (2012) Phase space approach for modeling of epileptic dynamics.
Phys Rev E 85(6):61918.
45.
46.
47.
48.
49.
50.
Stam CJ (2005) Nonlinear dynamical analysis of EEG and MEG: review of an emerging field. Clin Neurophysiol
116(10):2266–301.
Lehnertz K, Elger CE (1995) Spatio-temporal dynamics of the primary epileptogenic area in temporal lobe epilepsy
characterized by neuronal complexity loss. Electroencephalogr Clin Neurophysiol 95(2):108–17.
Frey U, et al. (2010) Switch-matrix-based high-density microelectrode array in CMOS technology. IEEE J Solid-
State Circuits 45(2):467–482.
Livi P, Heer F, Frey U, Bakkum DJ, Hierlemann A (2010) Compact voltage and current stimulation buffer for
high-density microelectrode arrays. IEEE Trans Biomed Circuits Syst 4(6 PART 1):372–378.
Lorenz EN (1969) Atmospheric predictability as revealed by naturally occurring analogues. J Atmos Sci 26:636–
646.
Kennel M, Abarbanel H (2002) False neighbors and false strands: A reliable minimum embedding dimension
algorithm. Phys Rev E 66(2):26209.
10
Figure 1.
Global bursts in cultured neural populations.
(a) CMOS-based recording apparatus.
(b) Spatial distribution of cells recorded in a representative preparation, Chip 1440.
(c) Spontaneous activity of neurons in the same preparation, Chip 1440. The color indicates the spike count within each 5-
ms time bin (white: 0 spikes; black >5 spikes). The abscissa represents the time. The majority of neurons fired together
within network bursts (columns of dark-colored bins).
(d) The temporal evolution of global mean field (normalized population firing rate) computed from c. The horizontal dashed
line indicates the threshold for burst detection for this preparation. (Inset) The distributions of inter-spike intervals (ISIs)
in five different preparations (solid lines) and the burst detection thresholds for the individual preparations (dashed lines).
The horizontal axis shows the normalized log-ISI (the logarithm of inverse global-activity; see Methods).
11
Figure 2.
Trace-back embedding (TBE) method in delay coordinates relates the global population activity to the local
activities.
(a) Delay reconstruction of the mean-field dynamics ("global states").
(b, c) Delay reconstructions of single neuron activities ("local states"). Two representative neurons (Cells 1 and 2) are
shown. Blue dots: the time points of peak global activities during the bursts; Red dots: the temporally preceding states from
the individual global activity peaks (here, shift time = -100 ms); magenta dots: the neighboring states within the
reconstructed state space for Cell 1 (10 nearest neighbors for each burst peak); cyan pluses: predicted states based on TBE
method. The protocol of TBE is as follows: (i) selecting a time point t on a local state trajectory (say, of neuron i) that
corresponds to a peak population activity found in the global state trajectory (which is an target bursting state to be
predicted), (ii) tracing-back the local state trajectory for a given time span, Δt, (iii) collecting "neighbor" time-points
corresponding to states nearby t-Δt in the local state trajectory (avoiding temporally too near points), (iv) mapping the
neighbor time-points back into the global state trajectory, and (v) proceeding the time with Δt on the global state trajectory
to examine whether the finally obtained states (cyan pluses) cluster within bursting state and how accurately they replicate
the target bursting state. The accuracy of burst prediction is higher when the traced-back states (red dotes) deviates more
from the ones in non-bursting period in the local trajectory. Note that the tracing time in the retrograde direction with Δt
(step ii) and mutual mapping between global and local states (steps i and iv) are newly introduced features in this study,
which extends the previous methods (24–26).
12
a
Figure 3.
Presages of global burst within single neurons.
(a) Burst detection with the TBE method based on a single representative neuron. The black traces represent global states
in individual trials around burst peaks. The blue curve is the burst detection threshold for each detection time-span, Δt;
global state exceeding this threshold provides burst predictor (or hallmarks of bursts after their occurrence). The threshold
(dashed green line) was defined by 95 percentile of global activity outside bursting period. The red curve shows the fraction
("hit rate") of successfully detected bursts in all the burst trials.
(b) Burst detection with TBE method based on the mean-field activity in the population.
(c) Burst detection with the momentary firing rate of the same representative single neuron as the one shown in panel a.
(d) Burst detection with the momentary mean-field firing-rate.
(e) Burst predictability is a robust property of individual neurons. Dividing the data into two halves (the odd and even trials
of spontaneous burst) shows that variability in burst-detection success rate is highly consistent over time. Each marker
corresponds to a neuron. The filled markers represent the results using the mean field activity. The success rate was
averaged over the range of time span -200 ms < Δt < -50 ms.
(f) Burst detection with TBE method is more accurate than that based on the momentary firing rate. The comparison of
success rates in burst detection with the TBE method and with the firing-rate-based method. Inset: the control analysis in
which we compared the success rates of the TBE method with that of the firing-rate-based method using multiple time bins,
by matching the number of bins to that used in the TBE method. The gray bars show the fraction of cells in all the
preparations. The arrowheads indicate the medians. The colored lines show the fractions in the individual preparations.
(g) Single cell can outperform the mean field at predicting the spontaneous network bursts, particularly when we utilize
the information in the temporal patterns of neural activity rather than the momentary activity. The light and dark blue
curves respectively show the TBE- and firing-rate-based success rates in burst detection with single neurons relative to that
with the mean field. For each method, the cells were sorted based on the success rate.
13
Figure 4.
The network structures in non-bursting periods explains the local burst predictability.
(a) Causal network analysis. The network depicts the directed pairwise interactions among neurons. The filled (orange)
and open (blue) markers represents the putative excitatory and inhibitory cells, respectively. The thickness of arrows
indicates the absolute strength of causal coupling ("causality"). For clarity of illustration, only the top 5% strong couplings
are shown, and the nodes are distributed by multi-dimensional scaling with the distances defined to be inversely
proportional to the causality. The figure shows the result for Chip 1440.
(b) Synaptically connected neuron pairs show larger causal interaction. The bar labels show the time span to be used for
the causality analysis. Bars in dark and light blue: average cross-embedding values for cells that showed short-latent (dark,
latency < 10 ms) and long (dark, latency > 10 ms) stimulus-triggered spike increase. The average causality index within
time span [-100, 100] ms dissociated the post-synaptic cells from other cells (P<0.005, Wilcoxon sign rank test,
independent samples). The error bar shows the s.e.m across neurons.
(c) Identification of synaptic connectivity by electrical stimulation experiment. The figure shows the peri-stimulus time
histograms of three representative neurons (from top to bottom, putative direct, indirect, and non-post cells, respectively).
The asterisks indicate the earliest significant increase (P>0.05) of spike count, compared to the spike count distribution in
the pre-stimulus period (from -500 to -100 ms). The figure shows the spike histogram smoothed with a boxcar kernel of 5-
ms width for visualization purpose.
(d) The relationship between burst predictability and the interaction strength in different neuron types (putative excitatory
and inhibitory neurons). The neurons were classified into two groups depending on the strength of input or output causal
couplings: the top half of neurons was labeled as "strong" whereas the bottom half was labeled as "weak." The error bar
shows the s.e.m across neurons.
(e) The same as panel d, but shows the individual neurons (the dots) with a 3D plot. The surfaces represent the fitted
piecewise linear functions for each neuron types. The color conventions follow that of panel d.
14
Supplementary Figure S1.
The population activity dynamics visualized with principal component analysis (PCA).
The trajectories of population activity were plotted within the space of top three principal components (PCs 1–3), by
applying PCA to the multi-neuron time series for each preparation. The results for three representative preparations are
shown.
15
Supplementary Figure S2.
The delay-embedding theorem (known as Takens' theorem).
Suppose that we have the time-series of three variables, (𝑥(𝑡), 𝑦(𝑡), 𝑧(𝑡)) , generated by some differential
equations: 𝑥 (𝑡) = 𝑓(𝑥(𝑡), 𝑦(𝑡), 𝑧(𝑡)), 𝑦 (𝑡) = 𝑔(𝑥(𝑡), 𝑦(𝑡), 𝑧(𝑡)) and 𝑧(𝑡) = ℎ(𝑥(𝑡), 𝑦(𝑡), 𝑧(𝑡)).
(a) The evolution of the global state (𝑥(𝑡), 𝑦(𝑡), 𝑧(𝑡)) is represented by a trajectory in the space of those three variables.
(b) Consider observing the temporal sequence of a single variable, e.g., 𝑥(𝑡).
(c) The attractor topology in the original global state space is fully recovered in a delay-coordinate of the observed variable,
(𝑥(𝑡), 𝑥(𝑡 − 𝜏), 𝑥(𝑡 − 𝜏)) with an arbitrary unit delay, 𝜏. Namely, we can construct a smooth one-to-one map from the
reconstructed attractor to the original attractor.
16
Supplementary Figure S3.
Identification of Excitatory and Inhibitory Neurons
(a) Representative neurons in immunostaining with MAP2 and GABA. Action potentials below the insets were obtained
at white rectangles, putatively from a neighboring neuron in a circle. The peak-to-peak time, Tpp, was defined as time
duration from negative peak to positive peak of action potential.
(b) Histogram of Tpp. Excitatory neurons (green) had larger Tpp than inhibitory neurons (magenta). K-means method to
Tpp was employed to separate excitatory and inhibitory neurons.
17
Supplementary Figure S4.
Microstimulation-based estimation of synaptic connectivity in a pair-wise manner.
(a) Axonal stimulation on an arbitrary neuron elicited bidirectional action potential propagation.
(b) Raw data of neural responses at a putative pre-synaptic neuron and post-synaptic neuron. Data from one hundred trials
are superimposed.
(c) Raster plot of (b). Antidromic, direct action potentials exhibited precise temporal response, while orthodromic, synaptic
action potentials were elicited stochastically with significant temporal jitters.
(d) Post-stimulus spike histograms of (c)
18
|
1708.00540 | 1 | 1708 | 2017-08-01T22:42:32 | A silicon-based microelectrode array with a microdrive for monitoring brainstem regions of freely moving rats | [
"q-bio.NC"
] | Objective. Exploring neural activity behind synchronization and time locking in brain circuits is one of the most important tasks in neuroscience. Our goal was to design and characterize a microelectrode array (MEA) system specifically for obtaining in vivo extracellular recordings from three deep-brain areas of freely moving rats, simultaneously. The target areas, the deep mesencephalic reticular-, pedunculopontine tegmental- and pontine reticular nuclei are related to the regulation of sleep-wake cycles. Approach. The three targeted nuclei are collinear, therefore a single-shank MEA was designed in order to contact them. The silicon-based device was equipped with 3*4 recording sites, located according to the geometry of the brain regions. Furthermore, a microdrive was developed to allow fine actuation and post-implantation relocation of the probe. The probe was attached to a rigid printed circuit board, which was fastened to the microdrive. A flexible cable was designed in order to provide not only electronic connection between the probe and the amplifier system, but sufficient freedom for the movements of the probe as well. Main results. The microdrive was stable enough to allow precise electrode targeting into the tissue via a single track. The microelectrodes on the probe were suitable for recording neural activity from the three targeted brainstem areas. Significance. The system offers a robust solution to provide long-term interface between an array of precisely defined microelectrodes and deep-brain areas of a behaving rodent. The microdrive allowed us to fine-tune the probe location and easily scan through the regions of interest. | q-bio.NC | q-bio | A robust silicon-based microelectrode array, mounted on a microdrive
A silicon-based microelectrode array with a microdrive for
monitoring brainstem regions of freely moving rats
G Márton1,2, P Baracskay3, B Cseri3, B Plósz4, G Juhász3, Z Fekete2,5, A Pongrácz2,5
1
Comparative Psychophysiology Department, Institute of Cognitive Neuroscience and
Physiology, Research Centre for Natural Sciences, Hungarian Academy of Sciences,
2 Magyar Tudósok Blvd., H-1117, Budapest, Hungary
2
MEMS Laboratory, Institute for Technical Physics and Materials Science, Centre for
Energy Research, Hungarian Academy of Sciences, 29-33 Konkoly Thege Miklós st.,
H-1121, Budapest, Hungary
Research Group of Proteomics, Institute of Biology, Faculty of Science, Eötvös Loránd
University, 1/C Pázmány P. Walkway, H-1117, Budapest, Hungary
Plósz Microengineering Office Ltd., 119 Üllői st., H-1091, Budapest, Hungary
MTA EK NAP B Research Group for Implantable Microsystems, 29-33 Konkoly Thege
Miklós st., H-1121, Budapest, Hungary
3
4
5
E-mail: [email protected]
1
A robust silicon-based microelectrode array, mounted on a microdrive
Abstract
Objective. Exploring neural activity behind synchronization and time locking in brain circuits is one of
the most important tasks in neuroscience. Our goal was to design and characterize a microelectrode
array (MEA) system specifically for obtaining in vivo extracellular recordings from three deep-brain
areas of freely moving rats, simultaneously. The target areas, the deep mesencephalic reticular-,
pedunculopontine tegmental- and pontine reticular nuclei are related to the regulation of sleep-wake
cycles.
Approach. The three targeted nuclei are collinear, therefore a single-shank MEA was designed in order
to contact them. The silicon-based device was equipped with 3×4 recording sites, located according to
the geometry of the brain regions. Furthermore, a microdrive was developed to allow fine actuation and
post-implantation relocation of the probe. The probe was attached to a rigid printed circuit board, which
was fastened to the microdrive. A flexible cable was designed in order to provide not only electronic
connection between the probe and the amplifier system, but sufficient freedom for the movements of
the probe as well.
Main results. The microdrive was stable enough to allow precise electrode targeting into the tissue via
a single track. The microelectrodes on the probe were suitable for recording neural activity from the
three targeted brainstem areas.
Significance. The system offers a robust solution to provide long-term interface between an array of
precisely defined microelectrodes and deep-brain areas of a behaving rodent. The microdrive allowed
us to fine-tune the probe location and easily scan through the regions of interest.
Keywords: Brainstem reticular formation; Freely moving rats; Single unit recording; Silicon-based
probe; extracellular microelectrode array; neural; microdrive
1. Introduction
Recording electrical activity in different brain areas simultaneously is increasingly attracting the
attention of the neuroscience community [1-3], which creates new approaches in the system design of
implantable devices as well. One of the main aims of researchers is to record as many neurons as
possible from different brain areas, searching for synchrony and time-locking of activities. The
applicable number of electrodes is limited by the tissue injury made by a certain electrode construction
and the size and distribution of recording sites. Up to now, plenty of different electrode designs have
been introduced such as multiple microwire arrays or bundles [4-7], micromachined silicon-based
microelectrode arrays (MEAs) [8], polymer- [9], ceramic-based [10] and titanium [11] probes. When
extracellular electrodes are chronically used, signal detection is hindered by glial scar formation around
the implants [12, 13]. This is one of the reasons why it is advantageous if the implanted probes can be
precisely moved postoperatively. This function can be enabled with the use of microdrives [14, 15].
Modern silicon-based microtechnology, fuelled by the electronic industry allows precise batch
fabrication of microelectromechanical systems (MEMS) with reproducible features and small
dimensions, including versatile and cost-effective ways of making neural MEAs [16]. Numerous
electrodes can be placed at different points along each probe, moreover, they offer the potential of
integrating signal processing circuits on them. Silicon, along with the usually applied silicon dioxide
2
A robust silicon-based microelectrode array, mounted on a microdrive
and silicon nitride insulator thin films [17], is a biocompatible material [18, 19]. Silicon can be tailored
e.g. with deep reactive ion etching (DRIE) in order to form versatile probe geometries [20]. Thin-film
microelectrodes, located on the silicon probes are suitable for in vivo local field potential and unit
recordings [21-23]. They are widely used in freely moving animals for measurements from different
brain regions, such as the neocortex or the hippocampus [24, 25]. Silicon-based microtechnology allows
one to equip neural probes with delivery channels for injecting chemical compounds close to the active
recording sites [26-28] or optical waveguides for optogenetic experiments [29]. According to these
considerations, silicon is an excellent candidate to be substrate material for neural MEAs. However,
mechanical properties of silicon microstructures are not always satisfactory. Not only the shafts of the
probes can break, but cracking and degradation of the electric wires can occur. Such failures were
detected in 15 µm thick MEAs, chronically implanted into mice [30]. Failures due to poor mechanical
properties might be the main reason why silicon-based MEAs are hardly used in deep-brain regions
even in studies involving rodents (with some exceptions [31]), while mechanical strains can be larger
in case of higher mammals [30]. Recent studies revealed that silicon probes with "ultra-long" shafts
can be successfully fabricated [32, 33], and these reinforced or relatively thick devices might be able to
withstand larger forces [34]. The wide-spread application of ultra-long" silicon-based MEAs for
discovering deep-brain regions might contribute to the progress of neurosciences.
Our neuroscientific aim was the simultaneous recording of neural activity from three nuclei of the
brainstem reticular formation: the deep mesencephalic reticular nucleus (DpMe), the pedunculopontine
tegmental nucleus (PPTg) and the oral part of the pontine reticular nucleus (PnO).
The targeted structures are part of the ascending reticular activating system (ARAS) and play a key role
in the regulation of different functional brain states such as the sleep-wake cycle, hypersynchronization,
desynchronization and arousal through integrating and transmitting different sensory inputs from the
spinal cord and information from other brainstem areas [35, 36]. The ARAS consists of neuron-poor
nuclei and a gigantocellular neural networks, called the reticular formation. Although these nuclei seem
separate, many of them are overlapping and they are technically one anatomical unit. These areas give
ascending projections to the cortex, the basal ganglia, and the thalamus, ascending descending
connections to the spinal chord, and they are richly interconnected with each other [37]. Selective lesion
of the PnO induces chronic comatic state, thus the PnO likely acts as an activating centre [36].
Because of the rich anatomical interconnections [38, 39], investigating functional correlation of
neuronal activity from these three nuclei is an intriguing question. Extracellular neural activity of the
selected areas has separately been measured in head-restrained or anaesthetized animals [40-44]. Nuclei
of the ascending reticular activating system were targeted by a number of experiments using microwires
in freely moving cats and rats [43, 45, 46]. However, these results are restricted to one area at a time.
To the authors' knowledge, no simultaneous extracellular unit recordings from several reticular
formation areas in freely moving rats have been presented. Insulated wire electrodes are frequently
employed for recordings from other deep brain areas as well [47].
In contrast with these techniques, we have decided to create a probe utilizing the above described
silicon-based MEMS technology, in order to interface with the deep brain structures in our interest. We
expected that the silicon shaft would allow sufficient, more precise targeting than microwires,
moreover, by implanting a specific formation of custom-designed array of electrodes at once, instead
of managing individual wire electrodes or bundles of them, time and effort could be saved. The brain
structures of interest are collinear, so we decided to create a single-shaft MEA with electrodes custom-
3
A robust silicon-based microelectrode array, mounted on a microdrive
arranged into three groups. We also aimed to construct a microdrive in order to finely relocate the sensor
and find spots for successful electrophysiological recordings, containing unit activities. Another trivial
way of interfacing with the three nuclei would have been the utilization three individual MEAs.
Mounting a microdrive system on the skull of a rat, which is suitable for handling the shafts of three
probes is probably a challenging, yet not impossible task. That would have allowed more freedom in
positioning, but caused more tissue damage and cost more preparation time and resources.
In summary, this work focuses on a system, with the help of which simultaneous recording of unit
activities of three deep brain structures was successfully performed in behaving rats for the first time.
It will describe the measurement methods, the neurobiological analysis of the obtained results exceeds
its scope.
2. Materials and methods
2.1. Probe design
A series of masks have been designed to fit the electrode array geometry to the anatomy of the brain
regions we proposed to investigate. Accordingly, the total length of the probe is 30 mm, the shaft length
is 25 mm, and the shaft thickness is the 200 μm initial thickness of the Si-wafer. The width of the probe
gradually increases from 60 μm up to 130 μm along the recording sites, and up to 400 μm at the base
of the probe to keep the robustness of the design. The distance between the tip and the middle of the
first recording site is 92 μm. The tip angle is 19°. This design contains square-shaped (12 m x 12 m)
platinum recording electrodes in three groups, located up to 3.2 mm from the probe tip. One group
consists of 4 electrodes with 150 µm site spacing, and the groups are 780 µm distant from each other.
The output platinum leads are 4 µm wide and 300 nm thick.
2.2. Microfabrication
The fabrication technology of the multielectrodes is based on a single-side, three-mask bulk
micromachining process performed in three phases: (a) bottom passivation layer deposition and lift-off
process for the formation of platinum electrodes, output leads and bonding pads, (b) top passivation
layer deposition and patterning, (c) etching of the passivation layers and (d) deep reactive ion etching
(DRIE) of silicon to form the probe body. The overall process flow is summarized in figure 1.
4
A robust silicon-based microelectrode array, mounted on a microdrive
Figure 1. Cross-sectional schematic diagram of the main fabrication steps. (a) SiO2 and SixNy layer
deposition and patterning of platinum wiring. (b) Deposition and selective opening of the SixNy
passivation layer. (c) Patterning of front-side Al mask and dry etching of the passivation and insulation
layers; evaporation of Al and spin-coating of photoresist to form the backside etch-stop layer. (d) Cross-
section of the ready-made probe after removing the masking and etch-stop layers.
2.2.1. Electrode array formation
200 μm thick, double-side polished (100) oriented 4-inch silicon wafers are used for probe fabrication.
In the first step a 100 nm thick SiO2 layer was thermally grown, followed by deposition of a 500 nm
thick LPCVD low-stress silicon nitride. The metal layer was then deposited and patterned by lift-off
process. The lift-off structure consisted of 1.8 μm thick photoresist (Microposit 1818) layer over
patterned 500 nm thick Al thin film. The metal layer consisted of a 15 nm thick adhesion layer of TiOx
and Pt. TiOx was formed by reactive sputtering of Ti in O2 (Ar/O2 ratio was 80:20) atmosphere. In the
same vacuum cycle 270 nm thick Pt was sputtered on top of TiOx. The lift-off was accomplished by
dissolving the photoresist pattern in acetone, this process was optimized by using water-cooled substrate
holder which diminished the resist distortion during TiOx/Pt sputtering. Subsequently, the removal of
Al patterns in a fourcomponent etching solution and low pressure chemical vapor deposition of 300 nm
thick SiNx insulating layer stack were performed. Contact holes were etched through the SiNx layer to
unblock the electrical recording sites at the probe tip and the pads at the base for wire bonding.
2.2.2 Probe shaping
A 500 nm thick Al layer was deposited on the front side and patterned by photolithography using 4 µm
thick SPR220 photoresist. The contour of the probe body was defined by Al wet etching. SPR220
photoresist was spun on the backside of the wafer acting as a stopping layer during the subsequent deep
reactive ion etching of silicon. The 3D micromachining process was performed in an Oxford Plasmalab
System 100 chamber using Bosch process.
2.3. Packaging
Rigid-flexible PCBs have been formed in order to provide sufficient support for the silicon probes and
enough freedom for the movements of the microdrive. The deformability of this component is illustrated
in figure 2. The substrate of its rigid parts was made of 1-mm thick FR-4. The substrate of the flexible
cable was manufactured from 50 µm thick polyimide. A conductor layer of copper has been formed on
5
A robust silicon-based microelectrode array, mounted on a microdrive
the top surface, plated with gold for the ease of wire bonding. The upper insulator layer on the PCB was
made of flexible solder-resist mask. The geometry of the device has been designed as follows. A 6 mm
x 9 mm rigid part was shaped for the support of the probe. A hole of 2 mm in diameter was designed
for fixing the PCB with a screw onto microdrive. The L-shaped flexible cable was 5 mm wide and 19.1
mm long at its midline. It provided electronic connection via 200-µm thick wires between the probe
support and the Preci-Dip connector part of the PCB. The connector part was rigid, suitable for soldering
a 12-pin section of a 90-degree angle pin connector (Preci-Dip 850-PP-N-050-20-001101).
Figure 2. The rigid-flexible PCB, containing a rigid part for probe support (1), the flexible polyimide
cable (2) and the Preci-Dip connector part (3).
2.4. Microdrive
The extended length of the probe requires stable and sufficiently vibration-free system to minimize
tissue damage along the probe track, while providing flexibility to move the implanted electrode shank
in the brain. A novel microdrive configuration was designed and is illustrated in figure 3. The 8 mm
wide, 7 mm thick, 25 mm high, body of the device weights 2.7 g and consists of three, half octagonal
prism-shaped blocks. The upper block (5), which contains the thread of the driving screw (5), and the
lower block (7) are made of aluminium. The electrode (4) can be fastened through the rigid part of the
PCB (3) to the microdrive by a small screw (9). The middle, electrode-carrier block (6) is held by two
tightly fit rods (8) and is driven by the large driving/positioning screw (10). This block is made of
polytetrafluorethylene (PTFE), as this self-lubricating type plastic renders the opportunity to highly
reduce the joint gap between the block and the rods. Thus vibration is reduced to the minimum, resulting
in a really robust and stable system. One 360° turn of the driving screw moves the probe tip with 300
µm, without any rotational movement.
6
A robust silicon-based microelectrode array, mounted on a microdrive
Figure 3. Layout of the Si-electrode - microdrive system. (a) Front (left) and rear (right) view of the
assembled system. 1: connector; 2: flexible cable; 3: the rigid part of the PCB; 4: the probe shaft;
5: microdrive upper block with the thread for the driving screw; 6: microdrive middle, electrode-carrier
block; 7: microdrive lower block; 8: two tightly fit rods; 9: fastening screw; 10: driving screw. (b)
Assembling of the driving screw, the microdrive and the electrode via the fastening screw.
(c) Schematic drawing of the sensor region of the shaft, with three groups of sites.
2.5. Electrochemical impedance spectroscopy
In order to characterize the electrodes, electrochemical impedance spectroscopy (EIS) measurements
were performed in physiological saline, using an Ag/AgCl reference electrode (Radelkis Ltd., Hungary)
and a counter electrode of platinum wire with relatively high surface area. The probe signal was
sinusoidal, 25 mV RMS. A Reference 600 instrument (Gamry Instruments, PA, USA) was used as
potentiostat and Gamry Framework 6.02 and Echem Analyst 6.02 software were used for experimental
control, data collection and analysis. Experiments were performed in a Faraday cage.
2.6. Animal surgery
Eleven adult Sprague-Dawley rats (Charles River Laboratories, from the breeding colony of ELU,
Hungary) were implanted with the probes under 4% isoflurane anaesthesia. The care and treatment of
all animals were conformed to Council Directive 86/609/EEC, the Hungarian Act of Animal Care and
Experimentation (1998, XXVIII), and local regulations for the care and use of animals in research. All
efforts were taken to reduce the number of animals used and to minimize the animals' pain and
suffering. Prior to surgery we chronically applied the antibiotic Enorofloxacin in 10 mg/kg dose to
prevent inflammatory reactions. We mounted the animals onto a stereotaxic instrument (David Kopf
Instruments, California, USA), and a longitudinal cut was made on the scalp. Skin and muscles were
retracted and the surface was cleaned. Four screw electrodes over the cortex were used to record the
electrocorticogram (ECoG): one over the right secondary motor cortex (AP: 4.5 mm, ML: 2.5 mm with
reference to the bregma), one over the right somatosensory cortex (AP: -2.9 mm, ML: 2.5 mm), two
7
A robust silicon-based microelectrode array, mounted on a microdrive
over the right and left visual cortex (AP: -8,3 mm, ML: 2,5 mm). Two additional screw electrodes over
the cerebellum served as grounding. After driving the screw electrodes into the cranium, the skull was
opened over the left hemisphere using a drill, and the dura mater was incised. The neural probe was
fastened to the microdrive that was adjusted to the stereotaxic instrument through a custom-made
adaptor, and the probe was inserted into the brain manually by oblique targeting (47° inclination to the
vertical), which made it possible to span all three target areas through a single electrode track at the
same time. After the initial penetration, silicone-vacuum grease (Beckman Coulter Inc., Fullerton, CA,
USA) was applied on the probe shank at the brain surface to provide smooth penetration further into
the tissue and prevent blood clot formation. Finally, SuperBond resin cement (Sun Medical Company
Ltd., Moriyama, Japan) was applied on the skull surface and the microdrive was fixed. A crown-like
structure was formed around the system, utilizing UniFast dentacrylic cement (GC America Inc.,
Chicago, USA) and sterile plastic components fabricated from syringes and Petri dishes, protecting its
external components. The structure typically weighted 12 g, was 28 mm long, 22 mm wide, and 30 mm
high. The microdrive – and thus the implant - was not sealed hermetically from the environment,
therefore proper hygienic conditions were in focus to avoid infection at the craniotomy. The
implantation is illustrated in figure 4.
2.7. Course of experiments
During recordings, the rat was kept in a cage of 40 cm × 60 cm, surrounded by a Faraday cage for electrical
shielding. The connecting cable was continuously held up by a light mechanical lever, which prevented the
rat from reaching the cable and damaging it, without impeding the animal in its movements.
Nine rats were allocated for gaining high amount of data considering the characteristics of the targeted
brain nuclei. To achieve this, we needed short, 1-2 hour sessions, during which the rats changed from
awakeness to sleeping state or vice versa. During these sessions, we measured firing patterns of
individual cells. After a successful session (or on the following day) we used the microdrive to relocate
the probe so that new groups of cells could be explored. During the initial implantation, the probe tip
was aimed at the area of the pedunculopontine tegmental nucleus. The Bregma was used as reference
point for targeting, Instead of surface coordinates, a depth coordinate was determined (AP -7.0 ; ML
1.5; dorsoventral 7.6), to which the tip was targeted in a 47° angle and from where the oblique path
down to deeper areas started. The tilting of the head in the sagittal plane was avoided by maintaining
the "flat skull position" by checking and refining the correctness of the lambda (DV = 0) and interaural
line (AP = 9.0; DV = 10.1) coordinates in reference to the bregma. During insertion, the probe was
slowly (approximately 2-3 mm/min), manually advanced, solely with the usually dorsovenrtral
directional drive of the stereotaxic apparatus (which was tilted so that the driving direction was parallel
to the direction of the probe shaft).
Thus each electrode group could record from different deep brain areas over time. At the deepest
penetration the tip reached the caudal pontine reticular nucleus. Altogether, the system allowed us to
record 12 unit channels (4 from each target area) and 4 ECoG channels from each animal. The electrode
array was slowly advanced (60-100 µm per day) during the recording period (minimum 22, maximum
125 days) and stopped if extracellular unit firing coud be recorded, until the electode tip reached the
PnC.
Two additional rats were sacrificed in order to gain some information about the chronic static
performance of the electrodes (without driving). In this case, we used the very same methodology for
8
A robust silicon-based microelectrode array, mounted on a microdrive
implantation, as previously described, but did not drive the probe so frequently, as instead, we kept
them at the same location for weeks.
Figure 4. Implantation surgery and histological evaluation of electrode position. (a), (b) Positioning
and inserting of the electrode into the brain tissue. (c) Implanted rat after surgery. (d) Illustration of the
implantation strategy. Figure based on Fig. 82. of [48]. PnC: pontine reticular nucleus, caudal part; PnO:
pontine reticular nucleus, oral part; PPTg: pedunculopontine tegmental nucleus
2.8. Recording system and data analysis methods
We used a multichannel amplifier (Multiamp SMA-1a, Supertech) and a data processing system (CED
Micro1401, Cambridge Electronics Design Ltd., Cambridge, UK) for data acquisition. Extracellular
action potentials were captured at 33 kHz on each electrode whenever their amplitude reached the 100
µV threshold. We used a 500 – 8000 Hz online filter and a 300 – 6000 Hz offline filter before spike
sorting. Electrocorticogram digitalization frequency was 5 kHz (offline filtering between 0.2 and 300
Hz). Data files were recorded with Spike2 (Cambridge Electronic Design Ltd., Cambridge, UK)
software. Spike waveforms were sorted by using a 3D off-line spike analysis program (OFS/v2, Plexon
Inc., Dallas, TX, USA). With the help of principal component analysis, we determined scatterplots
using the first two principal components. K-means clustering analysis was used to define cluster borders
of single units. ECoG spectrograms, autocorrelations and cross-correlations were calculated to validate
9
A robust silicon-based microelectrode array, mounted on a microdrive
the cluster verification with NeuroExplorer (Plexon Inc., Dallas, TX, USA). To assess the quality of
spike cluster separation, we investigated the spike contamination of the refractory period in the
autocorrelograms [49]. If such contaminations were present, additional efforts, such as manual
clustering were made to separate the waveforms.
In order to quantify the quality of the unit cluster separation, we employed three statistical methods: a
multivariate analysis of variance (MANOVA) test (F), a nonparametric test (J3) and the calculation of
the Davies-Bouldin validity index (DB), which is also nonparametric. These methods are described in
detail in the supporting text of a research paper by Nicolelis et al. [50]. The J3 index is calculated by
dividing the average distance between clusters (J2) by the average distance of points from their cluster
mean (J1). Therefore, higher J3 values indicate better separation rates. The DB index is defined as
𝐷𝐵 =
1
𝑛
𝑛
𝑘=1
∑ max
𝑘≠𝑙
{
𝑆𝑛(𝑄𝑘)+𝑆𝑛(𝑄𝑙)
𝑑(𝑄𝑘,𝑄𝑙)
} ,
(1)
where n is the number of clusters, Sn(Q) is the average distances of the elements of cluster Q to the
center of cluster Q and d(Qk,Ql) gives the distance between the center of the Qk and Ql clusters [50].
Good clustering is characterized by high F, J3 and small BD index values.
2.9. Histological evaluation
Electrode placement can be verified by electrical stimulation through the recording sites in rats deeply
anesthetized with urethane [51, 52]. We used 200–400 ms square wave current pulses with a current
strength of 1–10 μA. Animals were then transcardially perfused with 4% paraformaldehyde in sodium-
phosphate buffer. 60 µm longitudinal sections were cut from the brains using a freezing microtome, and
the tissue was stained with the Gallyas silver-staining method [53, 54]. Thus the electrode track of the
chronically inserted probe and acutely injured „dark" neurons caused by the electrical stimulation can
be seen at the same time.
3. Results and discussion
3.1. The realized system
The custom design technology allowed precise electrode array arrangement on the silicon probe, fitted
to the location of the three brain areas we intended to observe. Photographs and scanning electron
microscopic images of the MEA are shown in figure 5(a)-(c). The results of the EIS measurements are
shown in figure 5(d). At 1 kHz, 3.12 MΩ ± 0.21 MΩ average impedance magnitude was yielded for the
12 × 12 µm2 electrodes, which is in consistence with the literature [55]. The robustness of the device
has proven to be sufficient for implantation. We have observed neither breaking nor bending of the
shafts. The flexible connection to the socket and firm holding of the electrode in a hard printed circuit
board (PCB) attached to the microdrive kept the stability of the probe and provided freedom for driving.
While custom designed microdrives are commonly used for the manipulation of fine wire electrodes,
tetrodes or hexatrodes [56-58], it is less frequent in case of silicon-based MEAs. With a silicon-based
MEMS thermal actuator, driving of a microelectrode array with 8.8 µm step resolution was achieved
[59]. However, the construction of the device requires deposition of 11 layers onto the substrate wafer,
including 4 layers of polysilicon. In our case, the driving was achieved without involving subtle motors
10
A robust silicon-based microelectrode array, mounted on a microdrive
or active elements. Compared to the dDrive system of NeuroNexus (Ann Arbor, MI, USA), the system
presented here utilizes aluminium blocks instead of plastic, and two vertical aligners improve stability
of the vertically moving block rather than purely relying on the driving screw. The whole system met
the particular demands of the angled implantation and deep location of the targeted areas.
Figure 5. (a) The silicon probe mounted onto the rigid-flexible PCB. (b)-(c). Scanning electron
microscopic images of the probe tip and a recording site. (d) Average EIS spectrum of the electrodes of
a probe. Standard deviation values were less than 10% of the corresponding magnitude values on all
frequencies. At 1 kHz, the magnitude of the impedance was 3.12 MΩ ± 0.21 MΩ.
The microdrive was stable during the recording sessions. The stereotaxic instrument provided a specific
targeting angle and precise implantation during surgery, and the microdrive allowed uni-directional
positioning of the MEA even after six months. Figure 6 shows images of the histological findings. The
dark neurons of the Gallyas staining method indicated an approximately 4-600 µm wide trail of
perturbed tissue along the probe track and confirmed successful targeting of the deep-brain regions of
interest. No sign of bending was observed on the used, explanted probes.
11
A robust silicon-based microelectrode array, mounted on a microdrive
Figure 6. Electrode position evaluated with histology. Slices were stained with Gallyas method. Dark
neurons indicate the position of the probe shaft in the brain tissue. (a)-(b) Electrode track with a
schematic drawing of the inserted probe. Scale bars are respectively 1.5 mm and 0.8 mm. Abbreviations:
7n: facial nerve or its root; DpMe: deep mesencephalic nucleus; LC: locus coeruleus; lfp: longitudinal
fasciculus of the pons; MVePC: medial vestibular nucleus, parvicellular part; MVPO: medioventral
periolivary nucleus; Pn: pontine nuclei; PnC: pontine reticular nucleus, caudal part; PnO: pontine
reticular nucleus, oral part; (c)-(e) Zoomed images of the probe track in the targeted nuclei. Scale bars
are 400 µm.
3.2. Unit activities recorded from the targeted nuclei
The majority of the implanted MEAs (n = 9) were utilized in order to obtain single unit activities from
different cells in the nuclei of freely moving rats. As mentioned earlier, in section 2.7. Course of
experiments, these probes were driven further with the microdrive after successful, 1-2 hour recording
sessions, which contained transitions from awake/sleeping states. We have successfully recorded single
unit activity chronically from all target areas (DpMe, PpTg, PnO). Figure 7 shows examples of recorded
units during an epoch of 15 seconds in a rat, 25 days after surgery. Raw data and spectrogram of the
ECoG, raw data and unit firing from 4 multi-unit recording sites and the averaged waveforms of sorted
individual units are presented. The spectrogram of the ECoG reveal sleep spindles, starting at around t
= 2 and 14 seconds. Spike trains of cells are presented that fire frequently at the beginning of the spindle
(M1 units), and ones that cease to be active at the end of the spindle (B2, B3 units). The data gained by
the MEAs gave insight into the firing patterns of different cells that take part in sleep regulation in the
brainstem.
We were able to record approximately 200 extracellular units from this animal, of which 163 were
thoroughly analyzed after sorting. Units that fired for just a short period of time or had too low firing
frequency or high spike waveform variance were excluded. Of the 163 neurons, 15 were present on a
single electrode site only. Thus we recorded 148 neurons concurrently. The active contacts between
12
A robust silicon-based microelectrode array, mounted on a microdrive
parallel recordings were minimally 2, maximally 6 at the same time. The contacts showed unit activity
in ~27.5% of the cases with an average of 5 recorded cells. These results show that individual cell
activities in deep brain regions of a rodent can be recorded with relatively high yield using such silicon
probes.
The number of simultaneous recordings of sorted units from the target nuclei (DpMe, PPTg, PnO) was
achieved, however, sparsely: altogether 13 such sessions occurred in the 9 rats out of altogether 266
analyzed sessions (on the average, 1.4 ± 0.5 per rat).
Figure 7. Representative parallel recordings from different sites on an implanted probe in a freely
moving rat, 25 days after surgery. (a) Raw data and spectrogram of the ECoG channel, and unit channels
with single unit patterns (spike trains). Elevated power in the lower frequency range of the ECoG
channel indicates sleep spindles starting at around t = 2 and 14 seconds. Activities of cells that are in
connection with the spindles were recorded. (b) 0.24 second long recording sections (highlighted with
a red band on panel A) at a faster time-base after offline filtering (ECoG: 0.2 – 300 Hz, probe site
channels: 300-6000 Hz). (c) Different types of representative autocorrelograms show heterogeneous
spiking properties.
13
A robust silicon-based microelectrode array, mounted on a microdrive
3.3. Signal stability
Two additional rats were implanted in order to gain insight into the static chronic performance of the
electrodes. These MEAs were not driven after each successful recording session, as instead, we let them
stay in the same location for 2-7 days. Figure 8 shows the functionality of such a probe during a 33-day
period. The number of different single unit activities varied from 1 to 8, with an average of 3.9 ± 2.0.
Figure 8(a) shows the number of single unit activities on each day, when recordings were performed.
Black triangles mark the occasions when the probe was driven further into the tissue. The quality of
unit clustering is shown in figure 8(b). The mean ± standard error for the indices throughout the 33-day
period were the following: F = 8.23 ± 5.43, J3 = 2.76 ± 1.53, DB = 0.26 ± 0.1. For these values, the
measurements of Nicolelis et al. with microwire arrays in the cortex of rhesus monkeys yielded F =
11.5 ± 1.0, J3 = 3.8 ± 0.4, DB = 0.32±0.01 [50]. Figure 8(c)-(c') illustrate the results of cluster analysis
of three cells recorded from the same animal on day 21 and 22, respectively, showing that good quality
single units of the same clusters can be recorded on different days.
14
A robust silicon-based microelectrode array, mounted on a microdrive
Figure 8. (a) The number of single unit activities on each day, when recordings were performed from
a rat. Black triangles mark the occasions when the probe was driven further into the tissue. (b) The
quality of unit clustering. The mean ± standard error values considering the 33-day period:
F = 8.23 ± 5.43, J3 = 2.76 ± 1.53, DB = 0.26 ± 0.1. (c)-(c') Typical unit clusters on a channel during a
session and the same clusters obtained from the signals recorded on the following day. Horizontal and
vertical scale bars on the spike waveforms are 0.5 ms and 100 µV, respectively. The cluster plots contain
data of 120 minute long recordings.
Unit activities on 20-30% of the electrodes on 15 µm thick planar Michigan MEAs were reported 30
days post-implant, in the motor cortex of rats [60]. With such probes, 10-15% unit yield was measured
after 3-6 months, even though a significant portion of the electrodes was in white matter [13].
Comparing our results to those findings, the long-term performance of our system is slightly poorer.
The causes of this can be various, we provide three possibilities here. Firstly, highly different brain
15
A robust silicon-based microelectrode array, mounted on a microdrive
regions might respond differently to the implantation: the cells in our targeted nuclei might be more
sensitive compared to the cells of the cortex or the hippocampus. Such heterogeneity in unit stability
was even observed in different layers of the cortex: layers IV to VI were claimed to have the greatest
yield acutely, but layers II to IV the best yield in chronic time points [13]. Secondly, the much larger
(200 µm) shaft thickness of our probes is disadvantageous in this aspect, it can induce a more aggressive
immune response [61]. Finally, interfacing with the deep brain nuclei required a much longer
implantation track than the cortical areas would have required, which results in a bigger chance of
disrupting blood vessels, thus facilitating chronic signal degradation [12].
It was not our goal to measure the same unit activities for weeks or months (which is a common priority
e.g. for the development of brain-computer interfaces with invasive electrodes). Rather, we intended to
observe firing patterns of different cells of deep brain nuclei during 1-2 hour sessions. In the light of
our aims, the microdrive alleviated the problem of losing signals of unit activities, as slight motions of
the probe could bring the electrodes into less perturbed tissue regions and explore new cells.
4. Conclusions
We report here on a novel silicon-based MEA design, tailored to be used in a specific neuroscientific
experiment, which required interfacing with three deep brain nuclei related to the regulation of sleep-
awake cycles. The applied fabrication technology provided cost-effective batch production of the
MEAs. The stiff body of the probe allowed sufficient implantation, even into the brainstem, while the
custom-designed electrode array formation provided coverage of the investigated brain regions. The
probe was suitable for measuring unit activities in the targeted structures, with high yield, through
several recording sites of freely moving rats, fulfilling our goals. The main novelty of our system is the
utilization of a special microdrive, which enabled us to accurately relocate the electrodes after the
implantation, and explore new cells. Taking into account that the brainstem is one of the most difficult
targets for neural recording in rats, our system can probably be used in many different experiments,
when two or more brain structures can be exposed in an electrode track, even if the track is not
perpendicular to the cerebral surface. In a future study, we intend to investigate the usability of such
MEAs in the central nervous system of larger mammals, such as cats or primates.
Acknowledgements
We thank András Czurkó for critical reading of the manuscript. We are thankful for the laboratory
personnel of the MEMS Laboratory of the Institute for Technical Physics and Materials Science. This
work was supported by the KTIA_13_NAP-A-IV/1-4;6 research grants. Z. Fekete is grateful for the
support of the Hungarian Brain Research Program (KTIA NAP 13-2-2014-0022). A. Pongracz is
grateful for the Bolyai Janos scholarship of the Hungarian Academy of Sciences.
References
Fujisawa S and Buzsaki G 2011 A 4 Hz oscillation adaptively synchronizes prefrontal,
VTA, and hippocampal activities Neuron 72 153-65
Buzsaki G 2010 Neural syntax: cell assemblies, synapsembles, and readers Neuron 68
362-85
16
[1]
[2]
A robust silicon-based microelectrode array, mounted on a microdrive
[3]
[4]
Buzsaki G 2004 Large-scale recording of neuronal ensembles Nature neuroscience 7
446-51
Kubie J L 1984 A driveable bundle of microwires for collecting single-unit data from
freely-moving rats Physiol Behav 32 115-8
[5] Wilson M A and McNaughton B L 1994 Reactivation of hippocampal ensemble
[6]
[7]
memories during sleep Science 265 676-9
Rinaldi P, Juhasz G and Verzeano M 1976 Analysis of circulation of neuronal activity
in the waking cortex Brain Res Bull 1 429-35
Verzeano M and Calma I 1954 Unit activity in spindle bursts Journal of
neurophysiology 17 417-28
[8] Wise K D and Najafi K 1991 Microfabrication techniques for integrated sensors and
[9]
microsystems Science 254 1335-42
Lai H-Y, Liao L-D, Lin C-T, Hsu J-H, He X, Chen Y-Y, Chang J-Y, Chen H-F, Tsang
S and Shih Y-Y I 2012 Design, simulation and experimental validation of a novel
flexible neural probe for deep brain stimulation and multichannel recording Journal of
Neural Engineering 9 036001
[10] Moxon K A, Leiser S C, Gerhardt G A, Barbee K A and Chapin J K 2004 Ceramic-
based multisite electrode arrays for chronic single-neuron recording IEEE transactions
on bio-medical engineering 51 647-56
[11] McCarthy P T, Rao M P and Otto K J 2011 Simultaneous recording of rat auditory
cortex and thalamus via a titanium-based, microfabricated, microelectrode device
Journal of neural engineering 8 046007
[12] Kozai T D, Jaquins-Gerstl A S, Vazquez A L, Michael A C and Cui X T 2015 Brain
tissue responses to neural implants impact signal sensitivity and intervention strategies
ACS chemical neuroscience 6 48-67
[13] Kozai T D, Du Z, Gugel Z V, Smith M A, Chase S M, Bodily L M, Caparosa E M,
Friedlander R M and Cui X T 2015 Comprehensive chronic laminar single-unit, multi-
unit, and local field potential recording performance with planar single shank electrode
arrays J Neurosci Methods 242 15-40
[14] Otchy T M and Lveczky B P 2012 Design and assembly of an ultra-light motorized
microdrive for chronic neural recordings in small animals J Vis Exp 8
[15] Vandecasteele M, M S, Royer S, Belluscio M, Berenyi A, Diba K, Fujisawa S,
Grosmark A, Mao D, Mizuseki K, Patel J, Stark E, Sullivan D, Watson B and Buzsaki
G 2012 Large-scale recording of neurons by movable silicon probes in behaving
rodents J Vis Exp 4
[16] Fekete Z 2015 Recent advances in silicon-based neural microelectrodes and
microsystems: a review Sensors and Actuators B: Chemical 215 300-15
[17] Drake K L, Wise K D, Farraye J, Anderson D J and BeMent S L 1988 Performance of
planar multisite microprobes in recording extracellular single-unit intracortical activity
Biomedical Engineering, IEEE Transactions on 35 719-32
[18] Schmidt S, Horch K and Normann R 1993 Biocompatibility of silicon-based electrode
arrays implanted in feline cortical tissue J Biomed Mater Res 27 1393-9
[19] Neumann A, Reske T, Held M, Jahnke K, Ragoss C and Maier H R 2004 Comparative
investigation of the biocompatibility of various silicon nitride ceramic qualities in vitro
J Mater Sci Mater Med 15 1135-40
[20] Cheung K, Gun L, Djupsund K, Yang D and Lee L P 2000 A new neural probe using
SOI wafers with topological interlocking mechanisms. In: Microtechnologies in
Medicine and Biology, 1st Annual International, Conference On. 2000, pp 507-11
17
A robust silicon-based microelectrode array, mounted on a microdrive
[21] Grand L, Wittner L, Herwik S, Gothelid E, Ruther P, Oscarsson S, Neves H, Dombovari
B, Csercsa R, Karmos G and Ulbert I 2010 Short and long term biocompatibility of
NeuroProbes silicon probes Journal of neuroscience methods 189 216-29
[22] Nicolelis M A and Ribeiro S 2002 Multielectrode recordings: the next steps Current
Opinion in Neurobiology 12 602-6
[23] Szarowski D H, Andersen M D, Retterer S, Spence A J, Isaacson M, Craighead H G,
Turner J N and Shain W 2003 Brain responses to micro-machined silicon devices Brain
Research 983 23-35
[24] Csicsvári J, Henze D A, Jamieson B, Harris K D, Sirota A, Bartho P, Wise K D and
Buzsáki G 2003 Massively parallel recording of unit and local field potentials with
silicon-based electrodes Journal of neurophysiology 90 1314-23
[25] Bragin A, Hetke J, Wilson C L, Anderson D J, Engel Jr J and Buzsáki G 2000 Multiple
site silicon-based probes for chronic recordings in freely moving rats: implantation,
recording and histological verification Journal of neuroscience methods 98 77-82
[26] Fekete Z, Pongrácz A, Márton G and Fürjes P 2013b On the Fabrication Parameters of
Buried Microchannels Integrated in In-Plane Silicon Microprobes Materials Science
Forum 729 210-5
[27] Pongrácz A, Fekete Z, Márton G, Bérces Z, Ulbert I and Fürjes P 2013 Deep-brain
silicon multielectrodes for simultaneous in vivo neural recording and drug delivery
Sensors and Actuators B: Chemical
[28] Frey O, Wal P D v d, Spieth S, Brett O, Seidl K, Paul O, Ruther P, Zengerle R and
Rooij N F d 2011 Biosensor microprobes with integrated microfluidic channels for bi-
directional neurochemical interaction Journal of Neural Engineering 8 066001
[29] Wu F, Stark E, Im M, Cho I-J, Yoon E-S, Buzsáki G, Wise K D and Yoon E 2013 An
implantable neural probe with monolithically integrated dielectric waveguide and
recording electrodes for optogenetics applications Journal of Neural Engineering 10
056012
[30] Kozai T D, Catt K, Li X, Gugel Z V, Olafsson V T, Vazquez A L and Cui X T 2015
Mechanical failure modes of chronically implanted planar silicon-based neural probes
for laminar recording Biomaterials 37 25-39
[31] Mena-Segovia J, Sims H M, Magill P J and Bolam J P 2008 Cholinergic brainstem
neurons modulate cortical gamma activity during slow oscillations The Journal of
physiology 586 2947-60
[32] Márton G, Fekete Z, Fiath R, Baracskay P, Ulbert I, Juhász G, Battistig G and Pongrácz
A 2013 In vivo measurements with robust silicon-based multielectrode arrays with
extreme shaft lengths IEEE Sensors in press
[33] Hajj-Hassan M, Chodavarapu V P and Musallam S 2009 Microfabrication of ultra-long
reinforced silicon neural electrodes Micro & Nano Letters, IET 4 53-8
[34] Fekete Z, Hajnal Z, Márton G, Fürjes P and Pongrácz A 2013 Fracture analysis of
silicon microprobes designed for deep-brain stimulation Microelectronic Engineering
103 160-6
[35] Moruzzi G and Magoun H W 1949 Brain stem reticular formation and activation of the
EEG Electroencephalogr Clin Neurophysiol 1 455-73
[36] Camacho Evangelista A and Reinoso Suarez F 1964 Activating and Synchronizing
Centers in Cat Brain: Electroencephalograms after Lesions Science 146 268-70
[37] Villanueva L, Desbois C, Le Bars D and Bernard J F 1998 Organization of diencephalic
projections from the medullary subnucleus reticularis dorsalis and the adjacent cuneate
18
A robust silicon-based microelectrode array, mounted on a microdrive
nucleus: a retrograde and anterograde tracer study in the rat The Journal of comparative
neurology 390 133-60
[38] Winn P 2006 How best to consider the structure and function of the pedunculopontine
tegmental nucleus: evidence from animal studies J Neurol Sci 248 234-50
[39] Rodrigo-Angulo M L, Rodriguez-Veiga E and Reinoso-Suarez F 2005 A quantitative
study of the brainstem cholinergic projections to the ventral part of the oral pontine
reticular nucleus (REM sleep induction site) in the cat Exp Brain Res 160 334-43
[40] Nuñez A, de Andres I and Garcia-Austt E 1991 Relationships of nucleus reticularis
pontis oralis neuronal discharge with sensory and carbachol evoked hippocampal theta
rhythm Exp Brain Res 87 303-8
[41] Nuñez A, Moreno-Balandran M E, Rodrigo-Angulo M L, Garzon M and De Andres I
2006 Relationship between the perifornical hypothalamic area and oral pontine reticular
nucleus in the rat. Possible implication of the hypocretinergic projection in the control
of rapid eye movement sleep Eur J Neurosci 24 2834-42
[42] Roš H, Magill P J, Moss J, Bolam J P and Mena-Segovia J 2010 Distinct types of non-
cholinergic pedunculopontine neurons are differentially modulated during global brain
states Neuroscience 170 78-91
[43] Vertes R P and Miller N E 1976 Brain stem neurons that fire selectively to a conditioned
stimulus for shock Brain Research 103 229-42
[44] Steriade M, Dossi R C, Pare D and Oakson G 1991 Fast oscillations (20-40 Hz) in
thalamocortical systems and their potentiation by mesopontine cholinergic nuclei in the
cat Proceedings of the National Academy of Sciences of the United States of America
88 4396-400
[45] McCarley R W and Hobson J A 1971 Single neuron activity in cat gigantocellular
tegmental field: selectivity of discharge in desynchronized sleep Science 174 1250-2
[46] Siegel J M, Wheeler R L and McGinty D J 1979 Activity of medullary reticular
formation neurons in the unrestrained cat during waking and sleep Brain Research 179
49-60
[47] Tseng W T, Yen C T and Tsai M L 2011 A bundled microwire array for long-term
chronic single-unit recording in deep brain regions of behaving rats Journal of
neuroscience methods 201 368-76
[48] Paxinos G and Watson C 1998 The rat brain in stereotaxic coordinates: Academic
Press)
[49] Harris K D, Henze D A, Csicsvari J, Hirase H and Buzsaki G 2000 Accuracy of tetrode
spike separation as determined by simultaneous intracellular and extracellular
measurements Journal of neurophysiology 84 401-14
[50] Nicolelis M A L, Dimitrov D, Carmena J M, Crist R, Lehew G, Kralik J D and Wise S
P 2003 Chronic, multisite, multielectrode recordings in macaque monkeys Proceedings
of the National Academy of Sciences 100 11041-6
[51] Czurkó A, Hirase H, Csicsvári J and Buzsáki G 1999 Sustained activation of
hippocampal pyramidal cells by 'space clamping' in a running wheel Eur J Neurosci 11
344-52
[52] Lőrincz M, Oláh M, Baracskay P, Szilágyi N and Juhász G 2007 Propagation of spike
and wave activity to the medial prefrontal cortex and dorsal raphe nucleus of WAG/Rij
rats Physiol Behav 90 318-24
[53] Gallyas F, Guldner F H, Zoltay G and Wolff J R 1990 Golgi-like demonstration of
"dark" neurons with an argyrophil III method for experimental neuropathology Acta
Neuropathol 79 620-8
19
A robust silicon-based microelectrode array, mounted on a microdrive
[54] Gallyas F, Hsu M and Buzsáki G 1993 Four modified silver methods for thick sections
of formaldehyde-fixed mammalian central nervous tissue: 'dark' neurons, perikarya of
all neurons, microglial cells and capillaries Journal of neuroscience methods 50 159-
64
[55] Schuettler M 2007 Electrochemical Properties of Platinum Electrodes in Vitro:
Comparison of Six Different Surface Qualities. In: Engineering in Medicine and
Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE,
pp 186-9
[56] Toth A, Petyko Z, Mathe K, Szabo I and Czurko A 2007 Improved version of the
printed circuit board (PCB) modular multi-channel microdrive for extracellular
electrophysiological recordings J Neurosci Methods 159 51-6
[57] Otchy T M and Ӧlveczky B P 2012 Design and Assembly of an Ultra-light Motorized
Microdrive for Chronic Neural Recordings in Small Animals Journal of Visualized
Experiments : JoVE 4314
[58] Santos L, Opris I, Fuqua J, Hampson R E and Deadwyler S A 2012 A novel tetrode
microdrive for simultaneous multi-neuron recording from different regions of primate
brain Journal of neuroscience methods 205 368-74
[59] Muthuswamy J, Okandan M, Gilletti A, Baker M S and Jain T 2005 An array of
microactuated microelectrodes for monitoring single-neuronal activity in rodents IEEE
Trans Biomed Eng 52 1470-7
[60] Ward M P, Rajdev P, Ellison C and Irazoqui P P 2009 Toward a comparison of
microelectrodes for acute and chronic recordings Brain Res 28 183-200
[61] Seymour J P and Kipke D R 2007 Neural probe design for reduced tissue encapsulation
in CNS Biomaterials 28 3594-607
20
|
1704.02741 | 1 | 1704 | 2017-04-10T07:38:28 | Avalanches and Generalized Memory Associativity in a Network Model for Conscious and Unconscious Mental Functioning | [
"q-bio.NC",
"cond-mat.stat-mech"
] | We explore statistical characteristics of avalanches associated with the dynamics of a complex-network model, where two modules corresponding to sensorial and symbolic memories interact, representing unconscious and conscious mental processes. The model illustrates Freud's ideas regarding the neuroses and that consciousness is related with symbolic and linguistic memory activity in the brain. It incorporates the Stariolo-Tsallis generalization of the Boltzmann Machine in order to model memory retrieval and associativity. In the present work, we define and measure avalanche size distributions during memory retrieval, in order to gain insight regarding basic aspects of the functioning of these complex networks. The avalanche sizes defined for our model should be related to the time consumed and also to the size of the neuronal region which is activated, during memory retrieval. This allows the qualitative comparison of the behaviour of the distribution of cluster sizes, obtained during fMRI measurements of the propagation of signals in the brain, with the distribution of avalanche sizes obtained in our simulation experiments. This comparison corroborates the indication that the Nonextensive Statistical Mechanics formalism may indeed be more well suited to model the complex networks which constitute brain and mental structure. | q-bio.NC | q-bio | Avalanches and Generalized Memory Associativity in a Network
Model for Conscious and Unconscious Mental Functioning
Maheen Siddiqui∗
Centre for Brain and Cognitive Development,
Birkbeck College, University of London,
London WC1E 7HX, United Kingdom.
Roseli S. Wedemann, Corresponding Author†
Instituto de Matem´atica e Estat´ıstica,
Universidade do Estado do Rio de Janeiro,
Rua Sao Francisco Xavier 524, 20550-013, Rio de Janeiro, Brazil.
Henrik Jensen‡
Centre for Complexity Science and Department of Mathematics,
Imperial College London,
London SW7 2AZ, United Kingdom,
and
Institute of Innovative Research,
Tokyo Institute of Technology,
4259, Nagatsuta-cho,
Yokohama 226-8502, Japan.
(Dated: July 25, 2018)
7
1
0
2
r
p
A
0
1
]
.
C
N
o
i
b
-
q
[
1
v
1
4
7
2
0
.
4
0
7
1
:
v
i
X
r
a
1
Abstract
We explore statistical characteristics of avalanches associated with the dynamics of a complex-
network model, where two modules corresponding to sensorial and symbolic memories interact,
representing unconscious and conscious mental processes. The model illustrates Freud's ideas
regarding the neuroses and that consciousness is related with symbolic and linguistic memory
activity in the brain. It incorporates the Stariolo-Tsallis generalization of the Boltzmann Machine
in order to model memory retrieval and associativity. In the present work, we define and measure
avalanche size distributions during memory retrieval, in order to gain insight regarding basic aspects
of the functioning of these complex networks. The avalanche sizes defined for our model should be
related to the time consumed and also to the size of the neuronal region which is activated, during
memory retrieval. This allows the qualitative comparison of the behaviour of the distribution
of cluster sizes, obtained during fMRI measurements of the propagation of signals in the brain,
with the distribution of avalanche sizes obtained in our simulation experiments. This comparison
corroborates the indication that the Nonextensive Statistical Mechanics formalism may indeed be
more well suited to model the complex networks which constitute brain and mental structure.
PACS numbers:
Keywords: Consciousness-Unconsciousness, Neuroses, Self-organized Neural Networks, Boltzmann Machine,
Generalized Simulated Annealing, Avalanches
∗Electronic address: [email protected]
†Electronic address: [email protected]
‡Electronic address: [email protected]
2
I.
INTRODUCTION
The purpose of the present work is to define, measure and study statistical properties
of avalanches which occur during memory retrieval in an artificial neural network, devel-
oped to model neurotic phenomena and the associated conscious/unconscious interactions
in mental life [1].
It was famously reported by Freud [2, 3] that patients with neurotic
symptoms systematically repeated these symptoms in the guise of ideas and impulses. This
tendency was referred to as the compulsion to repeat. The patients' repetitive behaviour
suggested to Freud that they had ". . . the intention of correcting a distressing portion of the
past. . . " [4]. Although the patient is aware of the obsessional ideas and impulses, and of
the performance of the neurotic actions, the concomitant psychical predeterminants remain
unconscious.
In order to infer these predeterminants, and to bring them under the light
of consciousness, Freud developed the analytical treatment called working-through. This
treatment provides, through the interpretation of the predeterminants by freely talking in
psychoanalytical sessions, the connections into which they are inserted.
One of the early seminal findings of psychoanalytic research is that neurotic symptoms
arise from traumatic and repressed memories. In fact, the repressed material can be con-
strued as knowledge that, in spite of being present in the subject, is inaccessible to him
through symbolical representation. In other words, this knowledge is momentarily or per-
manently inaccessible to the patient's conscience, and is therefore regarded as unconscious
knowledge [4, 5]. In these considerations, by symbolic representation we mean the association
of symbols to meaning, as occurs in language, as well as in other ways of expressing thoughts
and emotions, such as artistic representations (e.g., a painting or musical composition), and
the recollection of dreams.
Consequently, an important part of the analytical treatment proposed by Freud consists
of a procedure for "bringing what is unconscious into consciousness". This is tantamount
to a basic technique aiming at "filling up the gaps in the patient's memories, to remove
his amnesias" [4]. Indeed, these amnesias are closely related to the origin of the neurotic
symptoms, that is, to the compulsion to repeat. An important purpose of the therapeu-
tic method called working-through, whereby neurotic analysands obtain relief and cure of
painful symptoms, is to develop knowledge regarding the causes of those symptoms. This is
achieved by accessing unconscious memories through free associative talking, which yields
3
understanding and a change in the analysand's compulsion to repeat [2, 3, 5]. The method is
mainly based on the analysis of free associative talking, symptoms, parapraxes (slips of the
tongue and pen, misreading, forgetting, etc.), dreams, and also on analyzing what is acted
out in transference. This procedure allows the patient to slowly symbolize his repressed
memories and to create new representations of his past experiences.
An illustrative, schematic model of this neurotic mechanism and the working-through
therapeutic method, was advanced in [1, 6]. It was there proposed that the neuroses can be
understood in terms of an associative memory process in neural networks, where the network,
when presented with an input pattern, retrieves a stored pattern which is most similar to the
one currently shown. In order to model the compulsion to repeat a neurotic symptom, it was
assumed that such a symptom is produced when the subject receives a stimulus resembling a
repressed or traumatic memory trace. The stimulus then contributes to stabilize the neural
net in a minimal energy state, corresponding to the trace that synthesizes the original
repressed experience, which in turn produces a neurotic response (an act). The neurotic
act does not result from the stimulus as a new situation, but as a response to the repressed
memory.
The model is based on the conception that the linguistic, symbolic, and associative pro-
cess associated with psychoanalytic working-through therapy is mapped onto a correspond-
ing scheme of reinforcing synapses among memory traces in the brain. These connections
should involve declarative memory, implying that repressed memories are, at least par-
tially, transformed into conscious ones. This takes into account the paramount importance
that language has in psychoanalytic sessions, and the idea that unconscious memories are
precisely those that cannot be expressed symbolically. It was thus proposed that, as the
analysand symbolically elaborates manifestations of unconscious material through transfer-
ence in psychoanalytic sessions, he creates new neural connections, and reinforces or inhibits
older ones, reconfiguring the topology of his neural net. The network topology arising from
this reconfiguration process stabilizes onto new energy minima, associated with new acts.
Following the memory organization advanced in [7] (see also [1]), it is assumed that neu-
rons belong to two hierarchically structured modules, respectively corresponding to sensorial
and symbolic memories. Mental images of stimuli received by sensory receptors, either from
the environment or from the body itself, are represented by memory traces stored in the sen-
sorial memory module. On the other hand, higher level representations of traces in sensorial
4
memory, i.e. symbols, are stored in the symbolic memory module. This module represents
brain structures corresponding to symbolic processing, language, and consciousness. Sen-
sorial and symbolic memories are not isolated from each other and indeed, they interact,
generating unconscious and conscious mental activity.
In the model, the unconscious compulsion to repeat in neuroses [2 -- 4] is interpreted as a
bodily response (an act) to an input stimulus (of any kind) that resonates with a pattern
in sensorial memory, without activating symbolic memory. In this sense, the compulsion to
repeat is akin to a reflexive act. This accounts for neurotic patients' claim that they cannot
explain their neurotic acts. A sensorial memory trace becomes conscious when its retrieval
also activates the retrieval of patterns in symbolic memory. If this happens, the resulting
output does not resemble reflexive behaviour, and another level of processing becomes rele-
vant. One may also have symbolic representations of emotions, as when someone says "I felt
a warm happiness when I embraced my young nephew". Sensorial information that is not
(or cannot be) associated to a symbol stays unconscious. The neurotic mechanism described
here is in line with the hypothesis that the emergence of conscious experience requires the
existence of a physical (neural) layer that is able to support metarepresentations [8].
A detailed description of the neural network model for these neurotic features can be found
in [1], which includes a full description of the relevant algorithms. Memory functioning was
first modelled by recourse to a Boltzmann Machine (BM), which is a stochastic extension of
the Hopfield model. However, it was later found that the node-degree distributions of the
hierarchically clustered network topologies generated by the model's clustering algorithm
(see section II), with long and short-range synapses, are well represented by the asymp-
totic power-law, q-exponential distributions. This behaviour indicates that the statistical
features of our model may not be well described by Boltzmann-Gibbs statistical mechan-
ics, but rather by Nonextensive Statistical Mechanics (NSM) [9 -- 11]. The NSM theoretical
framework, and its variegated applications to physics, biology, economics, and other areas,
have been the focus of an intense research activity in recent years [10]. The concomitant
formalism has been applied to the study of diverse types of complex systems, including sys-
tems with long-range interactions [12], systems exhibiting weak chaos [13], complex networks
[14], processes involving nonlinear diffusion [15], and many others. In recent work [16, 17]
(see also references therein), the authors have also used Hopfield neural networks to model
attachment in developmental psychology, as well as behavioural patterns in psychology and
5
psychotherapy. In the present work, we thus use the model described in [1], where memory
is simulated by a generalization of the BM inspired on NSM, called Generalized Simulated
Annealing (GSA) [1, 9], and this affects the sequence of associations of thoughts in the
mental processes we are illustrating.
The model is in good qualitative agreement with the main facts provided by psycho-
analytic experience. In particular, it is consistent with the (sometimes exasperating) slow
nature of the working-through process. The model's dynamics reproduces in a plausible
way the re-association of unconscious sensorial memory traces, and of new experiences, to
symbolic processing areas, mimicking the repetitive, adaptive, reinforcement learning in-
volved in the simulation of working-through. As a result of this self-reconfiguration process,
represented in the model by a change in network connectivity, the analysand is partially
freed from his/her original neurotic states and concomitant acts. The new network topology
evolves to, and stabilizes itself onto, new energy minima. These resulting network states
are associated with new conscious or unconscious acts. Evidently, the ultimate aim of both
the analysand and the analyst is to reach new states, and generate new acts, that are more
pleasant and comfortable to the analysand and his relations. As the therapy progresses, the
analysand rewrites his/her life history, through a process of new significations. Furthermore,
and perhaps more importantly, both the present and the future of the analysand are also
being rewritten, by creating new possibilities, and by opening new windows of opportunity
for the pursuit of happiness. Our central tenet is that this story is embedded, i.e. written,
in the individual's biological neural network.
It is clear that psychoanalytic theory is still far from having the rigorous quantitative
support required by modern science. However, we agree with some contemporary scien-
tists [18, 19], as well as early psychoanalysts [4, 20, 21] that, although this state of affairs
poses serious limitations and difficulties, it also constitutes a challenge and a stimulus for
further scientific research. These venues of enquiry are worth pursuing, since many find-
ings of psychodynamic theories have already contributed both to our understanding and
characterization of mental phenomena, as well as to the development of successful clinical
treatments for many mental disorders [18, 19].
In the present work, we study some additional properties of the aforementioned model,
which give us further knowledge on how basic microscopic and macroscopic features and
mechanisms influence emergent behaviour of the complex network structures proposed by
6
the model. In Section II, we give a brief overview of the basic algorithms that characterize
the model and which we use in our simulation experiments. We present a definition of
avalanche sizes during the memory retrieval mechanism and results of their measurements, in
simulation experiments in Section III. These avalanche sizes can be qualitatively compared
to brain imaging experiments [22], showing that the NSM memory retrieval mechanism
produces power-law behaviour, which does not emerge from BM functioning, i.e.
from
Boltzmann-Gibbs statistical mechanics.
II. NETWORK TOPOLOGY AND MEMORY ACCESS MECHANISMS
The topological structure of each of the two memory modules was generated by a clus-
tering algorithm proposed in [7] (see also [1]), which models the self-organizing process that
controls synaptic plasticity, resulting in a hierarchically structured neural network topol-
ogy. The algorithm is inspired on microscopic biological mechanisms, found in typical brain,
cellular processes of many animals [23, 24].
As an example we can mention the on-center/off-surround structure, characterized by
neurons that are in cooperation, through excitatory synapses, with other neurons in their
immediate neighbourhood, while they are in competition with neurons that lie outside these
surroundings. Competition and cooperation occur both in statically hardwired structures,
and as part of a variety of neuronal dynamical processes, where neurons compete for certain
chemicals [23, 24]. In synaptogenesis, for instance, stimulated neurons release neural growth
factors that spread through diffusion, reaching neighbouring cells and promoting synaptic
growth. The cells receiving neural growth factors make synapses and live, while those hav-
ing no contact with these substances die [23, 25]. A neuron releasing neural growth factors
ushers the process of synaptic formation in its tridimensional neighbourhood, and becomes
a center of synaptic convergence. Neighbouring neurons releasing different neural growth
factors at different rates give rise to several synaptic convergence centers. These centers
then compete through the new synapses being created in their surroundings. Through these
processes, a signaling network is established, that controls the development and plasticity
of the brain's neuronal circuits. This neural competition is started and controlled by envi-
ronmental stimulation and, consequently, it constitutes an important mechanism through
which features of the environment can be mapped onto brain structures.
7
A. Clustering Algorithm
We reproduce the algorithm proposed in [1, 7] here, to aid understanding of the mea-
surements and analysis which we introduce in this paper. This clustering algorithm models
the self-organizing process which controls synaptic plasticity, and results in a structured
hierarchical topology of each of the two memory modules. It consists of the following steps.
Step 1 The initial bidimensional positions of the neurons are randomly generated according
to a uniform probability density on a square sheet.
Step 2 To simulate synaptic growth, we assume a Gaussian solution of the equation gov-
erning the diffusion of the neural growth factors, thus avoiding the time-consuming
numerical treatment of this equation. Consequently, a synapse with strength wij (the
synaptic weight) is allocated to transmit the output signal from neuron nj to a neuron
ni, according to the Gaussian probability density
Pij = exp(−(rj − ri)2/(2σ2))/√2πσ2 ,
(1)
where rj and ri are the respective positions of nj and ni in the bidimensional sheet,
and σ is the standard deviation of the Gaussian distribution, which is here a model
parameter. If a synapse connecting nj to ni is generated, its strength wij is proportional
to Pij.
Step 3 It was verified in [26] that cortical maps representing different stimuli are formed,
such that each stimulus activates a group of neurons spatially close to each other, and
that these groups are uniformly distributed along the sheet of neurons representing
memory. So one now randomly selects m neurons which will each be a center of the
representation of a stimulus. To choose the value of m, one should take into account
the storage capacity of the BM [27].
Step 4 Reinforce synapses adjacent to each of the m centers chosen in Step 3, according
If ni is a center, define sumni = Pj wij. For each nj
to the following criteria.
adjacent to ni, increase wij by ∆wij, with probability P robnj = wij/sumni, where
∆wij = ηP robnj and η ∈ ℜ is a model parameter chosen in [0, 1]. After incrementing
wij, decrement ∆wij from the weights of all the other neighbours of ni, according to:
∀k 6= j,wik = wik − ∆wik, where ∆wik = (1 − wik/Pk6=j wik)∆wij.
8
Step 5 Repeat step 4 until a clustering criterion is met.
In the above clustering algorithm, Step 4 regulates the strength of synaptic connections,
i.e., plasticity, by intensifying synapses within a cluster and reducing synaptic strength
between clusters (disconnecting clusters). A cluster is therefore formed by a group of neu-
rons that are close to each other, with higher probability of being connected by stronger
synapses. This mechanism is akin to a preferential attachment criterion, constrained by an
energy conservation (neurosubstances) prescription, controlling synaptic plasticity. Neurons
that have been sensorially more stimulated and are therefore more strongly connected will
stimulate other neurons in their neighbourhoods and promote still stronger connections.
This is compatible with the biological mechanisms mentioned earlier.
The growth of long-range synapses in the brain occurs less frequently than the growth
of short-range ones. The reason for this is that the former are energetically more costly
than the latter. In order to allocate long-range synapses connecting clusters, one should
regard the basic learning scheme proposed by Hebb [27 -- 29], which is based on the idea that
synaptic growth among two neurons is promoted by their concomitant stimulation. Since
we are still not aware of the synaptic distributions that result in topologies which represent
the structure of associations of symbols in language and thought, as a first approximation,
we have allocated long-range synapses randomly among clusters of neurons (for a more
detailed discussion see [1, 7]). Within a randomly chosen cluster C, defined by one of
the m neurons which is the center of representation of a stimulus (step 3 of the clustering
algorithm), a neuron ni is chosen to receive a long-range connection with probability Pi =
Pj wij/Pnj ∈CPk wjk. If the long-range synapse connects clusters in different memory
sheets (sensorial and symbolic memories), its randomly chosen weight is multiplied by a
real number ζ in the interval (0, 1], reflecting the fact that, in neurotic patterns, sensorial
information is weakly accessible to consciousness, i.e., repressed.
B. Memory Retrieval
The topologies generated with the clustering algorithm have a hierarchical structure
and the average node-degree distributions present an asymptotic power-law behaviour [1].
The functioning of memory retrieval was originally modelled by a Boltzmann Machine
(BM) [27, 30]. There is no theoretical indication of the exact relation between network
9
topology and memory access dynamics. There have been indications that complex physical
systems characterized by spatial disorder and/or long-range interactions, often presenting
power-law behaviour (are asymptotically scale invariant) may be described by the Nonex-
tensive Statistical Mechanics (NSM) formalism [9 -- 15]. The power-law and generalized q-
exponential behaviour for the node-degree distributions of the network topologies generated
be the clustering algorithm [1] indicate that they may not be well described by Boltzmann-
Gibbs (BG) statistical mechanics, but rather by NSM [9]. Memory access was thus modelled
by a generalization of the BM called Generalized Simulated Annealing (GSA) [1, 9], and
this changes the chain of associations generated by the model.
In the BM [27, 30], the N nodes in the neural network are connected symmetrically
by weights wij = wji. The state Si of each unit ni takes output values in {0, 1}. As a
consequence of the symmetry of the connections, one can associate an energy functional
H({Si}) = −
wijSiSj ,
1
2Xij
(2)
to network state S = {Si} and, according to the BG distribution, the transition probability
(acceptance probability) from state S to S′, if H(S′) ≥ H(S), is given by
PBG(S → S′) = exp(cid:20) H(S) − H(S′)
T
(cid:21) ,
(3)
PGSA(S → S′) =
where T is the network "temperature" parameter. Pattern retrieval on the net is achieved
by a standard simulated annealing process, in which T is gradually lowered by a factor
α, according to the BG distribution [27]. In the NSM formalism, one uses a generalized
acceptance probability [9] for a transition from S to S′, if H(S′) ≥ H(S), given by
1
[1+(qA−1)(H(S ′)−H(S))/T ]1/(qA −1) ,
0 ,
if ϕ > 0 ,
if ϕ ≤ 0 ,
(4)
where qA is a parameter called q-acceptance and ϕ = 1 + (qA − 1)(H(S′) − H(S))/T . In
If one
the limit qA → 1, (4) reduces to the Boltzmann-Gibbs transition probability (3).
uses transition probability (4) in place of transition probability (3), in the standard BM
simulated annealing algorithm, the resulting procedure is called GSA (see [9] for a more
detailed discussion).
It is convenient that we define, for each transition from state S to S′ during annealing,
the quantity
∆E = H(S′) − H(S) .
10
(5)
Both the BM and GSA differ from a gradient descent minimization scheme since, besides
allowing state transitions that lower the total energy of the network (2), they also allow the
system to transition into a state with an increase in energy, depending on the values of T
and qA, according to (3) and (4). The BG transition probability (3) predominantly allows
changes of states with small increases in energy, and state transitions with higher energy
increases occur with almost negligible probabilities. The BM will thus strongly prefer visiting
state space within a nearby energy neighbourhood from the starting point (initial state of
the annealing process). The GSA transition probability (4) allows state transitions with
higher energy increases than the BM and, although the probabilities for these transitions
are very low, they are still considerable when compared to the BM. This allows the system to
transition into attractor states that are farther in state space from the initial network state
and also into basins of attraction corresponding to higher energy values [1, 31] (see Figure 2
in [31]). This implies that the hierarchically structured memory modules having both long
and short-range connections, with memory access modelled by GSA achieves associations
among memory traces that are not achieved by the BM [31]. This increase in associativity
observed in the GSA memory retrieval mechanism, when compared to the BM, suggests
that if memory functions according to the NSM theoretical framework, one will have a more
creative mode of memory and mental functioning.
In traditional neural network modelling, the temperature parameter is inspired by the fact
that biological neurons fire with variable strength, and that there are delays in synapses,
random fluctuations from the release of neurotransmitters, and so on. These effects can
be considered as noise in synaptic functioning [23, 27], and we may thus consider that
temperature in the BM and GSA and the qA parameter in GSA control noise. In the model
we are considering [1], non-zero temperature and qA 6= 1 values regulate associations among
memory traces, in an analogy with the concept that freely talking in analytic sessions and
stimulation from the psychoanalyst lower resistances and allow greater associativity and
creativity.
Once the network topology is generated by the clustering algorithm, one can find the
stored patterns by presenting many random patterns to both the BM and the GSA mech-
anism, with an annealing schedule α that allows stabilizing onto the many local minimum
values of the network energy function. Each of the minimum energy values corresponds to
a stable state of the network and is associated with a stored memory trace. These initially
11
stored patterns represent the neurotic memory attractors (as in the compulsion to repeat),
since they are associated with the two weakly linked sensorial and symbolic subnetworks.
In [26], Carvalho et al. proposed a neurocomputational model to describe how the original
memory traces are formed in cortical maps. It is important to note that in our experiments,
we are not aiming at finding a global minimum energy state, but at visiting the many local
minima of the energy landscape, which represent stored information in a cognitive network.
III. AVALANCHES IN MEMORY ACCESS
In the present work, we have defined and measured avalanche sizes during the memory
retrieval process. The size of an avalanche during the simulated annealing procedure, both in
the BM and in GSA, is defined to be the number of state changes that occur during annealing,
from the initial to the final network state, which corresponds to one of the minima in the
energy landscape of the network. This is equivalent to saying that the size of an avalanche
corresponds to the number of hops (steps) that the access mechanism performs on the
network energy landscape, from an initial state until it reaches one of the local minima.
The avalanche size should thus be proportional to the time for the propagation of a signal
(stimulus), during access of information in memory.
In other words, the avalanche size
should be related to the time associated with memory retrieval and also to the size of the
neuronal region activated during memory access.
In each of the simulation experiments we present here, we performed 2,048,000 minimiza-
tion (annealing) procedures, starting each one from a different random network configura-
tion. Since the simulations require much computational time (many days and even a few
weeks), we have analysed smaller networks with total number of neurons N = 32, such that
Nsens = Nsymb = 16 neurons belong to the sensorial and symbolic modules, respectively.
This is a small network size when one considers the brain, as there are billions of neurons
in the human brain. However, the brain is considered to be scale invariant [22] and because
the short-range mechanisms in the algorithms we use to create network topology are scal-
able, we expect that our experiments should, to some extent, qualitatively be comparable
to biological processes [1].
The annealing schedule was controlled so that the network stabilizes on the many local
minima of the energy landscape, and was the same for both machines in all experiments. We
12
conducted experiments with different values of the temperature T and qA parameters. Both
of these parameters model stochastic fluctuations in network functioning and we studied
how the change in T and qA values caused a change in the behaviour of the two machines,
thus affecting the distribution of avalanche sizes, the network energy loss and measurements
of correlations among energy increments along the path followed during annealing.
In Figure 1, we show the avalanche size distributions obtained with the two machines,
for T = 0.05 and qA = 1.3. In this experiment, when executing step 2 of the clustering
algorithm, 50% of the synapses of the network are excitatory (positive synaptic weights)
and 50% are inhibitory (negative synaptic weights). On the left, we see the avalanche size
distributions for the BM and on the right, for GSA. From the distribution in Figure 1-a, it
is clear that the BM has a preferred avalanche size of 125 with an associated probability of
occurrence of 0.96. This distribution also has a smaller peak at avalanche size 1325, which
occurs with probability 0.0022. The smaller peak has a slower decay before a sharp drop.
The distribution for GSA in Figure 1-b, on the other hand, behaves quite differently.
For the range of avalanche sizes we measured, we observe a monotonic drop in the fre-
quency for GSA. The points measured in this simulation can be approximately fitted by a
q-exponential [10]
expq(x) =
[1 + (1 − q)x]
0 ,
1
1−q ,
if 1 + (1 − q)x > 0 ,
if 1 + (1 − q)x ≤ 0 .
(6)
This function is at the core of nonextensive thermostatistics, being the result of the con-
strained optimization of the power-law, nonadditive entropic functional
Sq =
1
q − 1
[1 −Xi
pq
i ] ,
(7)
where pi is a normalized probability distribution and q is a real parameter. The q-exponential
asymptotically behaves as a power-law and, in Figure 1-b, the curve which fits the points
within the observed values of avalanche sizes, corresponding to q = qf = 1.19, does show an
asymptotic behaviour akin to a power-law.
The imaging technique in neuroscience referred to as functional magnetic resonance imag-
ing (fMRI) uses a method known as blood-oxygen-level dependent (BOLD) contrast imag-
ing [32]. Most cells in the body possess their own reservoirs of sugar which undergoes
respiration to produce energy, when the cells need it.
It is known that neurons do not
13
101
100
10-1
10-2
10-3
10-4
10-5
10-6
y
c
n
e
u
q
e
r
F
10-7
10
BM, T = 0.05
100
1000
10000
Avalanche Size
(a)
102
101
100
10-1
10-2
10-3
10-4
10-5
10-6
10-7
y
c
n
e
u
q
e
r
F
10-8
100
GSA, T = 0.05, qA = 1.3
0.09*(( 1-(1-1.19)*( (x - 360) /112.0 ) )**(1/(1-1.19)))
1000
Avalanche Size
(b)
10000
FIG. 1:
(a) Frequency of occurrence of avalanche sizes for the BM. (b) Frequency of avalanche
sizes with GSA. For both machines, T = 0.05 and for GSA, qA = 1.3. In both cases, 50% of the
synapses are inhibitive. All depicted quantities are dimensionless.
possess internal stores of energy and, when a neuron is in a firing state it requires energy
in the form of sugar and oxygen, which it obtains by means of a haemodynamic response,
whereby blood releases oxygen to the firing neurons at a greater rate in comparison to those
in resting state. This causes a difference in magnetic susceptibility between oxygenated and
deoxygenated blood, which results in a magnetic signal variation that may be detected on
an MRI scanner [32]. Studying changes in the brain BOLD signal allows observation of
different active areas of the brain to further understand certain aspects of brain functioning.
In 2003, the work of Beggs and Plenz [33] with cortical slices revealed the phenomena
which they called neuronal avalanches in the brain. In those observations, avalanches were
defined as short bursts of activity that last a few milliseconds, followed by several seconds
of inactivity. These phenomena can be observed in cortical slices of the neocortex, although
their relation to physiological processes in the brain is still unknown [33]. The avalanche
sizes detected by Beggs and Plenz follow a power-law-like distribution, with an exponent
of -3/2. In their observations, they recorded spontaneous local field potentials continuously
using a 60 channel multielectrode array, and avalanches are a measure of time duration and
spatial reach of the propagation of signals in the brain.
Tagliazucchi et al. [22] extended the work done by Beggs and Plenz [33], "by inspecting
only the relatively large amplitude BOLD signal peaks, suggesting that relevant information
14
can be condensed in discrete events". The authors present a spatiotemporal point process,
where the timing and location of these discrete events is obtained to study and capture the
dynamics of the brain in a resting state. When regions of the brain show activity above a
certain threshold level, the detected BOLD signal determines the location of the activity as
well as the duration. Figure 3-A in [22] shows examples of co-activated clusters, defined as
"... groups of contiguous voxels with signal above the threshold at a given time" (clusters
are measured in units of voxels).
In [22], Figure 3-D shows the distribution of average measured cluster sizes in fMRI im-
ages for ten individuals, which has a power-law (scale-free) behaviour spanning four orders
of magnitude, also with an exponent of approximately -3/2. The size of a cluster is propor-
tional to the duration of brain activity. The longer the duration for which an avalanche is
seen in the brain, the larger are the observed cluster sizes. This happens because when the
duration of activity is longer, clusters (voxels) cause the activation of neighbouring clusters,
before reducing activity and fading away. This behaviour is comparable to the avalanche size
distributions produced by our computational model, as in Figure 1-b, because an avalanche
size, in our simulation experiments, was defined in a way so that it should be proportional
to the time for propagation of a signal during the access of information in memory. This is
a reasonable assumption, since transition probabilities (3) and (4) which regulate network
functioning should mimic the way the memory network occupies phase-space, during the re-
trieval of a mnemenic trace. So the larger avalanche sizes observed in our simulations should
be associated with larger cluster sizes, as a larger avalanche size is associated with a longer
time for propagation of a signal, with access to a larger number of neighbouring neurons. We
therefore qualitatively compared the behaviour of the distribution of cluster sizes obtained
in [22] with the distribution of avalanche sizes obtained in our simulation experiments, for
some values of the T and qA parameters, and we found q-exponentials with a similar asymp-
totic power-law like behaviour, spanning similar orders of magnitude of avalanche sizes and
frequencies, only for the GSA machine, however with larger absolute values of the exponent
(exponent close to -5 in Figure 1-b and approximately -10 in Figure 4-b). The Boltzmann
machine did not present power-laws in our experiments. This qualitatively corroborates our
initial indication, that the power-law and generalized q-exponential behaviour of quantities
which describe the structure of the complex networks generated by the clustering algorithm
of our model suggest that they may not be well described by Boltzmann-Gibbs (BG) statis-
15
tical mechanics, but rather by NSM [1, 9].
For a sequence of values of energy increments ∆E given by (5) during an avalanche, we
also measured if the value ∆E(τ0) at step τ0 influences (is correlated to) the value at a later
step ∆E(τ0 + τ ). The temporal correlation function G(τ ) among elements in a sequence
separated by an interval τ is defined as [34]
G(τ ) =
m−τ
Xτ0=1
∆E(τ0)∆E(τ0 + τ )
m − τ
− m
Xτ0=1
∆E(τ0)
m !2
,
(8)
where m is the size of the sequence which is being considered, and should be much larger
than the largest value of τ in the domain of (8).
In Figure 2, we show correlations for one avalanche of the BM in Figure 2-a and for
another avalanche generated by GSA in Figure 2-b. Since avalanches in the BM and GSA
typically have different sizes, Figures 2-a and b correspond to different values of m. For
both machines we notice an approximate exponential fall of G(τ ) for smaller values of τ ,
with larger fluctuations around an average decay for GSA. These larger fluctuations for GSA
occur because GSA generates higher probabilities of making state transitions with positive
values of ∆E than the BM. After a rapid decrease, both figures converge to a situation
where they fluctuate around zero correlation values. A similar behaviour occurs for the
other avalanches which we measured. For larger values of T and qA, the fluctuations around
average correlation values are much larger, since the machines then have higher probabilities
of making state transitions with positive values of ∆E, and the first term in (8) is negative
more often than for lower values of T and qA.
We show avalanche size distributions obtained with the two machines, for T = 0.2 and
qA = 1.6, in Figure 3. As a consequence of the increase of the values of the two parameters
that control associativity and noise in the network, both machines have a higher probability
of increasing energy during the annealing process, and therefore achieve higher values of
avalanche sizes more frequently. The BM now has a preferred avalanche size of 1025, with
a corresponding probability of 0.31, as seen from Figure 3-a. The distribution of avalanche
sizes for GSA now loses the general q-exponential behaviour and has two preferred values
(the first is 4775 with 0.035 probability and the second is 6375 with probability 0.056), with a
very short power-law like behaviour after the second peak, before a sharp drop. The increase
in frequency of larger avalanche sizes for both machines is consistent with the increase in
associativity for larger values of T and qA found in [31], with GSA still producing more
16
)
τ
(
G
n
o
i
t
l
a
e
r
r
o
C
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
-0.005
0
GSA, Avalanche Size = 2873
BM, Avalanche Size = 1593
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
)
τ
(
G
n
o
i
t
l
a
e
r
r
o
C
5
10
τ
(a)
15
20
25
-0.01
0
5
10
15
25
30
35
40
20
τ
(b)
FIG. 2:
(a) Correlations among the sequence of ∆E values during an avalanche of the BM. The
avalanche size for this experiment is m = 1593 and the maximum value of τ is 25. (b) Correlations
among the ∆E for an avalanche generated by GSA. The avalanche size for this experiment is
m = 2873 and the maximum value of τ is 40. For both machines, T = 0.05 and for GSA, qA = 1.3.
50% of the synapses are inhibitive. All depicted quantities are dimensionless.
associativity than the BM.
There has been recent work regarding the study of the proportion of inhibitory synapses in
the brain (see [35] and references therein). In mammals, this proportion has been measured
to have values between 20 and 30%. We have thus conducted avalanche measurements in our
model so that, when executing step 2 of the clustering algorithm, 70% of randomly chosen
synapses of the network are excitatory (positive synaptic weights) and 30% are inhibitory
(negative synaptic weights). A distribution of avalanche sizes for this proportion of inhibitory
synapses, for both the BM and GSA with T = 0.2, and qA = 0.7 in GSA is shown in Figure 4.
Again we see a preferred avalanche size for the BM, in this case at 1025 with probability
0.32. We also observe once more for GSA in Figure 4-b, for the range of avalanche sizes
which we measured, a monotonic decrease in the distribution of avalanche sizes, so that
the experimental points resulting from this simulation can be approximately fitted by a
q-exponential, with q = qf = 1.098. This is also asymptotically in accordance with the
scale-free, power-law behaviour observed experimentally in [22].
Besides the avalanche size distributions and temporal correlations, we also measured the
17
100
10-1
10-2
10-3
10-4
10-5
10-6
y
c
n
e
u
q
e
r
F
10-7
10
GSA, T = 0.2, qA = 1.6
BM, T = 0.2
10-1
10-2
10-3
10-4
10-5
10-6
y
c
n
e
u
q
e
r
F
100
1000
10000
Avalanche Size
10-7
3000
4000
5000
6000
Avalanche Size
7000
8000
9000 10000
(a)
(b)
FIG. 3:
(a) Frequency of occurrence of avalanche sizes for the BM. (b) Frequency of avalanche
sizes with GSA. For both machines, T = 0.2 and for GSA, qA = 1.6. In both cases, 50% of the
synapses are inhibitive. All depicted quantities are dimensionless.
y
c
n
e
u
q
e
r
F
101
100
10-1
10-2
10-3
10-4
10-5
10-6
10-7
BM, T = 0.2
101
100
10-1
10-2
10-3
10-4
10-5
y
c
n
e
u
q
e
r
F
10-6
100
GSA, T = 0.2, qA = 0.7
0.35*(1-(1-1.098)*((x - 160)/161.0))**(1/(1-1.098))
1000
Avalanche Size
(b)
100
Avalanche Size
1000
(a)
FIG. 4:
(a) Frequency of occurrence of avalanche sizes for the BM. (b) Frequency of avalanche
sizes with GSA. For both machines, T = 0.2 and for GSA, qA = 0.7. In both cases, 30% of the
synapses are inhibitive. All depicted quantities are dimensionless.
distributions of the total energy lost by the network, during the annealing processes of each
set of simulation experiments, for both machines, i.e. H(SF inal) − H(SInit) , where SF inal
and SInit are respectively the final and initial states of the network, during one memory
retrieval procedure. These distributions are basically the same for both machines during the
18
same experiment, with just a very slight and negligible difference close to the tales of the
distributions, due to limited size statistics.
IV. CONCLUSIONS
We have further studied the memory mechanism in a schematic neural network model,
which illustrates conscious and unconscious mental processes in neurotic mental function-
ing [1 -- 5]. The model illustrates aspects of mental functioning related to memory associa-
tivity, creativity, consciousness and unconsciousness, which are present both in pathological
and normal mental processing. The model is based on a self-organizing, clustering algorithm
which generates a hierarchically structured bimodular neural network, based on basic biolog-
ical cellular mechanisms, which have aspects akin to a preferential attachment mechanism,
constrained by an energy conservation (conservation of neurosubstances) prescription, con-
trolling synaptic plasticity. The networks generated by the algorithm have a small-world-like
topology, with clusters of neurons strongly coupled by short-range connections (synapses)
and also less frequent long-range connections. The node-degree distributions of these net-
work topologies present a generalized q-exponential and asymptotic power-law behaviour.
This structure suggests that memory access may best be modelled by a generalization of the
Boltzmann Machine called Generalized Simulated Annealing, and this determines the chain
of associations generated by the model.
In the present work, we have introduced a definition of avalanche sizes during the simu-
lated annealing procedure of the BM and GSA machine, which model memory access. Both
the BM and GSA are regulated by parameters (temperature T and qA for the NSM formal-
ism) which control stochastic fluctuations in network functioning and model noisy behaviour
present in biological synaptic mechanisms. We conducted simulation experiments with dif-
ferent values of T and qA, and studied how this causes a change in the behaviour of the
two machines, thus affecting the distribution of avalanche sizes, the network energy loss and
measurements of correlations among energy increments along the path followed during an-
nealing. In these experiments, we found that for some values of T and qA, the avalanche size
distributions of the GSA machine may be approximately fitted by a q-exponential, with the
corresponding asymptotic power-law behaviour. The BM does not generate this behaviour
and produces smaller avalanche sizes, related to less associative memory functioning, when
19
compared to GSA.
We conjecture that avalanche sizes defined for our model should be related to the time
consumed and also to the size of the neuronal region which is activated, during access of
information in memory. This assumption is reasonable, since the transition probabilities (3)
and (4) which regulate the BM and GSA network functioning should mimic the way the
memory apparatus occupies phase-space, during the retrieval of a mnemenic trace. We may
therefore qualitatively compare the behaviour of the distribution of cluster sizes, obtained
during fMRI measurements of the propagation of signals in the brain reported in [22], with
the distribution of avalanche sizes obtained in our simulation experiments. For some values of
the T and qA parameters, we find a similar power-law like behaviour, spanning similar orders
of magnitude of avalanche sizes and frequencies, only when using the GSA machine, which
is based on the NSM formalism. The BM did not produce this pattern in our simulation
experiments. This result experimentally corroborates our original indication, that the q-
exponential and asymptotic power-law (scale-free) behaviour of macroscopic quantities which
describe the structure of the complex networks generated by the algorithm of our model
suggest that they may indeed be better described by the NSM [9] formalism, rather than by
Boltzmann-Gibbs (BG) statistical mechanics.
We have also considered recent indications that the brains of mammals are composed
of a proportion of inhibitive synapses, which has been measured to have values between
20 and 30% [35]. We therefore measured avalanche size distributions during memory ac-
cess, in networks generated by the clustering algorithm of our model with a proportion of
30% of inhibitive synapses.
In this case, we also found that, for some T and qA values,
the avalanche size distributions may be fitted by a q-exponential (with the corresponding
asymptotic power-law behaviour), only for the GSA machine. Indeed, the q-exponential fit
for the case we analysed seems even better with this proportion of inhibitive synapses than
with 50% of inhibitive synapses.
Although the fMRI measurements presented in [22], capture the dynamics of the brain
in a resting state, the authors argue that "... despite the fact that in resting data there are
not explicit inputs, the average BOLD signal around the extracted points ... still resembles
the hemodynamic response function (HRF) evoked by a stimulus". This further supports
the qualitative comparison of the results of our simulation experiments with their fMRI
measurements.
20
We may thus conclude that the computational model studied in this work is viable as an
illustrative schematic model of some basic mental mechanisms and in fact is qualitatively
comparable to fMRI data obtained from patients, demonstrating actual brain activity. This
comparison also reinforces the idea that complex systems such as the brain may well be
better described by Nonextensive Statistical Mechanics, which produces the asymptotic
power-law behaviour, for various temperatures and qA values, while Boltzmann-Gibbs sta-
tistical mechanic does not show such behaviour. In [33], the authors argue that in the critical
states characterized by power-laws, "the network may satisfy the competing demands of in-
formation transmission and network stability". As demonstrated in earlier work regarding
this model [1, 31], as well as in our simulations, memory retrieval governed by Nonex-
tensive Statistical Mechanics allows much more associativity in memory functioning than
Boltzmann-Gibbs statistical mechanics, while still achieving the stability necessary for infor-
mation storage (see also [36]). This results in a potentially more creative mental structure
with more capacity for the establishment of metaphors. From the perspective of psychother-
apy, patients with more creativity have a larger possibility of reassociating traumatic and
repressed experiences, and to construct new historical perspectives for their present and fu-
ture, during therapy. Further investigations along this line of research are important, since
they may reveal more basic structural features of brain and mental functioning.
Acknowledgments
We are grateful to Constantino Tsallis, Evaldo Curado, Angel R. Plastino and Abbas
Edalat for fruitful conversations. This research was developed with grants from the Brazil-
ian National Research Council (CNPq), the Rio de Janeiro State Research Foundation
(FAPERJ) and the Brazilian agency which funds graduate studies (CAPES).
[1] R.S. Wedemann, R. Donangelo, L.A.V. de Carvalho, Generalized Memory Associativity in a
Network Model for the Neuroses, Chaos 19 (2009) 015116-(1 -- 11).
[2] S. Freud, Beyond the Pleasure Principle, Standard Edition of The Complete Psychological
Works of Sigmund Freud, Vol. XI, First German edition in 1920, The Hogarth Press, London,
1974.
21
[3] S. Freud, Remembering, Repeating and Working-Through, Standard Edition of The Complete
Psychological Works of Sigmund Freud, Vol. XII, First German edition in 1914, The Hogarth
Press, London, 1953.
[4] S. Freud, Introductory Lectures on Psycho-Analysis, Standard Edition, First German edition
in 1917, W. W. Norton and Company, New York - London, 1966.
[5] S. Freud, The Unconscious, Standard Edition of The Complete Psychological Works of Sig-
mund Freud, First German edition in 1915, Vol. XIV, The Hogarth Press, London, 1957.
[6] R.S. Wedemann, R. Donangelo, L.A.V. de Carvalho, I.H. Martins, Memory Functioning in
Psychopathology, Lecture Notes in Computer Science 2329 (2002) 236 -- 245.
[7] R.S. Wedemann, L.A.V. de Carvalho, R. Donangelo, A Complex Neural Network Model for
Memory Functioning in Psychopathology, Lecture Notes in Computer Science 4131 (2006)
543 -- 552.
[8] A. Cleeremans, B. Timmermans, A. Pasquali, Consciousness and metarepresentation: A com-
putational sketch, Neural Networks 20 (2007) 1032 -- 1039.
[9] C. Tsallis, D.A. Stariolo, Generalized simulated annealing, Physica A 233 (1996) 395 -- 406.
[10] C. Tsallis, Introduction to Nonextensive Statistical Mechanics, Approaching a Complex World,
Springer, New York, 2009.
[11] C. Beck, Generalised information and entropy measures in physics, Contemporary Physics
50(4) (2009) 495 -- 510.
[12] C. Vignat, A. Plastino, A.R. Plastino, Entropic upper bound on gravitational binding energy,
Physica A 390 (2011) 2491 -- 2496.
[13] U. Tirnakli, E.P. Borges, The standard map: From Boltzmann-Gibbs statistics to Tsallis
statistics. Nature Sci. Rep. 6 (2016) 23644.1 -- 8.
[14] S. Brito, L.R. da Silva, C. Tsallis, Role of dimensionality in complex networks. Nature Sci.
Rep. 6 (2016) 27992.1 -- 8.
[15] M.S. Ribeiro, F.D. Nobre, E.M.F. Curado, Classes of N-Dimensional Nonlinear Fokker-Planck
Equations Associated to Tsallis Entropy, Entropy 13(11), (2011) 1928 -- 1944.
[16] A. Edalat, F. Mancinelli, Strong Attractors of Hopfield Neural Networks to Model Attachment
Types and Behavioural Patterns, in: P. Angelov, D. Levine (Eds.), Proceedings of The 2013
International Joint Conference on Neural Networks (IJCNN), IEEE - Curran Associates, Inc.,
New York, 2013, pp. 1-10.
22
[17] A. Edalat, F. Mancinelli, Introduction to self-attachment and its neural basis, in: D.S. Huang,
Y. Choe (Eds.), Proceedings of The 2015 International Joint Conference on Neural Networks
(IJCNN), IEEE - Curran Associates, Inc., New York, 2015, pp. 1-8.
[18] E. Kandel, Psychiatry, Psychoanalysis, and the New Biology of Mind, American Psychiatric
Publishing, Inc., Washington D.C., London, 2005.
[19] Shedler, J., The Efficacy of Psychodynamic Psychotherapy, American Psychologist 65(2)
(2010) 98 -- 109.
[20] Lacan, J.: On a Discourse that Might not be a Semblance. The Seminar of Jacques La-
can, Book XVIII. Translated by Cormac Gallagher from unedited French manuscripts,
http://www.lacaninireland.com, (accessed 07.04.17) Seminar in France (1971), French edi-
tion, Editions du Seuil, 2007.
[21] J. Lacan, O Semin´ario, Livro 8: A Transferencia, Jorge Zahar, Rio de Janeiro, 1992, First
French edition in 1991.
[22] E. Tagliazucchi, P. Balenzuela, D. Fraiman, D. R. Chialvo, Criticality in large-scale brain
fMRI dynamics unveiled by a novel point process analysis, Frontiers in Physiology -- Fractal
Physiology 3, (2012) Article 15.
[23] E.R. Kandel, J.H. Schwartz, T.M. Jessel (Eds.), Principles of Neural Science, MacGraw Hill,
New York, 2000.
[24] H. Hartline, F. Ratcliff, Inhibitory Interactions of Receptor Units in the Eye of Limulus,
Journal of General Physiology 40 (1957) 357 -- 376.
[25] W.F. Ganong, Review of Medical Physiology, MacGraw Hill, USA, 2003.
[26] L.A.V. de Carvalho, D.Q. Mendes, R.S. Wedemann, Creativity and Delusions: The Dopamin-
ergic Modulation of Cortical Maps, Lecture Notes in Computer Science 2657 (2003) 511 -- 520.
[27] J.A. Hertz, A. Krogh, R.G. Palmer, Introduction to the Theory of Neural Computation,
Perseus Books, Cambridge, Massachusetts, USA, 1991.
[28] G.M. Edelman, Wider than the Sky, a Revolutionary View of Consciousness, Penguin Books,
London, 2005.
[29] E.R. Kandel, I. Kupfermann, S. Iversen, Learning and Memory,
in: E.R. Kandel, J.H.
Schwartz, T.M. Jessel (Eds.), Principles of Neural Science, fourth ed., MacGraw Hill, New
York, 2000, pp. 1227-1246.
[30] W.C. Barbosa, Massively Parallel Models of Computation, Ellis Horwood Limited, England,
23
1993.
[31] R.S. Wedemann, L.A.V. de Carvalho, R. Donangelo, Access to Symbolization and Associativ-
ity Mechanisms in a Model of Conscious and Unconscious Processes, in: A. V. Samsonovich
and K. R. J´ohannsd´ottir (Eds.), Biologically Inspired Cognitive Architectures, Frontiers in
Artificial Intelligence and Applications, IOS Press, USA, 2011, pp. 444 -- 449.
[32] S.A. Huettel, A.W. Song, G. McCarthy, Functional Magnetic Resonance Imaging, second ed.,
Sinauer, Massachusetts, USA (2009).
[33] J.M. Beggs, D. Plenz, Neuronal avalanches in neocortical circuits, Journal of Neuroscience 23
(2003) 11167 -- 11177.
[34] H.J. Jensen, Self-Organized Criticality, Emergent Complex Behavior in Physical and Biological
Systems, Cambridge University Press, Cambridge, UK, 1998.
[35] V. Capano, H.J. Herrmann, L. de Arcangelis, Optimal percentage of inhibitory synapses in
multi-task learning, Nature Sci. Reps. 5 (2015) Article 9895.
[36] R.S. Wedemann, A.R. Plastino, Asymmetries in Synaptic Connections and the Nonlinear
Fokker-Planck Formalism. Lecture Notes in Computer Science 9886 (2016) 19 -- 27.
24
|
1608.01179 | 1 | 1608 | 2016-08-03T13:14:05 | Normalized neural representations of natural odors | [
"q-bio.NC",
"physics.bio-ph"
] | The olfactory system removes correlations in natural odors using a network of inhibitory neurons in the olfactory bulb. It has been proposed that this network integrates the response from all olfactory receptors and inhibits them equally. However, how such global inhibition influences the neural representations of odors is unclear. Here, we study a simple statistical model of this situation, which leads to concentration-invariant, sparse representations of the odor composition. We show that the inhibition strength can be tuned to obtain sparse representations that are still useful to discriminate odors that vary in relative concentration, size, and composition. The model reveals two generic consequences of global inhibition: (i) odors with many molecular species are more difficult to discriminate and (ii) receptor arrays with heterogeneous sensitivities perform badly. Our work can thus help to understand how global inhibition shapes normalized odor representations for further processing in the brain. | q-bio.NC | q-bio | Normalized neural representations of natural odors
David Zwicker1, 2, ∗
1School of Engineering and Applied Sciences, Harvard University, Cambridge, MA 02138, USA
2Kavli Institute for Bionano Science and Technology, Harvard University, Cambridge, MA 02138, USA
(Dated: July 3, 2018)
The olfactory system removes correlations in natural odors using a network of inhibitory
neurons in the olfactory bulb. It has been proposed that this network integrates the re-
sponse from all olfactory receptors and inhibits them equally. However, how such global
inhibition influences the neural representations of odors is unclear. Here, we study a
simple statistical model of this situation, which leads to concentration-invariant, sparse
representations of the odor composition. We show that the inhibition strength can be
tuned to obtain sparse representations that are still useful to discriminate odors that
vary in relative concentration, size, and composition. The model reveals two generic
consequences of global inhibition: (i) odors with many molecular species are more dif-
ficult to discriminate and (ii) receptor arrays with heterogeneous sensitivities perform
badly. Our work can thus help to understand how global inhibition shapes normalized
odor representations for further processing in the brain.
I. INTRODUCTION
Sensory systems encode information efficiently by re-
moving redundancies present in natural stimuli (Barlow,
2001, 1961). In natural images, for instance, neighbor-
ing regions are likely of similar brightness and the image
can thus be characterized by the regions of brightness
changes (Ruderman and Bialek, 1994). This structure
is exploited by ganglion cells in the retina that respond
to brightness gradients by receiving excitatory input from
photo receptors in one location and inhibitory input from
the surrounding (Demb and Singer, 2015). This typi-
cal center-surround inhibition results in neural patterns
that represent natural images efficiently (Carandini and
Heeger, 2012). Similarly, such local inhibition helps sepa-
rating sound frequencies in the ear and locations touched
on the skin (Isaacson and Scanziani, 2011). Vision, hear-
ing, and touch have in common that their stimulus spaces
have a metric for which typical correlations in natural
stimuli are local. Consequently, local inhibition can be
used to remove these correlations and reduce the high-
dimensional input to a lower-dimensional representation.
The olfactory stimulus space is also high-dimensional,
since odors are comprised of many molecules at different
concentrations. Moreover, the concentrations are also
often correlated, e. g., because the molecules originate
from the same source. However, these correlations are
not represented by neighboring neurons in the olfactory
system, since there is no obvious similarity metric for
molecules that could be used to achieve such an arrange-
ment (Nikolova and Jaworska, 2003). Because the olfac-
tory space lacks such a metric, local inhibition cannot be
used to remove correlations to form an efficient represen-
tation (Murthy, 2011; Soucy et al., 2009). Consequently,
∗ http://www.david-zwicker.de
the experimentally discovered inhibition in the olfactory
system (Yokoi, Mori, and Nakanishi, 1995) likely affects
neurons irrespective of their location. Such global inhi-
bition could for instance normalize the activities by their
sum, which has been observed experimentally (Olsen,
Bhandawat, and Wilson, 2010; Roland et al., 2016). This
normalization cannot reduce the correlation structure of
odors, but it could help separating the odor composition
(what is present?) from the odor intensity (how much is
there?) (Cleland, 2010; Laurent, 1999). This separation
is useful, since the composition identifies an odor source,
while the intensity information is necessary for finding
or avoiding it. However, how global inhibition shapes
such a bipartite representation of natural odors is little
understood.
In this paper, we study a simple model of the ol-
factory system that resembles its first processing lay-
ers, which transform the odor representation succes-
sively (Silva Teixeira, Cerqueira, and Silva Ferreira, 2016;
Wilson, 2013), see Fig. 1. Our model connects previous
results from simulations of the neural circuits (Cleland
and Sethupathy, 2006; Getz and Lutz, 1999; Li, 1990,
1994; Linster and Hasselmo, 1997; Zhang, Li, and Wu,
2013) to system-level descriptions of the olfactory sys-
tem (Hopfield, 1999; Koulakov, Gelperin, and Rinberg,
2007; Zwicker, Murugan, and Brenner, 2016). The main
feature of the model is global inhibition, which leads to
normalization. This separates the odor composition from
its intensity and encodes it in a sparse representation.
The inhibition strength controls the trade-off between
the sparsity and the transmitted information, which in-
fluences how well this code can be used to discriminate
odors in typical olfactory tasks. The model reveals two
generic consequences of global inhibition: (i) odors com-
prised of many different molecules exhibit sparser rep-
resentations and should thus be more difficult to distin-
guish and (ii) overly sensitive receptors could dominate
arXiv:1608.01179v1 [q-bio.NC] 3 Aug 2016
2
FIG. 1 Schematic picture of our model describing the signal processing in the olfactory bulb: An odor comprised of many
ligands excites the olfactory receptors and the signals from all receptors of the same type are accumulated in respective glomeruli.
Associated projections neurons receive excitatory input from a single glomerulus and are subject to global inhibition, mediated
by a network of local neurons. The activity of the projection neurons form a sparse, concentration-invariant odor representation.
the sparse responses and arrays with heterogeneous re-
ceptors should thus perform poorly.
II. SIMPLE MODEL OF THE OLFACTORY SYSTEM
Odors are blends of odorant molecules that are lig-
ands of the olfactory receptors. We describe an odor by
a vector c = (c1, c2, . . . , cNL ) that specifies the concen-
trations ci of all NL detectable ligands (ci ≥ 0). Gen-
erally, only a small subset of the NL ∼ 105 ligands are
present in natural odors, so most of the ci will typically
be zero. The ligands in an odor are detected by olfactory
receptor neurons, which reside in the nose in mammals
and in the antenna in insects (Kaupp, 2010). Each of
these neurons expresses receptors of one of NR genet-
ically defined types, where NR ≈ 50 for flies (Wilson,
2013), NR ≈ 300 for humans (Verbeurgt et al., 2014),
and NR ≈ 1000 for mice (Niimura, 2012). The excitation
of all receptor neurons of the same type is accumulated
in associated glomeruli (Su, Menuz, and Carlson, 2009),
whose excitation pattern forms the first odor representa-
tion, see Fig. 1. Here, the large number of ligands and
their possible mixtures are represented by a combina-
torial code, where each ligand typically excites multiple
receptor types (Malnic et al., 1999). It has been shown
experimentally that the excitation en of the glomerulus
associated with receptor type n can be approximated by
a linear function of the ligand concentrations c (Gupta,
Albeanu, and Bhalla, 2015; Silbering and Galizia, 2007;
Tabor et al., 2004),
en =
NLXi=1
Snici ,
(1)
where Sni denotes the sensitivity of glomerulus n to lig-
and i. We here consider a statistical description of com-
binatorial coding by studying random sensitivity matri-
ces with entries drawn independently from a log-normal
distribution. This distribution is parameterized by the
mean sensitivity ¯S and the standard deviation λ of the
underlying normal distribution. This choice is moti-
vated by experimental measurements, which also suggest
that λ ≈ 1 for flies and humans (Zwicker, Murugan,
and Brenner, 2016). We showed previously that such
random matrices typically decorrelate stimuli and thus
lead to near-optimal odor representations on the level of
glomeruli (Zwicker, Murugan, and Brenner, 2016).
In contrast to our previous model, we here consider
the odor representation encoded by projection neurons
(mitral and tufted cells in mammals), which constitute
the next layer after the glomeruli, see Fig. 1. Projec-
tion neurons typically receive excitatory input from a
single glomerulus (Jefferis et al., 2001) and inhibitory
input from many local neurons (granule cells in mam-
mals), which are connected to other projection neurons
and glomeruli (Cleland, 2010; Su, Menuz, and Carlson,
2009). The activity an of the projection neurons associ-
ated with receptor type n is a sigmoidal function of lig-
and concentrations (Bhandawat et al., 2007; Tan et al.,
2010). Additionally, all signals are subject to noise, both
from stochastic ligand-receptor interactions and from in-
ternal processing (Lowe and Gold, 1995), which limits the
number of distinguishable output activities. We capture
both effects by considering the simple case where only
two activities an can be distinguished. Here, the projec-
tion neurons are active when their excitatory input, the
respective excitation en, exceeds a threshold γ,
an =(0
1
en ≤ γ
en > γ .
(2)
Generally, γ could depend on the type n, but we here
consider a simple mean-field model, where all types ex-
hibit the same threshold. Nevertheless, this threshold
could still depend on global variables. Experimental
data (Asahina et al., 2009; Aungst et al., 2003; Banerjee
et al., 2015; Berck et al., 2016; Hong and Wilson, 2015;
Olsen, Bhandawat, and Wilson, 2010; Roland et al., 2016;
Silbering and Galizia, 2007) and modeling of the local
neurons (Cleland, 2010; Cleland and Sethupathy, 2006)
suggest that the total excitation of all glomeruli inhibits
all projection neurons. To capture this we postulate that
Nose (antenna)
Receptors
Olfactory bulb (antennal lobe)
Glomeruli
Projection neurons
Olfactory cortex
(mushroom body)
composition
intensity
Local neurons
YY Y
YYY
Y
Odor c
Excitation e
Activity a
3
is because both the excitations en and the threshold γ
are linear in c and Sni, see Eqs. 1 and 3, and the ac-
tivities an only depend on the ratio en/γ, see Eq. 2. In
fact, these equations can be interpreted as normaliza-
tion of the excitations by the total excitation followed by
thresholding with the constant threshold α/NR. Since
the representation a does not depend on ctot, it only en-
codes relative ligand concentrations, i. e., the odor com-
position. This property is called concentration invariance
and corresponds to the everyday experiences that odors
smell the same over many orders of magnitude in concen-
tration (Cleland et al., 2011; Uchida and Mainen, 2007;
Zhang, Li, and Wu, 2013). Indeed, experiments suggest
that the activity of projection neurons is concentration-
invariant (Cleland et al., 2007; Sachse and Galizia, 2002;
Sirotin, Shusterman, and Rinberg, 2015) and exhibits
more uniform distances between odors (Bhandawat et al.,
2007; Cleland et al., 2007), indicating that they encode
the odor composition efficiently.
To understand how odor compositions are encoded in
our model, we start with numerical simulations of Eqs. 1–
3 as described in the SI. Fig. 2A shows the excitations en
corresponding to an arbitrary odor. Here, the excitation
threshold is 1.4 times the mean excitation, and only three
channels are active (orange bars). The corresponding his-
togram in Fig. 2B shows that the number of active chan-
nels is typically small for this inhibition strength when
odors are presented with statistics Penv(c). Moreover,
the magnitude of the Pearson correlation coefficient be-
tween two channels is typically only 1 %, see SI. This
weak correlation is expected for the uncorrelated odors
and random sensitivity matrices that we consider here
and explains why the histogram in Fig. 2B is close to a
binomial distribution. The odor representations are thus
mainly characterized by the mean channel activity hani.
The mean channel activity hani depends on the in-
hibition strength α, the sensitivities Sni, and the odor
statistics Penv(c). To discuss these dependences, we next
introduce an approximation based on a statistical de-
scription of the associated excitation en. Here, we define
the normalized concentrations ci = ci/ctot and normal-
ized excitations en = en/(ctot ¯S), since an is independent
of ctot and ¯S. The statistics of ci can be estimated in
the typical case where odors are comprised of many lig-
ands, see SI. In the particular case where the ligands are
identically distributed the mean is hcii = NL−1 and the
2). Gen-
variance reads var(ci) ≈ (1 − p + σ2µ−2)/(pNL
erally, ci varies more if the underlying ci has higher coef-
ficient of variation σi/µi or if the mean odor size s =
pNL is smaller. The normalized excitation en is de-
fined such that its mean is 1 and the associated variance
can be written as a product of the external contribution
Vext =Pihc2
ii due to odors and the internal contribution
Vint = var(Sni)hSnii−2 due to sensitivities, see SI. In the
the threshold γ is a function of the total excitation, where
we for simplicity consider a linear dependence,
(3)
en .
NRXn=1
R
α N
γ =
Here, α is a parameter that controls the inhibition
strength.
Taken together, our model of the olfactory system com-
prises NR communication channels, each consisting of
receptors, a glomerulus, and projection neurons, which
interact via global inhibition, see Fig. 1. The Eqs. 1–3
describe how this system maps an odor c to an activ-
ity pattern a = (a1, a2, . . . , aNR). The amount of infor-
mation that can be learned about c by observing a is
quantified by the mutual information I, which reads
I = −Xa
P (a) log2 P (a) .
(4)
Here, the probability P (a) of observing output a is given
by P (a) = R P (ac)Penv(c) dc. The conditional proba-
bility P (ac) of observing a given c describes the process-
ing in the olfactory system and follows from the Eqs. 1–3.
In contrast, Penv(c) denotes the probability of encoun-
tering an odor c, which depends on the environment.
Consequently, the information I is not only a function of
the sensitivity matrix Sni and the inhibition strength α,
but also of the environment in which the receptors are
used (Zwicker, Murugan, and Brenner, 2016).
Natural odor statistics are hard to measure (Wright
and Thomson, 2005) and we thus cannot infer the distri-
bution Penv(c) from experimental data. Instead, we con-
sider a broad class of distributions parameterized by a few
parameters. For simplicity, we only consider uncorrelated
odors, where the concentrations ci of ligands are indepen-
dent. We denote by pi the probability that ligand i is part
of an odor. If this is the case, the associated ci is drawn
from a log-normal distribution with mean µi and stan-
dard deviation σi. This choice allows us to independently
adjust the mean odor size s = Pi pi, the mean of the
total concentration ctot = Pi ci, and the concentration
variations σi
. Averaged over all odors, ci then has mean
µi
i + piσ2
hcii = piµi and variance var(ci) = (pi − p2
i .
Note that typical odors can have hundreds of different
ligands (Wright and Thomson, 2005), but this is still well
below NL ∼ 105 and we thus have 1 (cid:28) s (cid:28) NL.
i )µ2
III. RESULTS
A. Global inhibition leads to concentration-invariant, sparse
representations
Our model has the interesting property that the odor
representation a does not change when the odor c or
the sensitivities Sni are scaled by a positive factor. This
4
FIG. 2 Global inhibition with thresholding leads to sparse odor representations a. (A) Excitations en for an arbitrary odor.
Active channels (orange) have an excitation above the threshold (red line, inhibition strength α = 1.4). The right axis
indicates the normalized excitation en = enNR/Pm em.
(B) Histogram of the number of active channels compared to a
binomial distribution (black line) with the same mean for α = 1.4. (C) Mean channel activity hani as a function of α. The
approximation given by Eq. 6 (solid line) is compared to numerical simulations (symbols, standard error of the mean smaller
than symbol size). The gray dotted line indicates a single expected active channel in humans, hani = 1
300 . (A–C) Additional
model parameters are NR = 32, NL = 256, pi = 0.1, µi = σi = 1, and λ = 1.
receptor sensitivities, which determine Vext and Vint, re-
spectively.
In particular, the width λ of the sensitiv-
ity distribution could also be under evolutionary control.
However, experimental data suggests that both flies and
humans exhibit λ ≈ 1 (Zwicker, Murugan, and Brenner,
2016). Additionally, we show in the SI that much smaller
or larger values lead to extremely sparse representations,
such that we will only consider λ = 1 in the following.
In this case, the inhibition strength α controls the spar-
sity of the odor representation in our simple model of the
olfactory system.
B. Sparse coding transmits useful information
One problem with sparse representations is that they
cannot encode as many odors as dense representations.
There is thus a maximal sparsity at which typical olfac-
tory tasks can still be performed. In general, the perfor-
mance of the olfactory system can be quantified by the
transmitted information I, which is defined in Eq. 4. If
we for simplicity neglect the small correlations between
channels, I can be approximated as (Zwicker, Murugan,
and Brenner, 2016)
I ≈ −
NRXn=1(cid:2)hani log2hani + (1 − hani) log2(1 − hani)(cid:3) .
(7)
A maximum of NR bits is transmitted when half the chan-
nels are active on average, hani = 1
2 . In our model, this
is the case for weak inhibition, α < 1, see Fig. 2C. In the
opposite case of significant inhibition, α > 1, few chan-
nels are typically active and the transmitted information
is smaller. In the limit hani (cid:28) 1, the information is ap-
proximately given by I ∼ 1
ln 2 NRhani· (1− lnhani), which
implies that even if only 10 % of the channels are active on
average, the information I is still almost half of the max-
imal value of NR bits. However, large information I does
not automatically indicate a good receptor array, since
only accessible information that can be used to solve a
− 1 ,
(5)
(cid:18)1 +
σ2
µ2(cid:19) Vint = eλ2
1 s
var(en) = VextVint Vext ≈
simple case of identically distributed ligands, we have
for 1 (cid:28) s (cid:28) NL, see SI. The normalized excitations
thus vary more if odors contain fewer ligands, concentra-
tions fluctuate stronger, or sensitivities are distributed
more broadly. Finally, the mean channel activity hani is
given by the probability that the excitation en exceeds
the threshold γ, see Eq. 2. This is equal to the probability
that the normalized excitation en exceeds the normalized
threshold γ = γ/( ¯Sctot). Replacing γ by its expectation
value hγi = α and using log-normally distributed en, we
obtain
ln(cid:0)1 + VextVint(cid:1)(6)
1 2
erfc(cid:18) ζ + ln α
2 (cid:19) with ζ =
2ζ 1
1 2
hani ≈
for log-normally distributed en, see SI. Fig. 2C shows
that this is a good approximation of the numerical re-
sults, which have been obtained from ensemble averages
of Eq. 2.
The mean activity hani can also be interpreted as the
mean fraction of channels that are activated by an odor,
such that small hani corresponds to sparse odor repre-
sentations. Fig. 2C shows that in our model this is the
case for large inhibition strength α, where hani ∼ e−ν
with ν ≈ (ln α)2/(4ζ), see SI. Since sparse representa-
tions are thought to be efficient for further processing
in the brain (Laurent, 1999; Olshausen and Field, 2004)
the inhibition strength α could be tuned, e. g., on evo-
lutionary time scales, to achieve an activity hani that
is optimal for processing the odor representation down-
stream. If the optimal value of hani is the same across
animals, our theory predicts that inhibition is stronger
in systems with more receptor types. However, this sim-
ple argument is not sufficient, since hani also depends
on the variations in the natural odor statistics and the
Numerics
Approximation, Eq. 6
0
1
2
Inhibition strength α
1
0.1
0.01
Activity(cid:31)an(cid:30)
C
0
5
Active channel count(cid:31)n an
0.3
0.2
0.1
0
B
Frequency
2
1
0
5
10
15
20
Channel n
25
30
30
25
20
15
10
05
A
Excitationen
5
considered. This qualitatively agrees with experiments
where humans are not able to identify all ligands in mix-
tures of more than three ligands (Goyert et al., 2007;
Jinks and Laing, 2001) and they fail to detect the pres-
ence or absence of ligands in mixtures of more then 15
ligands (Jinks and Laing, 1999).
Even if humans cannot identify ligands in large odors,
they might still be able to distinguish two such odors. To
study this, we next compare the representations of two
odors that each contain s ligands, sharing sB of them,
for the simple case where all ligands have the same con-
centration. Fig. 3C shows that the distance hdi between
the two odors decreases with larger sB, i. e., more sim-
ilar odors are more difficult to discriminate. However,
sB only has a strong effect if more than about 80 % of
the ligands are shared between odors. Conversely, the
inhibition strength α and the odor size s significantly
influence hdi for all values of sB. This agrees with the
results shown in Fig. 3B, where hdi exhibits a similar de-
pendence on α and s. While it is expected that the per-
formance decreases with large inhibition strength α since
fewer channels are active, the strong dependence on the
size s is surprising.
Why are odors with many ligands more difficult to dis-
criminate in our model? Since correlations between chan-
nels seem to be negligible, the most likely explanation is
that larger odors activate fewer channels. To test this
hypothesis, we determine the activity hani in the simple
case where all ligands in an odor have the same concen-
tration. Because of the normalization, the value of this
concentration does not matter and hani only depends on
In the
the inhibition strength α and the odor size s.
limit of large odors (s (cid:29) 1), the approximation given
in Eq. 6 yields hani ∼ e−βs with β ∼ (ln α)2, see SI. In
this case, the activity hani thus decreases exponentially
with s and this decrease is stronger for larger α. Con-
sequently, larger odors activate fewer channels and it is
thus less likely that a small change in such odors alters
the activation pattern a.
Larger odors activate fewer channels because the re-
spective excitations en have a smaller variability. For an
odor with s ligands of equal concentration, en is pro-
portional to the sum of s sensitivities Sni, see Eq. 1.
Consequently, en can be considered as a random vari-
able whose mean heni and variance var(en) scale with
s. The activity hani is given by the fraction of exci-
tations that exceed the threshold γ, which also scales
with s. This fraction typically scales with the coefficient
heni−1, which is proportional to s− 1
of variation var(en)
2
and is thus smaller for larger odors. Larger odors thus
activate fewer channels because there are fewer excita-
tions that are much larger than the mean, see Fig. 4A.
12
C. Larger odors have sparser representations
given task matters (Tikhonov, Little, and Gregor, 2015;
Tkacik and Bialek, 2016).
To test whether sparse representations are sufficient
to solve typical olfactory tasks, we next study how well
odors can be discriminated in our model. As a proxy
for the discriminability, we calculate the Hamming dis-
tance d between the odor representations, which is given
by the number of channels with different activity. In the
simple case of uncorrelated odors, which do not share
any ligands, the expected distance hdi is approximately
given by total number of active channels in both repre-
sentations. Consequently, uncorrelated odors can be dis-
tinguished even if their representations are very sparse.
However, realistic tasks typically require distinguishing
similar odors. We thus next study the discriminability
of odors that vary in the relative concentrations of their
ligands, their size, and their composition.
ct
cb
We start by determining the maximal dilution cb
ct
at
which a target odor at concentration ct can still be de-
tected in a background of concentration cb. We calcu-
late the expected difference hdi between the associated
representations from the probability that a given chan-
nel changes its activity when the target is added, see SI.
Since this probability is the same for all channels, hdi is
proportional to the number NR of channels. For the sim-
ple case where both the target and the background are a
single ligand, Fig. 3A shows that hdi decreases for smaller
target concentrations and is qualitatively the same for all
inhibition strengths α. For large dilutions cb
, hdi is in-
ct
versely proportional to the dilution, hdi ∝ NR
. Since
the addition of the target can only be detected reliably if
hdi > 2, which corresponds to a situation where one chan-
nel becomes inactive and another one active, our model
predicts that doubling the number NR of channels also
doubles the concentration sensitivity. Fig. 3A thus im-
plies that mice (NR ≈ 1000) should be able to detect the
addition of a target even if it is almost a hundred times
more dilute than the background, which is close to the
threshold that has been found experimentally (Mouret
et al., 2009). Conversely, flies (NR ≈ 50) should fail for
very small dilution factors.
We next study odors comprised of many ligands, since
typical odors are blends (Wright and Thomson, 2005).
For simplicity, we consider the detection of a single tar-
get ligand in a background mixture of varying size s when
the target ligand and the ligands in the background have
equal concentration, such that the target dilution is s.
Fig. 3B shows that the qualitative dependence of hdi on
the dilution is similar to the single ligand case in panel A,
but the maximal dilution for detecting the target is dif-
ferent. For instance, the model predicts that mice cannot
identify the addition of the target ligand to a background
consisting of more than ten ligands, while the maximal di-
lution was almost one hundred in the case of single back-
ground ligands. Consequently, the discrimination perfor-
mance seems to drop significantly when larger odors are
6
FIG. 3 Sparse coding is sufficient to distinguish odors with different relative ligand concentrations, size, and composition.
(A) Mean distance hdi between the representations of a background ligand at concentration cb and an odor with an additional
(B) Distance hdi
target ligand at concentration ct as a function of the dilution cb/ct for various inhibition strengths α.
resulting from adding a ligand to an odor comprised of s ligands as a function of s for various α. (C) Distance hdi between
the representations of two odors with s ligands, sharing sB of them, as a function of the similarity sB/s for small (s = 8, solid
lines) and large odors (s = 16, dashed lines). The colors indicate the same α as in the other panels. (A–C) The gray dotted
lines indicate the threshold hdi = 2 for NR = 50, 300, 1000 (corresponding to flies, humans, and mice; top to bottom). The
width of the sensitivity distribution is λ = 1.
FIG. 4 Larger odors activate fewer channels. (A) Comparison of the excitations en of small (odor size s = 8, upper panels)
and large odors (s = 64, lower panels) at α = 1.3. en for a single odor (left panels) and histograms for all odors (right panels)
are shown. Larger odors exhibit fewer active channels (dark blue), for which the excitations are above threshold (red line).
(B) Numerically determined hani as a function of s for various inhibition strengths α at small (σ/µ = 1, solid lines) and large
concentration variability (σ/µ = 10, dashed lines) at NL = 104. (C) Numerically determined hani as a function of σ/µ for
various α. (A–C) Additional model parameters are NR = 32, NL = 256, pi = 0.1, µi = σi = 1, and λ = 1. The gray dotted
line in B and C indicates a single expected active channel in humans, hani = 1
300 .
This is a direct consequence of the assumption that the
excitation threshold γ scales with the mean excitation
and this result does not depend on other details of the
model. Conversely, the dependence of hani on the inhi-
bition strength α is model specific, since it follows from
the shape of the tail of the excitation distribution.
In
particular, the influence of the odor size on hani is in-
significant for weak inhibition, α ≈ 1, because approxi-
mately half the channels are activated irrespective of the
variance var(en).
This qualitative explanation illustrates that depend-
ing on the variability of the excitations different odors
can have representations with very different sparsities.
Indeed, we find that the sparsity changes over several or-
ders of magnitude as a function of the odor size s in our
model, see Fig. 4B. Moreover, the concentration variabil-
ity σ
µ of the individual ligands also has a strong effect on
the sparsity, see Fig. 4C. This is because larger σ
µ implies
larger variations in the excitations, such that more chan-
nels exceed the threshold and become active.
In fact,
this dependence of hani on s and σ
µ is also qualitatively
captured by the analytical approximation given in Eq. 6,
which explicitly depends on the odor variability Vext de-
fined in Eq. 5. Taken together, our model shows that the
sparsity of the odor representations strongly depend on
the odor statistics Penv(c).
D. Effective arrays have similar receptor sensitivities
So far, we considered homogeneous receptor arrays,
where all receptor types have the same average sensitiv-
ity. However, realistic receptors vary in their biochem-
ical details and it might thus be difficult to have such
homogeneous arrays. We thus next consider the effect of
sensitivity variations between different receptors. This is
important, since a channel with overly sensitive receptors
will contribute significantly to the common threshold γ,
suppress the activity of other channels, and could thus
limit the coding capacity of the system, see Fig. 5A. To
study this, we consider sensitivity matrices Sni = ξnSiid
ni ,
where ξn denotes the mean sensitivity of receptor type n
and Siid
is the sensitivity matrix that we discussed so
ni
far, i. e., it is a random matrix where all entries are in-
dependently drawn from a log-normal distribution de-
scribed by the mean ¯S and width λ. Here, ξn captures
differences between receptor types, e. g., because of bio-
chemical differences or due to variations in copy number,
1
100
Concentration ratio cb/ct
10
1
10
Background mixture size s
0
0.2
0.8
Composition similarity sB/s
0.4
0.6
A
Relative ligand concentration
B
Odor size
C
Odor composition
s = 8
s = 16
0.1
0.01
10−3
Dist.(cid:31)d(cid:30)/NR
0.1
0.01
α = 1.5
α = 2
α = 3
10−3
Dist.(cid:31)d(cid:30)/NR
α = 1.5
α = 2
α = 3
Fly
Human
Mouse
0.1
0.01
10−3
Dist.(cid:31)d(cid:30)/NR
α = 1.5
α = 2
α = 3
0.1
Concentration variability σ/µ
10
1
1
0.1
0.01
10−3
10−4
10−5
Activity(cid:31)an(cid:30)
C
α = 1.5
α = 2
α = 3
1
100
10
Mixture size s
1000
1
0.1
0.01
10−3
10−4
10−5
Activity(cid:31)an(cid:30)
B
20
10
Channel n
30
0.1
0
Frequency
15
10
05
100
50
0
A
en
en
see SI. For this model, the mean excitation threshold is
hγi = α ¯ShctotiξtotNR−1 where ξtot = Pn ξn. The ex-
pected channel activity is approximately given by
Nrξn(cid:19) ,
hani ≈ 1 − F(cid:18) αξtot
(8)
where F (en) is the cumulative distribution function of
the normalized excitations en for ξn = 1, whose mean
is heni = 1 and whose variance is given by Eq. 5. Note
that hani does not change if all ξn are multiplied by the
same factor. In particular, the expression above reduces
to hani ≈ 1 − F (α) and thus Eq. 6 if all ξn are equal.
We first discuss the influence of the receptor sensitivi-
ties ξn by only varying one type, i. e., we change ξ1 while
setting ξn = 1 for n ≥ 2. Fig. 5B shows that for fixed
channel activity hani the transmitted information I is
maximal for a homogeneous receptor array (ξ1 = 1). I is
reduced for smaller ξ1 and for ξ1 = 0 it reaches the
value I0 of an array where the first receptor was removed.
Conversely, I can drop well below I0 when ξ1 is increased
above 1. In this case, the large excitation of the affected
channel not only leads to its likely activation, but it also
raises the threshold γ and thereby inhibits other chan-
nels, see Fig. 5A. In the extreme case of very large ξ1,
this channel will always be active while all other chan-
nels are silenced, which implies I = 0. There is thus a
critical value of ξ1 beyond which removing the receptor
from the array is advantageous for the overall perfor-
mance. Fig. 5B shows that increasing the sensitivity of a
receptor by only 40 % can make it useless in the context
of the whole array if representations are sparse.
So far, we only varied the sensitivity of a single re-
ceptor. To test how variations in the sensitivities of all
receptors affect the information I, we next consider log-
normally distributed ξn. Here, vanishing variance of ξn
corresponds to a homogeneous receptor array. Fig. 5C
shows that small variations in ξn can strongly reduce the
transmitted information I. Since I limits the discrimi-
native capability of the receptor array, this suggests that
receptor arrays with heterogeneous sensitivities perform
worse.
The simple model that we discuss here shows that
the excitation statistics of the different channels deter-
mine the properties of the resulting odor representation.
In particular, receptors that have lower excitations on
average might be suppressed often and thus contribute
less to the odor information. Since the excitation statis-
tics are influenced both by the sensitivities Sni and the
odor statistics Penv(c), this suggests that the sensitivities
should be adjusted to the odor statistics. In an optimal
receptor array, the sensitivities are chosen such that all
channels have the same probability to become active.
7
IV. DISCUSSION
We studied a simple model of odor representations,
which is based on normalization and a non-linear gain
function. This model separates the odor composition,
encoded in the activity a of the projection neurons, from
the odor intensity, which could be encoded by the to-
tal excitation etot or the threshold level γ (Mainland
et al., 2014). For significant inhibition the representa-
tion a is sparse and the set of active projection neurons
provides a natural odor 'tag' that could be used for iden-
tification and memorization in the downstream process-
ing (Stevens, 2015).
Sparse representations reduce the coding capacity and
transmit less information than dense ones. However,
even if the mean activity is hani = 0.01 and thus 50
times smaller than in maximally informative arrays with
hani = 0.5, the transmitted information I is only reduced
by a factor of 12, see Eq. 7. For humans with NR = 300,
this yields I ≈ 25 bits, allowing to encode 2I ≈ 107 dif-
ferent odor compositions. Note that the total informa-
tion Itot also includes information Iint about the odor
intensity, Itot = I + Iint. Here, Iint ≈ 10 bits would be
sufficient to encode the total concentration over a range
of 10 orders of magnitude with a resolution of 5 %, typ-
ical for humans (Cain, 1977).
In this case, our model
compresses the 300 bits of a maximally informative rep-
resentation on the level of glomeruli (Zwicker, Murugan,
and Brenner, 2016) to only Itot ≈ 35 bits on the level of
projection neurons.
The model discussed here is similar to our previous
model, where we discussed representations on the level
of the glomeruli (Zwicker, Murugan, and Brenner, 2016).
Both models use a maximum entropy principle to deter-
mine properties of optimal receptor arrays. To achieve
this, the receptor sensitivities must be tailored to the
odor statistics in both models. The main difference of the
models is the global inhibition discussed here, which sep-
arates the odor composition from its intensity and thus
removes the correlation between the glomeruli excitation
and the odor intensity (Haddad et al., 2010). Conse-
quently, odors can then be discriminated at all concen-
trations, while this was only possible in a narrow concen-
tration range in the glomeruli model (Zwicker, Murugan,
and Brenner, 2016). The additional normalization is thus
useful to separate odors, even if the projection neurons
encode less information than the respective glomeruli.
To estimate this information, we consider binary out-
puts in both models, which corresponds to very noisy
channels. However, the glomeruli model discusses arrays
of noisy receptor, while we here consider perfect recep-
tors whose signal is first normalized and then subjected
to noise. This additional processing reduces correlations
and leads to sparse representations, which might simplify
downstream computations. Consequently, this model is
suitable for describing natural olfaction, where the capac-
8
FIG. 5 Receptors with heterogeneous sensitivities make poor arrays. (A) Comparison of the excitations en for homogeneous
(ξ1 = 1, upper panels) and heterogeneous receptors (ξ1 = 2, lower panels). en for the same arbitrary odor (left panels) and
histograms for all odors (right panels) are shown for the first receptor (n = 1, orange) and all other receptors (n ≥ 2, blue).
Dark bars indicate excitations that are above the threshold (red line, inhibition strength α = 1.3). (B) Information I given
by Eq. 7 as a function of the sensitivity ξ1 of the first receptor. The channel activity hani calculated from Eq. 8 is set to
the given value by adjusting α. I is shown relative to the information I0 of a system without the first receptor (dotted line).
(C) Information I (line, mean; shaded area indicates standard deviation) of log-normally distributed ξn as a function of the
variation var(ξn)hξni−2 for various hani. (A–C) Remaining parameters are NR = 32, NL = 256, pi = 0.1, µi = σi = 1, and
λ = 1.
ity for the downstream computations is limited, while the
glomeruli model is relevant for artificial olfaction (Stitzel,
Aernecke, and Walt, 2011), since computers have enough
power to handle high-dimensional signals.
Sparse responses of projection neurons have been ob-
served in experiments (Davison and Katz, 2007; Rinberg,
Koulakov, and Gelperin, 2006). For instance, in mice
15 % of the projection neurons respond to a given single
ligand (Roland et al., 2016), suggesting significant inhi-
bition. However, in locust about two third of the pro-
jection neurons respond to any given odor (Perez-Orive
et al., 2002), which implies weak inhibition.
It is thus
conceivable that some animals exhibit sparse representa-
tions while others have maximally informative ones, al-
though additional experiments are needed to character-
ize the representations better. A direct experiment could
test whether the odor percept changes when the weakly
responding glomeruli are disabled artificially. Addition-
ally, it will be important to study the representations of
mono-molecular odors and mixtures at various concentra-
tion to better resemble the natural odor statistics. For
instance, our simple theory predicts that fewer than 15 %
of the projection neurons in mice respond when complex
mixtures are presented.
Indeed, experiments find that
only 3 to 10 % of the projection neurons in mice fire
for complex urine odors (Lin et al., 2005). Conversely,
the statistics of the activity of projection neurons in flies
seem to be independent of the stimulus (Stevens, 2016).
Our theory can also be tested by measuring how well
odors can be discriminated. For instance, odors are much
more difficult to distinguish if they contain more ligands
in our model, which has also been observed experimen-
tally (Weiss et al., 2012). Conversely, other experiments
indicate that the odor size only weakly influences the
odor discriminability (Bushdid et al., 2014). Taken to-
gether, there is some experimental evidence that the odor
representations and thus the discriminability change with
odor size, although there is also evidence to the contrary,
which could hint at mechanisms beyond global inhibition
that influence the odor representations.
The coding sparsity given by the mean channel ac-
tivity hani can be adjusted by changing the inhibition
strength α or the width λ of the receptor sensitivity dis-
tribution in our model. Additionally, hani is a function
of the natural odor statistics, i. e., the typical number
of ligands in odors and their concentration distribution.
Consequently, α or λ must be adjusted to keep hani con-
stant if the odor statistics change, e. g., because of sea-
sonal changes or migration to a different environment.
This adjustment could happen on multiple timescales,
reaching from evolutionary adaptations of the receptors
to near-instantaneous adjustments of the involved neu-
rons, and it is likely that the global inhibition is reg-
ulated on all levels (Wilson, 2013).
In this paper, we
investigated the simple case of constant α and λ, which
corresponds to slow regulation, but it is conceivable that
α could be regulated on short time scales. For instance,
the threshold could be lowered for larger odors to im-
prove their discriminability. Our model suggests that
such additional mechanisms are necessary to efficiently
discriminate odors of all sizes.
Our model also reveals that it is important to control
the properties of the individual communication channels
to have useful receptor arrays. For instance, increasing
the sensitivity of a given receptor by 40 % can be worse
then removing it completely, see Fig. 5A. Generally, a
receptor array is only effective if the different channels
have similar excitations on average. This suggests that
the sensitivities are tightly controlled and maybe even
adjusted to the odor statistics of the environment. On
evolutionary time scales, the sensitivities could be regu-
lated by point mutations of the receptors that change
how ligands bind (Yu et al., 2015). On shorter time
scales, the sensitivities could be regulated by changing
the receptor copy numbers, see SI. Since this is observed
experimentally (Yu and Wu, 2016), we predict that the
receptor copy numbers are adjusted such that the exci-
tations of all glomeruli are similar when averaged over
(cid:31)an(cid:30) = 0.5
(cid:31)an(cid:30) = 0.2
(cid:31)an(cid:30) = 0.1
0.2
0.4
0.6
0.8
1
30
20
10
C
Infor.I[bits]
0
0
(cid:31)an(cid:30) = 0.5
(cid:31)an(cid:30) = 0.2
(cid:31)an(cid:30) = 0.1
0.5
1
0
1.5
2
Sensitivity ξ1 of receptor 1
Sensitivity variation var(ξn)/(cid:31)ξn(cid:30)2
1
0.5
0
−0.5
−1
I−I0[bits]
B
ξ1 = 1
ξ1 = 2
5 10 15 20 25 30
0
0.03
Channel n
Frequency
60
40
20
0
60
40
20
0
A
en
en
natural odors. Alternatively, variations in the receptor
sensitivities could be balanced by more complex inhibi-
tion mechanism. For instance, experiments show that
different projection neurons have different susceptibili-
ties to inhibition (Hong and Wilson, 2015). Here, the
experimentally observed turnover of mitral cells and in-
terneurons (Lazarini and Lledo, 2011) could adjust the
inhibition mechanism locally, which could optimize the
olfactory system for a given environment (Mouret et al.,
2009). Such adaptation of the inhibition mechanism to
the current stimulus statistics and more complex mod-
els where the behavioral state of an animal could influ-
ence the olfactory bulb by top-down modulation (Wilson,
2013) will be interesting to explorer in the future.
Our simplified model neglects many details of the olfac-
tory system (Silva Teixeira, Cerqueira, and Silva Ferreira,
2016). For instance, we do not consider the dynamics of
inhalation and the odor absorption in the mucus (Pelosi,
2001; Schoenfeld and Cleland, 2005).
Instead, we here
directly parameterize the ligand distribution at the ol-
factory receptors, where we for simplicity neglect corre-
lations between ligands. It would be interesting to extend
the model for more complex stimuli and study how the
system decorrelates the input, identifies a target odor in
a background, and separates multiple odors from each
other. This likely involves many steps (Cleland et al.,
2011) and cannot be done perfectly with a single normal-
ization step and non-linear gain function. For instance,
it might be important to apply gain functions at the level
of receptors and the glomeruli to model finite sensitivity
and saturation effects. Additionally, it has been shown
that there is additional cross-talk on the level of recep-
tors (Ukhanov et al., 2010) and glomeruli (Aungst et al.,
2003; Silbering and Galizia, 2007), which could support
decorrelation. Generally, such cross-talk and the inhibi-
tion that we discussed here will be non-linear (Wilson,
2011). This could for instance be modeled by a divisive
normalization model that has been proposed for olfac-
tion (Olsen, Bhandawat, and Wilson, 2010).
It is also
likely that the inhibition of the projection neurons is not
driven by a single global variable. If glomeruli position-
ing carried some meaning (Murthy, 2011), local inhibi-
tion could help separating similar odors by enhancing
the contrast (Leon and Johnson, 2003). The discrimi-
nation of similar odors could also be improved if projec-
tion neurons had a larger output range, increasing the in-
formation capacity per channel. Finally, we completely
neglected the temporal dynamics of the olfactory sys-
tem, which play an important role for the adaptation be-
tween sniffs (Zufall and Leinders-Zufall, 2000) and might
also influence odor perception within a single sniff (Blau-
velt et al., 2013; Sirotin, Shusterman, and Rinberg, 2015;
Uchida, Poo, and Haddad, 2014).
9
ACKNOWLEDGMENTS
I thank Michael P. Brenner, Venkatesh N. Murthy,
Mikhail Tikhonov and Christoph A. Weber for helpful
discussions and a critical reading of the manuscript. This
research was funded by the Simons Foundation and the
German Science Foundation through ZW 222/1-1.
REFERENCES
Asahina, K., Louis, M., Piccinotti, S., and Vosshall, L. B., J
Biol 8, 9 (2009).
Aungst, J. L., Heyward, P. M., Puche, A. C., Karnup, S. V.,
Hayar, A., Szabo, G., and Shipley, M. T., Nature 426, 623
(2003).
Banerjee, A., Marbach, F., Anselmi, F., Koh, M. S., Davis,
M. B., Garcia da Silva, P., Delevich, K., Oyibo, H. K.,
Gupta, P., Li, B., and Albeanu, D. F., Neuron 87, 193
(2015).
Barlow, H., Network 12, 241 (2001).
Barlow, H. B.,
in Sensory Communication, edited by
W. Rosenblith (MIT press, 1961) pp. 217–234.
Berck, M. E., Khandelwal, A., Claus, L., Hernandez-Nunez,
L., Si, G., Tabone, C. J., Li, F., Truman, J. W., Fetter,
R. D., Louis, M., Samuel, A. D., and Cardona, A., Elife 5
(2016), 10.7554/eLife.14859.
Bhandawat, V., Olsen, S. R., Gouwens, N. W., Schlief, M. L.,
and Wilson, R. I., Nat. Neurosci. 10, 1474 (2007).
Blauvelt, D. G., Sato, T. F., Wienisch, M.,
V. N., Frontiers in neural circuits 7 (2013).
and Murthy,
Bushdid, C., Magnasco, M., Vosshall, L., and Keller, A.,
Science 343, 1370 (2014).
Cain, W. S., Science 195, 796 (1977).
Carandini, M. and Heeger, D. J., Nat Rev Neurosci 13, 51
(2012).
Cleland, T. A., Trends. Neurosci. 33, 130 (2010).
Cleland, T. A., Chen, S.-Y. T., Hozer, K. W., Ukatu, H. N.,
Wong, K. J., and Zheng, F., Front Neuroeng 4, 21 (2011).
Cleland, T. A., Johnson, B. A., Leon, M., and Linster, C.,
Proc. Natl. Acad. Sci. USA 104, 1953 (2007).
Cleland, T. A. and Sethupathy, P., BMC Neurosci 7, 7 (2006).
Davison, I. G. and Katz, L. C., J Neurosci 27, 2091 (2007).
Demb, J. B. and Singer, J. H., Annual Review of Vision Sci-
ence 1, 263 (2015).
Getz, W. M. and Lutz, A., Chem. Senses 24, 351 (1999).
Goyert, H. F., Frank, M. E., Gent, J. F., and Hettinger,
T. P., Brain Res Bull 72, 1 (2007).
Gupta, P., Albeanu, D. F., and Bhalla, U. S., Nat. Neurosci.
18, 272 (2015).
Haddad, R., Weiss, T., Khan, R., Nadler, B., Mandairon, N.,
Bensafi, M., Schneidman, E., and Sobel, N., J Neurosci
30, 9017 (2010).
Hong, E. J. and Wilson, R. I., Neuron 85, 573 (2015).
Hopfield, J., Proc. Natl. Acad. Sci. USA 96, 12506 (1999).
Isaacson, J. S. and Scanziani, M., Neuron 72, 231 (2011).
Jefferis, G. S., Marin, E. C., Stocker, R. F., and Luo, L.,
Nature 414, 204 (2001).
Jinks, A. and Laing, D. G., Perception 28, 395 (1999).
Jinks, A. and Laing, D. G., Physiol Behav 72, 51 (2001).
Kaupp, U. B., Nat Rev Neurosci 11, 188 (2010).
Koulakov, A., Gelperin, A., and Rinberg, D., J. Neurophysiol.
98, 3134 (2007).
Laurent, G., Science 286, 723 (1999).
Lazarini, F. and Lledo, P.-M., Trends Neurosci 34, 20 (2011).
Leon, M. and Johnson, B. A., Brain Res. Rev. 42, 23 (2003).
Li, Z., Biol Cybern 62, 349 (1990).
Li, Z., in Models of neural networks (Springer, 1994) Chap. 6,
pp. 221–251.
Lin, D. Y., Zhang, S.-Z., Block, E., and Katz, L. C., Nature
434, 470 (2005).
Stevens, C. F., Proc. Natl. Acad. Sci. USA 112, 9460 (2015).
Stevens, C. F., Proc. Natl. Acad. Sci. USA
(2016),
10.1073/pnas.1606339113.
Stitzel, S. E., Aernecke, M. J., and Walt, D. R., Annu. Rev.
Biomed. Eng. 13, 1 (2011).
Su, C.-Y., Menuz, K., and Carlson, J. R., Cell 139, 45 (2009).
Tabor, R., Yaksi, E., Weislogel, J.-M., and Friedrich, R. W.,
J. Neurosci. 24, 6611 (2004).
Linster, C. and Hasselmo, M., Behavioural brain research 84,
Tan, J., Savigner, A., Ma, M., and Luo, M., Neuron 65, 912
117 (1997).
(2010).
Lowe, G. and Gold, G. H., Proc. Natl. Acad. Sci. USA 92,
Tikhonov, M., Little, S. C., and Gregor, T., R Soc Open Sci
10
7864 (1995).
2, 150486 (2015).
Mainland, J. D., Lundstrom, J. N., Reisert, J., and Lowe,
Tkacik,
G.
and
G., Trends Neurosci. 37, 443 (2014).
Malnic, B., Hirono, J., Sato, T., and Buck, L. B., Cell 96,
713 (1999).
Mouret, A., Lepousez, G., Gras, J., Gabellec, M.-M., and
Lledo, P.-M., J Neurosci 29, 12302 (2009).
Murthy, V. N., Annu. Rev. Neurosci. 34, 233 (2011).
Niimura, Y., Curr Genomics 13, 103 (2012).
Nikolova, N. and Jaworska, J., QSAR & Combinatorial Sci-
ence 22, 1006 (2003).
Olsen, S. R., Bhandawat, V., and Wilson, R. I., Neuron 66,
287 (2010).
Olshausen, B. A. and Field, D. J., Curr Opin Neurobiol 14,
481 (2004).
Pelosi, P., Cellular and Molecular Life Sciences CMLS 58, 503
(2001).
Perez-Orive, J., Mazor, O., Turner, G. C., Cassenaer, S., Wil-
son, R. I., and Laurent, G., Science 297, 359 (2002).
Rinberg, D., Koulakov, A., and Gelperin, A., J Neurosci 26,
8857 (2006).
Roland, B., Jordan, R., Sosulski, D. L., Diodato, A., Fuku-
naga, I., Wickersham, I., Franks, K. M., Schaefer, A. T.,
and Fleischmann, A., Elife 5 (2016), 10.7554/eLife.16335.
Ruderman, and Bialek,, Phys. Rev. Lett. 73, 814 (1994).
Sachse, S. and Galizia, C. G., J Neurophysiol 87, 1106 (2002).
Schoenfeld, T. A. and Cleland, T. A., Trends in neurosciences
28, 620 (2005).
Silbering, A. F. and Galizia, C. G., J. Neurosci. 27, 11966
(2007).
Silva Teixeira, C. S., Cerqueira, N. M. F. S. A., and Silva Fer-
reira, A. C., Chem Senses 41, 105 (2016).
Sirotin, Y. B., Shusterman, R., and Rinberg, D., eNeuro 2
(2015), 10.1523/ENEURO.0083-15.2015.
Soucy, E. R., Albeanu, D. F., Fantana, A. L., Murthy, V. N.,
and Meister, M., Nat. Neurosci. 12, 210 (2009).
Bialek, W.,
Physics
Annual
89
7,
Review
(2016),
Condensed Matter
of
http://dx.doi.org/10.1146/annurev-conmatphys-031214-
014803.
Uchida, N. and Mainen, Z. F., Front Syst Neurosci 1, 3 (2007).
Uchida, N., Poo, C., and Haddad, R., Annu Rev Neurosci
37, 363 (2014).
Ukhanov, K., Corey, E. A., Brunert, D., Klasen, K., and
Ache, B. W., J Neurophysiol 103, 1114 (2010).
Verbeurgt, C., Wilkin, F., Tarabichi, M., Gregoire, F., Du-
and Chatelain, P., PLOS ONE 9, e96333
mont, J. E.,
(2014).
Weiss, T., Snitz, K., Yablonka, A., Khan, R. M., Gafsou, D.,
Schneidman, E., and Sobel, N., Proc. Natl. Acad. Sci. USA
109, 19959 (2012).
Wilson, R. I., Curr. Opin. Neurobiol. 21, 254 (2011).
Wilson, R. I., Annu. Rev. Neurosci. 36, 217 (2013).
Wright, G. A. and Thomson, M. G., in Integrative Plant Bio-
chemistry, Recent Advances in Phytochemistry, Vol. 39,
edited by J. Romeo (Elsevier, 2005) Chap. 8, pp. 191–226.
Yokoi, M., Mori, K., and Nakanishi, S., Proc. Natl. Acad.
Sci. USA 92, 3371 (1995).
Yu, C. R.
and Wu, Y., Exp Neurol
(2016),
10.1016/j.expneurol.2016.06.001.
Yu, Y., Claire, A., Ni, M. J., Adipietro, K. A., Golebiowski,
J., Matsunami, H., and Ma, M., Proc. Natl. Acad. Sci.
USA 112, 14966 (2015).
Zhang, D., Li, Y., and Wu, S., Comput Math Methods Med
2013, 507143 (2013).
Zufall, F. and Leinders-Zufall, T., Chem. Senses 25, 473
(2000).
Zwicker, D., Murugan, A., and Brenner, M. P., Proc. Natl.
Acad. Sci. USA 113, 5570 (2016).
Supporting Information: Normalized neural representations of natural odors
David Zwicker1, 2, ∗
1School of Engineering and Applied Sciences, Harvard University, Cambridge, MA 02138, USA
2Kavli Institute for Bionano Science and Technology, Harvard University, Cambridge, MA 02138, USA
(Dated: July 3, 2018)
(a variable) number of log-normally distributed random
variables, its distribution can be approximated by an-
other log-normal distribution (Wu, Mehta, and Zhang,
2005), which we parameterize by its mean µtot and vari-
ance σ2
tot. We consider the simple approximation where
these parameters are directly given by Eq. S2 (Fenton,
1960). This choice approximates the tail of the distri-
bution well, but leads to errors in the vicinity of the
mean (Wu, Mehta, and Zhang, 2005).
Since both ctot and ci are log-normally distributed
when ligand i is present in an odor (ci > 0), ci is also
log-normally distributed in this case and
hciici>0 =
var(ci)ci>0 =
χ
µi
µtot
µ2
i χ2
µ2
totµ−2
tot (cid:18) σ2
i
µ2
i
where χ = 1 + σ2
1 − pi, the statistics of ci read
χ + χ − 1(cid:19) ,
(S4a)
(S4b)
tot. Since ci = 0 with probability
hcii =
var(ci) =
χ
piµi
µtot
piµ2
µ2
i χ2
tot (cid:18) σ2
i
µ2
i
χ + χ − pi(cid:19) .
(S5a)
(S5b)
Note that the covariance cov(ci, cj) does not vanish since
the ci are not independent. In particular, var(Pi ci) = 0,
since Pi ci = 1 by definition. This condition is only
of the normalized excitations en = ¯S−1Pi Snici read
consistent with Eq. S5a if χ ≈ 1, which implies that ctot
must not vary much, σtot
µtot (cid:28) 1. Using χ = 1, the statistics
(S6a)
heni = 1
var(en) =
var(Sni)
¯
S2
DXi
c2
iE ,
(S6b)
where hPi c2
the statistics given in Eq. S5.
ii = hcii2 + var(ci) and
In the simple case where all ligands are drawn from the
ii with hc2
ii ≈Pihc2
same distribution (pi = p, µi = µ, σi = σ), we obtain
,
(S7)
,
(S8)
(cid:18)1 +
σ2
µ2(cid:19) var(Sni)
¯
S2
1 s
var(en) ≈
1 − p + σ2
µ2
sNL
,
var(ci) ≈
L
( σ2
µ2 + 1), such that
1 N
hcii ≈
ii ≈ 1
sNL
and hc2
which is equivalent to Eq. 5 in the main text.
CONTENTS
S1. Statistics of normalized concentrations and excitations
S2. Numerical simulations
S3. Approximate channel activity
S4. Odor discriminability
S5. Receptor binding model
1
2
2
2
3
S1. STATISTICS OF NORMALIZED CONCENTRATIONS
AND EXCITATIONS
Let pi be the probability that ligand i is present in an
odor. If it is present, its concentration ci is drawn from a
log-normal distribution with mean µi and standard devi-
ation σi, while ci = 0 if the ligand is not present. Hence,
hcii = piµi
var(ci) = (pi − p2
i )µ2
i + piσ2
i ,
(S1a)
(S1b)
while the covariances cov(ci, cj) = hcicji−hciihcji vanish
for i 6= j since the ligands are independent. The statistics
of the total concentration ctot =Pi ci read
NLXi=1
hctoti =
NLXi=1
var(ctot) =
hcii
and
The excitations en are given by en =Pi Snici, where the
sensitivities Sni are log-normally distributed with mean
hSnii = ¯S and variance var(Sni) = ¯S2(eλ2
− 1). Hence,
(S3a)
heni = ¯Shctoti
var(ci) .
(S2)
var(en) = ¯S2 var(ctot) + var(Sni)
NLXi=1
hc2
ii ,
(S3b)
i + σ2
ii = pi(µ2
where hc2
i ) and cov(en, em) = 0 for n 6= m.
We next determine the statistics of the normalized con-
centrations ci = ci/ctot. For simplicity, we consider large
odors, Pi pi (cid:29) 1, where ctot can be considered as an
independent random variable. Since ctot is the sum of
∗ [email protected]; http://www.david-zwicker.de
arXiv:1608.01179v1 [q-bio.NC] 3 Aug 2016
associated probability distribution function reads
2
(S10)
(S11)
#
2S2
n
exp"−(cid:0)Mn − ln(en)(cid:1)2
(cid:21) .
erfc(cid:20) Mn − ln(en)
√2Sn
and the cumulative distribution function is
1 2
F (en) =
f (en) =
1
√2πSnen
The parameters Mn and Sn can be determined from the
mean and variance
var(en) = e2Mn+S2
S2
n
2 (cid:19)
heni = exp(cid:18)Mn +
n − 1(cid:17) .
n(cid:16)eS2
Sn =p2ζ ,
and
Mn = lnheni − ζ
Solving these equations for Mn and Sn, we obtain
(S12a)
(S12b)
(S13)
hani ≈
where ζ = 1
2 ln(1 + var(en)heni−2). Eq. 6 of the main
text follows from this and Eq. 5. For small hani we have
2pζ/π
4ζ
/(x√π), valid for
which follows from erfc(x) ≈ e−x2
x (cid:29) 1. For small ζ, we obtain the approximate scal-
ing lnhani ∼ −(ln α)2/(4ζ), where ζ ∼ s−1 for s (cid:29) 1.
exp(cid:20)−
(ln(α) + ζ)2
(cid:21) ,
ln(α) + ζ
(S14)
S4. ODOR DISCRIMINABILITY
We quantify the discriminability of two odors by the
Hamming distance d of their respective representations a
for several different cases:
a. Uncorrelated odors The expected distance hdi be-
tween the activity patterns a(1) and a(2) of two inde-
pendent odors is
n i + ha(2)
hdi = NR(cid:16)ha(1)
n i and ha(2)
where ha(1)
n i denote the expected activities of
the two odors, averaged over sensitivity matrices, and we
neglect correlations cov(an, am) for simplicity.
n i(cid:17) ,
n iha(2)
n i − 2ha(1)
(S15)
FIG. S1 Influence of the width λ of the sensitivity distribu-
tion on the statistics of the odor representations.
(A) Ex-
pected channel activity hani as a function of λ for several
inhibition strengths α. Intermediated values, λ ≈ 1, lead to
larger activities. (B) Mean Pearson correlation coefficient hρi
calculated from an ensemble average of Eq. S9 as a function
of λ for several α. For small λ, hani was too small to esti-
mate ρ reliably. (A–B) Results are shown for small ( σ
µ = 1,
solid lines) and large ( σ
µ = 10, dashed lines) concentration
variability. Remaining parameters are NR = 32, NL = 256,
and pi = 0.1.
S2. NUMERICAL SIMULATIONS
We numerically calculated ensemble averages over
odors c and sensitivity matrices Sni. Here, we first
choose Sni by drawing all entries independently from
a log-normal distribution with mean ¯S = 1 and vari-
ance var(Sni) = eλ2
− 1. We then draw an odor c using
the following procedure: First, we determine which of
the NL ligands are present according to their probabili-
ties pi. Second, we draw the concentrations ci for each
ligand i that is present from a log-normal distribution
with mean µi and standard deviation σi. We then use
Eqs. 1–3 given in the main text to map the odor c to
a binary activity vector a, from which we can for in-
stance calculate the number of active channels. We ob-
tain ensemble averages of such quantities by repeating
these steps 105 times. This allows us to calculate the
mean activities hani, the covariances cov(an, am), and
the Pearson correlation coefficient ρ, which is defined as
ρ =
NR
1
2 − NR Xn6=m
cov(an, am)
(cid:2)var(an) var(am)(cid:3) 1
2
.
(S9)
Fig. S1 shows these quantities as a function of the width λ
of the sensitivity distribution. We also estimate P (a)
from an ensemble average to calculate the information I
from its definition given in Eq. 4 in the main text.
S3. APPROXIMATE CHANNEL ACTIVITY
We estimate the expected activity hani by the proba-
bility that the normalized excitations en exceed the ex-
pected normalized threshold α. Since both the sensitivi-
ties Sni and the normalized concentrations ci are approx-
imately log-normally distributed, en can also be approx-
imated by a log-normal distribution (Fenton, 1960). The
b. Adding target to background We calculate the ex-
pected change hdi of the representation when a target
odor ct is added to a background odor cb. Because the
odor concentrations are specified, we consider the actual
excitations en instead of the normalized quantities en.
Taking an ensemble average over sensitivity matrices,
the excitations associated with the two odors are charac-
terized by probability distribution functions f t
E(et) and
1 %
0 %
α = 1.5
α = 2
α = 3
0
3
Sensitivity distribution width λ
2
1
−1 %
−2 %
Correlation(cid:31)ρ(cid:30)
B
10−4
0.1
10
Sensitivity distribution width λ
1
1
0.1
0.01
α = 1.5
α = 2
α = 3
10−3
Activity(cid:31)an(cid:30)
A
f b
E(eb) for the target and the background, respectively.
We here consider log-normally distributed en, which are
parameterized by their mean and variance,
heni = ¯S
NLXi=1
ci
var(en) = var(Sni)
c2
i ,
(S16)
NLXi=1
where var(Sni) = ¯S2(eλ2
− 1).
tion heκi =R z f κ
When the target is added to the background, the
expected threshold hγi increases from γb = αhebi to
γs = α(hebi + heti), where heκi denotes the mean excita-
E(z) dz for κ = t, b. This increase in the
threshold can deactivate a channel if it was previously
active, i. e. if its excitation was larger than the thresh-
old associated with the background, eb > γb. For such
eb, the probability that the receptor gets deactivated by
adding the target is P (eb + et < γseb). Integrating over
all possible eb, we thus get the probability poff that a
channel becomes inactive,
poff =Z ∞
=Z ∞
γb
γb
P(cid:0)eb + et < γs(cid:12)(cid:12) eb(cid:1) f b
E(cid:0)eb(cid:1) deb ,
E(cid:0)γs − eb(cid:1)f b
F t
E(cid:0)eb(cid:1) deb
(S17)
E(et) is the cumulative distribution function as-
where F t
sociated with f t
E(et). Conversely, a channel becomes ac-
tive when the additional excitation by the target odor
brings it above the threshold γs. The associated proba-
bility pon reads
pon =Z γb
0
(cid:2)1 − F t
E(cid:0)γs − eb(cid:1)(cid:3) f b
E(cid:0)eb(cid:1) deb .
Taken together, the expected number hdi of channels that
change their state reads
hdi = NR ·(cid:0)pon + poff(cid:1) .
There are three simple limits that we can solve analyt-
If there is no target, heti = 0, the activation
ically:
pattern does not change and we have hdi = 0.
In the
opposing limit of a dominant target, heti → ∞, the acti-
vation patterns are independent and we recover the dis-
tance hdimax for uncorrelated odors, which is given by
Eq. S15. Lastly, in the case where the target and the
background are identically distributed, hebi = heti and
var(eb) = var(et), we have hdi = 1
2hdimax.
(S18)
(S19)
3
cause a baseline excitation eb, which is distributed ac-
E(eb). A channel is inactive for an odor with
cording to f b
E(hγi − eb), where F d
probability F d
E(ed) is the cumula-
tive distribution function of the excitation caused by the
sd = s − sb different ligands. Hence,
p = 2Z hγi
0
F d
E(z)(cid:2)1 − F d
E(z)(cid:3)f b
E(eb)deb ,
(S20)
where z = hγi− eb. Note that the upper bound of the in-
tegral is hγi since channels will be active for both odors if
eb ≥ hγi. The associated Hamming distance hdi between
the two odors is then given by hdi = pNR.
S5. RECEPTOR BINDING MODEL
We consider a simple model where receptors Rn get
activated when they bind ligands Li. This binding
is described by the chemical reaction Rn + Li (cid:10) RnLi,
where RnLi is the receptor-ligand complex. In equilib-
rium, the concentrations denoted by square brackets obey
[RnLi] = Kni·[Rn][Li], where Kni is the binding constant
of the reaction. Hence,
crec
n Knici
,
(S21)
[RnLi] =
1 +Pi Knici
n = [Rn]+Pi[RnLi] denotes
where we consider the case where multiple ligands com-
pete for the same receptor. Here, ci = [Li] is the concen-
tration of free ligands and crec
the fixed concentration of receptors, which is related to
the copy number of receptors of type n. We consider
a simple receptor model where the excitation is propor-
tional to the concentration of the bound ligands, such
that the excitation accumulated in glomerulus n reads
en = βn
N rec
n
crec
n
NLXi=1
[RnLi] = βnN rec
. (S22)
n Pi Knici
1 +Pi Knici
n
Here, N rec
is the copy number of receptors of type n
and βn characterizes their excitability, which could for in-
stance be modified by point mutations (Yu et al., 2015).
Defining Sni = βnN rec
n Kni, we recover Eq. 1 of the main
text in the limit of small concentrations,Pi Knici (cid:28) 1.
The sensitivities are thus proportional to the copy num-
ber N rec
and the biochemical details encoded in βnKni.
n
REFERENCES
c. Discriminating two odors of equal size We consider the
simple case of two odors that each contain s ligands at
equal concentration, sharing sb of them, such that the ex-
pected threshold hγi is the same for both odors. Similar
to the derivation above, we here calculate the probabil-
ity p that a channel is active for one odor, but not for
the other. The sb ligands that are present in both odors
Fenton, L. F., Communications Systems, IRE Transactions
on 8, 57 (1960).
Wu, J., Mehta, N. B., and Zhang, J., in GLOBECOM '05.
IEEE Global Telecommunications Conference, 2005., Vol. 6
(2005) pp. 3413–3417.
Yu, Y., Claire, A., Ni, M. J., Adipietro, K. A., Golebiowski,
J., Matsunami, H., and Ma, M., Proc. Natl. Acad. Sci.
USA 112, 14966 (2015).
|
1702.03579 | 1 | 1702 | 2017-02-12T21:18:10 | Autonomous line follower robot controlled by cell culture | [
"q-bio.NC"
] | Neuro-electronic hybrid promises to bring up a model architecture for computing. Such computing architecture could help to bring the power of biological connection and electronic circuits together for better computing paradigm. Such paradigms for solving real world tasks with higher accuracy is on demand now. A robot as a autonomous system is modeled here to navigate following a particular line. Sensory inputs from robot is directed as input to the cell culture in response to which motor commands are generated from the culture. | q-bio.NC | q-bio | Autonomous line follower robot controlled by cell
culture
Sayan Biswas
Department of Electrical Engineering
Jadavpur University, Kolkata, India
Email: [email protected]
7
1
0
2
b
e
F
2
1
]
.
C
N
o
i
b
-
q
[
1
v
9
7
5
3
0
.
2
0
7
1
:
v
i
X
r
a
Abstract-Neuro-electronic hybrid promises to bring up a
model architecture for computing. Such computing architecture
could help to bring the power of biological connection and
electronic circuits together for better computing paradigm. Such
paradigms for solving real world tasks with higher accuracy is
on demand now. A robot as a autonomous system is modeled
here to navigate following a particular line. Sensory inputs from
robot is directed as input to the cell culture in response to which
motor commands are generated from the culture.
Index Terms-Neuro-electronic hybrid, biological connection,
line follower, sensory
electronic circuits, autonomous system,
inputs, motor commands
I. INTRODUCTION
Computation carried by brain are fast and they are able
to take rapid decisions. At times brain can take really fast
decision at near zero reflex time. Although all kind of decision
making does not fall into this category, but obviously there
are instances where one can take decision in near zero reflex
time. In a visual search task to discriminate between a car
image and animal image, near zero reflex time is expected.
Decision making regarding colours whether is black or whites
takes almost zero reflex time. Such computation are carried
by brain real fast. Such abilities of the brain have inspired
researchers. Many strategies have been developed inspired
from such abilities of brain. Some of those strategies are
neuromorphic devices [1] [2] [3], artificial neural network [4],
fuzzy systems [5] [6].
An ongoing research is being carried out for making hardwares
suitable to mimic neuronal systems [1]. There are challenges
including versatility connectivity which limits the usage of cul-
tures like brain. Neuro electronic hybrid systems can prove to
be a possible computing architecture. On connecting cultures
of neurons properly with the real world such abilities of the
brain can be used to solve real world problems. This might
prove to be a essential platform to use biological network for
computing. Neurons are basic computation and functional unit
of brain which are versatile in terms of behavior. The neurons
in the brain form inter connected set of networks which are
dynamic in nature. Such dynamic nature of neuron in network
gives them learning abilities and makes them powerful enough
to serve varied types of function. Hence making the brain a
robust computational unit of brain. A hybrid system attempts
to solve real world problems by implementing this abilities of
brain.
MEA dish proves advantageous and promising for under-
standing network activity and properties of neuronal network.
Robotic control through such culture [7] [8] and training [9] of
neuronal network is an ongoing work at various labs. There are
several challenges required to be resolved for a practical and
real world application of such systems [10]. Decoding output
from an input to a network is open problem. To understand
such decoding pattern a deeper understanding of dynamic
network system [11] [12] [13] [14] [15] is essential.
The description here deals with a framework for using cell
cultures as a information processing tool, and to implement
the ability for the purpose of classification of sensory inputs
received and accordingly take decision of generating correct
motor commands. The framework shall elaborately explain a
proposal for implementation in real life.
II. LINE FOLLOWER
A line follower robot is an autonomous body expected to
navigate in a network by following a specific line. The track on
which the robot is expected to navigate is coloured black, and
the background is white (figure 2). Such devices are controlled
by a person using a visual feedback from the scene of the
track which is used to generate respective motor commands
on the controller to control the robot. This involves multi
sensory perception to compute the action required to be taken
from the visual input and act accordingly by sending motor
commands to the fingers to produce the desired effect on the
robot navigating on the track. As this requires intervention in
control from a human this can't be refereed to as a automated
system. The process is as shown in the figure 1.
Fig. 1: Robot Navigation by Human
Several computer vision [16] [17] along with machine lean-
ing [18] paradigm has been successful to implement the same
Fig. 2: TRACK
by making silicon chips to perform the computation that are
to be performed by the brain. There are cameras on the robot
which are performing the role of human eyes. The images from
the camera gives some features regarding the path on which
the robot is navigating. The direction on which the robot must
move is classified based on the features extraction method and
using a machine learning classifier. Once this is known, motor
commands to make the robot move in desired direction is sent
to the wheels. Hence the silicon chips and camera do the work
of eye and multi sensory perception [19], of relating visual
feedback to generate motor commands. Using computer vision
techniques for extracting visual features followed by machine
learning classifier to generate motor commands is shown in
the figure 3.
As specified the aim is to implement cell culture to make the
controlling of robot automated.
III. CELL CULTURES AS CONTROLLING UNIT
Multi electrode array - MEA [20] recording gives a great
opportunity to study network topology as it provides popu-
lation recording. It provides data from multiple electrodes or
channels which is thus effective for understanding network
events. A MEA has culture on its surface from which electrical
activity is recorded. A framework is modeled for using the
electrical activity recording to control a robot to navigate
following a line.
Involvement of cell as a contriving unit will be done by
following:
• By photo receptors cells
• By neural cells.
These two shall be dealt in following section.
IV. PHOTO RECEPTOR CELLS
Photo receptor cells are specialized cells in retina which
performs photo transduction. Visual photo transduction is
the method of converting light into electrical signals. Photo
receptors are of great importance as they convert light into sig-
nals which stimulate biological processes. These cells absorb
photons that causes a change in the cell's membrane potential.
The model proposes usage of such cell culture on MEA dish
for controlling a line follower.
The robot would be a simple system equipped with a camera
to serve the purpose of visual feedback. As the camera takes
up the current position of the robot, it is to be projected on
the MEA dish containing photo receptor cell cultures. The
usage of camera serves the purpose of eye. The projection
of the visual field is shown in the figure 5. The track (figure
2) if followed carefully one can find there are three possible
different visual field as shown in figure 6. The photo recep-
tors respond by different electrical activity to different light
intensities. Hence the projection of different visual field on
the culture is expected to generate different electrical activity
MEA dish. Hence to achieve control using photo receptor cell
culture [21] it requires a mapping of electrical activity to visual
field projected. This mapping would help to understand the
visual field or current location of robot with the electrical
activity recorded which would in turn help in generating motor
commands. After the culture is done three of this possible
visual field would be projected one by one, and corresponding
electrical activity generated would be recorded. This will help
Fig. 3: Robot Navigation by Machine Learning classifier
TABLE I: Motor Commands as per the Visual Field
Visual Field
Visual Field A
Visual Field B
Visual Field C
Motor Commands
Cover offset and take 90 ◦ left
Cover offset and take 90 ◦ right
Go straight
in achieving the required mapping of visual field and electrical
activity mapping. Once the mapping is achieved the robot
equipped with camera is left to be navigate on path. It would
take visual field project it on the culture followed by which
electrical activity shall be recorded based on which proper
motor command would be generated and sent to the robot.
The methodology is as follows:
• Visual fields are projected on to the cell culture and
mapping of electrical activity and visual field is obtained
• During test, the images are projected which will generate
motor commands as per obtained mapping
As per positioning of the camera in the experiment an offset
has to determined. The offset would prevent the robot to take
desired motor action as soon as visual field is detected. Offset
is shown in figure 4. If the robot would take say left turn on
attending visual field A or electrical activity corresponding
to visual field A, it would go off track out of line. This
demands necessity of having a pre determined offset. Hence
motor commands required to be generated as per the visual
field is given in table I. The procedure is illustrated in figure
7.
V. NEURON CELL
Neuron cultures obtained generally from rat brain are cul-
tured on MEA dish can be implemented as a control unit
for making autonomous navigator. Framework that will be
discussed here was proposed earlier [10] for a autonomous
Fig. 4: Offset
Fig. 5: Projection of the visual field captured through camera
on top of the MEA dish containing photoreceptor cells
(a) Visual field - A
(b) Visual field - B
(c) Visual field - C
Fig. 6: Visual field projected on MEA dish
Fig. 7: Automated robot navigation by photo receptor cell
(a) Sensor position - A
(c) Sensor position - C
Fig. 8: Sensor position of robot as it navigates. Blue circles denotes location of sensor.
(b) Sensor position - B
Fig. 9: Automated robot navigation by neural cell
TABLE II: Motor Commands as per the Sensor position
Sensor Position
Sensor Position A
Sensor Position B
Sensor Position C
Motor Commands
Take 90 ◦ left and a little forward drift
Take 90 ◦ right and a little forward drift
Go straight
Fig. 10: Result of left turn from sensor position A - figure 8a.
Blue circles denotes location of sensor.
system of navigating by avoiding obstacles. The proposed
framework had a accuracy about 98%. It would be discussed
how already existing framework could be utilized for modeling
a line follower system.
The robot which shall navigate must have two infrared - IR
sensors positioned pointing downward. As the robot navigates
there could the only three possible states of view as shown
in figure 8. A IR sensor contains a emitter and detector. The
emitter would emit infrared waves and purpose of detector
would be to detect presence of any IR wave. As white reflects
light and black absorbs it, same would be true with infrared
wave. Hence if IR sensor detect (1) IR waves that means
the colour is white and if it does not detect (0) the colour
must be black. Hence {1,1} {0,1} {1,0} corresponds to go
straight, go left, go right respectively (figure 8 gives a pictorial
representation). Hence knowing the output of IR sensor could
predict the desired output motor command. Activity of MEA
dish is used to understand the output of IR sensor. It was
found that probability of spike occurrence at some particular
electrode in response to stimulus sequence at some other elec-
trode depends on electrode stimulation the order of stimulation
and timing between the pulses [22]. A network of neuron
is well able to distinguish between various inputs based on
electrodes stimulated, the stimulating order and timing [10].
This ability of neuronal network was used to encode inputs to
a culture and decode output responses from the culture. From
electrodes that excited the network to highest level stimulating
electrode is chosen. Stimulating and recording electrodes are
so chosen that by input pattern could be predicted from activity
at recording electrodes [10]. Hence the input pattern of {1,1}
{0,1} {1,0} could be predicted from activity patter of neuron
cultures. As this prediction could be made foreseen motor
commands could be generated making it possible for the robot
navigate following a line.
Suppose the sensor position is as in figure 8a. If then robot
takes a 90 ◦ left turn it results to a position described in figure
10 which is a undesired position. Similarly the system would
attain a undesired position if robot takes a 90 ◦ right turn from
position C (figure 8c). This is undesired as one of the sensor
is lying on white region where as the other is on black region.
This issue could be resolved by generating a motor command
to take a 90 ◦ turn on the desired direction followed by a
automated little forward push so that both sensor come into
the white region.
Hence automated navigator using neuron culture is developed.
The corresponding motor commands are shown in table II. The
procedure is illustrated in figure 9.
VI. DISCUSSION AND CONCLUSIONS
This was a implementation model using two type of cell
culture in performing real world task of navigating following a
line. It was shown how could the system be made autonomous
by using two different type of cell cultures. The techniques for
usage of neuron culture have been implemented practically
and was found to have a accuracy of 98% [10]. It is expected
that if the same methodology is extended in autonomous line
follower system, as proposed, it would have a good accuracy
too. The strong idea in this approach is using cell cultures
as information processing tool, and to implement the ability
for the purpose of classification of sensory inputs received
and accordingly take decision of generating correct motor
commands. Driving can be looked as a process of locomotion
by obstacle avoidance and path following. A amalgamation of
object avoiding paradigm [10] and line follower techniques, as
suggested in thid work, could pave way for technological ad-
vancement of building new efficient and accurate autonomous
driving systems.
VII. ACKNOWLEDGMENT
Author would like to thank Shefali, Department of Food
Technology and Biochemical Engineering, Jadavpur Univer-
sity, Kolkata for the valuable comments in improving the
manuscript. Author would also like to thank Department of
Electrical Engineering, Jadavpur University, Kolkata.
REFERENCES
[1] M. Suri, O. Bichler, D. Querlioz, O. Cueto, L. Perniola, V. Sousa,
D. Vuillaume, C. Gamrat, and B. DeSalvo, "Phase change memory as
synapse for ultra-dense neuromorphic systems: Application to complex
visual pattern extraction," in Electron Devices Meeting (IEDM), 2011
IEEE International.
IEEE, 2011, pp. 4–4.
[2] S. Yu, Y. Wu, R. Jeyasingh, D. Kuzum, and H.-S. P. Wong, "An
electronic synapse device based on metal oxide resistive switching
memory for neuromorphic computation," IEEE Transactions on Electron
Devices, vol. 58, no. 8, pp. 2729–2737, 2011.
[3] G. Indiveri, "A neuromorphic vlsi device for implementing 2d selective
attention systems," IEEE Transactions on Neural Networks, vol. 12,
no. 6, pp. 1455–1463, 2001.
[4] J. E. Dayhoff and J. M. DeLeo, "Artificial neural networks," Cancer,
vol. 91, no. S8, pp. 1615–1635, 2001.
[5] D. Nauck, F. Klawonn, and R. Kruse, Foundations of neuro-fuzzy
systems.
John Wiley & Sons, Inc., 1997.
[6] C.-T. Lin and C. Lee, Neural fuzzy systems: a neuro-fuzzy synergism to
intelligent systems. Prentice-Hall, Inc., 1996.
[7] D. J. Bakkum, Z. C. Chao, and S. M. Potter, "Spatio-temporal electrical
stimuli shape behavior of an embodied cortical network in a goal-
directed learning task," Journal of neural engineering, vol. 5, no. 3,
p. 310, 2008.
[8] A. Novellino, P. D'angelo, L. Cozzi, M. Chiappalone, V. Sanguineti, and
S. Martinoia, "Connecting neurons to a mobile robot: an in vitro bidirec-
tional neural interface," Computational Intelligence and Neuroscience,
vol. 2007, 2007.
[9] M. E. Ruaro, P. Bonifazi, and V. Torre, "Toward the neurocomputer:
image processing and pattern recognition with neuronal cultures," IEEE
Transactions on Biomedical Engineering, vol. 52, no. 3, pp. 371–383,
2005.
[10] J. B. George, G. M. Abraham, B. Amrutur, and S. K. Sikdar, "Robot
navigation using neuro-electronic hybrid systems," in 2015 28th Inter-
national Conference on VLSI Design.
IEEE, 2015, pp. 93–98.
[11] S. Biswas, "Proposal for ranking nodes of neural network using activity
index," in Research in Computational Intelligence and Communication
Networks (ICRCICN), 2016 Second International Conference on.
IEEE,
2016, pp. 224–228.
[12] S. Biswas, "Extraction of network information - quality and quantity -
from nodes of neuronal network." Accepted at 2016 fifteenth IEEE
International Conference on Information Technology, 2016.
[13] S. Biswas, "How diverse is the network information obtained from the
nodes of a biological neural network ?" Accepted at 2017 Eleventh
IEEE International Conference on Intelligent Systems and Control, 2017.
[14] S. Biswas, "Novel algorithm to spatially locate information and activity
hub in biological neuronal network." Accepted at 2017 Fourth IEEE
International Conference on Signal Processing and Integrated Networks,
2017.
[15] E. Bullmore and O. Sporns, "Complex brain networks: graph theoretical
analysis of structural and functional systems," Nature Reviews Neuro-
science, vol. 10, no. 3, pp. 186–198, 2009.
[16] D. A. Forsyth and J. Ponce, Computer vision: a modern approach.
Prentice Hall Professional Technical Reference, 2002.
[17] S. E. Umbaugh, Computer Vision and Image Processing: A Practical
Approach Using Cviptools with Cdrom. Prentice Hall PTR, 1997.
[18] C. M. Bishop, "Pattern recognition," Machine Learning, vol. 128, 2006.
[19] B. Thakur, A. Mukherjee, A. Sen, and A. Banerjee, "A dynamical
framework to relate perceptual variability with multisensory information
processing," Scientific Reports, vol. 6, 2016.
[20] M. E. J. Obien, K. Deligkaris, T. Bullmann, D. J. Bakkum, and U. Frey,
"Revealing neuronal function through microelectrode array recordings,"
Frontiers in neuroscience, vol. 8, p. 423, 2015.
[21] T. Watanabe and M. C. Raff, "Rod photoreceptor development
in
vitro: intrinsic properties of proliferating neuroepithelial cells change
as development proceeds in the rat retina," Neuron, vol. 4, no. 3, pp.
461–467, 1990.
[22] J. B. George, G. M. Abraham, K. Singh, S. M. Ankolekar, B. Amrutur,
and S. K. Sikdar, "Input coding for neuro-electronic hybrid systems,"
Biosystems, vol. 126, pp. 1–11, 2014.
|
1210.7165 | 1 | 1210 | 2012-10-26T15:15:08 | Cellular Adaptation Accounts for the Sparse and Reliable Sensory Stimulus Representation | [
"q-bio.NC",
"physics.bio-ph"
] | Most neurons in peripheral sensory pathways initially respond vigorously when a preferred stimulus is presented, but adapt as stimulation continues. It is unclear how this phenomenon affects stimulus representation in the later stages of cortical sensory processing. Here, we show that a temporally sparse and reliable stimulus representation develops naturally in a network with adapting neurons. We find that cellular adaptation plays a critical role in the transient reduction of the trial-by-trial variability of cortical spiking, providing an explanation for a wide-spread and hitherto unexplained phenomenon by a simple mechanism. In insect olfaction, cellular adaptation is sufficient to explain the emergence of the temporally sparse and reliable stimulus representation in the mushroom body, independent of inhibitory mechanisms. Our results reveal a computational principle that relates neuronal firing rate adaptation to temporal sparse coding and variability suppression in nervous systems with a sequential processing architecture. | q-bio.NC | q-bio |
1
Cellular Adaptation Accounts for the Sparse and Reliable
Sensory Stimulus Representation
Farzad Farkhooi1,∗, Anja Froese2, Eilif Muller3, Randolf Menzel2, Martin P. Nawrot1
1 Neuroinformatics & Theoretical Neuroscience, Freie Universitat Berlin, and Bernstein
Center for Computational Neuroscience Berlin, Berlin, Germany.
2 Institute fur Biologie-Neurobiologie, Freie Universitat Berlin, Berlin, Germany.
3 Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne, Lausanne, Switzerland.
∗ Corresponding author: Konigin-Luise-Strasse 1-3, 14195 Berlin, Germany
[email protected]
Abstract
Most neurons in peripheral sensory pathways initially respond vigorously when a preferred stimulus
is presented, but adapt as stimulation continues.
It is unclear how this phenomenon affects stimulus
representation in the later stages of cortical sensory processing. Here, we show that a temporally sparse
and reliable stimulus representation develops naturally in a network with adapting neurons. We find
that cellular adaptation plays a critical role in the transient reduction of the trial-by-trial variability of
cortical spiking, providing an explanation for a wide-spread and hitherto unexplained phenomenon by a
simple mechanism. In insect olfaction, cellular adaptation is sufficient to explain the emergence of the
temporally sparse and reliable stimulus representation in the mushroom body, independent of inhibitory
mechanisms. Our results reveal a computational principle that relates neuronal firing rate adaptation
to temporal sparse coding and variability suppression in nervous systems with a sequential processing
architecture.
Introduction
The phenomenon of spike-frequency adaptation (SFA) [1] (also known as spike-rate adaptation) is a
fundamental process in nervous systems, which attenuates the neuronal stimulus responses to a lower level
following an initial high firing. This process can be mediated by different cell-intrinsic mechanisms that
involve a spike-triggered self-inhibition, and which can operate in a wide range of time scales [3,25]. These
processes are probably related to the early evolution of the excitable membrane [36,38] and are common to
vertebrate and invertebrate neurons, both in the peripheral and central nervous system [49]. Nonetheless,
the functional consequences of SFA in peripheral stages of sensory processing on the representation of
stimuli in later network stages remain unclear. For instance, while olfactory stimuli generally display a
slow kinetics, the early olfactory system shows temporally adaptive and precise onset responses [19, 40],
that may translate into narrow integration windows for the principal neurons in the next processing stage,
the piriform cortex pyramidal cells in rodents [35] and the mushroom body Kenyon cells in insects [34].
However, it remains unclear how sparse single-neuron responses can reliably encode information in face
of the response variability [41] and the sensitivity of cortical networks to small perturbations [24].
Here, we show that the SFA mechanism introduces a dynamical non-linearity in the transfer function
of neurons. Subsequently, the response onset becomes progressively sparser when transmitted across
successive processing stages. Additionally, the self-regulating effect of SFA causes a stimulus-triggered
reduction of firing variability by modulating the average inhibition in the balanced cortical network. In
this manner the temporally sparse representation is accompanied by an increased response reliability. Us-
ing this theoretical framework on data obtained from the olfactory system of the honeybee, we show that
2
Figure 1. Neuronal adaptation in the multi-stage processing network. (A) Schematic
illustration of a three-layered model of an adaptive pathway of sensory processing. The network consists
of three consecutive adaptive populations. Each population receives sensory input from an afferent
source (black arrows) and independent constant background excitation (dashed arrow). Input is
modeled by a Gaussian density and a sensory stimulus presented to the first population is modeled by
an increase in the mean input value. (B) Response profiles. The evoked state consists of a phasic-tonic
response in all populations. The tonic response level is decremented across the consequtive populations.
(C) Temporal sparseness Si
average spike count at t = 500 ms in the first population. Responses become progressively sparser as the
stimulus propagates into the network. (D) Secondary response profiles. The additional jump increase in
stimulus strength at t = 1000 ms during the evoked state of the first stimulus results in a secondary
phasic response in all populations with an amplitude overshoot in the 2nd and 3rd population.
r is measured by the integral over the firing rate and normalized by the
the sequential effect of the SFA shapes the Kenyon cells' temporal sparseness independent of inhibitory
synaptic mechanisms.
Our results reveal a generic, biophysically plausible mechanism that can explain the emergence of a
reliable and sparse stimulus representation, indicating the importance of cellular adaptation in sensory
computation at the network level.
RESULTS
Temporal sparseness emerges in successive adapting populations
To examine how successive adapting populations can achieve temporal sparseness, first we mathematically
analyzed a sequence of neuronal ensembles (Figure 1A), where each ensemble exhibits a generic model
of mean firing rate adaptation by means of a slow negative self-feedback [3, 22, 29] (Materials and
Pop. 1Pop. 2Pop. 3ABCD 3
Methods). This sequence of neuronal ensembles should be viewed as caricature for distinct stages in
the pathway of sensory processing. For instance in the mammalian olfactory system the sensory pathway
involves several stages from the olfactory sensory neurons to the olfactory bulb, the piriform cortex, and
then to higher cortical areas (Figure 1A).
The mean firing rate in the steady-state of a single adaptive population can be obtained by solving
the rate consistency equation, ¯ri = f (µi − τsqs¯riσ2
i ), where ¯r is the equilibrium mean firing rate of the
population i, µi and σ2
i are mean and variance of the total input into the population, respectively, f
is the response function (input-output transfer function, or f I-curve) of the population mean activity,
qs is the quantal conductance of the adaptation mechanism per unit of firing rate, and τs is the firing
rate relaxation time constant [3, 22, 29]. It is known that any sufficiently slow modulation (1/rmax < τs)
linearizes the steady-state solution, ¯ri, due to the self-inhibitory feedback being proportional to the firing
rate [10].
Here, for simplicity, we studied the case where all populations in the network exhibit the same initial
steady-state rate. This is achieved by adjustment of a constant background input to population i (doted
arrows, Figure 1A), resembling the stimulus irrelevant interactions in the network. All populations are
coupled by the same strength g. First, we calculated the average firing rate dynamics of the populations'
responses following a step increase in the mean input to the first layer (black arrow, Figure 1A). By
solving the dynamics of the mean firing rate and adaptation level concurrently, we obtained the mean-
field approximation of the populations' firing rates (Materials and Methods). The responses of each
population consisted of a phasic and a tonic part, typical for adapting neurons. Here, we plotted the mean
firing rate of three consecutive populations (Figure 1B). The phasic response to a step increase in the
input is presented across stages. However, the tonic response becomes increasingly suppressed in the later
stages (Figure 1B). This phenomenon is a general feature of successive adaptive neuronal populations.
To understand why the tonic response part, after the adaptation was reinstalled, is increasingly suppressed
in later stages we used the linearity of the adaptive transfer function and its stability conditions at the
steady-state [10]. Thus, a change in the mean rate of population i is the solution of ∆¯ri = f (g∆¯ri−1 −
∆¯riτsqs), and since the transfer function, f , is assumed to be increasing and the slope of ¯r is less
than unity, the solution can only existe for ∆¯ri−1 > ∆¯ri (Materials and Methods). The result in
this sub-section (Figure 1B-C) was established with a current based leaky integrated-and-fire response
function. However, the analysis presented here extends to the majority of neuronal transfer functions
since the stability and linearity of the adapted steady-states are granted [10, 22]. This simple effect
leads to a progressively sparser representation across successive stage of a generic feed-forward adaptive
processing. We quantified the temporal sparseness by Si
0 ri(t)dt, where ri(t) is the mean firing
rate of population i and T is the observation time window. Normalization of this measure by the first
population S1
r (T = 500) indicates sparser responses in later stages of the adaptive network (Figure 1C).
Does the suppression of the adapted response level impair the information about the presence of the
stimulus? To explore this, we studied a secondary increase in stimulus strengths of equal magnitude after
1 second when the network has relaxed to the evoked equilibrium (Figure 1D). The secondary stimulus
jump induced a secondary phasic response of comparable magnitude in the first population (Figure 1D).
However, in the later populations this jump evoked an icreased peak rate in the phasic response (Figure
1D). Notably, the coupling factor g between the populations shapes this phenomena. Here, we adjusted
g to achieve an equal onset response magnitude across the populations for the first stimulus jump at
t = 0, and a slight increase in the population onset response in the first population is amplified in the
later stages. This is due to the fact that the later stages accumulated less adaptation in their evoked
steady-state (level of adaption is proportional to mean firing rate).
Importantly, this result confirms
that the sustained presence of the stimulus is indeed stored in the level of cellular adaptation [32], even
though it is not reflected in the firing rate of the last population. Therefore, regardless of the absolute
amplitude of responses, the relative relation between secondary and initial onset keeps increasing across
layers. This type of a secondary overshoot is also experimentally known as sensory sensitization, where
r(T ) =(cid:82) T
4
an additional increase in the stimulus strength significantly enhances the responsiveness of later stages
after the network converged to an adapted steady-state [18].
Adaptation increases response reliability in the cortical network
The mean firing rate approach as above is insufficient to determine how reliable the observed response
transients are across repeated stimulations. In the prevailing spiking network model of the cortex, the bal-
ance of excitation and inhibition is quickly reinstalled within milliseconds after the onset of an excitatory
input [47] and self-generating recurrent fluctuations strongly dominate the dynamics of the interactions.
It has thus been questioned whether a sparse response of only very few action potentials could reliably
encode the presence of a stimulus [24].
To investigate the reliability of adaptive mapping from a dense stimulus to a sparse cortical spkie
response across successive processing stages we employed an adaptive population density formalisms
[11, 29] (Materials and Methods) along with numerical network simulations. We embeded a two-
layered sensory network with an afferent ensemble projecting to a cortical network (Figure 2A). The
afferent ensemble consisted of 4,000 adaptive neurons that included voltage dynamics, conductance-based
synapses, and spike-induced adaptation [29]. It resembles the sub-cortical sensory processing and each
neuron in the afferent ensemble projects to randomly 1% of the neurons in the cortical network. This is
a large circuit of the balanced network (Figure 2A) with 10,000 excitatory and 2,500 inhibitory neurons
with a typical random diluted connectivity of 1%. The spiking neuron model in the cortical network
again includes voltage dynamics, conductance-based synapses, and spike-induced adaptation [29]. All
neurons are alike and parameters are given in Table.3 in Muller et al. [29]. With appropriate adjustment
of the synaptic weights, the cortical network operates in a globally balanced manner, producing irregular,
asynchronous activity [20, 47, 48]. The distribution of firing rates for the network approximates a power-
law density [37] with an average firing rate of ≈ 3.0 Hz (Figure 2B) and the coefficients of variation (Cv)
for the inter-spike intervals is centered at a value slightly greater than unity (Figure 2C) indicating the
globally balanced and irregular state of the network [48]. Noteworthy, the activity of neurons in both
stages is fairly incoherent and spiking in each sub-network is independent. Therefore, one can apply an
adiabatic elimination of the fast variables and formulate a population density description where a detailed
neuron model reduces to a stochastic point process [11,29] that provides an analytical approximation and
understanding of the simulation results in this section (Materials and Methods).
The background input is modeled as a set of independent Poisson processes that drive both sub-
networks (dashed arrows, Figure 2A). The stimulus dependent input is an increase in the intensity of
the Poisson input into the afferent ensemble (solid arrow, Figure 2A). Before the stimulus became active
at time t = 0, a typical neuron showed an irregular spiking activity in both network stages (Figure 2B-
D). However, when a sufficiently strong stimulus is applied, neurons in both stages exhibited a transient
response before the population mean firing rate converges back to a new level of steady-state (Figure
2D,E). The population firing rate in the numerical simulations (solid line, Figure 2E) follow well the
adaptive population density treatment (filled circles, Figure 2E).
To measure the effect of neuronal adaptation on the temporal sparseness, we again computed the
number of spikes per neuron after the stimulus onset, Se(T ). We compare our standard adaptive network
with an adaptation time constant of τs = 110ms (solid lines, Figure 2F) with an only weakly-adaptive
control network (τs =30ms; dashed lines, Figure 2F). Note, that the adaptation time constant in the
weakly adaptive network is about equal to the membrane time constant and therefore plays a minor role
for the network dynamics. It showed that both sub-networks generated sharp phasic response, which in
the case of the cortical network evoked a single sharply timed spike within the first ≈ 10 − 20 ms in a
subset of neurons (Figure 2B,F). In the control case, the response is non-sparse and response spikes are
distrubuted throughout the stimulus period (dashed lines, Figure 2F). Overall, strong adaptation reduces
the total number of stimulus-induced action potentials per neuron and concentrates their occurrence
within an initial brief phase following fast changes in the stimulus. Thus, in accordance with the results
5
of the rate-based model in the previous section, one can conclude that the sequential stimulation of
adaptive network stages can account for the emergence of a sparse stimulus representation in a cortical
population.
To reveal the effect of adaptation on the response variability, we employed the time-resolved Fano
factor [31], F , which measures the spike-count variance divided by the mean spike count across 200
repeated stimulations. Spikes were counted in a 50ms time window and a sliding of 10ms [9]. As before, we
compared our standard adaptive network (Figure 2G,I; τs = 110ms) with the control network (Figure
2H,J; τs=30ms). Since the Fano factor is known to be strongly dependent on the firing rate, we adjusted
the input to the latter such that the steady state firing rates in both networks were mean-matched [9].
The input Poisson spike trains (F = 1) translated into slightly more regular spontaneous (t < 0) activity
in the afferent ensemble (Figure 2G), as neuronal membrane filtering and refractoriness reduced the
output variability. After the stimulus onset (t = 0), due to the increase in the mean input rate, the
average firing rate increased, however the variance of the number of events per trial did not increase
proportionally. Therefore, we observed a reduction in the Fano factor (Figure 2G). This phenomena
is independent of the adaptation mechanism in the neuron model and a quantitatively similar reduction
can be observed in the weakly adaptive afferent ensemble (Figure 2H). A comparison between our
standard adaptive and the control case reveals that the adaptive network is generally more regular in the
background and in the evoked state (Figure 2G,H). This is due to the previously known effect, where
adaptation induces negative serial correlations in the inter-spike intervals [5, 11] and as a result reduces
the Fano factor [7, 11].
In the next stage of processing, the distribution of F across neurons during spontaneous activity is
high due to the self-generated noise of the balanced circuits [23, 47]. This closely follows a wide spread
experimental finding where F > 1 in the spontaneous cortical activity [9] ( Figure 2G,H). Whenever
a sufficiently strong stimulus was applied, the internally generated fluctuations in the adaptive balanced
network were transiently suppressed, and as a result the Fano factor dropped sharply (Figure 2I).
However, this reduction of the Fano factor is a temporary phenomenon and F converges back to slightly
above the baseline variability (Figure 2I). At the same time, the evoked steady-state firing returned back
to the irregular and asynchronous state (Figure 2D). Indeed, this effect corresponds to a temporally
mis-match in the balanced input conditions to the cortical neurons since the self-inhibitory and slower
adaption effect prevents a rapid adjustment to the new input regime. This can be observed in the time
course of variability suppression that closely reflects the time constant of adaptation (Figure 2I). The
afferent ensemble structured the input to the cortical ensemble, determining the magnitude of the observed
reduction effect, and under the control condition where a pure Poisson input is provided to the cortical
balanced network, the reduction in F is reduced but the time scale of recovery remains unaltered (crosses,
Figure 2I). We contrast this adaptive behavior with the variability dynamics in the weakly adaptive
balanced network (Figure 2J). In this case there is no reduction in F , because for a short adaptation
time constant the convergence to the balanced state is very rapid [47]. The small increase in the input
noise strength leads to an increase of the self-generated randomness of the balanced network [23, 28].
To further understand the effect of the recurrent connectivity on the variability, one can compare
the evoked variability between the afferent ensemble and the cortical network. In the latter, fluctuation
intrinsic to the cortical networks dynamically bring back the cortical circuits to the spontaneous level
of high variability within the time scale of adaption relaxation (Figure 2I). This suggests the transient
suppression of variability during a stimulus response as a cortical-wide effect [9] could be caused by slow
adjustment of self-regulatory feedback.
6
Adaptive networks generate sparse and reliable responses in the insect olfac-
tory system
As a case study, we investigated the role of adaptation for the emergence of the sparse temporal code
in the insects olfactory system, which is analogoue to the mammalian olfactory system. We simulated a
reduced generic model of olfactory processing in insects using the adaptive neuron model [29]. The model
network consisted of an input layer with 1,480 olfactory sensory neurons (OSNs), which project to the
next layer representing the antennal lobe circuit with 24 projection neurons (PNs) and 96 inhibitory local
inter-neurons (LNs) that form a local feed-forward inhibitory micro-circuit with the PNs. The third layer
holds 1,000 Kenyon cells (KCs) receiving divergent-convergent input from PNs. The relative numbers
for all neurons approximate the anatomical ratios found in the olfactory pathway of the honeybee [27]
(Figure 3A). We introduced heterogeneity among neurons by randomizing their synaptic time constants
and the connectivity probabilities are chosen according to anatomical studies. Synaptic weights were
adjusted to achieve spontaneous firing statistics that match the observed physiological regimes. The SFA
parameters were identical throughout the network with τs = 110 ms (see Materials and Methods for
details).
Using this model, we sought to understand how adaptation contributes to temporally sparse odor
representations in the KC layer in a small sized network and under highly fluctuating input conditions. We
simulated the input to each OSN by an independent Poisson process, which is thought to be reminiscent
of the transduction process at the olfactory receptor level [30]. Stimulus activation was modeled by a step
increase in the Poisson intensity with uniformly jittered onset across the OSN population (Materials
and Methods). Following a transient onset response the OSNs adapted their firing to a new steady-state
(Figure 3B,C). The pronounced effect of adaptation becomes apparent when the adaptive population
response is compared to the OSN responses in the control network without any adaptation (τs=0; dashed
line, Figure 3C). In the next layer, the PN population activity is reflected in a dominant phasic-tonic
response profile (green line, Figure 3D), which closely matches the experimental observation [19]. This
is due to the self-inhibitory effect of the SFA mechanism, and to the feedforward inhibition received
from the LNs (magenta line, Figure 3D). Consequently, the KCs in the third layer produced only very
few action potentials following the response onset (red line , Figure 3E,G). The average number of
emitted KC response spikes per neuron, Se, is small in the adaptive network (average < 2) whereas
KCs continue spiking throughout stimulus presentation in the non-adaptive network (Figure 3G). This
finding closely resembles experimental findings of temporal sparseness of KC responses in different insect
species [6,34,43] and quantitatively matches the KC response statistics provided by Ito and colleagues [15].
The simulation results obtained here confirm the mathematically derived results in the first results section
(Figure 1) and show that neuronal adaption can cause a temporal sparse representation even in a fairly
small and highly structured layered network where the mathematical assumptions of infinite network size
and fundamentally incoherent activity are not full filled.
To test the effect of inhibition in the LN-PN micro-circuitry within the antennal lobe layer on the
emergence of temporal sparseness in the KC layer, we deactivated all LN-PN feedforward connections
and kept all other parameters fix. We found a profound increase in the amplitude of the KC popu-
lation response, both in the adaptive (red line,Figure 3F) and the non-adaptive network (dashed red
line,Figure 3F). This increase in response amplitude is carried by an increase in the number of respond-
ing KCs due to the increased excitatory input from the PNs, implying a strong reduction in the KC
population sparseness. Importantly, removing local inhibition did not alter the temporal profile of the
KC population response in the adaptive network (cf. red lines in Figure 3E,F), and thus temporal
sparseness was independent of inhibition in our network model.
How reliable is the sparse spike response across trials in single KCs? To answer this question, we
again measured the robustness of the stimulus representation by estimating the Fano factor across 200
simulation trials (Figure 3H). The network with adaptive neurons and inhibitory micro-circuitry ex-
hibited a low Fano factor (median ≈ 0.3) and a narrow distribution across all neurons. This follows the
7
experimental finding that the few spikes emitted by KCs are highly reliably [15] (network 1, Figure 3H).
Turning off the inhibitory micro-circuitry did not significantly change the response reliability (Wilcoxon
rank sum test, p-value = 0.01; network 2, Figure 3H). However, both networks that lacked adaptation
exhibited a significantly increased variability with a median Fano factor close to one (Wilcoxon rank sum
test, p-value = 0.01; networks 3 and 4, Figure 3H), independent of the presence or absence of inhibition
within the antennal lobe.
To explore whether neuronal adaptation could be the responsible mechanism for temporal sparseness
in the biological network, we performed a set of Calcium imaging experiments, monitoring Calcium
responses in the KC population of the honeybee mushroom body [43] (Materials and Methods). Our
computational model (Figure 3) predicted that blocking of the inhibitory microcircuit would increase
the population response amplitude but should not alter the temporal dynamics of the KC population
response. In a set of experiments, we tested this hypothesis by comparison of the KCs' evoked activity
in the presence and absence of GABAergic inhibition (Materials and Methods). First, we analyzed
the normalized Calcium response signal within the mushroom body lip region in response to a 3s, 2s,
1s and 0.5s odor stimulus (Figure 4A). We observed the same brief phasic response following stimulus
onset in all four cases with a characteristic slope of Calcium response decay that has been reported
previously [43]. Bath application of the GABAA antagonist picrotoxin (PTX) did not change the time
course of the Calcium response dynamics (Figure 4B-D). The effectiveness of the drug was verified by
the increased population response amplitude in initial phase (Figure 4C). Next, we tested the GABAB
antagonist hydrochloride (CGP) using the same protocol and again found an increase in the response
magnitude but no alteration of the response dynamics (Figure 4E,F).
DISCUSSION
We here propose that a simple neuron-intrinsic mechanism of spike-triggered adaptation can account for
a reliable and temporally sparse sensory stimulus representation across stages of sensory processing. At
the single neuron level SFA is known to induce the functional property of a fractional differentiation with
respect to the temporal profile of the input and thus offers the possibility of tuning the neuron's response
properties to the relevant stimulus time-scales at the cellular level [3, 4, 25]. Our results indicate that
a sensory processing in a feedforeward network with adaptive neurons focuses on the temporal changes
of the sensory input in a precise and temporally sparse manner (Figure 1B; Figure 2E and Figure
3E). At the same time the constancy of the stimulus is memorized in the cellular level of adaptation [32]
(Figure 1D). We hypothesize that at later processing stages in the sensory pathway, notably in the
sensory cortices and in the insect mushroom body, it is most relevant to encode and process relevant
dynamic changes in the sensory environment and to neglect the static stimulus features as well as the
dynamic fluctuations of receptor sensations present at very fast time scales. One prominent example is
primate vision where, in the absence of the self-generated dynamics of retina input due to microsaccades,
observers become functionally blind to stationary objects during fixations [26].
The emergence of a sparse representation has been demonstrated in various cortical sensory areas, for
example in visual [44], auditory [14], somatosensory [16], and olfactory [35] cortices, and thus manifests
a principle of sensory computation across sensory modalities and independent of the natural stimulus
kinetic. However, it has been repeatedly questioned whether a few informative spikes can survive in
the cortical network, which is highly sensitive to small perturbations [24, 50]. Our results show that
a biologically realistic cellular mechanism implemented at successive network stages can transform a
dense and highly variable Poisson input at the periphery into a sparse and highly reliable ensemble
representation in the cortical network. These results reflect previous theoretical evidence that SFA has
an extensive synchronizing-desynchronizing effect on population responses in a balanced network [46].
However, cellular adaptation has largely been neglected in theory and simulation of cortical networks,
with very few exceptions [46], although it facilitates a transition from a dense rate code to a temporal
8
code expressed in the concerted spiking of cortical cell assemblies [21] (Figure 2D).
Importantly, the adaption level adjusts with a dynamics that is slow compared to the dynamics
of excitatoary and inhibitory synaptic inputs. This circumstance allows for a transient mismatch of
the balanced state in the cortical network and thus leads to a transient reduction of the self-generated
(recurrent) noise (Figure 2I). This, in turn, explains why the temporally sparse representation can
be highly reliable. This result seems counter-intuitive in the light of previous studies that consistently
stressed the persistence of noise in the balanced network [23, 24, 28]. However, cellular adaptation has
largely been neglected in theory and simulation of cortical networks with very few exceptions [46].
Our results provide an understanding of the role of slow cellular adaptation dynamics for the transient
suppression of the trial-by-trial response variability. We thus arrived at a coherent model explanation
for the previously unexplained and wide-spread experimental phenomenon of a stimulus-driven transient
reduction of response variability in cortical neurons, which has been observed in different species and
across many cortical areas, from occipital to frontal, as e.g. reported by Churchland and colleagues for
14 independent datasets [9].
The insect olfactory system is experimentally well investigated and examplifies a pronounced sparse
coding scheme at the level of the mushroom body KCs. The olfactory system is analog in invertebrates
and vertebrates and the sparse stimulus representation is likewise observed in the pyramidal cells of the
piriform cortex [35], and the rapid responses in the mitral cells in the olfactory bulb [40] compare to those
of projection neurons in antennal lobe [19]. Our adaptive network model, designed in coarse analogy to the
insect olfactory system, produced increasingly phasic population responses as the stimulus-driven acitivty
propagated through the network. Our model results closely match the repeated experimental observation
of temporally sparse KC responses in extracellular recordings from the locust [34] and manduca [6, 15],
and in Calcium imaging in the honeybee [43].
In our experiments we could show that systemic blocking of GABAergic transmission did not affect the
temporal sparseness of the KC population response in the honeybee (Figure 4). This result might seem to
contradict former studies that stressed the role of inhibitory feed-forward [2] or feedback inhibition [13,33]
for the emergence of KC sparseness. However, the suggested inhibitory mechanisms and the sequential
effect in the adaptive network proposed here are not mutually exclusive and may act in concert to establish
and maintain a temporal and spatial sparse code in a rich and dynamic natural olfactory scene.
The adaptive network model manifests a low trial-to-trial variability of the sparse KC responses that
typically consist of only 1-2 spikes. In consequence, a sparsly activated KC ensemble is able to robustly
encode stimulus information. The low variability at the single cell level (Figure 3H) carries over to a low
variability of the population response [11]. This benefits downstream processing in the mushroom body
output neurons that integrate converging input from many KCs [27], and which were shown to reliably
encode odor-reward associations in the honeybee [42].
Our results here bear consequences of general importance for sensory coding theories. A mechanism of
self-inhibition at the cellular level can facilitate a temporally sparse ensemble code but does not require a
well adjusted interplay between the excitatory and inhibitory circuitry at the network level. This network
effect is robust due to the distributed nature of the underlying mechanism, which acts independently in
each single neuron. The regularizing effect of self-inhibition increases the signal-to-noise ratio not only of
single neuron responses but also of the neuronal population activity [11] that is postsynaptically integrated
in downstream neurons.
Materials and Methods
Rate model of generic feedforward adaptive network
To address analytically the sequential effect of adaptation in a feedforward network, we consider a model
in which populations are described by their firing rates. Although firing rate models typically provide a
fairly accurate description of network behavior when the neurons are firing asynchronously [45], they do
not capture all features of realistic networks. Therefore, we verify all of our predictions with a population
density formalism [11] as well as a large-scale simulation of realistic spiking neurons. To determine the
mean activity dynamics of a consecutive populations, we employed an standard mean firing rate model
of population i as
(cid:40)
ri = −ri + f (gri−1 + µi − aiσi)
ai = − 1
ai + riqs
τs
9
(1)
¯ri = f (mi − τsqs¯riσi).
where f is the transfer input-output function, τs is the adaptation time-scale, g is the coupling factor
between two populations and a is the adaptive negative feedback for the population i with qs strength
and σ is the standard deviation of the input. For simplicity we define mi = gri−1 + µi. Given f the
equilibrium can be determined by
(2)
This fix point is always stable [22] and the condition for the stability reads ∂rif (m − τsqs¯riσi)¯ri < 1
and ∂mi f (m − τsqs¯riσi)¯r < 0. It is also known that the ¯ri is a linear function, given a sufficiently slow
adaptation τs and a non-linear shape of f [10]. Thus, if ri−1 = ri in two subsequent populations, and
increased input to population i − 1 leads to smaller increase for next population i and the adapted level
of responses satisfy ∆ri−1 > ∆ri.
The stimilus onset response lies between the adapted steady-state and not adapted responses function
f given the level of new input and can be analytical calculated accordingly [3].
Population density approach to the adaptive neuronal ensemble
In Muller et al. [29] it is shown that by an adiabatic elimination of fast variables a detailed neuron
model including voltage dynamics, conductance-based synapses, and spike-induced adaptation reduces
to a stochastic point process. Thus, we define an orderly point process with a hazard function argument
with state variable x as
Pr(N [t, t + ∆t) > 0x, t)
hx(x, t) := lim
∆t→0+
∆t
.
(3)
where N [t, t + ∆t) is the number of events in ∆t. We assume the dynamic of the adaptation variable is
x = − 1
τs
x +
δ(t − tk)qs,
(4)
(cid:88)
k
(cid:21)
(cid:20) x
Pr(x, t)
− hx(x, t) Pr(x, t).
τs
where tk is the time of kth spike in the ensemble. Thus, the state variable distribution at time t in the
ensemble is governed by a master equation of the form
∂t Pr(x, t) = −∂x
+ hx(x − qs, t) Pr(x − qs, t)
(5)
We solve the Eq. 5 with the help of the transformations ts = η(x) := −τ log(x/q) and ψ(ts) = η(η−1(ts +
q)) numerically [29].
It turns out that indeed hx is the input-out transfer function of neurons in the
network where its instantaneous parameters are give by the input statistics [29]. Here, we used the mean-
field formalism developed by Lechner et al. [23] to approximately determine the averaged input within a
standard balanced network, as the parameters of the hazard function hx. Therefore, the functional form
of its solution in a compact form is
Pr(tst) = ktΩ(tst)
(6)
s is the initial condition state of the system and kt is a constant defined by (cid:82) Pr(tst)dts = 1.
where t0
Hence, it can be shown that the firing rate and the consistency equation of the ensemble is
rt =
hts (tst)Pr(tst)dts.
10
(7)
(8)
(cid:90) ∞
−∞
(cid:90) (cid:90)
Now, by applying the techniques in Farkhooi et al [11], we define a joint probability density as
ρn(tn, tn
xt0
x, t)
where an nth event occurs at time tn > t and the state variable is tn
and state of adaptation joint density recursively
x. We can write n + 1th event time
ρn+1(tn+1, tn+1
x
t0
x, t) =
ρ(tn+1 − tn, tn+1
x
tn
x, t)ρn(tn, tn
xt0
x, t)dtndtn
x
(9)
To simplify the integral equations, we use Bra-Ket notation and thereafter, we derive the Laplace trans-
form of the joint density by
where ρ1(t1, t1
x, t) = ρ(t1, t1
st0
t0
s, t) = ρn(s, tn
tn
x, t)(cid:105)
x, t) Next, we define the operator Pn(st),
ρn+1(s, tn+1
xt0
x, t)ρ(s, tn+1
xt0
x
x
ρn(st) = (cid:104)1 Pn(st) Pr(ts, t)(cid:105) = (cid:104)1 Pn(st) Pr(tx, t)(cid:105).
Now, by employing P (1, st) = 1/(µ1s2)[ρ2(st) − 2ρ1(st) + 1] as in [11], we derive
(10)
(11)
P (n, st) = 1/(µ1s2)[ρn+1(st) − 2ρn(st) + ρn−1(st)],
(12)
where P (n, st) is the Laplace transform of the probability density of observing n events in a given time
window. Now we derive
k ρk(st) = (cid:104)1 P(st)/(I − P(st)) Pr(st)(cid:105)
(13)
Ast =(cid:80)
(cid:90) T
where I is the identity operator. This equation represents the Laplace transfom of the autocorrelation
function. Thus, the Fano factor is Jst = 1/µ1s2 − (1 + 2 Ast) and the inverse Laplace transform is
JTt = 1 + (2/T )
(T − u)A(ut)du − T rt,
(14)
where A(ut) = L−1[ Ast].
0
Computational model of insect olfactory coding
Our model neuron is a general conductance-based integrate-and-fire neuron with spike-frequency adap-
tation as it is proposed in Muller et al. [29]. The model phenomenologically captures a wide array of
biophysical spike-frequency mechanisms such as M-type current, afterhyperpolarization (AHP-current)
and even slow recovery from inactivation of the fast sodium current [29]. The model neuron is also known
to perform high-pass filtering of the input frequencies following the universal model of adaptation [3].
Neuron parameters used follow the Table.3 in Muller et al. [29] The conductance model used for the static
synapses between the neurons is alpha-shaped with gamma distributed time constants from Γ(2.5, 4) and
Γ(5, 4) for excitatory and inhibitory synapses, respectively. All simulations were performed using the
NEST simulator [12] version 2.0beta and the Pynest interface.
11
The network connectivity is straight forward: each PN and LN receives excitatory connections from
20% randomly chosen OSNs [8, 39]. Additionally, every PN receives input from 50% randomly chosen
inhibitory LNs [8, 39]. In our model the PNs do not excite one another and each PN output diverges to
50% randomly chosen KCs [2, 17, 43].
We tuned the simulated network by adjusting the synaptic weights to achieve the same spontaneous
firing rate as reported experimentally: OSNs 15-25Hz [30], LNs 4-10Hz [8], PNs 3-10 [19] and KCs
0.3-1.0Hz [15].
Experimental methods
Experiments were performed following the methods published in Szyszka et al [43]. In summary, foraging
honeybees (Apis mellifera) were caught at the entrance of the hive, immobilized by chilling on ice, and
fixed in a plexiglas chamber before the head capsule was opened for dye injection. We retrogradely stained
clawed Kenyon cells (KC) of the median calyx, using the calcium sensor FURA-2 dextran (Molecular
Probes, Eugene, USA) with a dye loaded glass electrode, which was pricked into KC axons projecting to
the ventral median part of the α-lobe [43]. After dye injection the head capsule was closed, bees feed and
kept in a dark humid chamber for several hours.
The processing of imaging data was performed with custom written routines in IDL (RSI, Boulder,
CO, USA). In summary, changes in the calcium concentration were measured as absolute changes of
fluorescence: a ratio was calculated from the light intensities measured at 340nm and 380nm illumination
and the background fluorescence before odor onset was subtracted leading to ∆F with F = F340/F380.
Odor stimulation was preformed under a 20x objective of the microscope, the naturally occurring plant
odor octanol (Sigma Aldrich, Germany), diluted 1:100 in paraffine oil (FLUKA, Buchs, Switzerland), was
delivered to both antennae of the bee using a computer controlled, custom made olfactometer. To this,
odor loaded air was injected into a permanent airstream resulting in a further 1:10 dilution. Stimulus
duration was 3 seconds if not mentioned otherwise. The air was permanently exhausted.
For GABA blockage, a solution of 150 µl GABA receptor antagonist dissolved in ringer for final
concentration (10−5M picrotoxin (PTX, Sigma Aldrich, Germany) or 5x10−4M CGP54626 (CGP, Tocris
Bioscience, USA)) was bath applied to the brain after pre-treatment measurements. Measurements
started 10 min after drug application.
Acknowledgments
We wish to thank Peter Latham and Yifat Prut for helpful comments on this manuscript. Generous
funding was provided by the Bundesministerium fur Bildung und Forschung (Grant No.01GQ0941) to
the Bernstein Focus Neuronal Basis of Learning (BFNL) and by the Deutsche Forschungsgemeinschaft
(DFG) to the Collaborative Research Center for Theoretical Biology (SFB 618) and the DFG grant to
R.M. and A.F. (Me 365/31-1).
References
1. E. D. Adrian. The impulses produced by sensory nerve endings. The Journal of physiology,
61(1):4972, 1926.
2. Collins Assisi, Mark Stopfer, Gilles Laurent, and Maxim Bazhenov. Adaptive regulation of sparse-
ness by feedforward inhibition. Nat Neurosci, 10(9):1176 -- 84, September 2007.
3. Jan Benda and Andreas V. M. Herz. A universal model for spike-frequency adaptation. Neural
Computation, 15(11):2523 -- 2564, November 2003.
12
4. Jan Benda, Andre Longtin, and Len Maler. Spike-frequency adaptation separates transient com-
munication signals from background oscillations. J. Neurosci., 25(9):2312 -- 2321, March 2005.
5. Jan Benda, Leonard Maler, and Andr Longtin. Linear versus nonlinear signal transmission in
neuron models with adaptation currents or dynamic thresholds. Journal of Neurophysiology,
104(5):2806 -- 2820, November 2010.
6. Bede M Broome, Vivek Jayaraman, and Gilles Laurent. Encoding and decoding of overlapping
odor sequences. Neuron, 51(4):467 -- 82, August 2006.
7. Maurice J. Chacron, Leonard Maler, and Joseph Bastian. Electroreceptor neuron dynamics shape
information transmission. Nature Neuroscience, 8(5):673 -- 678, 2005.
8. Ya-Hui Chou, Maria L Spletter, Emre Yaksi, Jonathan C S Leong, Rachel I Wilson, and Liqun
Luo. Diversity and wiring variability of olfactory local interneurons in the drosophila antennal
lobe. Nature Neuroscience, 13(4):439 -- 449, February 2010.
9. Mark M Churchland, Byron M Yu, John P Cunningham, Leo P Sugrue, Marlene R Cohen, Greg S
Corrado, William T Newsome, Andrew M Clark, Paymon Hosseini, Benjamin B Scott, David C
Bradley, Matthew A Smith, Adam Kohn, J Anthony Movshon, Katherine M Armstrong, Tirin
Moore, Steve W Chang, Lawrence H Snyder, Stephen G Lisberger, Nicholas J Priebe, Ian M
Finn, David Ferster, Stephen I Ryu, Gopal Santhanam, Maneesh Sahani, and Krishna V Shenoy.
Stimulus onset quenches neural variability: a widespread cortical phenomenon. Nat Neurosci,
13(3):369 -- 378, March 2010.
10. B. Ermentrout. Linearization of FI curves by adaptation. Neural computation, 10(7):17211729,
1998.
11. Farzad Farkhooi, Eilif Muller, and Martin P. Nawrot. Adaptation reduces variability of the neuronal
population code. Physical Review E, 83(5):050905, May 2011.
12. Marc-Oliver Gewaltig and Markus Diesmann. NEST (NEural simulation tool). Scholarpedia,
2(4):1430, 2007.
13. N. Gupta and M. Stopfer. Functional analysis of a higher olfactory center, the lateral horn. Journal
of Neuroscience, 32(24):8138 -- 8148, June 2012.
14. Tom Hromdka, Michael R DeWeese, and Anthony M Zador. Sparse representation of sounds in
the unanesthetized auditory cortex. PLoS Biol, 6(1):e16, January 2008.
15. Iori Ito, Rose Chik-ying Ong, Baranidharan Raman, and Mark Stopfer. Sparse odor representation
and olfactory learning. Nature Neuroscience, 11(10):1177 -- 1184, September 2008.
16. Shantanu P Jadhav, Jason Wolfe, and Daniel E Feldman. Sparse temporal coding of elementary
tactile features during active whisker sensation. Nature Neuroscience, 12(6):792 -- 800, May 2009.
17. Ron A. Jortner, S. Sarah Farivar, and Gilles Laurent. A simple connectivity scheme for sparse
coding in an olfactory system. The Journal of Neuroscience, 27(7):1659 -- 1669, February 2007.
18. Mikiko Kadohisa and Donald A. Wilson. Olfactory cortical adaptation facilitates detection of odors
against background. Journal of Neurophysiology, 95(3):1888 -- 1896, March 2006.
19. Sabine Krofczik, Randolf Menzel, and Martin P. Nawrot. Rapid odor processing in the honey-
bee antennal lobe network. Frontiers in Computational Neuroscience, 2, 2008. PMID: 19221584
PMCID: 2636688.
13
20. A. Kumar, S. Schrader, A. Aertsen, and S. Rotter. The high-conductance state of cortical networks.
Neural Computation, 20(1):143, 2008.
21. Arvind Kumar, Stefan Rotter, and Ad Aertsen. Spiking activity propagation in neuronal networks:
reconciling different perspectives on neural coding. Nat Rev Neurosci, 11(9):615 -- 627, 2010.
22. Giancarlo LaCamera, Alexander Rauch, Hans-R Luscher, Walter Senn, and Stefano Fusi. Min-
imal models of adapted neuronal response to in vivo-like input currents. Neural Comput,
16(10):21012124, October 2004.
23. Alexander Lerchner, Cristina Ursta, John Hertz, Mandana Ahmadi, Pauline Ruffiot, and Sren
Enemark. Response variability in balanced cortical networks. Neural Computation, 18(3):634 -- 659,
2006.
24. Michael London, Arnd Roth, Lisa Beeren, Michael Hausser, and Peter E. Latham. Sensitivity to
perturbations in vivo implies high noise and suggests rate coding in cortex. Nature, 466(7302):123 --
127, July 2010.
25. Brian N. Lundstrom, Matthew H. Higgs, William J. Spain, and Adrienne L. Fairhall. Fractional
differentiation by neocortical pyramidal neurons. Nature Neuroscience, 11(11):1335 -- 1342, 2008.
26. Susana Martinez-Conde, Stephen L. Macknik, Xoana G. Troncoso, and David H. Hubel. Microsac-
cades: a neurophysiological analysis. Trends in Neurosciences, 32(9):463 -- 475, September 2009.
27. R. Menzel and Larry R. Squire. Olfaction in invertebrates: Honeybee. In Encyclopedia of Neuro-
science, pages 43 -- 48. Academic Press, Oxford, 2009.
28. Michael Monteforte and Fred Wolf. Dynamical entropy production in spiking neuron networks in
the balanced state. Physical Review Letters, 105(26):268104, December 2010.
29. Eilif Muller, Lars Buesing, Johannes Schemmel, and Karlheinz Meier. Spike-frequency adapting
neural ensembles: Beyond mean adaptation and renewal theories. Neural Comp., 19(11):2958 -- 3010,
2007.
30. Katherine I Nagel and Rachel I Wilson. Biophysical mechanisms underlying olfactory receptor
neuron dynamics. Nature Neuroscience, 14(2):208 -- 216, January 2011.
31. Martin P. Nawrot, Clemens Boucsein, Victor Rodriguez Molina, Alexa Riehle, Ad Aertsen, and
Stefan Rotter. Measurement of variability dynamics in cortical spike trains. Journal of Neuroscience
Methods, 169(2):374 -- 390, April 2008.
32. William H. Nesse, Leonard Maler, and Andr Longtin. Biophysical information representation in
temporally correlated spike trains. Proceedings of the National Academy of Sciences, 107(51):21973
-- 21978, December 2010.
33. Maria Papadopoulou, Stijn Cassenaer, Thomas Nowotny, and Gilles Laurent. Normalization for
sparse encoding of odors by a wide-field interneuron. Science, 332(6030):721 -- 725, May 2011.
34. Javier Perez-Orive, Ofer Mazor, Glenn C Turner, Stijn Cassenaer, Rachel I Wilson, and Gilles
Laurent. Oscillations and sparsening of odor representations in the mushroom body. Science,
297(5580):359 -- 65, July 2002.
35. Cindy Poo and Jeffry S. Isaacson. Odor representations in olfactory cortex: Sparse coding, global
inhibition, and oscillations. Neuron, 62(6):850 -- 861, June 2009.
14
36. R. Ranganathan. Evolutionary origins of ion channels. Proceedings of the National Academy of
Sciences of the United States of America, 91(9):3484, 1994.
37. Alex Roxin, Nicolas Brunel, David Hansel, Gianluigi Mongillo, and Carl van Vreeswijk. On the dis-
tribution of firing rates in networks of cortical neurons. The Journal of Neuroscience, 31(45):16217
-- 16226, November 2011.
38. B. Rudy. Diversity and ubiquity of k channels. Neuroscience, 25(3):729 -- 749, June 1988.
39. Silke Sachse and C. Giovanni Galizia. Role of inhibition for temporal and spatial odor representation
in olfactory output neurons: A calcium imaging study. J Neurophysiol, 87(2):1106 -- 1117, February
2002.
40. Roman Shusterman, Matthew C Smear, Alexei A Koulakov, and Dmitry Rinberg. Precise olfactory
responses tile the sniff cycle. Nature Neuroscience, 14(8):1039 -- 1044, July 2011.
41. Richard B. Stein, E. Roderich Gossen, and Kelvin E. Jones. Neuronal variability: noise or part of
the signal? Nat Rev Neurosci, 6(5):389 -- 397, May 2005.
42. Martin Fritz Strube-Bloss, Martin Paul Nawrot, and Randolf Menzel. Mushroom body output
neurons encode odor-reward associations. J. Neurosci., 31(8):3129 -- 3140, February 2011.
43. Paul Szyszka, Mathias Ditzen, Alexander Galkin, C. Giovanni Galizia, and Randolf Menzel. Spars-
ening and temporal sharpening of olfactory representations in the honeybee mushroom bodies.
Journal of Neurophysiology, 94(5):3303 -- 3313, November 2005.
44. David J. Tolhurst, Darragh Smyth, and Ian D. Thompson. The sparseness of neuronal responses
in ferret primary visual cortex. The Journal of Neuroscience, 29(8):2355 -- 2370, February 2009.
45. A. Treves. Mean-field analysis of neuronal spike dynamics. Network: Computation in Neural
Systems, 4(3):259284, 1993.
46. C. van Vreeswijk. Analysis of the asynchronous state in networks of strongly coupled oscillators.
Phys Rev Lett, 84(22):51105113, May 2000.
47. C. van Vreeswijk and H. Sompolinsky. Chaotic balanced state in a model of cortical circuits. Neural
Comput, 10(6):1321 -- 71, August 1998.
48. Tim P Vogels and L F Abbott. Gating multiple signals through detailed balance of excitation and
inhibition in spiking networks. Nat Neurosci, 12(4):483 -- 491, April 2009.
49. Barry Wark, Brian Nils Lundstrom, and Adrienne Fairhall. Sensory adaptation. Current opinion
in neurobiology, 17(4):423 -- 429, August 2007. PMID: 17714934 PMCID: PMC2084204.
50. Jason Wolfe, Arthur R Houweling, and Michael Brecht. Sparse and powerful cortical spikes. Current
Opinion in Neurobiology, 20(3):306 -- 312, June 2010.
15
Figure 2. Reliability of a temporally sparse code in the balanced cortical network. (A)
Schematic of a two-layer model of sub-cortical and early cortical sensory processing. The afferent
ensemble (blue) consists of 4,000 independent neurons, and each neuron projects to 1% of the neurons
in the cortical network (green). The cortical network is a balanced network in the asynchronous and
irregular state with random connectivity. In both populations black circles represent excitatory neurons
and magenta circles represent inhibitory cells. (B) The distribution of firing rates across neurons in the
cortical network is fat-tailed and the average firing rate is approximately 3Hz. (C) The distribution of
the coefficient of variation (Cv) across neurons in the balanced cortical network confirms irregular
spiking. (D) Spike raster plot for a sample set of 30 afferent neurons (blue dots) and 30 excitatory
cortical neurons (green dots). At t = 0 (gray triangle) the stimulus presentation starts. (E) Population
averaged firing for both network stages. The simulation (solid lines) follows the calculated ensemble
average predicted by the adaptive density treatment (black circles). The firing rate in the simulated
network is estimated with a 20ms bin size. (F) Number of spikes per neuron Se after stimulus onset
(t = 0) for the adaptive network (solid lines) and the weakly adaptive control network (dashed lines).
Cortical excitatory neurons (green) produced less spikes than neurons in the earlier stage (afferent
ensemble, blue). The shaded area indicates the standard deviation across neurons. (G,H) Fano factor
dynamics of the afferent ensemble in the network with strongly adapting neurons (G, τs = 110 ms) and
in the weakly-adaptive (h, τs = 30 ms) network, estimated across 200 trials in a 50ms window and a
sliding of 10ms for the ensemble network with adaptation. The black circles indicate the theoretical
value of the Fano factor computed by adaptive density treatment and shaded area is the standard
deviation of the Fano factor across neurons in the network. (I) The Fano factor of strongly adaptive
neurons in cortical balanced network reduced transiently during the initial phasic response part. The
crosses show the adaptive cortical ensemble Fano factor for the case where the afferent ensemble neurons
were modeled as a Poisson process with the same steady-state firing rate and without adaption. (J)
The Fano factor in the weakly adaptive cortical network did not exhibit a reduction during stimulation.
Poisson ProcessesAfferent ensembleCortical networkABCDEFGHIJ 16
Figure 3. Neuronal adaptation generates temporal sparseness in a generic model of the
insect olfactory network. (A) Schematic drawing of a simplified model of the insect olfactory
network for a single pathway of odor coding. Olfactory receptor neurons (OSNs, first layer, n=1,480)
project to the antennal lobe network (second layer) consisting of projection neurons (PNs, n=24) and
local neurons (magenta, n=96), which make inhibitory connections with PNs. PNs project to the
Kenyon cells (KCs) in the mushroom body (third layer). (B) Spike raster plot of randomly selected
OSNs (blue), LNs (magenta), PNs (green) and KCs (red) indicates that spiking activity in the network
became progressively sparser as the Poisson input propagated into the network. (C) Average
population rate of OSNs in the adaptive network (blue solid line) and the non-adaptive control network
(dashed blue lines). The shaded area indicates the firing rate distribution of the neurons. The firing
rate was estimated with 20ms bin size. (D) Average response in the antennal lobe network. PNs
(green) and LNs (magenta) exhibited the typical phasic-tonic response profile in the adaptive network
(solid lines) but not in the non-adaptive case (dashed lines). (E) Kenyon cell activity. In the adaptive
network the KC population exhibits a brief response immideately after stimulus onset, which quickly
returns close to baseline. This is contrasted by a tonic response profile throughout the stimulus in the
non-adaptive case. (F) Effect of the inhibitory micro-circuit. By turning off the inhibitory LN-PN
connections the population response amplitude of the KCs was increased, while the population response
dynamics did not change.(G) Sparseness of KCs. The average number of spikes per neuron emitted
since stimulus onset indicates that the adaptive ensemble encodes stimulus informatoin with only very
few spikes. (G) Reliability of KCs responses. The Fano factor of the KCs in different network scenarios
is estimated across 200 trials in a 100ms time window after stimulus onset. Network 1: (+)Adaptation
(+)Inhibition, network 2: (+)Adaptation (-)Inhibition, network 3: (-)Adaptation (-)Inhibition, network
4: (-)Adaptation (+)Inhibition. Both networks with SFA are significantly more reliable in their
stimulus encoding than the non-adaptive networks.
Poisson ProcessesReceptor Neurons Antennal lobeKenyon CellsABCDEFGH 17
Figure 4. Blocking GABAergic transmission in the honeybee changes amplitude but not
duration of the KC population response. (A) Temporal response profile of the Ca signal imaged
in the mushroom body lip region of one honeybee for different stimulus durations as indicated by color.
(B) Temporal response profiles as in (A) in one honeybee after application of PTX. (C) Response
profiles imaged from 6 control animals (gray) and their average (black) for a 3s stimulus as indicated by
the stimulus bar. The responses measured in 6 animals in which GABAA transmission was blocked with
PTX (red) shows a considerably higher population response amplitude. The shaded area indicated the
standard deviation of responses across bees. (D) Average amplitude-normalized responses are highly
similar in animals treated with PTX and control animals. (E) Blocking GABAB transmission with
CGP in 6 animals (blue) again results in an increased response amplitude compared to 6 control
animals (black). The shaded area indicates the standard deviation across individuals. (F) Average
normalized response profiles are highly similar in the CGP-treated and control animals.
ABCDEF |
1704.08669 | 1 | 1704 | 2017-04-27T17:33:32 | Evolution of moments and correlations in non-renewal escape-time processes | [
"q-bio.NC",
"math.PR",
"physics.comp-ph",
"physics.data-an"
] | The theoretical description of non-renewal stochastic systems is a challenge. Analytical results are often not available or can only be obtained under strong conditions, limiting their applicability. Also, numerical results have mostly been obtained by ad-hoc Monte--Carlo simulations, which are usually computationally expensive when a high degree of accuracy is needed. To gain quantitative insight into these systems under general conditions, we here introduce a numerical iterated first-passage time approach based on solving the time-dependent Fokker-Planck equation (FPE) to describe the statistics of non-renewal stochastic systems. We illustrate the approach using spike-triggered neuronal adaptation in the leaky and perfect integrate-and-fire model, respectively. The transition to stationarity of first-passage time moments and their sequential correlations occur on a non-trivial timescale that depends on all system parameters. Surprisingly this is so for both single exponential and scale-free power-law adaptation. The method works beyond the small noise and timescale separation approximations. It shows excellent agreement with direct Monte Carlo simulations, which allows for the computation of transient and stationary distributions. We compare different methods to compute the evolution of the moments and serial correlation coefficients (SCC), and discuss the challenge of reliably computing the SCC which we find to be very sensitive to numerical inaccuracies for both the leaky and perfect integrate-and-fire models. In conclusion, our methods provide a general picture of non-renewal dynamics in a wide range of stochastic systems exhibiting short and long-range correlations. | q-bio.NC | q-bio | Evolution of moments and correlations in non-renewal escape-time processes
Wilhelm Braun,1, ∗ Rüdiger Thul,2 and André Longtin1
1Department of Physics and Centre for Neural Dynamics,
University of Ottawa, 598 King Edward, Ottawa K1N 6N5, Canada
2Centre for Mathematical Medicine and Biology, School of Mathematical Sciences,
University of Nottingham, Nottingham, NG7 2RD, UK
(Dated: September 28, 2018)
7
1
0
2
r
p
A
7
2
]
.
C
N
o
i
b
-
q
[
1
v
9
6
6
8
0
.
4
0
7
1
:
v
i
X
r
a
The theoretical description of non-renewal stochastic systems is a challenge. Analytical results are
often not available or can only be obtained under strong conditions, limiting their applicability. Also,
numerical results have mostly been obtained by ad-hoc Monte -- Carlo simulations, which are usually
computationally expensive when a high degree of accuracy is needed. To gain quantitative insight
into these systems under general conditions, we here introduce a numerical iterated first-passage
time approach based on solving the time-dependent Fokker -- Planck equation (FPE) to describe
the statistics of non-renewal stochastic systems. We illustrate the approach using spike-triggered
neuronal adaptation in the leaky and perfect integrate-and-fire model, respectively. The transition
to stationarity of first-passage time moments and their sequential correlations occur on a non-trivial
timescale that depends on all system parameters. Surprisingly this is so for both single exponential
and scale-free power-law adaptation. The method works beyond the small noise and timescale
separation approximations.
It shows excellent agreement with direct Monte Carlo simulations,
which allows for the computation of transient and stationary distributions. We compare different
methods to compute the evolution of the moments and serial correlation coefficients (SCC), and
discuss the challenge of reliably computing the SCC which we find to be very sensitive to numerical
inaccuracies for both the leaky and perfect integrate-and-fire models. In conclusion, our methods
provide a general picture of non-renewal dynamics in a wide range of stochastic systems exhibiting
short and long-range correlations.
I.
INTRODUCTION
A general property of diverse systems, ranging from su-
perconducting quantum interference devices (SQUIDs)
[1], to lasers [2] to excitable cells [3 -- 6] is that time in-
tervals between specific events are not statistically inde-
pendent. The theoretical description of such non-renewal
stochastic processes [7] poses a significant challenge, as it
implies that the present state of the system depends, in
general, on the whole past evolution or parts of it, and not
just on the previous state. Analytical approximations to
tackle such memory effects have included the assumption
of stationarity [8], small stochasticity [9] and time-scale
separation [10 -- 12] between stochastic and deterministic
parts of the dynamics.
Even if these approximations allow for some insight into
the parameter dependence of e.g. serial correlations and
can be used to understand experimental data, as exem-
plified in [9, 13 -- 16] in the context of excitable systems, it
is desirable to understand the statistics of model systems
without making simplifying assumptions. Regarding sta-
tionarity, real systems rarely operate in a stationary state
due to transients that arise from deterministic or ran-
dom perturbations. A prominent example are cortical
neurons. An average cortical neuron receives random in-
puts from approximately 104 other neurons, whose ac-
tivity is modulated by non-stationary sensory and other
inputs, resulting in transient neuronal dynamics [17] that
∗ [email protected]
only become stationary after a certain time. It is there-
fore important to understand how statistical properties of
inter-event times evolve and become invariant following
a transient regime due to internal dynamics and external
inputs. Keeping with illustrations from neural dynam-
ics, it is well known that physiologically relevant pro-
cesses underlying neural coding rarely only have one well-
defined timescale [18, 19]. This has lead researchers in
various theoretical fields to consider multiple time-scale
dynamics [20]. An important example of a system with
multiple time scales is neuronal adaptation, where a neu-
ron's firing rate adjusts in response to a stimulus. Adap-
tation with multiple timescales, or even no time scale as
in the case of power-law adaptation [21 -- 23], is now known
to be biophysically relevant, and even optimal for some
tasks [24]. Recently, it was also shown that a neuron
model with adaptive firing thresholds exhibiting multi-
ple timescales is the optimal choice for the prediction
of spike times in cortical neurons [25, 26]. Therefore,
a theoretical description of adaptation without a single
well-defined timescale is an important goal.
In this paper, we show how to describe two-dimensional
non-renewal dynamics by an iterated first-passage time
(iFPT) approach. This approach allows us to determine
stationary statistical properties of the system as well as
providing a description of the transition to stationarity.
We furthermore show how to compute serial correlations
in the time series generated by the firing times of the
system. While our approach is general and applicable
to any system where first-passage times [27] play a role,
we illustrate its versatility with two important examples,
namely spike-triggered neuronal adaptation with a sin-
gle exponential current and a power-law current without
an intrinsic timescale, respectively. Using the underly-
ing time-dependent FPE to describe the system, we only
need to apply mathematically convenient standard ab-
sorbing boundary conditions to obtain stationary distri-
butions, e.g. that of the adaptation current upon firing.
Moreover, the methods developed here can easily be ex-
tended to models with correlation-generating determin-
istic input currents as recently considered in [28].
II. MODEL
We consider a stochastic differential equation (SDE)
driven by an external signal s(t):
dX(t) = µ(X(t))dt + φ(X(t))dW (t) − s(t)dt .
(1)
X is defined on the domain (−∞, xth]. If X reaches xth,
the system is said to have generated and event, and X is
instantaneously reset to 0. For all examples in this study,
we chose the Ornstein -- Uhlenbeck process (OUP) given
its prominence in the field of stochastic systems. For the
OUP, we fix the correlation time τm = 1
γ , bias current I0
and noise intensity σ as follows: µ(X(t)) = γ(I0 − X(t)),
φ(X(t)) = σγ. W (t) is a standard Brownian motion and
we set xth = 1. Given that the OUP is the basis for
integrate-and-fire (IF) neuron models, which are among
the most popular neuron descriptions [29], we refer to
events as spikes and to s(t) as a time-dependent adapta-
tion current in the present study. The general dynamics
of s(t) obeys a single autonomous ordinary differential
equation (ODE)
s = ω(s) ,
(2)
and s is increased by a fixed amount κ when X = xth:
s → s + κ, which is the mechanism for spike-triggered
adaptation [30]. When s(t) is also reset to its starting
value s(0), the model is a renewal model and its firing
statistics may be studied using standard techniques, see
e.g. [31] for a recent review.
Here we focus on two forms of the adaptation current.
The first one is power-law adaptation, for which
ω(s) = −
1
α
s2(t) .
(cid:16) t
(3)
(cid:17)−1
.
This ODE has the general solution s(t) =
Therefore, the current s in this case has a power-law time
dependence with no intrinsic time scale [21].
The second adaptation current is given by a single expo-
nential decay with time scale τa:
α + 1
s(0)
which has the general solution s(t) = s(0)e
The time to the first spike event is the following first-
passage time (FPT):
τa .
− t
2
T1 = inf(t > 0 : X(t) > xthX(0) = 0, s(t = 0) = s(0)) .
FIG. 1. Sample paths of the model for a power-law adaptation
current given by Eq. 3. Top: X(t) (Eq. 1), the horizontal
dashed lines are at X = 0 and xth = 1. Bottom: s(t) (Eq. 2)
When X reaches xth, s undergoes a jump of size κ. The
subsequent ISIs Tk have distributions Fk(t), and the starting
have distributions Gk(s) for k ≥ 1. Parameter
values s(k)
values are α = 3.0, γ = 1.0, σ = 0.8, I0 = 4.0, κ = 3.0.
0
In the non-renewal case we are studying here, subsequent
firing times will in general not have the same distribution
as T1. We define the kth interspike interval (ISI) as
Tk = inf
t −
Ti
: X(t) ≥ xth, t >
Ti
.
(5)
The first moment of the kth ISI is given by τ 1
k = E(Tk).
The second moment of the kth ISI will be denoted by
k = E((Tk)2) and the kth firing rate is given by the
τ 2
inverse of the corresponding mean ISI rk = 1
. The kth
τ 1
k
standard deviation m2(k) is then given by
m2(k) =
k − (τ 1
τ 2
k )2 .
(6)
The values of the peak adaptation current after the kth
firing are defined for k ≥ 1 as
(cid:32)
k−1(cid:88)
i=1
(cid:33)
k−1(cid:88)
i=1
(cid:113)
(cid:32)
(cid:33)
k(cid:88)
i=1
ω(s) = −
1
τa
s(t) ,
(4)
s(k)
0 =
s(t−) + κ : t =
Ti
,
(7)
0.00.51.01.52.02.53.0−1.0−0.50.00.51.01.5X0.00.51.01.52.02.53.0t01234567ss(0)0s(1)0∼G1(s)T1∼F1(t)s(2)0∼G2(s)T2∼F2(t)s(3)0∼G3(s)T3∼F3(t) where t− indicates that we take the left-sided limit.
For simplicity, we choose s to be started from a point
(s(0)
0 = κ), instead of from a biophysically more realistic
initial distribution. However, the methods we are going
to describe in the following are also valid when s is ini-
tially started from a distribution.
The central challenge is to obtain the distributions Gk
and Fk for k ≥ 1, which are the distributions of s(k)
and
Tk defined by Eqs. 7 and 5, respectively. An example
realization for the case of power-law adaptation is shown
in Fig. 1. The knowledge of these distributions is key to
understanding the non-renewal dynamics, as they form a
hidden Markov model of the underlying non-Markovian
dynamics [15, 16, 32]. Therefore, once these distributions
are known, the non-renewal dynamical system breaks up
into coupled renewal dynamical systems, which are much
more tractable mathematically. This gives rise to the
iFPT approach which we now explain.
0
III. THE IFPT APPROACH
Being a diffusion process, the system given by Eqs. 1
and 2 is governed by a two-dimensional time-dependent
FPE [33]. The FPE for the probability density func-
tion p(t; x, s)dxds = P(X(t) ∈ (x, x + dx), s(t) ∈ (s, s +
ds)X(0) = x0, s(0) = s0) reads (we use x(cid:62) = (x, s) for
brevity)
∂tp(t; x) = ∇ · (A(x)∇p(t; x)) − ∇ · (F(x)p( t; x)) , (8)
where A is the diffusion matrix and F the drift vector,
which can be obtained in a straightforward way from the
SDE for X, Eq. 1, and the ODE for s, Eq. 2. Explicitly,
we have
(cid:18) φ(x)2
2
0
(cid:19)
0
0
,
(9)
A(x) =
and
F(x) = (µ(x) − s, ω(s))
(cid:62)
.
time T1 , CDF1(t) = (cid:82) t
The IF property of X entails that we have an absorbing
boundary at X = xth for all times t: p(t; x = xth, s) = 0.
With this boundary condition, we can compute the cu-
mulative distribution function (CDF) of the first-passage
0 F1(λ)dλ, by time evolution of
the FPE on a computational domain Ω ⊂ R2, which we
choose to be a rectangle extending to sufficiently negative
values in the x-direction [34]:
(cid:90)
CDF1(t) = 1 −
Ω
p1(t; z)dz ,
(10)
3
0
0
, the value of s(k)
where p1 is the solution to Eq. 8 with the initial condition
p1(0; x, s) = δ(x − x0), where x0 = (0, κ)(cid:62). For the
computation of F1, we then only need to differentiate
Eq. 10.
To describe adaptation, one needs to compute the statis-
tics of the peak adaptation currents, as defined by Eq. 7.
Hence, we need to characterize the hidden Markov model
generated by the ISIs Tk and the peak adaptation cur-
rents s(k)
0 . Whereas the dynamics of s and X as a whole
is non-Markovian, the distribution of the ISI Tk is com-
pletely determined by the distribution Gk−1, and the se-
quence {Tk, s(k−1)
∞
k=1 is therefore Markovian. Knowing
}
the values of Tk and s(k−1)
is fixed,
and the distribution of the ISI Tk+1 can be obtained by
solving an FPT problem with s(k)
as initial condition for
s.
The central observation now is that for the second ISI,
X again starts from 0, whereas s starts from a distribu-
tion G1. This is because X evolves stochastically, and
therefore reaches the threshold xth at different times T1,
corresponding to different values of s(t = T1) + κ imme-
diately after the firing event. To compute the second ISI,
we therefore need to know G1. This can be iterated: to
compute the distribution Fk of the kth ISI, we need the
distribution Gk−1. This is the central idea of the iFPT
approach. To set up the iFPT approach, we first observe
that between threshold crossings of X, s evolves deter-
ministically. Therefore, when we know the PDF of the
first FPT, by conservation of probability, we also know
the distribution of s after the first firing event:
0
0
(cid:12)(cid:12)(cid:12)(cid:12) dt(s)
ds
(cid:12)(cid:12)(cid:12)(cid:12)F1(t(s)) ,
G1 (s − κ) =
(11)
where t(s) is the inverse function of s. The support of G1
is shifted, because of the jump of size κ that s undergoes
when X reaches its threshold xth. For the second ISI
T2, s is started from the distribution G1(s) instead of a
point, whereas X is started from a point again (Fig. 1).
This means that to obtain F2, the FPE is started from
a distribution: p2(0; x) ∝ G1(s)δ(x). This generalizes to
values of k larger than 1. For the kth ISI distribution
Fk(t), we therefore must choose
pk(0; x) ∝ Gk−1(s)δ(x) .
(12)
We show how to obtain the distributions Gk for k > 1
in Section V. Linear splines are used to create a mesh
be normalized appropriately, so that (cid:82)
function approximating Eq. 12 on the computational do-
main Ω. Due to this approximation, Eq. 12 then has to
Ω pk(0; z)dz = 1.
(cid:82)
The FPE is then solved again, and the distributions
Fk(t) are obtained analogously to Eq. 10: CDFk(t) =
Ω pk(t; z)dz, i.e. by timestepping the FPE to obtain
1−
the CDF of the kth ISI followed by a numerical differen-
tiation. This constitutes the iFPT approach.
To quantify the accuracy of our numerical methods, we
also compute the relative disagreement ∆ between results
obtained by the iFPT approach and direct MC simula-
tions. It is defined for a quantity Z by
∆(Z) = ZiFPT − ZMC
ZiFPT
.
(13)
We performed MC simulations for two different simula-
tion setups, the first one without any boundary correc-
tion (plain MC), and the second one with a boundary
correction according to Giraudo and Sacerdote (MC-GS)
[31, 35, 36]. This boundary correction is applied to reduce
the systematic overestimation of FPTs when using the
Euler -- Maruyama scheme. We compute the relative dis-
agreement given by Eq. 13 using either MC simulations
with or without boundary correction; we observed that
the order of the relative disagreement is unchanged, but
in general, the plain MC algorithm gives rise to larger dis-
agreements than MC-GS. A decrease in the relative dis-
agreement is expected, because the GS correction method
should yield an improved weak error of O(h) [37], in con-
trast to the plain MC simulation, which has a weak error
of O(h 1
2 ) [38], where h is the time step for the discretiza-
tion of the SDE, Eq. 1. In the following, the timestep
for MC simulations is chosen to be h = 10−3 and we
choose M = 106 independent realizations. The plain
Euler-Maruyama scheme then gives rise to a weak error
2 ) ≈ 3 · 10−2 when estimating moments of first
of O(h 1
passage times, which is one order of magnitude larger
= 10−3. There-
than the MC error proportional to
fore, in our simulations, the plain MC error is negligible
in comparison to the error introduced by the finite-time
discretization of the SDE (Eq. 1).
For the numerical solution of the FPE (Eq. 8), we choose
a finite element discretization method [39] and evolve the
system using either a stabilized Crank -- Nicolson (CN)
scheme [40] in Fig. 2 or an Euler timestepping scheme
[39] in Fig. 3.
The relative disagreement between MC simulations and
finite-element solutions stays largely constant across dif-
ferent lags k when we use the CN scheme instead of the
Euler scheme as can be seen by comparing the lower pan-
els of Figs. 2 and 3; the sizes of the disagreement are com-
parable in magnitude. This suggests that the remaining
small discrepancy between MC simulations and PDE re-
sults can be largely explained with the errors associated
with the MC simulation method. In particular, note that
for the examples we show, the MC-GS weak error of size
O(h) is comparable in magnitude to the numerator of
∆(Z) (Eq. 13), i.e. the absolute disagreement. We will
see what effects this has on the computation of correla-
tions in Section VI.
1√
M
4
FIG. 2. Top: Evolution of the rate rk = 1
(left) and stan-
τ 1
k
dard deviation given by Eq. 6 (right) of ISIs as a function of
k. Empty circles: Plain MC simulations of Eqs. 1 and 2. Tri-
angles: MC simulations with GS boundary correction. Filled
circles: moments obtained from numerical solution of FPE
using a CN timestepping scheme. M = 106 independent MC
realizations for each value of k. Power-law adaptation (Eq. 3)
with α = 5.5, I0 = 6.0, σ = 1.3, γ = 1.0, κ = 5.5. Bottom:
Relative disagreements defined by Eq. 13, where an MC-GS
algorithm was used.
IV. TRANSITION TO STATIONARITY
We show the evolution of the rate and standard deviation
of ISIs in Figs. 2 and 3. The rates decreases, whereas the
standard deviation increases until both quantities reach a
stationary value. Given that these quantities are derived
from moments of the ISI distributions, the distributions
also converge towards a stationary form. The conver-
gence towards stationarity of ISI and peak adaptation
current distributions (Fk and Gk, respectively) is shown
in Figs. 4 and 5. As expected for adapting models, the
ISIs (whose distributions are shown in Fig. 4) increase
with higher k, which means that the rate rk decreases.
This is well captured by the iFPT approach, with a max-
imal relative disagreement smaller than 3% in Fig. 2 and
smaller than 2% in Fig. 3. Also, the width of both the ISI
distributions and the peak adaptation current distribu-
tions (Fig. 5) increases, which is reflected by the increase
of the variances of Fk and Gk shown in Fig. 2 and 3.
The mean of the peak adaptation currents shifts to the
right as stationarity is reached. Moreover, stationarity
is reached with varying speed, i.e.
for different values
of the lag k (compare Fig. 2 with Fig. 3). Generally,
123450.91.01.11.21.31.41.51.61.71.8rk123450.320.340.360.380.400.420.44m2(k)12345k0.0040.0060.0080.0100.0120.0140.0160.0180.0200.022∆(rk)12345k0.0040.0060.0080.0100.0120.0140.016∆(m2(k)) 5
FIG. 4. PDF Fk of the kth ISI Tk (Eq. 5). The symbols
are MC simulations (M = 106 MC realizations) as indicated
in the legends. Solid black lines: PDF obtained by numerical
solution of the FPE, Eq. 8. In the left panel, the distributions
are practically indistinguishable after k = 2. Left: Power-law
adaptation (Eq. 3). Right: Single exponential adaptation
(Eq. 4). Parameter values as in Fig. 2 for power-law adapta-
tion and as in Fig. 3 for exponential adaptation. Both panels
show results for plain MC simulations.
value that is larger than those typically found around the
mode of the stationary distribution) can build up. The
system reaches stationarity rapidly because it is quasi-
deterministic as the dynamics of s dominates the sys-
tem, and the stochastic fluctuations of X will only cause
small perturbations. For both power-law and single ex-
ponential adaptation currents, it is possible to reach the
stationary regime already after one or two firing events as
in Fig. 2, or to have a long transient regime as in Fig. 3.
The initial condition s(0)
for the adaptation current can
0
also be chosen to control the speed of the transition. If it
is placed far away from the mean of the stationary distri-
bution, the transition will take a longer time; also, it is
possible to obtain a non-monotonic behaviour of the rate
as a function of the interval number when s(0)
is placed
0
far above the aforementioned mean. The first mean ISI
will then be the longest statistically, in contrast to the
examples we show in Figs. 2 and 3.
V. STATISTICS OF THE kTH ISI
The iFPT approach can be iterated beyond the first two
firing events to obtain the distribution for the third ISI,
F3(t). However, for the computation of the third ISI,
no equation similar to Eq. 11 can be used to obtain G2
because s was started from a distribution to obtain F2.
Indeed, for one fixed time T2, there are many different
starting values s(1)
0 due to the stochastic dynamics of X.
Importantly, the ISI T2 and the initial condition s(1)
are
0
not independent random variables (for a large value of
s(1)
0 , a large ISI T2 is more probable and vice versa) so
that we can obtain the value that s reaches after the
second firing by the following observation (focusing on
power-law adaptation): given that T2 = λ and s(1)
0 = ν,
FIG. 3. Top: Evolution of the rate rk = 1
(left) and stan-
τ 1
k
dard deviation given by Eq. 6 (right) of ISIs as a function of
k. Empty circles: Plain MC simulations of Eqs. 1 and 2. Tri-
angles: MC simulations with GS boundary correction. Filled
circles: moments obtained from numerical solution of FPE us-
ing an Euler timestepping scheme. M = 106 independent MC
realizations for each value of k. Single exponential adaptation
(Eq. 4) with τa = 1.0, I0 = 5.0, σ = 1.0, γ = 1.0, κ = 1.0.
Bottom: Relative disagreements defined by Eq. 13, where
an MC-GS algorithm was used.
the speed of adaptation can be controlled by adjusting
the bias current I0 and the noise level σ as well as the
adaptation strength (size of the kick κ and, in the case
of single exponential adaptation, the timescale τa). We
have carried out additional MC simulations (not shown)
to obtain insight into how these parameters influence the
speed of the transition to stationarity. A higher bias cur-
rent and a higher noise level will in general lead to a
less rapid transition to stationarity. This can be under-
stood as follows: X is driven to threshold more rapidly,
causing the inhibition to build up quickly, reaching val-
ues that are higher than those typically found around
the peak of the stationary distribution. This slows down
the transition to stationarity, because s needs to decay
first. Moreover, a large kick size κ and a rapidly de-
caying adaptation current will cause a quick transition.
For the latter case, this is easily understood as we are
then nearly dealing with a renewal system: after a short
initial period, the effect of the adaptation current on the
firing time statistics is negligible. For the former case, we
note that the larger the kick size κ, the more pronounced
the inhibitory effect of adaptation within one ISI, which
means that X takes longer to reach threshold before a
large out-of-equilibrium average value of the current s (a
123456782.22.42.62.83.03.23.43.63.8rk123456780.120.140.160.180.200.220.24m2(k)12345678k0.0000.0020.0040.0060.0080.0100.0120.014∆(rk)12345678k0.0020.0040.0060.0080.0100.0120.0140.0160.0180.020∆(m2(k))0.00.51.01.52.02.53.0t10−210−1100Fk(t)k=1k=2k=30.00.20.40.60.8t10−210−1100k=1k=2k=3 6
0
FIG. 5. PDF Gk of s(k)
(Eq. 7). The symbols are MC simula-
tions (M = 106 MC realizations) as indicated in the legends.
The black solid line is G1 obtained from Eq. 11 or G2 (G3)
obtained using Eq. 18. In the left panel, the distributions are
practically indistinguishable after k = 2. Left: Power-law
adaptation (Eq. 3). Right: Single exponential adaptation
(Eq. 4). Parameter values as in Fig. 2 for power-law adapta-
tion and as in Fig. 3 for exponential adaptation. Both panels
show results for plain MC simulations.
FIG. 6. Conditional FPT densities H(λ, ν) given by Eq. 17
relating the initial value of the adaptation current to the dis-
tribution of the following ISI. Computed using numerical so-
lutions to the FPE Eq. 8. Left: Power-law adaptation, com-
puted using a CN timestepping scheme. Right: Exponential
adaptation, computed using an Euler timestepping scheme.
Parameter values as in Fig. 2 for power-law adaptation and
as in Fig. 3 for exponential adaptation.
ν
λ
0 = 1
α + 1
we have s(2)
+ κ. We have included the jump of
size κ due to the definition of s(2)
(see Fig. 1). This em-
0
phasizes that once two values in the triplet (T2, s(1)
0 , s(2)
0 )
are fixed, the third one is determined. In the following,
we again use λ to denote a fixed FPT and ν to denote
a fixed initial value of the adaptation current s. Anal-
ogously, we generally have for s(k)
0 : given Tk = λ and
s(k−1)
0 = f (λ, ν) is determined. The function f
0
determining the subsequent value of the peak adaptation
current given the previous ISI λ and the previous peak
value of the adaptation current ν reads for power-law
adaptation (Eq. 3):
= ν, s(k)
f (λ, ν) = κ +
1
λ
α + 1
ν
,
(14)
whereas for exponential adaptation (Eq. 4), we have
f (λ, ν) = κ + ν exp
−
.
(15)
An alternative way is to fix the value of s(k)
and then
put a constraint on the time Tk when s(k−1)
is fixed:
given s(k−1)
0 = θ, Tk = h(ν, θ) is deter-
mined, where, for power-law adaptation, we have the ISI
as a function of the previous and subsequent adaptation
values:
= ν and s(k)
0
0
0
(cid:18)
(cid:19)
λ
τa
(cid:18) 1
θ − κ −
(cid:19)
1
ν
h(ν, θ) = α
.
(16)
The function h is defined by solving the equation
f (λ, ν) = θ for λ. To actually compute the density Gk,
(cid:16)
(cid:17)
we need to relate the above observations to densities that
we can compute with the iFPT approach. To that end,
we now define the conditional density H:
H(λ, ν)dλ ≡ P
T1 ∈ (λ, λ + dλ)s(0)
0 = ν
.
(17)
We have used T1 and s(0)
in the definition Eq. 17 to
0
stress that, for the purpose of the computation of H, we
only need to solve the FPT problem for T1 using differ-
ent values of the initial condition s(0)
It will become
0 .
apparent below that we only need to compute H once,
because it does not depend on the firing index k. For
a fixed value of ν, H(λ, ν) is an FPT probability den-
sity. Our notation emphasizes that H is a function of
two variables. ν sets the level of initial inhibition, i.e.
the starting value of s. We show the function H for both
power-law and exponential adaptation in Fig. 6. We see
that with increasing starting value ν, the mode of the
FPT distribution shifts to larger times. For power-law
adaptation, the shape of the FPT distributions does not
change much, whereas for exponential adaptation, the
distributions become broader with increasing ν.
With this at hand, we now show how to practically com-
pute the distributions Gk for k > 1. We can obtain the
CDF of s(k)
0 by observing that:
(cid:17)
(cid:90)
D(k−1)(θ) H(λ, ν)Gk−1(ν)dλdν ,
s(k)
0 ≤ θ
=
(18)
(cid:16)
P
with
D(k−1)(θ) =
(λ, ν > 0 ν ∈ supp (Gk−1) , h(ν, θ) ≤ λ ≤ Tmax) .
The function h defined in Eq. 16 ensures that for a fixed
value of ν, we collect all times λ so that s(k)
0 ≤ θ, which
(19)
7891011s10−310−210−1Gk(s)k=1k=2k=31.52.02.53.03.5s10−210−1100k=1k=2k=3 ensures that f (λ, ν) ≤ θ for fixed values of θ and ν. Tmax
is chosen so that H(Tmax, ν) ≈ 0 ∀ν ∈ supp (Gk−1). This
means that Tmax should be chosen in the tail of the FPT
distribution. Note that for the iFPT approach, one only
has to compute H(λ, ν) over the support of Gk for k ≥ 1
once [41], and then multiply it with the adaptation cur-
rent distribution of the previous iteration Gk−1. This
function then needs to be integrated according to Eq. 18,
and the PDF Gk can be obtained by numerical differ-
entiation. We show results for G2 and G3 using Eq. 18
in Fig. 5. The agreement between MC simulations and
Eq. 18 is excellent.
Iterating these ideas into the stationary regime, the ideas
of the previous paragraph can be used to directly com-
pute the stationary density of the peak adaptation cur-
rent. We assume that stationarity is reached after k∗
− 1
firing events. Then s(k∗)
have the same dis-
tribution. Consequently, the stationary density of the
peak adaptation current after firing satisfies the two-
dimensional integral equation (cf. Eq. 18)
and s(k∗+1)
0
0
7
notion of stationarity, so that the SCC depends on both
the position n of the ISI in the spike train as well as on
the lag k between ISIs. Since we have already computed
the distributions of the kth ISI, we can readily compute
the variances and means in Eq. 21, i.e. the terms given by
Eqs. 22 and 23. It is slightly more complicated to com-
pute the first term in the numerator, E (TnTn+k), because
we need the joint density p2(Tn, Tn+k) of Tn and Tn+k.
In the present study, we focus on k = 1. By definition,
we have
E(TnTn+1) =
dTndTn+1 Tn Tn+1 p2(Tn, Tn+1),
=
dTn dTn+1 Tn Tn+1p1(Tn+1Tn)Fn(Tn),
where as previously Fn(Tn) is the density of the nth ISI
and p1(Tn+1Tn) is the conditional density of Tn+1 given
Tn. Because Tn+1 is statistically determined only by s(n)
0 ,
we can define this conditional density as
Q(θ) =
D(θ) H(λ, ν)Q
(cid:48)(ν)dλdν ,
(20)
p1(Tn+1Tn) =
dy p(Tn+1, yTn),
(cid:90)
(cid:90)
(cid:16)
(cid:17)
s(k∗)
0 ≤ θ
denotes the CDF of the sta-
where Q(θ) ≡ P
tionary peak value for the adaptation current s (Eq. 2).
We have checked that this equation is indeed satisfied
by the stationary distributions for the peak adaptation
current obtained from MC simulations (data not shown).
Therefore, Eq. 20 can serve as a tool to check whether
a given distribution for the peak adaptation current is
stationary, or alternatively as a way to compute Q(θ)
directly if the function H is known.
VI. CORRELATIONS BETWEEN INTERSPIKE
INTERVALS
We now show how to compute serial correlations with the
iFPT approach. We define the SCC [15, 42] between the
nth ISI Tn and the (n + k)th ISI Tn+k according to
where p(Tn+1, yTn) denotes the joint density of Tn+1 and
s(n)
0 = y conditioned on the previous ISI Tn. We can
rewrite this as follows:
p(Tn+1, yTn) =
p3(Tn+1, y, Tn)
p(Tn)
,
=
=
p(Tn+1y, Tn)p(y, Tn)
p(Tn)
,
p(Tn+1y, Tn)p(yTn)p(Tn)
p(Tn)
,
= p(Tn+1y, Tn)p(yTn).
Now, as we have previously shown, the statistics of Tn+1
is completely determined when s(n)
0 = y is fixed, hence
p(Tn+1y, Tn) = p(Tn+1y) ≡ H(Tn+1, y). Therefore, we
have
(cid:90)
(cid:90)
(cid:90)
(cid:90)
SCC(n, k) = E (TnTn+k) − Q1(n, k)
Q2(n, k)
,
(21)
E(TnTn+1) =
where
(22)
Q1(n, k) = E(Tn)E(Tn+k) ,
and
Q2(n, k) = m2(k)m2(n + k) =(cid:112)Var(Tn)Var(Tn+k) .
(23)
Here, Var(Tn) denotes the variance of the nth ISI distri-
bution, and m2 is the standard deviation given by Eq. 6.
Note that the definition Eq. 21 does not make use of the
dTn dTn+1 dy Tn Tn+1H(Tn+1, y)p(yTn)Fn(Tn) .
This can be further simplified by noting that p(yTn) =
Fn(Tn) and therefore
p(y,Tn)
E(TnTn+1) =
dTn dTn+1 dy Tn Tn+1 H(Tn+1, y)p(y, Tn) .
(24)
For n = 1, Eq. 24 can be simplified because s(1)
0
deterministic function of T1 (see Eq. 11), so that p(s(1)
is a
0 =
This results in an alternative, approximative expression
for the expectation E (TnTn+1):
8
(cid:90)
E (TnTn+1) =
dTndTn+1dy Tn Tn+1 H(Tn+1, y)Gn(y)Fn(Tn) .
(27)
Eq. 27 is therefore equivalent to Eq. 24 if p(y, Tn) =
Gn(y)Fn(Tn). Although Fig. 7 demonstates that p does
not factorise (the joint density is negatively sloped), we
show below that Eq. 27 approximates Eq. 24 very well.
Given that Eq. 27 does not require additional MC simu-
lations, the small error introduced by Eq. 27 is well offset
by the large reduction in computational cost.
There exists a third alternative expression for the expec-
tation of the product of ISIs suggested for a different, but
related, model, in [43]. It reads in our notation [44]
(cid:90)
E (TnTn+1) =
dTn dTn+1 dy Tn Tn+1 H(Tn+1, f (Tn, y))H(Tn, y)Gn−1(y) ,
(28)
where f is given by Eq. 14 or Eq. 15. The term
H(Tn, y)Gn−1(y) that appears in Eq. 28 is the same as
the one on the right-hand side in Eq. 18, which we used
to obtain the CDF of s(n)
0 .
For n = 1, Eq. 28 reads
(cid:90)
E (T1T2) =
dT1 dT2 T1 T2 H
(cid:16)
T2, f
(cid:16)
T1, s(0)
0
(cid:17)(cid:17)
(cid:124)
H
(cid:16)
(cid:123)(cid:122)
T1, s(0)
0
=F1(T1)
(cid:17)
(cid:125)
,
(cid:16)
(cid:17)
0
y − s(0)
because s is started from a point s(0)
0 , so that formally
, which collapses the integration over
G0(y) = δ
y in Eq. 28. Thus, for n = 1, Eq. 24 and 28 coincide. This
is also true for higher values of n. A proof for this equiva-
lence is presented in Appendix A. Eq. 28 only makes use
of the quantities H and G, which can be computed using
the iFPT approach as explained in the previous section.
We show comparisons between MC simulations and the
three expressions for E (TnTn+1), Eq. 24, Eq. 27 and
Eq. 28, in Fig. 8. The results presented in Fig. 8 are
in agreement with the observation that the two expres-
sions Eq. 24 and Eq. 28 are equivalent. We find that the
agreement of Eq. 28 with MC simulations is comparable
to Eq. 27, particularly for exponential adaptation. Inter-
estingly, the formally correct Eq. 24 and the approximate
Eq. 27 give comparable results; Eq. 24 slightly deviates
from MC simulations and Eq. 27 when n gets larger. The
maximal relative disagreement between MC and iFPT re-
sults is less than 2% (Fig. 8, bottom panels). We will see
below that the SCC is best approximated by using the ex-
act result Eq. 24 (or equivalently Eq. 28), as we expect.
0
FIG. 7. MC simulations with GS boundary correction for
the joint density of s(n)
(i.e. nth peak value of adaptation
current following the nth FPT) and Tn for different values of
n. Left: Power-law adaptation, n = 2. Right: Exponential
adaptation, n = 3. M = 106 MC realizations. Parameter
values as in Fig. 2 for power-law adaptation and as in Fig. 3
for exponential adaptation.
y, T1 = x) = δ
by Eqs. 14 and 15. Hence, we have for n = 1
y − f (x, s(0)
0 )
F1(x), where f is defined
(cid:17)
(cid:16)
(cid:90)
E(T1T2) =
dT1 dT2 T1 T2 H(T2, f (T1, s(0)
0 ))F1(T1).
0
(cid:17)
s(n)
0 = y, Tn
We have shown in Section V how to obtain the con-
(cid:16)
ditional FPT density H. To evaluate Eq. 24 for gen-
eral n, we still need to compute the joint density
. This can be achieved by means of an
p
MC simulation, where we fix a value of n and then record
the frequency with which pairs of s(n)
and Tn are gener-
ated by the system. We show an example of these den-
sities in Fig. 7. The most notable feature is an inverse
proportionality between s(n)
and Tn. The longer e.g. T2,
the less likely it is for the value of s after the second
firing, s(2)
Eq. 24 is formally correct, but not very practical for ac-
tual computations. This is because to apply the iFPT
approach, it is desirable to obtain all quantities needed
for the SCC using solutions of the FPE only, and no
MC simulations. These, however, are required to ob-
tain an approximation for the joint density p(s(n)
0 , Tn) in
Eq. 24. We therefore propose an approximation to com-
pute E (TnTn+1) using the available densities F, G and
H only. To that end, we note that
0 , to attain a high value.
0
p3(Tn+1, Tn, y) = p(Tn+1, Tny)p(y) .
(25)
If we now assume that Tn and s(n)
can approximate this as follows:
0
are independent, we
p3(Tn+1, Tn, y) ≈ p(Tn+1y)p(y)p(Tn) =
H(Tn+1, y)Gn(y)Fn(Tn) .
(26)
We attribute the discrepancy between MC simulations
and the exact result Eq. 24 to the error caused by the
numerical integration over the MC approximation of the
joint density p(y, Tn). We checked that applying a kernel
density estimation [45] to the MC results for p(y, Tn) did
not alter these results.
Similar results for Q1 and Q2 (Eqs. 22 and 23) are shown
in Figs. 9 and 10. The agreement is good, with the maxi-
mal relative disagreement always less than 5%. The rela-
tive disagreement for the statistics of the product of two
adjacent ISIs, Q1(n, 1), is in general larger than the error
for the moments, as can be seen by comparing Figs. 2 and
3 with Fig. 9. Indeed, for the case of power-law adapta-
tion, we observe an increase of roughly one order of mag-
nitude in the relative error even when the more accurate
CN scheme is used (see e.g.
left panels of Fig. 9). An
exception is the computation of the joint expectation,
shown in Fig. 8, where, depending on which methods
are compared, the relative disagreement is comparable
in size to the one for the computation of the moments
shown in Figs. 2 and 3. This increase of the relative dis-
agreement makes the computation of the SCC using the
iFPT approach hard, because the two expressions in the
numerator of Eq. 21 are quite close to one another for the
parameter values we have chosen here, meaning that the
numerator is small and indeed of the same magnitude or
even smaller as the relative disagreement, e.g. −2.6·10−2
in the left panel and −8.6 · 10−4 in the right panel of
Fig. 8 for n = 1 for the MC-GS simulation method. The
increase in the relative discrepancy is caused by error
propagation, because for the second-order statistics, one
has to multiply two quantities that both come with an
individual error. Therefore, whereas the iFPT approach
can in principle also be used to compute serial correla-
tions present in the spike train, obtaining reliable results
can in general be a computational challenge. When the
negative serial correlations are stronger, so that the dif-
ference in the numerator of Eq. 21 is larger, the iFPT
approach should give more accurate results. We stress
that the dominant source of error is not the computation
of the joint expectation E (TnTn+1) of ISIs, but the prod-
uct of the expectation of ISIs and the variances, which
can be seen by comparing the lower panels of Fig. 8 with
those of Figs. 9 and 10.
Finally, we show the SCC at lag 1 obtained by MC simu-
lations and PDE numerics in Fig. 11. The agreement
is worse than for all previously considered quantities,
but still reasonable. To verify the MC simulations, we
checked that our MC simulation setup was able to re-
produce known analytical results for the SCC obtained
in [15] for certain limiting cases. The anti-correlations
between adjacent ISIs (SCC(n, 1) < 0) strengthen until
they reach a stationary value.
Thus, we see that MC and PDE results for the SCC in
general do not agree as well as one would expect from the
good agreement of the expecations E(TnTn+1) in Fig. 8.
The deviation is likely more pronounced for parameters
that lead to small negative SCCs, which we have for both
9
FIG. 8. Top: Expectation E (TnTn+1) of adjacent ISIs for
different values of n. Left: power-law adaptation. Right:
exponential adaptation. M = 106 MC realizations. Empty
circles: Plain MC simulations of Eqs. 1 and 2. Triangles:
MC simulations with GS boundary correction. Pentagons:
Eq. 24. For power-law adaptation, the pentagons are on top
of the empty triangles and filled circles and hence not visi-
ble. Filled circles: Eq. 27. Diamonds: Eq. 28. The diamonds
are nearly on top of the filled circles and hence not visible.
The PDE results were obtained using a CN scheme (power-
law adaptation) or an Euler timestepping scheme (exponential
adaptation). The vertical error bars show the MC error for
a > 99.99% confidence interval. Bottom: Relative disagree-
ments defined by Eq. 13, where for power-law adapation, the
GS boundary corrected MC algorithm was used, and for ex-
ponential adaptation, the plain MC algorithm was used. For
the iFPT quantity, Eq. 24 was used. Parameter values as in
Fig. 2 for power-law adaptation and as in Fig. 3 for exponen-
tial adaptation.
models considered in this section. In the next Section, we
will compare this with results for the perfect integrate-
and-fire model, where parameter values are chosen so
that the SCCs are more negative and hence the agree-
ment is better. This is because the two terms in the nu-
merator of Eq. 21 are close to each other for small SCCs,
and hence a small error in them impacts the accuracy of
the SCC computation quite dramatically.
We show in the next section that our methods reproduce
known stationary analytical results for the SCC when we
consider the perfect integrate-and-fire model with single
exponential adaptation in a parameter regime where we
have large negative correlations, thus demonstrating that
our methods are sound, but SCC calculations are very
sensitive to numerical inaccuracies.
123450.50.60.70.80.91.0E(TnTn+1)12345n0.0000.0020.0040.0060.0080.0100.0120.014∆(E(TnTn+1))12345670.080.100.120.140.160.181234567n0.0020.0040.0060.0080.0100.0120.0140.016 10
FIG. 9. Top: Q1(n, 1) (defined in Eq. 22) of adjacent ISIs
for different values of n. Left: power-law adaptation. Right:
exponential adaptation. Empty circles: MC simulations of
Eqs. 1 and 2. Triangles: MC simulations with GS boundary
correction. Filled circles: Eq. 22, CN timestepping scheme
for power-law adaptation and Euler timestepping scheme for
exponential adaptation. Bottom: Relative disagreement de-
fined by Eq. 13, a plain MC algorithm was used to obtain
the relative disagreement. Parameter values as in Fig. 2 for
power-law adaptation and as in Fig. 3 for exponential adap-
tation.
FIG. 10. Top: Q2(n, 1) (defined in Eq. 23) for adjacent ISIs
for different values of n. Left: power-law adaptation. Right:
exponential adaptation. Empty circles: MC simulations of
Eqs. 1 and 2. Triangles: MC simulations with GS boundary
correction. Filled circles: Eq. 23, CN timestepping scheme
for power-law adaptation and Euler timestepping scheme for
exponential adaptation. Bottom: Relative disagreement de-
fined by Eq. 13, a plain MC algorithm was used to obtain
the relative disagreement. Parameter values as in Fig. 2 for
power-law adaptation and as in Fig. 3 for exponential adapta-
tion. Parameter values as in Fig. 2 for power-law adaptation
and as in Fig. 3 for exponential adaptation.
A. The perfect integrate-and-fire model
The adapting perfect integrate-and-fire (PIF) model
driven by white Gaussian noise and a single exponen-
tial adaptation current is one of the simplest models for
spike-triggered adaptation. For small noise intensity, an-
alytical expressions for the stationary SCC exist. We here
study this stationary limit case and compare analytical
formulas to results obtained with the iFPT approach.
The model reads (we follow the notation of [46] and [47])
dX = (I0 − s)dt + √2DdW (t) ,
ds
dt
= −
s
τa
.
(29)
(30)
The adaptation mechanism works in analogy to the pre-
vious model (Eq. 2): whenever X reaches the threshold
X = 1, s receives a kick of size ∆ ≡
and X is instan-
taneously reset to 0.
The stationary SCC at lag 1 for this model under the
assumption of small noise (i.e. D (cid:28) 1) reads [46]
(cid:101)∆
τa
SCC(k = 1) = −
α(1 − θ)(1 − α2θ)
1 + α2 − 2α2θ
,
(31)
where
α =
T ∗ =
s∗
− ∆
s∗
1 +(cid:101)∆
I0
,
θ =
,
s∗ =
,
I0 − s∗
I0 − s∗ + ∆
(cid:16)
− T ∗
∆
1 − exp
τa
(cid:17) .
Thus, we can compute the SCC in closed analytical form
as a function of the system parameters. This formula
serves as an important benchmark for our numerical re-
sults. In particular, we expect that after the described
transition to stationarity, the SCC given by Eq. 21 will
approach the stationary SCC given by Eq. 31. This is
confirmed in Fig. 12.
In particular, the agreement be-
tween MC simulations and the exact formula Eq. 24 is
very good (the relative disagreement between PDE nu-
merics and the analytical result is less than 6% for the
12340.50.60.70.80.91.01.1Q1(n,1)1234n0.0180.0190.0200.0210.0220.0230.0240.0250.026∆(Q1(n,1))12345670.080.100.120.140.160.180.201234567n0.0370.0380.0390.0400.0410.0420.0430.0440.04512340.1400.1450.1500.1550.1600.1650.1700.1750.1800.185Q2(n,1)1234n0.0020.0040.0060.0080.0100.0120.014∆(Q2(n,1))12345670.0200.0250.0300.0350.0400.0450.0501234567n0.0000.0050.0100.0150.0200.025 11
FIG. 11. Serial correlation coefficient at lag n = 1 (defined in
Eq. 21) for adjacent ISIs for different values of n for the LIF
model (Eq. 1). Left: power-law adaptation. Right: expo-
nential adaptation. Empty circles: Plain MC simulations of
Eqs. 1 and 2. Triangles: MC simulations with GS boundary
correction. Filled circles: PDE results, SCC computed using
Eq. 27. Pentagons: PDE results, SCC computed using Eq. 24.
CN timestepping scheme for power-law adaptation and Euler
timestepping scheme for exponential adaptation. Diamonds:
PDE results, SCC computed using Eq. 28. Parameter val-
ues as in Fig. 2 for power-law adaptation and as in Fig. 3
for exponential adaptation. Even if the joint expectations
shown in Fig. 8 agree well, this does not imply that the SCC
will be well approximated; the small correlation values (i.e.
the difference between E(TnTn+1) and Q1(n, 1)) for the two
examples lead to large discrepancies in the SCCs, which are
especially significant for the case of power-law adaptation.
stationary value); the agreement of MC simulations with
the approximation Eq. 27 is a bit worse, but still rea-
sonable. Thus, we conclude that our methodology can
be used more generally to compute the evolution of the
moments and SCCs. However, as seen in the previous
section, to obtain a good agreement between MC simula-
tions and PDE numerics, the computational effort might
be rather large. In particular, we note that the PIF ex-
ample shown in Fig. 12 gives rise to stronger negative
SCCs, which means that the error propagation has less
of an effect, but is still present, even when moments of
firing times between MC and PDE numerics disagree by
less than 1% (data not shown). We finally note that it is
also possible to analytically compute the stationary SCC
at higher lags and for different models (e.g. the leaky
integrate-and-fire model in the presence of weak noise
or for small adaptation currents) using the approach de-
scribed in [47], or, using a different approach, in [15].
FIG. 12. Serial correlation coefficient at lag n = 1 (defined in
Eq. 21) for different values of n for the PIF model (Eq. 30).
The dashed horizontal line is the stationary SCC given by
Eq. 31. Empty circles: Plain MC simulations of Eq. 30. Pen-
tagons: Eq. 24. Filled circles: Eq. 27. The PDE results
were obtained using a CN timestepping scheme. M = 106
MC realizations. Timestep h = 10−3. Parameter values:
D = 0.1, τa = 5.0, (cid:101)∆ = 10, I0 = 5.5, s(0)
0 = 5.0.
VII. SUMMARY AND CONCLUSIONS
In this paper, we have developed a numerical method
for the computation of moments and correlations in gen-
eral two-dimensional non-renewal escape time processes.
Our approach relies on the numerical solution of a two-
dimensional time-dependent FPE with initial conditions
obtained from marginal distributions of previous states of
the system. Crucially, the computation scheme presented
in this study is general insofar as it can be applied to any
stochastic process with a known reset condition (Eq. 1)
and any deterministic signal (Eq. 2). As an important
application, we have described the transition to station-
arity in a stochastic IF neuron model with spike-triggered
adaptation, which causes non-trivial ISI correlations. A
different mechanism for introducing positive correlations
between ISIs has recently been reported in [28] and can
equally well be analyzed with the presented methodol-
ogy. Moreover, our approach enables us to determine the
non-trivial timescale of transition to a stationary adapted
state by counting the number of intervals needed for this
transition.
Experimentally, the transition to stationarity is often
characterized by the behaviour of the instantaneous firing
rate [17, 18] [48]. The instantaneous firing rate is usually
obtained by averaging the neuronal activity bin-wise for
a fixed time. This differs from the firing rate used here
as given by the inverse of the mean ISI (Eq. 5). In other
words, while the instantaneous firing rate is measured in
real time, our firing rate relates to interval numbers. This
entails that for a given time t, the firing rate contains con-
tributions from, in general, past firing events that may
have occurred at any point k in the spike train. Knowing
the joint distributions of all ISIs Tk, it is at least in princi-
ple possible to reconstruct the instantaneous firing rate,
whereas given the instantaneous firing rate, we cannot
1234n−0.35−0.30−0.25−0.20SCC(n,1)1234567n−0.20−0.15−0.10−0.050.000.050.100.15123n−0.80−0.75−0.70−0.65−0.60−0.55−0.50SCC(n,1) reconstruct the joint distributions of the individual ISIs
Tk. Despite the difference in the definition of the firing
rate, it might be an interesting topic for further study to
classify the time scales of the transition to stationarity
both experimentally and based on the theory presented
here.
The computation of ISI moments using the iFPT ap-
proach is computationally inexpensive, giving rise to
small relative disagreements between solutions of the
FPE and direct MC simulations. In contrast, the com-
putation of correlations is harder. We observed that we
lost one order of magnitude in accuracy compared to the
simulation of the moments for the quantities Q1 (Eq. 22)
and Q2 (Eq. 23), which makes the reliable computation
of SCCs a computationally challenging task. We con-
clude that even a relative disagreement of ISI moments
between Monte Carlo estimations and PDE solutions of
the order 10−3 is not enough to reliably estimate the
SCC using PDE numerics only (but this might be spe-
cific for the examples we have considered), indicating that
more refined numerical methods or larger computational
ressources, or indeed both, are needed. When the differ-
ence between the two terms in the numerator of Eq. 21 is
large, the small error made by the numerical solution of
the PDE should have a less detrimental influence on the
final result. The need for more refined numerical meth-
ods is further substantiated by the fact that the more
accurate asymptotically stable CN timestepping scheme
did not result in a significant decrease in the relative dis-
agreement between PDE results and both plain MC and
MC-GS simulations, for both moments of firing intervals
and the SCC. In this paper, we have only discussed the
error associated with MC simulations, because it is read-
ily available. The numerical solution of the FPE is of
course also subject to numerical errors and future work
will likely benefit from a discussion about how to sys-
tematically reduce these errors. In this context, it might
be beneficial to compare the finite-element methods used
here to other methods for solving PDEs, such as finite
difference and finite volume methods [49]. A systematic
error estimation study might be made more difficult by
the fact that the diffusion matrix (Eq. 9) is not positive
definite [50, 51].
There is an alternative method to compute the ISI dis-
tributions given the distributions of the peak adaptation
currents using the formula
(cid:90)
supp(Gk−1) H(t, y)Gk−1(y)dy .
Fk(t) =
This is an integral equation frequently used in the con-
text of randomized FPT problems [52, 53], where usu-
ally Fk and the kernel H are given, and one tries to find
a matching distribution Gk−1 of starting points. Using
Eq. 32, we do not have to solve a time-dependent PDE
for each ISI, but must compute H once as the solution
of a time-dependent FPE with varying initial conditions
for s, similar to the computation of F1. The averaging
(32)
E(TnTn+1) =
(A1)
dTn dTn+1 ds(n)
0 Tn Tn+1 H(Tn+1, s(n)
0 )p(s(n)
0 , Tn) .
(cid:90)
(cid:90)
12
that the iFPT approach amounts to is particularly clear
in this formulation. The densities Gk are obtained as
discussed above (see Eq. 18). The approach relying on
Eq. 32 might be computationally less expensive, but we
found that it is not as exact as solving a time-dependent
FPE for each ISI, especially at larger times. This is likely
caused by errors when computing H, as the numerical
integration in Eq. 32 can be performed accurately and
efficiently. However, Eq. 32 could be useful for analytical
explorations when H is known.
We finally emphasize that our approach did not use the
complicated boundary conditions for stationary IF mod-
els, where the probability flux at threshold gives rise to
a discontinuity of the probability flux at reset [9, 54].
In contrast, our approach allows for the computation
of transient and stationary distributions of the adapta-
tion dynamics in an iterative fashion, requiring the solu-
tion of a two-dimensional time-dependent PDE. The only
boundary condition that has to be taken into account
is an absorbing boundary condition for the probability
density at the threshold xth. This makes the problem
tractable using finite-element approximation methods for
time-dependent PDEs, resulting in a general description
of two-dimensional IF models with spike-triggered adap-
tation. The approach we have described in this paper
can in principle also be used to gain analytical insight
into these system, however, quantities such as H and the
solution of a two-dimensional time-dependent PDE seem
to be unavailable in closed analytical form except in the
most simple cases.
ACKNOWLEDGMENTS
We would like to thank Alexandre Payeur for insightful
discussions and helpful comments (especially about the
equivalence of Eqs. 24 and 28) and NSERC Canada for
funding this work. Furthermore, we would like to thank
an anonymous referee for comments that helped us to
improve the manuscript.
Appendix A: Equivalence of Eqs. 24 and 28
We here show that Eq. 24 and Eq. 28 are equivalent.
We recall Eq. 24:
We re-write Eq. 28 as follows:
E (TnTn+1) =
dTn dTn+1 dy Tn Tn+1 H (Tn+1, f (Tn, y)) p(y, Tn) ,
(A2)
s(n−1)
0
y
=
have
and we
where
replaced
H(Tn, y)Gn−1(y) = p(y, Tn). Note that in Eq. A1,
p is the joint density of s(n)
and Tn, whereas p is the
joint density of s(n−1)
By inspection, the two expressions are identical if we can
show that
and Tn in Eq. A2.
0
0
13
0
to s(n)
), we have
0 = f (Tn, s(n−1)
for Tn and Tn+1 fixed.
Starting from the second line in Eq. A3, we change the
integration variable from s(n−1)
0 by observing that
ds(n)
from s(n)
and
ds(n−1)
therefore ds(n−1)
. We need to as-
sume that f is invertible with respect to the second ar-
gument, which is the case for both power-law (Eq. 14)
and exponential adaptation (Eq. 15) considered in this
paper. The integral then becomes
(cid:19)−1
= ds(n)
= ∂f
∂s(n−1)
∂s(n−1)
(cid:18)
∂f
0
0
0
0
0
0
0
(cid:90)
(cid:90)
(cid:104)
0 )p(s(n)
0 , Tn) =
H(Tn+1, f (Tn, s(n−1)
0
))p(s(n−1)
0
, Tn) ,
(A3)
ds(n)
ds(n−1)
0 H(Tn+1, s(n)
(cid:90)
0
dTn dTn+1 ds(n)
0 Tn Tn+1 H
(cid:16)
(cid:17)
(cid:104)
p
Tn+1, s(n)
0 )
f−1(Tn, s(n)
0 ), Tn
(cid:105)(cid:32)
(cid:33)−1
.
∂f
∂s(n−1)
0
(cid:105)(cid:18)
(cid:19)−1
is nothing but
so that
, Tn) to p(s(n)
0 , Tn). In-
f−1(Tn, s(n)
0 ), Tn)
But p
∂s(n−1)
the transformation from p(s(n−1)
deed, we have (fixing Tn)
∂f
0
0
(cid:104)
f−1(Tn, s(n)
0 ), Tn
(cid:105)(cid:32)
∂f
∂s(n−1)
0
p(s(n)
0 , Tn) = p
(A4)
(cid:33)−1
.
(A6)
p(s(n)
0 , Tn)
0
∂s(n)
∂s(n−1)
0
= p(s(n−1)
0
, Tn) ,
(A5)
Therefore, Eq. 24 and Eq. 28 are equivalent.
[1] A. Palacios, J. Aven, P. Longhini, V. In, and A. R. Bul-
J. Benda, J Neurosci. 32, 17332 (2012).
sara, Phys. Rev. E 74, 021122 (2006).
[2] A. Aragoneses, L. Carpi, N. Tarasov, D. V. Churkin,
M. C. Torrent, C. Masoller, and S. K. Turitsyn, Phys.
Rev. Lett. 116, 033902 (2016).
[3] M. J. Chacron, B. Lindner, and A. Longtin, Phys. Rev.
Lett. 92 (2004).
[4] J. A. Reinoso, M. C. Torrent, and C. Masoller, Phys.
Rev. E 94, 032218 (2016).
[5] S. Coombes, R. Thul, J. Laudanski, A. R. Palmer, and
C. J. Sumner, Frontiers in Life Science 5, 91 (2011).
[6] W. H. Nesse, L. Maler, and A. Longtin, Proc. Nat. Acad.
Sci. USA 107, 21973 (2010).
[7] D. R. Cox and P. A. W. Lewis, The Statistical Analysis
of Series of Events, Methuen's Monographs on Applied
Probability and Statistics (John Wiley, London, 1966).
[8] R. Rosenbaum, Front. Comput. Neurosci. 10 (2016).
[9] T. Schwalger and B. Lindner, Phys. Rev. E 92, 062703
(2015).
[15] E. Urdapilleta, Phys. Rev. E 84, 041904 (2011).
[16] E. Urdapilleta, Europhysics Letters 115 (2016).
[17] R. Naud and W. Gerstner, PLOS Computational Biology
8, 1 (2012).
[18] G. La Camera, A. Rauch, D. Thurbon, H.-R. Lüscher,
W. Senn, and S. Fusi, Journal of Neurophysiology 96,
3448 (2006).
[19] N. Ulanovsky, L. Las, D. Farkas, and I. Nelken, Journal
of Neuroscience 24, 10440 (2004).
[20] C. Kuehn, Multiple Time Scale Dynamics, 1st ed., Ap-
plied Mathematical Sciences No. 191 (Springer-Verlag,
2015).
[21] P. J. Drew and L. Abbott, J Neurophysiol 96, 826 (2006).
[22] B. N. Lundstrom, M. H. Higgs, W. J. Spain, and A. L.
Fairhall, Nat. Neurosci. 11, 1335 (2008).
[23] C. Pozzorini, R. Naud, S. Mensi, and W. Gerstner, Nat.
Neurosci. 16, 942 (2013).
[24] S. E. Clarke, R. Naud, A. Longtin, and L. Maler, Proc.
[10] Y.-H. Liu and X.-J. Wang, Journal of Computational
Nat. Acad. Sci. USA 110, 13624 (2013).
Neuroscience 10, 25 (2001).
[25] R. Kobayashi, Y. Tsubo, and S. Shinomoto, Front. Com-
[11] E. Muller, L. Buesing, J. Schemmel, and K. Meier, Neu-
put. Neurosci. 3, 121 (2009).
ral Computation 19, 2958 (2007).
[12] L. Hertäg, D. Durstewitz,
and N. Brunel, Frontiers
in Computational Neuroscience 8 (2014), 10.3389/fn-
com.2014.00116.
[13] T. Schwalger, K. Fisch, J. Benda, and B. Lindner, PLoS
Comput. Biol. 6 (2010).
[26] W. Gerstner and R. Naud, Science 326, 379 (2009).
[27] S. Redner, A Guide to First-Passage Processes, 1st ed.
(Cambridge University Press, 2010).
[28] G. D'Onofrio, E. Pirozzi,
and M. Magnasco, in EU-
ROCAST 2015, LNCS 9520, edited by R. M.-D. et al.
(Springer-Verlag, Berlin Heidelberg, 2015) pp. 166 -- 173.
[14] K. Fisch, T. Schwalger, B. Lindner, A. Herz,
and
[29] A. N. Burkitt, Biol. Cybern. 95, 1 (2006).
[30] J. Benda and A. V. M. Herz, Neural Computation 15,
(2016).
14
[43] M. J. Chacron, K. Pakdaman, and A. Longtin, Neural
[44] Note
Computation 15, 253 (2003).
3.17
(cid:82) dTndTn+1dy TnTn+1H(Tn+1, f (Tn, y))H (Tn, y)G∗(y).
E (TnTn+1)
stationary
state:
that
Eq.
the
in
[43]
is
in
=
estimation,"
[45] "Scipy v0.18.1 reference guide for Gaussian kernel
http://docs.scipy.org/doc/
density
scipy-0.18.1/reference/generated/scipy.stats.
gaussian_kde.html, accessed: 2017-01-31.
interspike-interval
statistics of
non-renewal neuron models, Ph.D. thesis, Humboldt-
Universität zu Berlin (2013).
[46] T. Schwalger, The
[47] T. Schwalger and B. Lindner, Frontiers in Computational
Neuroscience 7 (2013).
[48] Note that in [18], the timescale for single exponential
adaptation was inferred from the time course of the
numerically obtained instantaneous firing rate, showing
that the time course of the rate is not described by the
same single exponential time course of the adaptation
current. A similar observation is made in [30].
[49] R. Rosenbaum, F. Marpeau, J. Ma, A. Barua,
and
K. Josić, Journal of Mathematical Biology 65, 1 (2012).
[50] S. Salsa, Partial Differential Equations in Action, 1st ed.
(Springer-Verlag, 2008).
[51] A. Ern and J.-L. Guermond, Theory and Practice of Fi-
nite Elements, 1st ed., Applied Mathematical Sciences,
Vol. 159 (Springer-Verlag New York, 2004) p. 526.
[52] K. Jackson, A. Kreinin, and W. Zhang, Statistics and
Probability Letters 79, 2422 (2009).
[53] S. Jaimungal, A. Kreinin, and A. Valov, SIAM Theory
Probabl. Appl. 58, 493 (2014).
[54] N. Brunel and S. Sergi, J. Theor. Biol. 195, 87 (1998).
2523 (2003).
[31] L. Sacerdote and M. T. Giraudo, in Stochastic Biomath-
ematical Models, (M.Bachar et al. (eds.)), Lecture Notes
in Mathematics 2058, Chapter 5. , 99 (2013).
[32] C. van Vreeswijk, in Analysis of Parallel Spike Trains,
Springer Series in Computational Neuroscience, Vol. 7,
edited by S. Grün and S. Rotter (Springer Verlag, 2010)
pp. 3 -- 20.
[33] H. Risken, The Fokker-Planck equation: Methods of so-
lution and applications, 2nd ed. (Springer, 1989).
[34] B. F. Spencer Jr. and L. Bergman, Nonlinear Dynamics
4, 357 (1993).
[35] M. T. Giraudo and L. Sacerdote, Communications in
Statistics - Simulation and Computation 28, 1135 (1999).
[36] B. Lindner and A. Longtin, Journal of Theoretical Biol-
ogy 232, 505 (2005).
87, 167 (2000).
[37] E. Gobet, Stochastic Processes and their Applications
[38] D. Higham, X. Mao, M. Roj, Q. Song,
and G. Yin,
SIAM/ASA Journal on Uncertainty Quantification 1, 2
(2013).
[39] A. Logg, K.-A. Mardal, and G. Wells, eds., Automated
Solution of Differential Equations by the Finite Element
Method, Lecture Notes in Computational Science and En-
gineering, Vol. 84 (Springer Verlag, 2012).
[40] P. Knabner and L. Angermann, Numerical methods for
elliptic and parabolic partial differential equations, Texts
in Applied Mathematics, Vol. 44 (Springer Verlag, New
York Berlin Heidelberg, 2003).
0 + k · κ).
[41] The support of Gk is the open interval (0, s(0)
[42] R. Mankin and A. Rekker, Phys. Rev. E 94, 062103
|
1603.00397 | 1 | 1603 | 2016-02-29T18:44:23 | A Langevin model for complex cardiological time series | [
"q-bio.NC",
"cond-mat.stat-mech"
] | There has been considerable efforts to understand the underlying complex dynamics in physiological time series. Methods originated from statistical physics revealed a non-Gaussian statistics and long range correlations in those signals. This suggests that the regulatory system operates out of equilibrium. Herein the complex fluctuations in blood pressure time series were successful described by physiological motivated Langevin equation under a sigmoid restoring force with multiplicative noise. | q-bio.NC | q-bio |
A Langevin model for complex cardiological time series
C.E.C. Galhardo1, B. C. Coutinho2, T.J.P.Penna3, M.A. de Menezes3 and P.P.S. Soares4
1 Materials Metrology Division, National Institute of Metrology,
Quality and Technology, Duque de Caxias, RJ ,Brazil
2 Petrobr´as, Cenpes, Rio de Janeiro, RJ, Brazil
3 Instituto de F´ısica, Universidade Federal Fluminense, Niteroi, RJ ,Brazil and
4 Instituto Biom´edico, Universidade Federal Fluminense, Niteroi, RJ, Brazil
There has been considerable efforts to understand the underlying complex dynamics in physiolog-
ical time series. Methods originated from statistical physics revealed a non-Gaussian statistics and
long range correlations in those signals. This suggests that the regulatory system operates out of
equilibrium. Herein the complex fluctuations in blood pressure time series were successful described
by physiological motivated Langevin equation under a sigmoid restoring force with multiplicative
noise.
PACS numbers: 05.40.-a,87.80.Vt,87.19.L-,87.19.Hh
Introduction - The autonomic nervous system is able
to maintain life signals at safe levels by the action of a
pair of nerve branches, called sympathetic and parasym-
pathetic. While the sympathetic prepares our body for
"flight or fight" the parasympathetic (or vagal) is con-
sidered as a "rest and digest" system.
In many cases,
to achieve the homeostatic optimal levels, these systems
have a competitive approach: while one start up an phys-
iological reaction the other one suppress it [1].
Homeostasis depends on the blood flow according to
the metabolic demands of each body part. The exchange
of nutrients and metabolites occurs when blood flows
through capillary channels. The perfusion either into or
out the capillary depends on the blood pressure. Ade-
quate levels of blood pressure are controlled by several
mechanisms that can be classified according to the re-
sponse delay: the long term and short term control.
In order to maintain homeostasis the body automat-
ically responds to changes. These responses are called
reflexes. The principal short term reflex regulation of ar-
terial blood pressure is the baroreflex. Stretch-sensitive
mechanoreceptors are located in the carotid sinus and
aortic arch connected to the brainstem, or nucleus trac-
tus solitarii, by the glossopharyngeal and vagal nerves.
After integrating the afferent signals the central nervous
system, in turn, excite/inhibit the vagal branch if the
pressure is high/low enough, closing the circuit for what
can be regarded as a self-inhibitory feedback [2].
In addition, blood pressure may vary to adapt different
physiological conditions such as exercise [3] or pregnancy
[4] and in certain disease states such as hypertension [5].
The optimal level of blood pressure must be risk adjusted:
cannot be high enough to cause structural damage and
cannot be low enough to hinder the nutrient flow.
It has been suggest that the underlying cardiac con-
trol system can be characterized as a complex system.
Indeed, heart and blood pressure present fractal time se-
ries with long range correlations and non-Gaussian dis-
tributions [6] and undergoes a breakdown of critical char-
acteristics like a continous second ordem phase transtion
[7].
In this letter, we propose an analytically tractable
stochastic model for the baroreflex that capture the frac-
tality and the non-Gaussian behavior in the blood pres-
sure time series. This model leads to solutions in terms of
the so called q-Gaussians, which is a well known function
in the framework of Nonextensive Statistical Mechanics.
This theory have shown excellent results describing time
series systems with fractal structure and an introduction
and many examples of the theory can be found in [8].
Experiment - Two groups of adult male Wistar rats
were analyzed: control rats and chronic sinoaortic dener-
vated rats, animals surgically denervated 20 days before
measurements. Sinoaortic denervated was performed us-
ing the methods described by Krieger et al [9], and ba-
sically consists of full disruption of the nerve fibers con-
necting the the afferent signals from the baroreceptors to
the brainstem, leading to hypertension, tachycardia, and
an increase in blood pressure lability. Twenty days after
the surgery, only the increase of blood pressure lability
is usually observed [10]. This adaptation had been pre-
viously investigated in the light of detrended fluctuation
analysis [11].
Blood pressure was recorded from the left femoral
artery for 90 minutes in conscious rats. Before the ana-
log to digital conversion, blood pressure was low-pass
filtered (fc= 50 Hz) for high-frequency noise removal,
and recorded with a 2kHz sampling frequency. Diastolic
(minimum) values were detected after parabolic interpo-
lation and signal artifacts were visually identified and
removed. A more detailed account of this experiment
can be found in [10]. Since the measurements were done
in awake conscious unrestrained rats some distortions in
the blood pressure signal arise due to their movements.
To reduce this problem we discard series that show any
kind of discontinuities.
y =x0.9
y =x0.5
101
102
n
103
c)
b)
)
n
(
F
104
103
102
101
100
10-1
105
100
a)
)
g
H
m
m
(
P
B
A
95
90
85
80
75
70
65
60
55
0
2000
4000
6000
Counts
8000
10000
10-4 10-3 10-2 10-1
Normalized frequency
FIG. 1. Representative data of diastolic blood pressure from
one animal. (a) a typical diastolic blood pressure record is
presented. (b) The histogram for the blood pressure values.
A Gaussian distribution (black dash-line) with sample mean
and sample variance was plotted with the histogram, high-
lighting the non-Gaussian behavior.
(c) DFA for diastolic
blood pressure series. For long time scales the DFA expo-
nent is α = 0.93. We also applied DFA to shuffled data (red
up triangles) obtaining α ≈ 0.5. The curves α = 0.5 (red full
line), 0.9 (blue full line) are plotted as guides to the eye.
Results and Discussion - Figure 1a a typical diastolic
blood pressure record from a control animal is presented.
The histogram for the blood pressure values is presented
in figure 1c. A Gaussian distribution with sample mean
and sample variance was plotted with the histogram,
highlighting the non-Gaussian behavior. Furthermore,
it clearly presents a positive skew. Despise the fact that
the time series has a constant mean of 81.00mmHg over
a range of 104 inter-beats, a complex pattern of fluc-
tuations were observed. To characterize the stochastic
dynamics of the blood pressure a detrended fluctuation
analysis (DFA) was performed. The log-log plot of the
fluctuation function F (n) is presented in figure 1b. A
crossover around n = 35 separate two different behav-
iors. For long time scales the DFA exponent α = 0.93
while a for short time scales α = 1.18. This crossover
had been discussed previously and it could be associated
with autonomic nervous system control of arterial blood
pressure [11, 12]. To confirm the presence of long range
correlations, the time series were randomly shuffled and
DFA was applied. The shuffled time series present DFA
exponent α ≈ 0.5, as typical white noise, show at figure
1b in red, which demonstrates that long range correla-
tions arises from the blood pressure control system.
External perturbations are continuously disrupting the
cardiac system. In such noisy environment the autonomic
2
nervous system must keep the blood pressure at accept-
able levels by integrating chemical and mechanical input
from afferents to regulate the blood pressure. However,
the neural transmission itself is noisy [13]. Information
transmission along the axons is made by electrical signals.
Those signals, called action potentials, are created by ion
channels in a cell membrane. They travel down the axon
to its end where the neurotransmitters are release in the
synaptic gap. Those transmitters activate receptors in
the post synaptic neuron [14]. The noise source are di-
verse, for example, they could have physical background
in thermodynamic or quantum mechanics, as it happens
with sensor neurons, or could be built up in the cellular
network [15].
In this sense, the cardiac neural control
system is under intrinsic and extrinsic noise.
100
10-1
10-2
10-3
)
.
u
.
a
(
y
c
n
e
u
q
e
r
f
d
e
z
i
l
a
m
r
o
N
100
10-1
10-2
10-3
)
.
u
.
a
(
y
c
n
e
u
q
e
r
f
d
e
z
i
l
a
m
r
o
N
−30
−20
−10
0
ps
10
20
30
−30
−20
−10
0
ps
10
20
30
(a) Diastolic pressure distribution
control rats
(b) Diastolic pressure distribu-
tion denervated rats
FIG. 2. Semi log histogram of rescaled experimental dias-
tolic pressure data. Figure 2a present the data of four con-
trol rats while figure 2b shows the data of four denervated
animals. All data where rescaled ps = k0(p − p)/σ, where
k0 = 4, p is the blood pressure data, p is the median and
σ is the sample standard deviation. Despite the fact that
the animals were unrestrained during the measurement, the
rescaled data collapse. In each figure the stable solution of
the model where plotted. For the control animals, figure 2a,
a q-Gaussian distribution with q = 11/9 were found. For
the denervated rats,figure 2b, a almost Gaussian distribution
were found with q = 1.04.
Synaptic signal transmission is found to be modeled
as a diffusive process wherein additive and multiplicative
noise plays relevant role to describe the neuronal response
[16, 17]. Let p = P − p where P is the blood pressure
and p is a measure of central tendency, like the mean or
the median. Then the neural control of blood pressure
dynamics could be modeled as a Brownian particle under
a restoring force of the baroreflex:
dp
dt
= f (p) + g(p)ξ(t) + η(t).
(1)
where f (p) is the restoring force, g(p) is the diffusion co-
efficient, ξ(t) is the multiplicative noise and η(t) is the ad-
ditive noise. They both have zero mean and show Marko-
vian correlations, hξtξt′ i = M δtt′ and hηtηt′i = Rδtt′.
Physiologically η(t) could be interpreted as external per-
turbations on cardiac system while g(p)ξ(t) is a noise
arising from the neural transmission.
Harris and Wolpert [18] propose an unifying optimal
control theory for information process in motor systems.
The theory is based on a single physiological assump-
tion: the neural noise increases in variance with the size
of the control signal. This assumption is then made by
the model presented in this paper. Here the noise am-
plitude is signal dependent and it is proportional to the
control system. Several complex systems, like financial
and turbulence [19, 20], has a linear drift coefficient and a
quadratic state-dependent diffusion coefficient. In those
system f (p) = − dU(p)
and g2(p) ∝ U (p). Analogously,
dp
the restoring force of the baroreflex could be expressed
as above with U (p) = τ
2 g(p)2.Generally the baroreflex is
represented by a sigmoid function [21], f (p) = A− C
1+e−Bp
where A, B, C are the model parameters. Integrating the
τ h−Ap + C ln(1+eBp)
i
sigmoid one can found g2(p) = 2
where A < C, A > 0, and B > 0 must hold that g2(p)
be strictly positive.
B
Equation (1) could be rewritten as Fokker-Planck
equation using Ito calculus:
∂F (p)
∂t
= −
∂[f (p)F (p)]
∂p
+
∂[G′(p)F (p)]
∂p
+ G(p)
∂2[F (p)]
∂p2
(2)
where G(p) = R+M g2(p), G′(p) = ∂G(p)
and F (p) is the
arterial blood pressure distribution. The physiologically
viable solution for equation 2 must have F (p → ∞) = 0
as boundary condition. In those conditions, equation 2
has a q-Gaussian distribution as stationary solution [22]:
∂p
F (p) = N (cid:2)1 − (1 − q)βg(p)2(cid:3)
1
1−q
(3)
By replacing equation 3 in equation 2, we obtain:
gg ′{(τ +M )[2+βτ (q−1)g2 +2]−βτ (R+M g2)} = 0. (4)
B ln( C
R
If gg ′ = 0 implies f (p) = 0 for some p∗. As A < C and
B > 0 then p∗ = − 1
A − 1) > 0. The positive value of
p∗ is associated with the continuous action of the sympa-
thetic activity when not inhibited. In other words, when
the blood pressure is near the central value there is an
continuous activity, called sympathetic tone, of the ner-
vous system to increase the blood pressure. Otherwise, if
gg ′ 6= 0, equation 4 must hold for every p, which implies
q = 2 − 1
. If coupling constant
τ is larger than multiplicative noise amplitude such as
M/τ → 0 then q → 1 and F (p) converge to Gaussian
distribution. The multiplicative noise becomes too small
and the long tail vanish. On the other hand, if M/τ → ∞
q → 2 and F (p) converges to the Lorentz distribution [23]
1+M/τ and β = 2(1+M/τ )
3
destroying homeostasis. In spite of q-Gaussian distribu-
tion are defined for q < 3, to describe the non-Gaussian
behavior observed in figure 1c and keep the physiological
feasibility, q must stick in between 1 and 2.
To compare the proposed model with the experimen-
tally observed diastolic blood pressure data, a rescale
transformation where performed: ps = k0(p−p)/σ, where
k0 = 4, p is the blood pressure, p is it median and σ
is it standard deviation. Once the animals were unre-
strained, each BP series could shown different central
values and variability. However, when the rescale where
performed the data collapse as figure 2 shows.
In the
same figure the stable solution of the model, equation
3, where plotted. For control animals, figure 2a, the
following parameters were used: τ = 21/5, M = 6/5,
A = 7/5, B = 2/5, and C = 9/5. These values imply-
ing q = 11/9 ≈ 1.22. A value of q = 1.26 ± 0.1 where
observed in heart rate variability [24]. Several other non-
equilibrium systems presents q = 1.22,for example, fina-
cial markets [25], hadron-hadron collisions [26], and geo-
logical faults [27]. For surgically disrupted animals, pre-
sented at figure 2b, the following parameters were used:
τ = 26/5, M = 1/5, A = 1, B = 1/5, and C = 9/5.
These values implying q = 1.04, very close to the Gaus-
sian distribution (q = 1.0), showing that the non Gauss-
sian fluctuations are intrinsicly associated with the con-
trol feedback loop at short time scale.
To characterize quantitatively the dynamics, a Monte
Carlo simulation of the equation 1 were performed for
both group. The same parameters of the analitical curves
in figure 2 were used. To understand how the model could
reproduce the underlying dynamics a DFA of the simu-
lated series were compared with the blood pressure real
data. As recently discussed DFA could present a biased
estimator for the Hurst exponent [28]. Nevertheless it
still holding as good methodology to work on real data
[29], and any bias that could be introduced in the fluctu-
ation function will happens on both time series, the real
and the simulated one, being no hindrance to the analy-
sis. The results of the DFA for synthetic series and real
data is presented in figure 3. The time scales of the blood
pressure control feedback loop is commonly analyzed in
three different ranges: high , low and very low frequency
. The baroreflex plays a significant role at high and low
frequency control [30].
After 20 days, since major damages were inflicted on
the neural circuit responsible for the baroreflex, the con-
trol were achieved by some other redundant mechanism
but with larger response delay [11]. Figure 3 shows the
model captures the fluctuations behavior in the barore-
flex control range for both groups. Time scales larger
than n > 500 are in the very low frequency range and
the model developed here plays no role. However, except
for long time scales, the fluctuation functions reveals a
very similar pattern between synthetic and experimen-
tal time series. Based on this analysis, we conclude the
1e+06
100000
10000
1000
100
10
1
)
n
(
F
0.1
1
10
100
n
1000
10000
(a) DFA control rats
1e+07
1e+06
100000
10000
)
n
(
F
1000
100
10
1
0.1
1
10
100
n
1000
10000
(b) DFA denervated rats
FIG. 3. Comparison between DFA analysis of the experimen-
tal data and DFA of the time series generated by equation
1. The experimental data for four different animals of each
group are plotted with open symbols (purple (cid:3), green ◦, black
△ and blue ▽). The DFA from simulated series are plotted
with red ×. The curves were translated for better visibility,
and values of F (n) are vertically shifted.
In figure 3a the
control group is presented while in 3b the denervated group
is presented. The model capture the fluctuations behavior
until the n ≈ 500 for both groups. The error bars represent
the standard deviation over 1000 simulations.
model presented here captures the complex behavior of
blood pressure control.
A langevin model based on a sigmoidal restoring force
with multiplicative noise for the diastolic blood pressure
time series was discussed. The stationary solution was
a q-Gaussian distribution with q = 11/9 that describes
remarkably well the blood pressure time series recorded
from femoral artery for 30 minutes in conscious uncon-
strained rats. To investigate the model dynamics a syn-
thetic time series was generated by a monte carlo simula-
tion of the equivalent Langevin equation. The synthetic
time series and the experimental data was characterized
using DFA and the fluctuation functions were compared.
The fluctuation functions reveals a very similar pattern
in the synthetic and experimental time series for high and
low frequency.
[1] L. K. McCorry, Am. J. Pharm. Educ. 71, 78 (2007).
4
[2] A.
C.
Guyton
Hall,
Textbook of Medical Physiology (W.B. Saunders Com-
pany, 2000).
and
J.
E.
[3] K. M. Gallagher, P. J. Fadel, S. A. Smith, M. Strømstad,
K. Ide, N. H. Secher, and P. B. Raven, Exp. Physiol. 91,
79 (2006).
[4] M. E. Crandall and C. M. Heesch, Am. J. Physiol. Regul.
Integr. Comp. Physiol. 258, R1417 (1990).
[5] G. Grassi, B. M. Cattaneo, G. Seravalle, A. Lanfranchi,
and G. Mancia, Hypertension 31, 68 (1998).
[6] C. K. Peng, J. Mietus, J. M. Hausdorff, S. Havlin, H. E.
Stanley, and A. L. Goldberger, Phys. Rev. Lett. 70, 1343
(1993).
[7] K. Kiyono, Z. R. Struzik, N. Aoyagi, F. Togo, and Y. Ya-
mamoto, Phys. Rev. Lett. 95, 058101 (2005).
[8] C. Tsallis et al., Introduction to nonextensive statistical
mechanics (Springer, 2009).
[9] E. M. Krieger, Circ. Res. 15, 511 (1964).
[10] P. P. d. S. Soares, C. S. Port, F. M. F. Abdalla, R. N.
De La Fuente, E. D. Moreira, E. M. Krieger, and M. C.
Irigoyen, J. Cardiovasc. Pharmacol. 47, 331 (2006).
[11] C. E. C. Galhardo, T. J. P. Penna, M. A. de Menezes,
and P. Soares, New J. Phys. 11, 103005 (2009).
[12] K. Fuchs, A. Y. Schumann, A. Kuhnhold, P. Guzik,
J. Piskorski, G. Schmidt,
and J. W. Kantelhardt, in
Proc. 6th Conference of the European Study Group on
Cardiovascular Oscillations, Vol. 56 (2010).
[13] R. B. Stein, E. R. Gossen, and K. E. Jones, Nat. Rev.
Neurosci. 6, 389 (2005).
[14] J. A. White, J. T. Rubinstein, and A. R. Kay, Trends
Neurosci. 23, 131 (2000).
[15] A. A. Faisal, L. P. Selen, and D. M. Wolpert, Nat. Rev.
Neurosci. 9, 292 (2008).
[16] C. Boucsein, T. Tetzlaff, R. Meier, A. Aertsen,
and
B. Naundorf, J. Neurosci. 29, 1006 (2009).
[17] M. J. Richardson and R. Swarbrick, Phys. Rev. Lett. 105,
178102 (2010).
[18] C. M. Harris and D. M. Wolpert, Nature 394, 780 (1998).
[19] J. Peinke, R. Friedrich, et al., Phys. Rev. Lett. 83, 5495
(1999).
[20] S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner, and
Y. Dodge, Nature 381, 767 (1996).
[21] B. Kent, J. Drane, B. Blumenstein, and J. Manning,
Cardiology 57, 295 (1972).
[22] L. Borland, Physics Letters A 245, 67 (1998).
[23] S. Umarov, C. Tsallis, and S. Steinberg, Milan J. Math.
76, 307 (2008).
[24] G. Pavlos, L. Karakatsanis, M. Xenakis, E. Pavlos, A. Il-
iopoulos, and D. Sarafopoulos, Physica A 395, 58 (2014).
[25] M. Ausloos and K. Ivanova, Phys. Rev. E 68, 046122
(2003).
[26] G. Barnafoldi, K. Urmossy, and T. Bir´o, in Journal of
Physics: Conference Series, Vol. 270 (IOP Publishing,
2011) p. 012008.
[27] F. Vallianatos, Europhys. Lett. 102, 28006 (2013).
[28] R. Bryce and K. Sprague, Sci. Rep. 2 (2012).
[29] Y.-H. Shao, G.-F. Gu, Z.-Q. Jiang, W.-X. Zhou, and
D. Sornette, Sci. Rep. 2 (2012).
[30] C. Cerutti, C. Barres, and C. Paultre, Am. J. Physiol.
Heart Circ. Physiol. 266, H1993 (1994).
|
1607.01959 | 1 | 1607 | 2016-07-07T10:54:53 | The Energy Landscape of Neurophysiological Activity Implicit in Brain Network Structure | [
"q-bio.NC"
] | A critical mystery in neuroscience lies in determining how anatomical structure impacts the complex functional dynamics of human thought. How does large-scale brain circuitry constrain states of neuronal activity and transitions between those states? We address these questions using a maximum entropy model of brain dynamics informed by white matter tractography. We demonstrate that the most probable brain states -- characterized by minimal energy -- display common activation profiles across brain areas: local spatially-contiguous sets of brain regions reminiscent of cognitive systems are co-activated frequently. The predicted activation rate of these systems is highly correlated with the observed activation rate measured in a separate resting state fMRI data set, validating the utility of the maximum entropy model in describing neurophysiologial dynamics. This approach also offers a formal notion of the energy of activity within a system, and the energy of activity shared between systems. We observe that within- and between-system energies cleanly separate cognitive systems into distinct categories, optimized for differential contributions to integrated v.s. segregated function. These results support the notion that energetic and structural constraints circumscribe brain dynamics, offering novel insights into the roles that cognitive systems play in driving whole-brain activation patterns. | q-bio.NC | q-bio |
The Energy Landscape of Neurophysiological Activity
Implicit in Brain Network Structure
Shi Gu1,2, Matthew Cieslak3, Benjamin Baird4, Sarah F. Muldoon5, Scott T. Grafton3,
Fabio Pasqualetti6, Danielle S. Bassett2,7,*
1 Applied Mathematics and Computational Science, University of Pennsylvania,
Philadelphia, PA, 19104 USA
2 Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, 19104
USA
3 Department of Psychological and Brain Sciences, University of California, Santa
Barbara, CA, 93106 USA
4 Center for Sleep and Consciousness, University of Wisconsin - Madison, Madison, WI
53706
5 Department of Mathematics and CDSE Program, University at Buffalo, SUNY,
Buffalo, NY 14260
6 Department of Mechanical Engineering, University of California, Riverside, CA, 92521
USA
7 Department of Electrical & Systems Engineering, University of Pennsylvania,
Philadelphia, PA, 19104 USA
* To whom correspondence should be addressed: [email protected].
Abstract
A critical mystery in neuroscience lies in determining how anatomical structure impacts
the complex functional dynamics of human thought. How does large-scale brain
circuitry constrain states of neuronal activity and transitions between those states? We
address these questions using a maximum entropy model of brain dynamics informed by
white matter tractography. We demonstrate that the most probable brain states --
characterized by minimal energy -- display common activation profiles across brain areas:
local spatially-contiguous sets of brain regions reminiscent of cognitive systems are
co-activated frequently. The predicted activation rate of these systems is highly
correlated with the observed activation rate measured in a separate resting state fMRI
data set, validating the utility of the maximum entropy model in describing
neurophysiologial dynamics. This approach also offers a formal notion of the energy of
activity within a system, and the energy of activity shared between systems. We
observe that within- and between-system energies cleanly separate cognitive systems
into distinct categories, optimized for differential contributions to integrated versus
segregated function. These results support the notion that energetic and structural
constraints circumscribe brain dynamics, offering novel insights into the roles that
cognitive systems play in driving whole-brain activation patterns.
Author Summary
How does the complex interconnection structure between large-scale brain regions
impact how we think? Does this structure inform how we move between different
PLOS
1/21
thoughts or actions? We address these questions using a simple mathematical model of
brain dynamics that is informed by empirical estimates of anatomical interconnection
structure between brain regions. Our results suggest that while interconnection
structure plays an important role in constraining and predicting brain dynamics, an
additional layer of explanatory power is offered by considering energetic constraints on
those dynamics. These data offer a richer landscape of mechanisms that enhance our
understanding of how we may move from one thought to the next.
Introduction
A human's adaptability to rapidly changing environments depends critically on the
brain's ability to carefully control the time within (and transitions among) different
states. Here, we use the term state to refer to a pattern of activity across neurons or
brain regions [1]. The recent era of brain mapping has beautifully demonstrated that
the pattern of activity across the brain or portions thereof [2] differs in different
cognitive states [3]. These variable patterns of activity have enabled the study of
cognitive function via the manipulation of distinct task elements [3], the combination of
task elements [4, 5], or the temporal interleaving of task elements [6, 7]. Such methods
for studying cognitive function are built on the traditional view of mental
chronectomy [8], which suggests that brain states are additive and therefore separable in
both space and time (although see [9] for a discussion of potential caveats).
Philosophically, the supposed separability and additivity of brain states suggests the
presence of strong constraints on the patterns of activations that can be elicited by the
human's environment. The two most common types of constraints studied in the
literature are energetic constraints and structural constraints [10]. Energetic constraints
refer to fundamental limits on the evolution [11] or usage of neural systems [12], which
inform the costs of establishing and maintaining functional connections between
anatomically distributed neurons [13]. Such constraints can be collectively studied
within the broad theory of brain function posited by the free energy principal -- a notion
drawn from statistical physics and information theory -- which states that the brain
changes its state to minimize the free energy in neural activity [14, 15]. The posited
preference for low energy states motivates an examination of the time within and
transitions among local minimums of a predicted energy landscape of brain
activity [16, 17].
While energetic costs likely form critical constraints on functional brain dynamics, an
arguably equally important influence is the underlying structure and anatomy linking
brain areas. Intuitively, quickly changing the activity of two brain regions that are not
directly connected to one another by a structural pathway may be more challenging
than changing the activity of two regions that are directly connected [13, 18]. Indeed,
the role of structural connectivity in constraining and shaping brain dynamics has been
the topic of intense neuroscientific inquiry in recent years [19 -- 23]. Evidence suggests
that the pattern of connections between brain regions directly informs not only the ease
with which the brain may control state transitions [24], but also the ease with which one
can externally elicit a state transition using non-invasive neurostimulation [25].
While energy and anatomy both form critical constraints on brain dynamics, they
have largely been studied in isolation, hampering an understanding of their collective
influence. Here, we propose a novel framework that combines energetic and structural
constraints on brain state dynamics in a free energy model explicitly informed by
structural connectivity. Using this framework, we map out the predicted energy
landscape of brain states, identify local minima in the energy landscape, and study the
profile of activation patterns present in these minima. Our approach offers fundamental
insights into the distinct role that brain regions and larger cognitive systems play in
PLOS
2/21
distributing energy to enable cognitive function. Further, the results lay important
groundwork for the study of energy landscapes in psychiatric disease and neurological
disorders, where brain state transitions are known to be critically altered but
mechanisms driving these alterations remain far from understood [26, 27].
Fig 1. Conceptual Schematic. (A) A weighted structural brain network represents
the number of white matter streamlines connecting brain regions. (B) We model each
brain region as a binary object, being either active or inactive. (C) Using a maximum
entropy model, we infer the full landscape of predicted activity patterns -- binary
vectors indicating the regions that are active and the regions that are not active -- as
well as the energy of each pattern (or state). We use this mathematical framework to
identify and study local minima in the energy landscape: states predicted to form the
foundational repertoire of brain function.
Materials and Methods
Human DSI Data Acquisition and Preprocessing
Diffusion spectrum images (DSI) were acquired for a total of 48 subjects (mean age
22.6±5.1 years, 24 female, 2 left handed) along with a T 1 weighted anatomical scan at
each scanning session [28]. Of these subjects, 41 were scanned ones, 1 was scanned
twice, and 6 were scanned three times, for a total of 61 scans.
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 written consent in
accordance with the Institutional Review Board/Human Subjects Committee,
University of California, Santa Barbara.
DSI data were reconstructed in DSI Studio (www.dsi-studio.labsolver.org) using
PLOS
3/21
State 1State 2State 3Brain NetworkAnteriorPosteriorInferiorSuperior1: Active0: Not activeEdge: Structural ConnectionActivityState EnergyACBActivation Magnitude (a.u.)01 q-space diffeomorphic reconstruction (QSDR) [29]. 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.
Structural Network Construction
Anatomical scans were segmented using FreeSurfer [30] and parcellated according to the
Lausanne 2008 atlas included in the connectome mapping toolkit [31]. A parcellation
scheme including 234 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.
From these data, we built structural brain networks from each of the 61 diffusion
spectrum imaging scans. Consistent with previous work [13, 24, 25, 32 -- 37], we defined
these structural brain networks from the streamlines linking N = 234 large-scale cortical
and subcortical regions extracted from the Lausanne atlas [31]. We summarize these
estimates in a weighted adjacency matrix A whose entries Aij reflect the structural
connectivity between region i and region j (Fig. 1A).
Following [24], here we use an edge weight definition based on the quantitative
anisotropy (QA). QA is described by Yeh et. al (2010) as a measurement of the signal
strength for a specific fiber population a in an ODF Ψ(a) [38, 39]. QA is given by the
difference between Ψ(a) and the isotropic component of the spin density function (SDF,
ψ) ISO (ψ) scaled by the SDF's scaling constant. Along-streamline QA was calculated
based on the angles actually used when tracking each streamline. Although
along-streamline QA is more specific to the anatomical structure being tracked, QA is
more sensitive to MRI artifacts such as B1 inhomogeneity. QA is calculated for each
streamline. We then averaged values over all streamlines connecting a pair of regions,
and used this value to weight the edge between the regions.
Resting state fMRI data
As an interesting comparison to the computational model, we used resting state fMRI
data collected from an independent cohort composed of 25 healthy right-handed adult
subjects (15 female), with a mean age of 19.6 years (STD 2.06 year). All subjects gave
informed consent in writing, in accordance with the Institutional Review Board of the
University of California, Santa Barbara. Resting-state fMRI scans were collected on a
3.0-T Siemens Tim Trio scanner equipped with high performance gradients at the
University of California, Santa Barbara Brain Imaging Center. A T2*-weighted
echo-planar imaging (EPI) sequence was used (TR=2000 ms; TE=30 ms; flip angle=90°;
acquisition matrix=64×64; FOV=192 mm; acquisition voxel size = 3×3×3.5 mm; 37
interleaved slices; acquisition length=410s).
PLOS
4/21
We preprocessed the resting state fMRI data using an in-house script adapted from
the workflows described in detail elsewhere [40, 41]. The first four volumes of each
sequence were dropped to control for instability effects of the scanner. Slice timing and
motion correction were performed in AFNIusing 3dvolreg and FreeSurfer's BBRegister
was used to co-register functional and anatomical spaces. Brain, CSF, and WM masks
were extracted, the time series were masked with the brain mask, and grand-mean
scaling was applied. The temporal derivative of the original 6 displacement and rotation
motion parameters was obtained and the quadratic term was calculated for each of
these 12 motion parameters, resulting in a total of 24 motion parameters which were
regressed from the signal. The principal components of physiological noise were
estimated using CompCor (aCompCor and tCompCor) and these components were
additionally regressed from the signal. The global signal was not regressed. Finally,
signals were low passed below 0.1 Hz and high passed above 0.01 Hz in AFNI. To
extract regional brain signals from the voxel-level time series, a mask for each brain
region in the Lausanne2008 atlas was obtained and FreeSurfer was used to individually
map regions to the subject space. A winner-takes-all algorithm was used to combine
mapped regions into a single mask. The resulting signal for each region was then
extracted in FreeSurfer using mrisegstats.
Following data preprocessing and time series extraction, we next turned to
extracting observed brain states. Importantly, physiological changes relating to neural
computations take place on a time scale much smaller than the time scale of BOLD
image acquisition. Thus, we treat each TR as representing a distinct brain state. To
maximize consistency between the model-based and data-based approaches, we
transformed the continuous BOLD magnitude values into a binary state vector by
thresholding regional BOLD signals at 0. From the set of binary state vectors across all
TRs, we defined activation rates in a manner identical to that described for the
maximum entropy model data.
Defining an Energy Landscape
We begin by defining a brain state both intuitively and in mathematical terms. A brain
state is a macroscopic pattern of BOLD activity across K regions of the brain (Fig. 1A).
For simplicity, here we study the case in which each brain region i can be either active
(σi = 1) or inactive (σi = 0). Then, the binary vector σ = (σ1, σ2, . . . , σK) respresents a
brain state configuration.
Next, we wish to define the energy of a brain state. We build on prior work
demonstrating the neurophysiological relevance of maximum entropy models in
estimating the energy of brain states in rest and task conditions [42, 43]. For a given
state σ, we write its energy in the second order expansion:
Jiσi,
(1)
(cid:88)
i(cid:54)=j
Jijσiσj −(cid:88)
i
E(σ) = − 1
2
where J represents an interaction matrix whose elements Jij indicate the strength of the
interaction between region i and region j. If Jij > 0, this edge (i, j) decreases the
energy of state σ, while if Jij < 0, this edge (i, j) increases the energy of state σ. The
column sum of the structural brain network, Ji =(cid:80)
K, is the strength of
√
j Jij/
region i. In the thermodynamic equilibrium of the system associated with the energy
defined in Eqn [1], the entropy of the system is maximized and the probability of the
configuration σ is P (σ) ∝ e−E(σ).
The choice of the interaction matrix is an important one, particularly as it tunes the
relative contribution of edges to the system energy. In this study, we seek to study
structural interactions in light of an appropriate null model. We therefore define the
PLOS
5/21
interaction matrix to be equal to the modularity matrix [44] of the structural brain
network:
for i (cid:54)= j, where A is the adjacency matrix, pi =(cid:80)K
(A − ppT/2m)ij
i=1 Aij, and 2m =(cid:80)K
1
2m
Jij =
choice ensures that any element Jij measures the difference between the strength of the
edge Aij in the brain and the expected strength of that edge in an appropriate null
model (here given as the Newman-Girvan null model [45]). If the edge is stronger than
expected, it will decrease the energy of the whole system when activated, while if the
edge is weaker than expected, it will increase the energy of the whole system when
activated.
(2)
j=1 pj. This
Discovering Local Minima
The model described above provides an explicit correspondence between a brain's state
and the energy of that state, in essence formalizing a multidimensional landscape on
which brain dynamics may occur. We now turn to identifying and characterizing the
local minima of that energy landscape (Fig. 1C). We begin by defining a local minimum:
K) is a local minimum if E(σ) ≥ E(σ∗) for all vectors σ
a binary state σ∗ = (σ∗
satisfying σ − σ∗1 = 1, which means that the state σ∗ realizes the lowest energy
among its neighboring states within the closed unit sphere. We wish to collect all local
minima in a matrix Σ∗ with
1, . . . , σ∗
Σ∗ =
1,2
2,2
1,1
2,1
σ∗
σ∗
σ∗
σ∗
...
...
K,1 σ∗
σ∗
K,2
···
···
. . .
···
1,N
2,N
σ∗
σ∗
...
σ∗
K,N
(3)
K×N
where N is the number of local minima and K is the number of nodes in the structural
brain network (or equivalently the cardinality of the adjacency matrix A).
Now that we have defined a local minimum of the energy landscape, we wish to
discover these local minima given the pattern of white matter connections represented
in structural brain networks. To discover local minima of E(σ), we first note that the
total number of states σ = (σ1, . . . , σK) is 2K, which -- when K = 234 -- prohibits an
exhaustive analysis of all possibilities. Moreover, the problem of finding the ground
state is an NP-complete problem [46], and thus it is unrealistic to expect to identify all
local minima of a structural brain network. We therefore choose to employ a clever
heuristic to identify local minima. Specifically, we combine the Metropolis Sampling
method [47] and a steep search algorithm using gradient descent methods. We identify a
starting position by choosing a state uniformly at random from the set of all possible
states. Then, we step through the energy landscape via a random walk driven by the
Metropolis Sampling criteria (see Algorithm 1). At each point on the walk, we use the
PLOS
6/21
steep search algorithm to identify the closest local minimum.
Algorithm 1: Heuristic Algorithm to Sample the Energy Landscape to Identify
Local Minima.
1 Let σj be the vector obtained by changing the value of the j-th entry of σ;
2 for t = 1 : N do
3
t−1 with probability p = min(1, e−β∗(E( σ)−E(σt−1))) and
Randomly select an index j ∈ {1, ..., K};
Set σt = σt = σj
σt = σt = σt−1 otherwise;
while σt is not a local minimum do
Set j∗ = arg minj E( σj
t );
Set σt = σj∗
t ;
4
5
6
7
8
end
Set σ∗
t = σt;
9
10 end
1, σ∗
2, . . . , σ∗
Here, σ1, σ2, . . . , σN are the sampled states, σ∗
N are the sampled local
minima, and β is the temperature parameter which can be absorbed in E(σ). In the
context of any sampling procedure, it is important to determine the number of samples
necessary to adequately cover the space. Theoretically, we wish to identify a number of
samples N following which the distribution of energies of the local minima remains
stable. Practically speaking, we choose 4 million samples in this study, and demonstrate
the stability of the energy distribution in the Supplement. A second important
consideration is to determine the initial state, that is the state from which the random
walk begins. Here we choose this state uniformly at random from the set of all possible
states. However, this dependence on a uniform probability distribution may not be
consistent with the actual probability of states in the energy landscape. We therefore
must ensure that our results are not dependent on our choice of initial state. To ensure
independence from the initial state, we dismiss the first 30, 000 local minima identified,
and we demonstrate in the Supplement that this procedure ensures our results are not
dependent on the choice of initial state.
Characterizing Local Minima
Following collection of local minima, we wished to characterize their nature as well as
their relationships to one another. First, we estimated the radius of each local minimum
as the Hamming distance from the minimum to the closest sampled point on energy
landscape. Next, calculated the Hamming distance from each local minimum to the first
sampled local minimum, a second quantification of the diversity of the energy lanscape
that we traverse in our sampling. Finally, we quantify how diverse the observed local
minima are by calculating the pairwise normalized mutual information [48] of each pair
of local minima.
Next, we wished to understand the role of different regions and cognitive systems in
the minimal energy states. Cognitive systems are sets of brain regions that show
coordinated activity profiles in resting state or task-based BOLD fMRI data [9]. They
include the visual, somatosensory, motor, auditory, default mode, salience,
fronto-parietal, cingulo-opercular, dorsal and ventral attention systems, as well as
subcortical areas. Here, the specific association of regions of interest to cognitive
systems are exactly as listed in [24] and based originally on results described in [49]. We
characterize the roles of these systems in the local minima by assessing their activation
rates, as well as the utilization energies required for communication within and between
systems.
PLOS
7/21
Activation Rates
Intuitively, we define the activation rate of a node i as the average activation state of
that node over all the local minima. Formally, the activation rate for region i is defined
as
ri =
,
(4)
(cid:80)N
l=1 σ∗
N
il
where l indexes over states, and recall N is the number of local minima. The computed
activation rate offers a prediction of which regions are more versus less active across the
local minima (that is, the brain's locally "stable " states), and can be directly compared
with the resting state activation rate estimated from empirical fMRI data.
Utilization Energies
To complement the information provided by the activation rates, we also defined the
energetic costs associated with utilizing within- and between-systems interactions. We
note that each cognitive system is a subnetwork of the whole brain network. We use the
index set I to indicate the set of nodes associated with the cognitive system, and thus
I gives the number of nodes in the system. Then, for a given state σ, the
within-system energy measures the cost associated with the set of interactions
constituting the subnetwork. The between-system energy measures the cost associated
with the set of interactions between the subnetwork and all other nodes in the whole
network. Formally, we define
EW (σ) = −
1
2I(I − 1)
EB(σ) = −
1
2I(K − I)
(cid:88)
(cid:88)
i(cid:54)=j,i,j∈I
i∈I,j /∈I
Jijσiσj
Jijσiσj
where EW measures the within-system energy, EB measures the between-system energy,
and the normalization coefficients 1/(II − 1), 1/(I(K − I)) are chosen by
considering the number of the corresponding interactions.
Permutation Tests for State Association
For a given local minimum configuration σ∗, we associate it with system iσ∗ ,
iσ∗ = arg max
i
NMI(σ∗, σsys
i
),
i
where σsys
is the configuration pattern of system i such that the corresponding regions
for system i are activated and the others not and where "NMI" refers to the Normalized
Mutual Information [48], which is used to measure the similarities between the two
states. To obtain the null distribution, for each local minimum configuration σ∗ in the
collection Σ∗, we permute the configuration at each position of σ∗ to achieve a random
configuration with the same activation rate, and we then compute the associated
percentage in each system. Then we repeat this procedure to generate N samples and
construct the null distribution of the probability of being configured as each system
pattern. Considering the large size of the state collection, the variance of the samples in
the null distribution will be small. We pick N = 50 here. See Fig. 4 D for the results.
PLOS
8/21
Results
Local Minima in the Brain's Energy Landscape
By sampling the energy landscape of each structural connectivity matrix, we identified
an average of approximately 450 local minima or low energy brain states: binary vectors
indicating the pattern of active and inactive brain regions (see Methods). On average
across the 61 scans, 144 brain regions were active in a given local minimum, representing
61.70% of the total (standard deviation: 6.04%). This large percentage suggests that
the brain may utilize a broad set of rich and diverse activations to perform cognitive
functions [50], rather than activation patterns localized to small geographic areas.
To quantify this diversity, we examined the location of minima on the energy
landscape, the size of the basins surrounding the minima, and the mutual information
between minima. First, we estimated the distance from the first local minima identified
to all subsequent minima (see Methods; Fig. 2A). We observe an order of magnitude
change in the distance between the first and second local minima, and the first and last
local minima, suggesting that local minima span a broad geographic domain in the
energy landscape. Interestingly, these minima differ not only in their location on the
energy landscape, but also in the size of the basins surrounding them. We estimate
basin size by calculating the radius of each local minimum (see Methods) and show that
the distribution of radii follows a power-law, with the majority of minima displaying a
small radius, and only a few minima displaying a large radius (Fig. 2B). More
specifically, we fit the function P (r) = Cr−α -- where C is a constant -- to the data
using a statistically principled approach [51, 52]. We identified an α = 2.6300 for r > 6,
and the p-value for the goodness of fit was p < 1 × 10−5 indicating that the power law
was an accurate fit to the data. As a final quantification of minima diversity, we
estimated the normalized mutual information between every pair of local minima, as an
intuitive measurement of the similarity between anatomical compositions of the minima.
We observe that the probability distribution of normalized mutual information between
minima pairs is heavy tailed, indicating that most minima pairs display very dissimilar
anatomical compositions, and only a few minima pairs display similar anatomical
compositions (Fig. 2C).
From a neurophysiological perspective, it is also important to note that these local
minima displayed significant local structure. Specifically, we found that regions within
known cognitive systems tended to be active together. The probability that regions
were co-active is 48.22%, which was significantly greater than that expected in a null
distribution (associated p − value was p < 1 × 10−5; see Methods). These results
indicate that the structural connectivity between brain regions, and the assumption of
energy conservation, predict that regions that belong to the same cognitive system will
tend to be co active with one another during diverse cognitive functions. Indeed, these
predictions are consistent with previous studies of functional neuroimaging data
demonstrating that groups co-active regions tended to align well with known cognitive
systems [53, 54].
Activation Rates of Cognitive Systems
Given the alignment of activation patterns with cognitive systems, we next asked
whether certain cognitive systems were activated more frequently than others. To
address this question, we studied the activation rate of each cognitive system, which
measures how frequently the regions in the cognitive system participated in the set of
states identified as local minima. Intuitively, if the activation rate is high, the system is
more likely to be active in diverse brain states. We observed that systems indeed
showed signficantly different activation rates (Fig. 2D). Sensorimotor systems (auditory,
PLOS
9/21
Fig 2. Simulated Activation Rates.(A) The distribution of distances from the first
local minimum to other local minima. Each point and error-bar is calculated across a
bin of 30 minima; error bars indicate standard error of the mean over the 30 minima.
(B) The probability distribution of the radius of each local minimum is well-fit by a
power-law. The radius of a local minimum is defined as its distance to the closest
sampled point on the energy landscape. (C) The distribution of the pairwise normalized
mutual information between all pairs of local minima. (D) Average activation rates for
all 14 a priori defined cognitive systems [49]. Error bars indicate standard error of the
mean across subjects.
visual, somatosensory) tended to display the lowest activation rates, followed by higher
order cognitive systems (salience, attention, fronto-parietal, and cingulo-opercular), and
subcortical structures. The system with the largest activate rate was the default mode
system, suggesting that activation of this system is particularly explicable from
structural connectivity and the assumption of energy conservation. The unique role of
the default mode system is consistent with predictions from network control theory that
highlight the optimal placement of default mode regions within the network to
maximize potential to move the brain into many easily reachable states with minimal
energetic costs [24].
It is important to determine whether this activation rate is driven by simple
properties of the structural connectivity matrix that do not depend on assumptions of
energy conservation. To address this question, we next assessed the relationship
between a simple summary statistic of the structural connectivity matrix -- the strength,
or weighted degree, of a brain region -- and the predicted activation rate drawn from the
maximum entropy model. We observed that the activation rate was not well predicted
by the weighted degree on average over brain regions (see Supplement). These data
suggest that the additional assumption of energy conservation produces a set of brain
states that cannot be predicted from simple statistics of structural connectivity alone.
PLOS
10/21
x100101102Pr(X ≥ x)10-510-410-310-210-1100The n-th local Minimum105045165225285345405465525585s-s020406080100120140160180200Normal Mutual Information00.20.40.60.81Probability00.050.10.15ABCActivation Rate0.500.550.600.650.700.750.80Ventral AttentionCingulo-OpercularFronto-ParietalDefault ModeSomatosensoryDorsal AttentionVisualOtherAuditorySubcorticalD Relating Predicted Activation Rates to Rates Observed in
Functional Neuroimaging Data
Before exercising the model further to explore how energy is utilized in the brain, we
wished to quantify the relationships between the theoretically predicted activation rates,
and activation rates observed empirically in functional neuroimaging data. Specifically,
we studied fMRI data acquired in a separate set of healthy adult human subjects.
Next, we wished to directly quantify similarities between the predicted activation
rates and those observed empirically in resting state fMRI. In the resting state data, we
observed that highly active regions were located in broad swaths of frontal and parietal
cortex, as well as medial prefrontal, precuneus, and cingulate (Fig. 3A). This pattern of
high activation is consistent with the so-called "default-mode" of resting state brain
function [55]. In our maximum entropy model, we observed that the areas predicted to
have high activation rates show a broad similarity to those observed empirically in the
resting state (Fig. 3B). Indeed, we observed that the empirical resting activation rate of
brain regions is significantly correlated with the activation rate predicted from the
maximum entropy model (Fig. 3C; Pearson correlation coefficient r = 0.18, p = 0.0046).
These results suggest that the modeling framework we use here has significant
similarities to observable features of resting state brain dynamics. However, it is
important to mention that there are also noticeable differences between the two maps:
the predicted activation rates are strong along the medial wall, while the resting state
activation rates extend to larger sections of lateral cortices.
Fig 3. Validating Predicted Activation Rates in Functional Neuroimaging
Data. (A) From resting state BOLD data acquired in an independent cohort, we
estimated the true activation rate by transforming the continuous BOLD magnitudes to
binary state vectors by thresholding the signals at 0 (see Methods). We use these binary
state vectors to estimate the activation rates of each brain region across the full resting
state scan. Here we show the mean activation rate of each brain region, averaged over
subjects. (B) For comparison, we also show the mean predicted activation rate
estimated from the local minima of the maximum entropy model, as defined in
Equation [4], and averaged over subjects. (C) We observe that the activation rates
estimated from resting state fMRI data are significantly positively correlated with the
activation rates estimated from the local minima of the maximum entropy model
(Pearson's correlation coefficient r = 0.18, p = 0.0046). Each data point represents a
brain region, with either observed or predicted activation rates averaged over subjects.
Utilization Energies of Cognitive Systems
We next turned to exercising our model to further understand the potential constraints
on brain state dynamics. Specifically, we asked how cognitive systems utilized the
PLOS
11/21
Model Activation RateSimulated Activation RateReal Activation Rate0.40.50.60.70.80.91.0AReal Activation Rate0.480.4850.490.4950.50.5050.510.5150.52r = 0.1845p = 0.0046BC0101 minimal energy presumably available to them. Intuitively, this question encompasses
both how energy is utilized by within-system interactions, and how energy is utilized by
between-system interactions. We therefore defined the within-system energy, which
measures the cost associated with the set of interactions constituting the cognitive
system, and the between-system energy, which measures the cost associated with the set
of interactions between cognitive systems. We observed a fairly strong dissociation
between these two variables: cognitive systems that display a large within-system
energy are not necessarily those that display a large between-system energy (see Fig. 4A
and B). Indeed, within- and between-system energies are not significantly correlated
across systems (Pearson's correlation coefficient r = 0.2287 and p = 0.5250), suggesting
that these two variables offer markers of distinct constraints.
Moreover, we observed that the 2-dimensional plane mapped out by the within- and
between-system energies of all brain regions revealed the presence of 4 surprisingly
distinct clusters (Fig. 4C) that are not explicable by simple statistics such as network
degree (see Supplement). Each cluster represents a unique strategy in energy utilization
that is directly reflected in its activation pattern; in other words, each cluster offers a
distinct balance between the energetic costs of within-system interactions and the
energetic costs of between-system interactions. The central cluster, displaying high
within-system energies but low between-system energies, is composed of subcortical and
fronto-parietal systems. A high between-system energy cone emanating from this
central cluster is composed of predominantly primary and secondary sensorimotor
cortices in somatosensory, visual, and auditory systems. A second cone emanating from
the central cluster with a slightly lower between-system energy is composed
predominantly of regions in the default mode system. The final cone emanating from
the central cluster with between-system energies near zero is composed predominantly
of regions in the dorsal and ventral attention systems. These results suggest that
sensorimotor, default mode, attention, and cognitive control circuits display differential
preferences for energy utilization: regions in attentional systems share less energy with
other networks than regions in sensorimotor systems, while the default mode maintains
an intermediate balance.
The clear differences in the energies utilized by different cognitive systems and by
between-system interactions begs the question of whether the brain cares about these
energies. Does the brain prefer smaller within-system energies, smaller between-system
energies, or some balance between the two? To address this question, we studied the
ensemble of local minima, and asked which systems were commonly expressed.
Specifically, for each local minimum, we determined which system was most activated,
assigned the minimum to that system, and performed this assignment for all minima.
We observed that 3 systems were represented at higher percentages than expected in a
permutation-based null model (see Methods): the default mode system, the visual
system, and the somatosensory system (Fig. 4D). Importantly, these three systems
represent the systems with the highest between system energies (Fig. 4C), but are
indistinguishable from other cognitive systems in terms of within-system energy. These
results suggest that the brain may prefer high integration between systems over low
integration, and that the constraint of between-system energies is more fundamental to
brain function than the constraint of within-system energies.
Discussion
In this paper, we address the question of how large-scale brain circuitry and distinct
energetic constraints produce whole-brain patterns of activity. We build our approach
on a maximum entropy model of brain dynamics that is explicitly informed by estimates
of white matter microstructure derived from deterministic tractography algorithms. The
PLOS
12/21
Fig 4. Utilization Energies of Cognitive Systems. (A) Average within-system
energy of each cognitive system; error bars indicate standard error of the mean across
subjects. (B) Average between-system energy of each cognitive system; error bars
indicate standard error of the mean across subjects. (C) The 2-dimensional plane
mapped out by the within- and between-system energies of different brain systems.
Each data point represents a different brain region, and visual clusters of regions are
highlighted with lightly colored sectors. The sector direction is determined by
minimizing the squared loss in point density of the local cloud and the width is
determined by the orthogonal standard derivation at the center along the sector
direction. In this panel, all data points represent values averaged across subjects. (D)
The percentages of minima displaying preferential activation of each system; each
minima was assigned to the system which whom it shared the largest normalized mutual
information. Errorbars indicate the differences between the observed percentages and
those of the null distribution with random activation patterns across regions.
model allows us to study minimal energy states, which we observe to be composed of
co-activity in local spatially-contiguous sets of brain regions reminiscent of cognitive
systems. These systems are differentially active, and activity patterns are significantly
correlated with the observed activation rate measured in a separate resting state fMRI
data set. Finally, we exercise this model to ask how cognitive systems utilize the
minimal energy presumably available to them. We find that the energy utilized within
and between cognitive systems distinguishes 4 classes of energy utilization dynamics,
corresponding to sensorimotor, default mode, attention, and cognitive control functions.
These results suggest that diverse cognitive systems are optimized for differential
contributions to integrated versus segregated function via distinct patterns of energy
utilization. More generally, the results highlight the importance of considering energetic
constraints in linking structural connectivity to observed dynamics of neural activity.
PLOS
13/21
-0.4-0.200.20.40.60.81.01.2-1.2-1.0-0.8-0.6-0.4-0.20Ventral AttentionCingulo-OpercularFronto-ParietalDefault ModeSomatosensoryDorsal AttentionVisualOtherAuditorySubcorticalA-0.4-0.200.20.40.60.81.01.2Ventral AttentionCingulo-OpercularFronto-ParietalDefault ModeSomatosensoryDorsal AttentionVisualOtherAuditorySubcorticalBx10-5x10-6-2-1.5-1.0-0.500.51-1.5-1.0-0.500.51.01.52.02.5x10-6x10-5CDWithin-System EnergyWithin-System EnergyBetween-System EnergyBetween-System EnergyVentral AttentionCingulo-OpercularFronto-ParietalDefault ModeSomatosensoryDorsal AttentionVisualAuditorySubcorticalOther-0.4-0.3-0.2-0.100.10.2Relative Activation Rate The Role of Activation vs. Connectivity in Understanding
Brain Dynamics
As the interest in understanding structural brain connectomics has blossomed in the last
several years [56, 57], it has not been accompanied by an equally vivid interest in linking
subsequent insights to the more traditional notions of brain activation profiles [58].
Indeed, the fields of systems and cognitive neuroscience have instead experienced a
pervasive divide between the relatively newer notions of connectome mapping and the
relatively traditional yet highly effective notions of brain mapping [59], which have led
to powerful insights into neural function in the last quarter century [60]. This divide is
at least in part due to the fact that graph theory and network-based methods on which
the field of connectomics is based have few tools available to link node properties
(activity) with edge properties (connectivity) [61]. While technically explicable, however,
the conceptual divide between these fields can only lead to their detriment, and
synergistic efforts are necessary to develop a language in which both activity and
connectivity can be examined in concert [58]. This study offers one explicit
mathematical modeling framework in which to study the relationships between
activation profiles across the whole brain and underlying structural connectivity linking
brain regions. Complementary approaches include the model-based techniques
formalized in the publicly available resource The Virtual Brain [62, 63].
Co-activation Architecture
In this study, we observed that brain regions affiliated with known cognitive systems --
including somatosensory, visual, auditory, default mode, dorsal and ventral attention,
fronto-parietal, and cingulo-opercular -- also tend to be active together with one another
in low energy brain states. Indeed, these theoretical results are consistent with previous
studies of functional neuroimaging data demonstrating that groups of co-active regions
tended to align well with known cognitive systems [53, 54]. This correspondence is
particularly interesting when one considers how these cognitive systems were initially
defined: and that is, based on strong and dense functional connectivity [49]. Thus, our
results point to a fundamental mapping between activity and connectivity: regions that
are active together tend to be functionally connected together. The presence of such a
map has been empirically observed in the resting state (though not in task [58]), in both
healthy controls and in people with schizophrenia where the map appears to be
fundamentally altered in its nature [64 -- 66]. Here we offer a mechanism for this mapping
based on a combined consideration of energy- and connectivity-based constraints.
Critical Importance of Energy Constraints
The quest to understand and predict brain dynamics from the architecture of
underlying structural connectivity is certainly not a new one. In fact, there have been
concerted efforts over the last decade and more to identify structural predictors of the
resting state BOLD signal. Seminal contributions have included the observations of
statistically significant correlations between structural connectivity estimated from
diffusion imaging data and functional connectivity estimated from fMRI [67], as well as
extensions of these correlations that account for long distance paths along white matter
tracts [22] and spectral properties of structural matrices [68]. The question of how brain
structure constrains a wide range of brain states (beyond simply the resting state) is a
very open area of inquiry. Moreover, this question is particularly challenging to address
with empirical data because it is difficult to obtain data from humans in more than a
handful of task states [69]. For this reason, computational models play a very important
role in offering testbeds for the development of theories linking structure to ensembles of
PLOS
14/21
brain states, which can in turn offer testable predictions. Our results suggest that an
understanding of the relationship between brain structure and function is perhaps
ill-constrained when examining connectivity alone. The additional assumption of energy
conservation produces a set of brain states that cannot be simply predicted from
statistics of structural connectivity, perhaps offering a mechanism for the large amount
of unexplained variance in prior predictions [22, 67].
Methodological Considerations
Our results are built on the formalism of the maximum entropy model, which is
predicated on pairwise statistics [70]. However, emerging evidence suggests that some
neurophysiological phenomenon are better studied in the framework of simplicial
complexes rather than dyads [71]. For example, integrate and fire neurons exposed to
common fluctuating input display strong beyond-pairwise correlations that cannot be
captured by maximum entropy models [72]. Similar arguments can also be made for
co-activation patterns in BOLD fMRI [53, 54, 73]. It will be interesting in future to
determine the role energetic and structural constraints on observed higher-order
functional interactions during human cognitive function.
A second important consideration is that the maximum entropy model is appropriate
for systems at equilibrium. Therefore, the local minima identified may not accurately
represent the full class of states expected to be elicited by daily activity. Instead, the
local minima identified here are expected to more accurately represent the set of states
expected to appear as a person rests in the so-called default mode of brain function,
which is thought to lie near a stable equilibrium [21]. Such an interpretation is consistent
with our findings that the activation rates predicted by the maximum entropy model are
strongly correlated with the activation rates observed in resting state fMRI data.
Conclusion
The analyses presented in this study produce information regarding an underlying
energy landscape through which the brain is predicted to move. The existence of such a
landscape motivates the very interesting question of how the brain transitions between
states. In sampling this landscape, we have used a simple random walk in an effort to
extract a large ensemble of possible brain states, measured by local minima. However,
the question of which walks a healthy (or diseased) brain might take through this
landscape remains open. Such walks or dynamical trajectories may be determined by
energetic inputs to certain regions of the brain [24], either by external stimuli or by
neuromodulation [25]. In this context, network control theory may offer explicit
predictions regarding the optimal dynamic trajectories that the brain may take through
a set of states to move from an initial state to a target state with little energetic
resources [24, 74]. In addition to inputs to single regions, changes in a cognitive task --
for example elicited by task-switching paradigms -- may also drive a specific trajectory
of brain states. Indeed, it is intuitively plausible that the asymmetric switch costs
observed between cognitively effortful and less effortful tasks [75, 76] may be explained
by characteristics of the energy landscape defined by structural connections between
task-activated brain regions.
PLOS
15/21
Supporting Information
S1 Fig. Stability of the Energy Distribution with respect to the Number
of Samples. We plot the probability distribution of the energy for the first 2000, 4000,
6000 and 8000 samples. We observe that the shapes of the probability distributions are
qualitatively consistent. We confirm this qualitative observation with
Kolmogorov-Smirnov tests (see Supplemental text).
S2 Fig. Activation Rate is Poorly Predicted by Regional Degree and
Energy. (A) Scatterplot of weighted regional degree versus activation rate. (B)
Scatterplot of regional energy versus activation rate. We observe that the activation
rate is not well predicted by either regional energy or weighted degree.
S3 Fig. Relationship Between Utilization Energies and Degree. (A)
Scatterplot of within-system connectivity versus between-system connectivity, for
individual brain regions that are color-coded by cognitive system. (B) The same data
presented in panel (A) except only for the default mode, attention, and task control
systems, demonstrating the indistinguishability of default mode and attention systems.
(C) Scatterplot of the between- and within- system energy. (D) The same data
presented in panel (C) except only for the detault mode, attention, and task control
systems. We observe that cognitive systems are more clearly separated in the
2-dimensional space of the within- and between-system energies than in the
2-dimensional space of the within- and between-system connectivity. Across all four
panels, data points indicate brain regions, and color of data points indicates which
cognitive system that region is affiliated with (see legend for map from color to
cognitive system).
S4 Fig. Simulated Activation Rate is Significantly Correlated with the
Rates Observed in Resting State Functional Neuroimaging Data. We observe
that the activation rates estimated from resting state fMRI data are significantly
positively correlated with the activation rates estimated from the local minima of the
maximum entropy model (Pearson's correlation coefficient r = 0.17, p = 0.0094). Each
data point represents a brain region, with either observed or predicted activation rates
averaged over subjects.
S1 Appendix. S1 Appendix. Supplementary Information for "The Energy
Landscape of Neurophysiological Activity Implicit in Brain Network Structure".
Acknowledgments
We thank Ankit Khambhati, Lia Papadopoulus, and Evelyn Tang for helpful comments
on an earlier fersion of this manuscript. D.S.B., S.G., and R.F.B. acknowledge support
from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan
Foundation, the Army Research Laboratory and the Army Research Office through
contract numbers W911NF-10-2-0022 and W911NF-14-1-0679,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, BCS-1430087, and PHY-1554488). FP
acknowledges support from BCS-1430280. B.B. was supported by the National
Institutes of Health under Ruth L. Kirschstein National Research Service Award
F32NS089348 from the NINDS.
PLOS
16/21
References
1. Tang YY, Rothbart MK, Posner MI. Neural correlates of establishing,
maintaining, and switching brain states. Trends in cognitive sciences.
2012;16(6):330 -- 337.
2. Mahmoudi A, Takerkart S, Regragui F, Boussaoud D, Brovelli A. Multivoxel
pattern analysis for fMRI data: a review. Comput Math Methods Med.
2012;2012:961257.
3. Gazzaniga MS, editor. The cognitive neurosciences. MIT Press; 2013.
4. Szameitat AJ, Schubert T, Muller HJ. How to test for dual-task-specific effects in
brain imaging studies: an evaluation of potential analysis methods. Neuroimage.
2011;54(3):1765 -- 1773.
5. Alavash M, Hilgetag CC, Thiel CM, Giessing C. Persistency and flexibility of
complex brain networks underlie dual-task interference. Hum Brain Mapp.
2015;36(9):3542 -- 3562.
6. Ruge H, Jamadar S, Zimmermann U, Karayanidis F. The many faces of
preparatory control in task switching: reviewing a decade of fMRI research. Hum
Brain Mapp. 2013;34(1):12 -- 35.
7. Muhle-Karbe PS, De Baene W, Brass M. Do tasks matter in task switching?
Dissociating domain-general from context-specific brain activity. Neuroimage.
2014;99:332 -- 341.
8. Donders FC. On the speed of mental processes. Acta Psychol. 1969;30:412 -- 431.
9. Mattar MG, Cole MW, Thompson-Schill SL, Bassett DS. PLoS Comput Biol.
2015;11(12):e1004533.
10. Bullmore E, Sporns O. The economy of brain network organization. Nature
Reviews Neuroscience. 2012;13(5):336 -- 349.
11. Niven JE, Laughlin SB. Energy limitation as a selective pressure on the evolution
of sensory systems. J Exp Biol. 2008;211:1792 -- 1804.
12. Attwell D, Laughlin SB. An energy budget for signalling in the grey matter of
the brain. J Cereb Blood Flow and Metab. 2001;21:1133 -- 1145.
13. Bassett DS, Greenfield D, Meyer-Lindenberg A, Weinberger DR, Moore SW,
Bullmore ET. Efficient physical embedding of topologically complex information
processing networks in brains and computer circuits. PLoS Comput Biol.
2010;6(4):e1000748.
14. Friston K, Kilner J, Harrison L. A free energy principle for the brain. Journal of
Physiology-Paris. 2006;100(1):70 -- 87.
15. Friston K. The free-energy principle: a unified brain theory? Nature Reviews
Neuroscience. 2010;11(2):127 -- 138.
16. Moreno-Bote R, Rinzel J, Rubin N. Noise-induced alternations in an attractor
network model of perceptual bistability. Journal of neurophysiology.
2007;98(3):1125 -- 1139.
17. Tsodyks M, Pawelzik K, Markram H. Neural networks with dynamic synapses.
Neural computation. 1998;10(4):821 -- 835.
PLOS
17/21
18. Achard S, Bullmore E. Efficiency and cost of economical brain functional
networks. PLoS computational biology. 2007;3(2):e17.
19. Honey C, Sporns O, Cammoun L, Gigandet X, Thiran JP, Meuli R, et al.
Predicting human resting-state functional connectivity from structural
connectivity. Proceedings of the National Academy of Sciences.
2009;106(6):2035 -- 2040.
20. Honey CJ, Thivierge JP, Sporns O. Can structure predict function in the human
brain? Neuroimage. 2010;52(3):766 -- 776.
21. Deco G, Jirsa VK. Ongoing cortical activity at rest: criticality, multistability, and
ghost attractors. J Neurosci. 2012;32(10):3366 -- 3375.
22. Goni J, van den Heuvel MP, Avena-Koenigsberger A, Velez de Mendizabal N,
Betzel RF, Griffa A, et al. Resting-brain functional connectivity predicted by
analytic measures of network communication. Proc Natl Acad Sci U S A.
2014;111(2):833 -- 838.
23. Becker CO, Pequito S, Pappas GJ, Miller MB, Grafton ST, Bassett DS, et al.
Accurately Predicting Functional Connectivity from Diffusion Imaging.
arXiv;1512:02602.
24. Gu S, Pasqualetti F, Cieslak M, Telesford QK, Yu AB, Kahn AE, et al.
Controllability of structural brain networks. Nat Commun. 2015;6:8414.
25. Muldoon SF, Pasqualetti F, Gu S, Cieslak M, Grafton ST, Vettel JM, et al.
Stimulation-based control of dynamic brain networks. arXiv. 2016;In Prep.
26. Ravizza SM, Moua KC, Long D, Carter CS. The impact of context processing
deficits on task-switching performance in schizophrenia. Schizophr Res.
2010;116(2 -- 3):274 -- 279.
27. Wylie GR, Clark EA, Butler PD, Javitt DC. Schizophrenia patients show task
switching deficits consistent with N-methyl-d-aspartate system dysfunction but
not global executive deficits: implications for pathophysiology of executive
dysfunction in schizophrenia. Schizophr Bull. 2010;36(3):585 -- 594.
28. Cieslak M, Grafton ST. Local termination pattern analysis: a tool for comparing
white matter morphology. Brain Imaging Behav. 2014;8(2):292 -- 299.
29. Yeh FC, Tseng W. NTU-90: a high angular resolution brain atlas constructed by
q-space diffeomorphic reconstruction. Neuroimage. 2011;58(1):91 -- 99.
30. Dale AM, Fischl B, Sereno MI. Cortical surface-based analysis. I. Segmentation
and surface reconstruction. NeuroImage. 1999;9(2):179 -- 194.
31. Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, Wedeen VJ, et al.
Mapping the structural core of human cerebral cortex. PLoS Biology.
2008;6(7):e159.
32. Bassett DS, Brown JA, Deshpande V, Carlson JM, Grafton ST. Conserved and
variable architecture of human white matter connectivity. Neuroimage.
2011;54(2):1262 -- 1279.
33. Hermundstad AM, Bassett DS, Brown KS, Aminoff EM, Clewett D, Freeman S,
et al. Structural foundations of resting-state and task-based functional
connectivity in the human brain. Proc Natl Acad Sci U S A.
2013;110(15):6169 -- 6174.
PLOS
18/21
34. Hermundstad AM, Brown KS, Bassett DS, Aminoff EM, Frithsen A, Johnson A,
et al. Structurally-constrained relationships between cognitive states in the
human brain. PLoS Comput Biol. 2014;10(5):e1003591.
35. Biol PC. Resolving structural variability in network models and the brain.
10;(3):e1003491.
36. Small-World Propensity in Weighted, Real-World Networks. arXiv.
2015;1505:02194.
37. Sizemore A, Giusti C, Bassett DS. Classification of weighted networks through
mesoscale homological features. arXiv. 2015;1512:06457.
38. Yeh FC, Wedeen VJ, Tseng WY. Generalized Q-Sampling Imaging. Medical
Imaging, IEEE Transactions on. 2010;29(9):1626 -- 1635.
39. Tuch DS. Q-ball imaging. Magnetic Resonance in Medicine.
2004;52(6):1358 -- 1372.
40. Baird B, Smallwood J, Gorgolewski KJ, Margulies DS. Medial and lateral
networks in anterior prefrontal cortex support metacognitive ability for memory
and perception. The Journal of Neuroscience. 2013;33(42):16657 -- 16665.
41. Satterthwaite TD, Elliott MA, Gerraty RT, Ruparel K, Loughead J, Calkins ME,
et al. An improved framework for confound regression and filtering for control of
motion artifact in the preprocessing of resting-state functional connectivity data.
Neuroimage. 2013;64:240 -- 256.
42. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, et al. A pairwise
maximum entropy model accurately describes resting-state human brain networks.
Nat Commun. 2013;4:1370.
43. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, et al. Energy
landscapes of resting-state brain networks. Front Neuroinform. 2014;8:12.
44. Newman MEJ. Modularity and community structure in networks. Proceedings of
the National Academy of Sciences of the United States of America.
2006;103(23):8577 -- 8696.
45. Clauset A, Newman ME, Moore C. Finding community structure in very large
networks. Physical review E. 2004;70(6):066111.
46. Cipra BA. The Ising model is NP-complete. SIAM News. 2000;33(6):1 -- 3.
47. Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E. Equation of
state calculations by fast computing machines. The journal of chemical physics.
1953;21(6):1087 -- 1092.
48. Manning CD, Raghavan P, Schutze H, et al. Introduction to information retrieval.
vol. 1. Cambridge university press Cambridge; 2008.
49. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, et al.
Functional network organization of the human brain. Neuron. 2011;72(4):665 -- 678.
50. Evolution of brain network dynamics in neurodevelopment. arXiv. 2016;.
51. Clauset A, Shalizi CR, Newman MEJ. Power-law distributions in empirical data.
SIAM Review. 2009;51(4):661 -- 703.
PLOS
19/21
52. Y V, Clauset A. Power-law distributions in binned empirical data. Annals of
Applied Statistics. 2014;8(1):89 -- 119.
53. Crossley NA, Mechelli A, Vertes PE, Winton-Brown TT, Patel AX, Ginestet CE,
et al. Cognitive relevance of the community structure of the human brain
functional coactivation network. Proc Natl Acad Sci U S A.
2013;110(28):11583 -- 11588.
54. Crossley NA, Mechelli A, Scott J, Carletti F, Fox PT, McGuire P, et al. The
hubs of the human connectome are generally implicated in the anatomy of brain
disorders. Brain. 2014;137(Pt 8):2382 -- 2395.
55. Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard DA, Shulman GL.
A default mode of brain function. Proceedings of the National Academy of
Sciences. 2001;98(2):676 -- 682.
56. Sporns O, Tononi G, Kotter R. The human connectome: A structural description
of the human brain. PLoS Comput Biol. 2005;1(4):e42.
57. Sporns O. The human connectome: a complex network. Ann N Y Acad Sci.
2011;1224:109 -- 125.
58. Bassett DS, Yang M, Wymbs NF, Grafton ST. Learning-induced autonomy of
sensorimotor systems. Nat Neurosci. 2015;18(5):744 -- 751.
59. Sporns O. Cerebral cartography and connectomics. Philos Trans R Soc Lond B
Biol Sci. 2015;370(1668):20140173.
60. Zeki S. Introduction: cerebral cartography 1905 -- 2005. Phil Trans R Soc B.
2005;360:651 -- -652.
61. Newman MEJ, Clauset A. Structure and inference in annotated networks. arXiv.
2015;1507:04001.
62. Schirner M, Rothmeier S, Jirsa VK, McIntosh AR, Ritter P. An automated
pipeline for constructing personalized virtual brains from multimodal
neuroimaging data. Neuroimage. 2015;117:343 -- 357.
63. Ritter P, Schirner M, McIntosh AR, Jirsa VK. The virtual brain integrates
computational modeling and multimodal neuroimaging. Brain Connect.
2013;3(2):121 -- 145.
64. Bassett DS, Nelson BG, Mueller BA, Camchong J, Lim KO. Altered resting state
complexity in schizophrenia. Neuroimage. 2012;59(3):2196 -- 2207.
65. Zalesky A, Fornito A, Egan GF, Pantelis C, Bullmore ET. The relationship
between regional and inter-regional functional connectivity deficits in
schizophrenia. Hum Brain Mapp. 2012;33(11):2535 -- 2549.
66. Yu Q, Sui J, Liu J, Plis SM, Kiehl KA, Pearlson G, et al. Disrupted correlation
between low frequency power and connectivity strength of resting state brain
networks in schizophrenia. Schizophr Res. 2013;143(1):165 -- 171.
67. Honey CJ, Sporns O, Cammoun L, Gigandet X, Thiran JP, Meuli R, et al.
Predicting human resting-state functional connectivity from structural
connectivity. Proc Natl Acad Sci U S A. 2009;106(6):2035 -- 2040.
PLOS
20/21
68. Becker C, Pequito S, Pappas GJ, Miller MB, Grafton ST, Preciado V. Accurately
Predicting Functional Connectivity from Diffusion Imaging. arXiv.
2016;1512:02602.
69. Cole MW, Bassett DS, Power JD, Braver TS, Petersen SE. Intrinsic and
task-evoked network architectures of the human brain. Neuron.
2014;83(1):238 -- 251.
70. Bialek W, Cavagna A, Giardina I, Mora T, Silvestri E, Viale M, et al. Statistical
mechanics for natural flocks of birds. Proc Natl Acad Sci U S A.
2012;109(13):4786 -- 4791.
71. Giusti C, Ghrist R, Bassett DS. Two's company, three (or more) is a simplex:
Algebraic-topological tools for understanding higher-order structure in neural
data. Trends in Cognitive Sciences. 2016;Under Consideration.
72. Leen DA, Shea-Brown E. A Simple Mechanism for Beyond-Pairwise Correlations
in Integrate-and-Fire Neurons. J Math Neurosci. 2015;5(1):30.
73. Fox PT, Lancaster JL, Laird AR, Eickhoff SB. Meta-analysis in human
neuroimaging: computational modeling of large-scale databases. Annu Rev
Neurosci. 2014;37:409 -- 434.
74. Pasqualetti F, Zampieri S, Bullo F. Controllability Metrics, Limitations and
Algorithms for Complex Networks. IEEE Transactions on Control of Network
Systems. 2014;1(1).
75. Wu S, Hitchman G, Tan J, Zhao Y, Tang D, Wang L, et al. The neural dynamic
mechanisms of asymmetric switch costs in a combined Stroop-task-switching
paradigm. Sci Rep. 2015;5:10240.
76. Davidson MC, Amso D, Anderson LC, Diamond A. Development of cognitive
control and executive functions from 4 to 13 years: evidence from manipulations
of memory, inhibition, and task switching. Neuropsychologia.
2006;44(11):2037 -- 2078.
PLOS
21/21
|
1904.06712 | 1 | 1904 | 2019-04-14T15:46:37 | Bilateral symmetry strengthens the perceptual salience of figure against ground | [
"q-bio.NC"
] | Although symmetry has been discussed in terms of a major law of perceptual organization since the early conceptual efforts of the Gestalt school (Wertheimer, Metzger, Koffka and others), the first quantitative measurements testing for effects of symmetry on processes of Gestalt formation have seen the day only recently. In this study, a psychophysical rating study and a "foreground" versus "background" choice response time experiment were run with human observers to test for effects of bilateral symmetry on the perceived strength of figure against ground in triangular Kanizsa configurations. Displays with and without bilateral symmetry, identical physically specified to total contour ratio and constant local contrast intensity within and across conditions, but variable local contrast polarity and variable orientation in the plane were presented in a random order to human observers. Configurations with bilateral symmetry produced significantly stronger figure against ground percepts reflected by greater subjective magnitudes and consistently higher percentages of "foreground" judgments accompanied by significantly shorter response times. These effects of symmetry depend neither on the orientation of the axis of symmetry, nor on the contrast polarity of the physical inducers. It is concluded that bilateral symmetry, irrespective of orientation, significantly contributes to the, largely sign invariant, visual mechanisms of shape segregation that determine the salience of figure against ground in perceptually ambiguous image configurations. | q-bio.NC | q-bio | Bilateral symmetry strengthens the perceptual
salience of figure against ground
Birgitta Dresp-Langley 1*
1 Centre National de la Recherche Scientifique, ICube Lab UMR 7357 CNRS and Strasbourg University,
Strasbourg, FRANCE; [email protected]
* Correspondence: [email protected]
configurations. Displays with
Abstract: Although symmetry has been discussed in terms of a major law of perceptual
organization since the early conceptual efforts of the Gestalt school (Wertheimer, Metzger, Koffka
and others), the first quantitative measurements testing for effects of symmetry on processes of
Gestalt formation have seen the day only recently. In this study, a psychophysical rating study and
a "foreground"-"background" choice response time experiment were run with human observers to
test for effects of bilateral symmetry on the perceived strength of figure-ground in triangular
Kanizsa
identical
physically-specified-to-total contour ratio and constant local contrast intensity within and across
conditions, but variable local contrast polarity and variable orientation in the plane were presented
in a random order to human observers. Configurations with bilateral symmetry produced
significantly stronger figure-ground percepts reflected by greater subjective magnitudes and
consistently higher percentages of "foreground" judgments accompanied by significantly shorter
response times. These effects of symmetry depend neither on the orientation of the axis of
symmetry, nor on the contrast polarity of the physical inducers. It is concluded that bilateral
symmetry, irrespective of orientation, significantly contributes to the, largely sign-invariant, visual
mechanisms of figure-ground segregation that determine the salience of figure-ground in
perceptually ambiguous configurations.
and without bilateral
symmetry,
Keywords: bilateral symmetry; physically-specified-to-total contour ratio; Kanizsa's triangle;
figure-ground segregation; law of good Gestalt; subjective magnitude estimation; choice response
time
1. Introduction
The Gestalt psychologists Max Wertheimer and Wolfgang Metzger [1, 2] formulated and discussed
"laws of perception" to predict how perceptual grouping operates under specific conditions of visual
configuration. Their important work was translated into the English language in 2012 and 2009
respectively by Lothar Spillmann and colleagues [1, 2], making this important early conceptual work
available to a broader audience. In physical science, a law is a prediction that can be proven true and,
ideally, the limits of which can also be clearly determined. In perceptual science, the Gestalt laws are
used to express principles or conditions of visual configuration to explain why we see the world as
we do. It is argued that specific principles of, or conditions for, "good Gestalt" need to be fulfilled to
enable what is called perceptual grouping, i.e. a perceptual solution that yields the most plausible
interpretation of a given physical configuration. Since all physical stimuli are by nature ambiguous
to our perception, they need to be interpreted by the brain to produce coherent and unambiguous
percepts that allow us to act on the physical world effectively. The "Law of Symmetry" is a major
Gestalt law. It predicts that visual elements that are symmetrical would be more likely to form a
group, i.e. to be perceived as a "good Gestalt", in comparison with asymmetrical ones. Visual
symmetry has, indeed, proven a determining factor in shape perception [3-6]. In particular, vertical
mirror symmetry has proven an important cue to shape extraction from abstract, non-familiar visual
elements presented in conditions of heightened ambiguïty (noise). Across different noise levels,
symmetric elements form perceptually more salient shapes than asymmetric ones and are, therefore,
more readily detected [5].
The Italian Gestalt psychologist Gaetano Kanizsa [7] discussed a series of ambiguous planar
configurations that give rise to powerful figure-ground percepts, with apparent shapes emerging in
the foreground, delineated by contours that are perceptually completed beyond physically specified
contrast edges. The Gestalt school and Kanizsa himself considered these phenomena as marginal
cases of perception ("margini quasi-percettivi"), and argued that these latter provide insight into the
fundamentals of perceptual organization because they put underlying processes to the test under
extreme conditions at the capacity limits of the perceptual system. Later-on, the figures seen in such
configurations were termed "illusory" by cognitive psychologists; the Gestalt psychologists
themselves never used this term, which is, of course, misleading. An illusion, by definition, cannot
be verified by independent observation - it only exists in the mind of the person experiencing it. The
phenomena described by Kanizsa have clearly defined physical correlates, with measurable
systematic effects on perception. One of these configurations is the famous Kanizsa triangle (Figure
1). The Kanizsa figures have been studied extensively to single out factors of physical variation that
affect the subjective brightness or darkness of the figures and/or the figure contours. The results
from these studies, based on a variety of different experimental measures, are reviewed in sufficient
detail elsewhere [8-13]. They are not the object of this study here. Here, we measured the perceptual
salience of the figure-ground percept irrespective of the relative darkness, brightness, or clarity of
either the induced surfaces or their boundaries, as is made perfectly clear in the instructions given to
subjects. As raised previously by others, the response criterion of the subjects in judgment tasks
using this type of ambiguous figure here [8] is directly dependent on the semantic precision of
instructions given. Formulating these latter appropriately is to make sure the perceptual
phenomenon under study, not a related one, is reflected by the psychophysical data.
The influence of variations in the intensity of local luminance contrast of the physical inducers
that produce a perceptual filling in [14] of bright or dark surfaces, leading to figure-ground percepts
at the centre of configurations, was demonstrated for the first time by Heinemann's pioneering
studies [15]. These were extended later on by others [12] to various configurations of the Kanizsa
type, where observers had to adjust the luminance of the central figure region until it matched the
phenomenal appearance of the general background. This is equivalent to a cancellation of the
phenomenal appearance of figure and ground. The configurations produce filling-in consistent with
classic simultaneous contrast effects where the central figure appears darker than the general
background when surrounded by phenomenally white inducers, and brighter when surrounded by
phenomenally black inducers. The perceptual salience of the central figures increases consistently
with the luminance contrast intensity of the inducers, up to some optimal limit. When that optimal
limit is reached and the contrast intensity of inducers increases further, the figure intensity is
diminished again and may be annulled completely at the highest physical contrast levels [15]. The
simultaneous contrast filling-in that leads to the figure-ground percepts in Kanizsa configurations is
thus predicted by specific physical parameters of the inducers, which were all controlled
experimentally to keep them constant across symmetry conditions in this study here.
When the physical contrast intensity of the configurational elements is optimal [12-15] and not
varied between displays, the next most important physical parameter that straightforwardly
determines the salience or subjective strength of the figure-ground percept in Kanizsa configurations
is the physically specified-to-total contour ratio, or support ratio. This was proven in a series of
experiments by Shipley and Kellmann [11] using subjective magnitude estimation, a classic
psychophysical rating procedure similar to the one applied in this study. Here, the Kanizsa triangle
is exploited to probe for effects of symmetry on figure-ground from occlusion cues. The Kanizsa
triangle is one of the most cited examples of a specific class of Gestalt configurations where
perceived surface depth arises from the local occlusion cues. In this specific shape class,
figure-ground results directly from a process of surface completion through boundary interpolation
across the physically specified edge contours of the inducers providing the occlusion cues [16-20].
Occluded object completion thus reflects the workings of fundamental visual mechanisms for
recovering object percepts from fragmented input, and the ability of human perception to read
structure into an apparently chaotic physical world [21, 22]. The functional interactions between
configurational symmetry and other structural factors in this important perceptual process are still
unknown. The dependency of figure-ground salience on the support ratio (Figure 1), a scale
invariant metric, is associated with the ecologically desirable consequence that perceptual salience
will not change with variations in viewing distance [11].
At constant physical contrast intensity of configurations with a constant support ratio, the contrast
polarity of inducers, i.e. whether they are dark on lighter backgrounds, bright on darker
backgrounds or a mixture of both on a medium grey background, does not affect the salience or
subjective strength of the resulting figure-ground percept, provided the contrast polarity is
homogenous within each configurational element [12, 23, 24, 25, 28, 29]. When contrast polarity is
not homogenous within the configurational elements, then, and only then, may the perceptual
salience of the figure-ground percept be reduced [26], or not [8] depending on task instructions.
Perceptual figure -- ground organization is thus determined by visual mechanisms that integrate the
contrast intensity and the spatial information carried by the configurational elements while mostly
discarding information on contrast polarity. This is predicted by sign-invariant models based on the
functional properties of cortical neurons of the complex type, which are orientation selective but
insensitive to contrast polarity [14, 23, 24, 25, 29, 35]. Such sign-invariant visual mechanisms have
the ecologically desirable consequence that the simplest plausible representation of figure and
ground can be achieved when the signal input from local contrast regions is ambiguous.
The classic version of the Kanizsa triangle is a configuration with perfect vertical bilateral, or
mirror symmetry (Figure 1, top, and bottom left). Whether physical display variations producing
asymmetric configurations (Figure 1, bottom right) would affect figure-ground salience in this
specific case is not known. The motivation of this study was, therefore, to test whether symmetry
contributes to figure-ground strength in one of the most classic Gestalt configurations where surface
depth results from visual interpolation across fragments. Two variations of the Kanizsa triangle with
identical support ratio as defined by Shipley and Kellmann [11] and identical triangular area size
were generated; one with perfect bilateral symmetry (Figure 1, bottom left), the other asymmetric
(Figure 1, bottom right). To test for possible interactions between symmetry and the orientation of
the configurations in the plane, presentations were varied and bilateral symmetry was not always
vertical but could be vertical or horizontal, in a random order. In the light of earlier findings, with
abstract shapes presented in noisy contexts [5], vertical mirror symmetry significantly increased the
probability that a shape was seen as figure against the ground. Thus, the bilateral symmetry of
vertical orientations may also generate stronger effects on the figure-ground salience of surfaces
completed by interpolation. The physical inducers, either dark on grey, light on grey, or light and
dark on grey, displayed variations in contrast polarity across inducers, but never within, in both
types of configuration, symmetric and asymmetric. In the light of previous findings, these variations
should not affect figure-ground strength, given that the polarity of contrast was always homogenous
within inducers in the different configurations [8, 10, 12, 26, 29].
In a first experiment, psychophysical magnitude estimation was used to measure the salience, or
subjective strength, of the figure-ground percepts in the different conditions. In a second
experiment, a choice response time was run, with a selected set of configurations, to test whether
more salient figure-ground percepts consistently produce, as would be expected, shorter response
times with consistent "foreground" response probabilities.
Figure 1. Variations of the Kanizsa triangle with phenomenally black inducers on a grey background.
The subjective strength of the triangular figure-ground percept emerging in the center of the
configuration critically depends on the physically specified-to-total contour, or support ratio.
Stronger surface percepts are produced by higher support ratios (top). To test for effects of symmetry
on the salience of figure-ground, configurations with bilateral symmetry (bottom left) and without
(bottom right) were generated. All the configurations had identical support ratio and therefore
identical area size. All physically specified elements in the configurations (inducers) were of identical
size and contrast intensity.
2. Materials and Methods
Triangular Kanizsa configurations with and without bilateral symmetry (Figure 1 bottom left and
right respectively, for illustration), identical support ratio and surface area, variable orientation
(vertical base down, as in Figure 1 bottom left, vertical base up, sideways base left, sideways base
right) and variable inducer polarity (three black inducers on grey or '- - -', three inducers white on
grey or '+++', two black and one white inducer on grey or '- - +', two white and one black inducer on
grey or '++ -') were presented in random order to ten human observers in a single-presentation-per-
figure subjective magnitude estimation (rating) experiment with moduli. Four of these configurations
with a single orientation (vertical base down, as in Figure 1 bottom left) and uniformly positive (+++)
or uniformly negative (---) contrast polarity, two with vertical symmetry and two without, were
presented to six additional human observers in a repeated measures (four presentations per
configuration) choice response time experiment with two additional control stimuli (triangles with
minimally visible line contours ("ghosts") only, no surface contrast).
Stimuli
The image configurations were computer generated using an HP Zbook 15 G2 Mobile
Workstation equipped with a 4th generation Intel Core i7-6700 processor and an NVIDIA Quadro
K5100 graphic card.
Table 1. Figure dimensions in centimeters (cm) with the overall support ratio and surface area as a
function of configuration (symmetric versus asymmetric). The symmetry factor only varies
systematically between configurations, the shape interpretation ("triangle") is the same and so are all
relevant physical parameters.
Symmetric Asymmetric
Triangle base (b)
9 cm
13 cm
Triangle side 1
Triangle side 2
Triangle height (h)
Triangle surface area (1/2bxh)
Physical inducer radius
Support ratio
12 cm
12 cm
11 cm
49.5 cm
2 cm
0.36
11 cm
9 cm
7.62 cm
49.5 cm
2 cm
0.36
Configurational dimensions in terms of size (in pixels and in cm on screen) of triangle base and
triangle sides, the inducer radius, which was identical in all the configurations, and the overall
physical-to-total contour ratio, also identical in all the configurations, are summarized in Table 1.
Luminance values of the different configural elements were determined photometrically using an
OPTICAL photometer (Cambridge Research Systems). The ADOBE RGB coordinates of the
phenomenally grey background (RGB: 140, 140, 140) yield a background luminance of 55 cd/m2. The
phenomenally black inducers (RGB: 5, 5, 5) a luminance of 4 cd/m2, and the phenomenally white
inducers (RGB: 240, 240, 240) a luminance of 98 cd/m2. The moduli from the sujective rating task (RGB:
135, 135, 135 for the phenomenally darker ones and RGB: 145, 145, 145 for the phenomenally lighter
ones) had a luminance of 52 cd/m2 or 58 cd/m2 respectively. The line contour control configurations
(RGB: 120, 120, 120) from the choice response time task had a luminance of 48 cd/m2. The physically
specified contrast intensities with positive and negative signs may be calculated using the Weber
Contrast (Weber Ratio, W) formula:
W = (Lconfig-Lbackground)/Lbackground (1).
As a consequence, we have a positive W of +0.92 for the phenomenally white inducers, a
negative W of -.78 for the phenomenally black inducers, a positive and a negative W of +.09 and -.09
respectively for the minimal-contrast moduli from the subjective rating task, and a negative W of -.13
for the minimal contrast line contour control configurations from the choice response time
experiment.
Presentation of configurations
The configurations were presented in random order on the screen of the HP Zbook 15 G2
Mobile Workstation, which has a pixel resolution of 1920x1080 and a 60 Hz refresh rate. Random
selection, presentation, and response coding were computer controlled using Python for Windows.
The duration of presentation of each single configuration was observer controlled in both
experiments, a subsequent presentation always initiated 800 milliseconds after the observer had
typed his/her response on
the
single-presentation subjective magnitude estimation (rating) task, with the different variations in
orientation and in local contrast polarity, are shown in Figure 2 a and b, for illustration. Illustrations
of the 6 configurations from the repeated measures choice response time task are shown in Figure 3.
the computer keyboard. The 32 configurations
from
Experimental procedure
Subjects were seated in front of the workstation at a distance of about 90 cm from the screen in a
semi-dark room. In the subjective magnitude estimation task, they were shown a set of moduli
consisting of minimal-contrast triangular surfaces of the same spatial dimensions as the symmetric
and asymmetric triangular centers of the test configurations. These moduli are shown in Figure 2c for
illustration. Subjects were told to associate the phenomenal strength of the moduli with a rating score
of '11'. It was then made clear to them that they would be shown different triangular configurations,
with black and white patches around them. They were then asked to rate the subjective strength of
the figure-ground percept at the centre of the test configurations, in terms of the strength of the
segregation into foreground and background, regardless of the direction of the perceived contrast (i.
e. subjectively darker or subjectively lighter), by a number between '0' and '10', bearing in mind that
the highest number was to reflect a figure salience closest to that of the real-contrast moduli they had
seen just before. Each of the 16 asymmetric and the 16 symmetric configurations (Figure 2a and b,
respectively, for illustration) was presented only once to each of the subjects in a single random
order session. In the choice response time task, subjects were asked to judge as swiftly as possible
whether the triangle displayed on the screen seemed to stand out as foreground against the grey
general background, or to lie behind the general grey background. In this experiment, the outlined
triangular shape control configurations (Figure 3, on the right) were presented at a minimal, just
visible negative line contrast intensity and no surface contrast. This renders a highly ambiguous (one
subject mentioned "ghost-like") appearance on the screen with no clear figure-ground assignment.
Each of the six configurations (Figure 3, for illustration) was shown four times, in random order, to
each of six subjects in a single individual session.
Subjects
Ten individuals (six men, four women) between 20 and 31 years old, all of them with normal or
corrected to normal vision, participated in the subjective magnitude estimation experiment. Six
further individuals (five men, one woman), also young and with normal or corrected to normal
vision, participated
time experiment. Participants were mostly
undergraduates involved in medical or language studies. None of them was familiar with the
configurations presented to them, and all of them were naïve to the purpose of the study. The
experiments were conducted in accordance with the Declaration of Helsinki (1964) and in full
conformity with the author's host institution's (CNRS) ethical standards committee. Informed
consent was obtained from each of the participants.
the choice response
in
a)
b)
c)
Figure 2. The configurations from the subjective rating experiment (a) The asymmetric Kanizsa
triangles, with varying orientation and inducer contrast polarities (b) The Kanizsa triangles with
axial symmetry c) the moduli for benchmarking the subjective rating scale (0-10). Subjects were told to
associate the moduli with a figure-ground strength rating of '11', regardless of the direction of the
perceived contrast.
Figure 3. The test and control configurations ("ghosts") from the choice response time experiment. Six
subjects were asked to judge as swiftly as possible whether the triangle displayed on the screen
seemed to stand out as foreground against the grey general background, or to lie behind the general
grey background.
Data analysis
The data from the subjective rating experiment, with a Cartesian design plan written in terms of
Subject10xSymmetry2xOrientation4xPolarity4, produced a total of 320 subjective ratings. These data
were fed into a Three-Way ANOVA. Means, standard errors, effect sizes, and F statistics with
probability limits were determined. The data from the choice response time experiment, with a
Cartesian design plan written in terms of Subject6xSymmetry2xPolarity3xRepeatedMeasures4,
produced a total of 144 choice data and a total of 144 response times. In the experimental design
plan, the control configuration represents the third modality of the "polarity" factor, with the three
factor levels "positive" or '+++', "negative" or '- - -', and "control". The response times were fed into a
Two-Way Repeated Measures ANOVA with individual data averaged over the four levels of the
repetition factor R4 and without the third level of the "polarity" factor, i.e. the analysis plan therefore
reads Subject6xSymmetry2xPolarity2.
3. Results
The results of the analyses on the data from the subjective rating experiment and the choice
response time experiment are shown here below in the form of graphs and tables.
3.1. Subjective magnitude estimation task
The individual data from this experiment are made available in Table S1 of the Supplementary
Materials section. The data show a good consistency between participants, within and across
conditions, with a moderate amount of inter-individual variability. The main effect of symmetry on
the subjective magnitude of the figure-ground percept in the Kanizsa configuration is highlighted by
the two graphs in Figure 4. The average magnitude of figure salience in terms of average subjective
ratings produced by symmetric and asymmetric configurations is plotted as a function of the
orientation of the configurations in the plane (Figure 4, top), and as a function of the contrast polarity
of the inducers (Figure 4, bottom). While the subjective ratings display no major variations with
either orientation of the configurations or the contrast polarity of the inducers, they are consistently
and noticeably affected by lack of symmetry. Subjective ratings are found to be markedly stronger
for all the configurations with bilateral symmetry, irrespective of orientation and/or contrast
polarity. This effect is highlighted further by the statistics shown in Table 2 here below, which
summarizes the observations from Figure 4 in terms of results from the Three-Way ANOVA on the
subjective ratings of the ten subjects. The effect of symmetry is statistically significant, the effects of
orientation and inducer polarity are not.
Table 2. Three-Way ANOVA results with the means (average subjective magnitudes), standard
errors (SEM), and F statistics for effects of main factors and their interactions from the analysis of the
subjective rating data
Level
Mean SEM
F
3.8
6.1
4.7
5.3
4.8
4.7
4.8
5.2
4.9
4.7
_
_
_
F(3, 319)=1.24; NS
0.15 F (1, 319)=108.8; p<.001
0.13
0.21
0.22
0.20
0.19
0.20
0.23
0.21
0.22
F(3, 319)=1.17; NS
_
_
_
F(3, 319)=1.83; NS
F(3, 319)=0.41; NS
F(9, 319)=0.67; NS
Factor
Symmetry (S2)
Polarity (P4)
asymmetric
symmetric
- - -
+++
- - +
++ -
Orientation (O4)
vertical base bottom
vertical base top
sideways base left
sideways base right
Symmetry x Polarity
Symmetry x Orientation
Polarity x Orientation
interaction
interaction
interaction
Choice response time task
The raw data from this experiment are made available in Table S2 of the Supplementary
Materials section. The perceptual judgments from the choice task, shown here below in Table 3 in
terms of the percentage of "foreground" responses as a function of configuration (asymmetric vs
symmetric) and the contrast polarity of the inducers (negative vs positive vs control), expressed in terms
of the phenomenal appearance of the inducers here, display a consistent majority of "foreground"
responses in the ambiguous Kanizsa configurations, with noticeably higher percentages of
"foreground" in the configurations with bilateral symmetry, irrespective of inducer appearance (or
polarity). In response to the control configurations, the percentages of "foreground" responses show
no such clear trend. This is explained by the fact that the outlined shape control configurations
(Figure 3, on the right) were presented at a minimal, just visible negative line contrast intensity,
which made them particularly ambiguous with respect to figure-ground organization in the plane.
The outlines are not perceived as clearly belonging to a specific depth level, which is reflected in the
results here by a near random distribution of "foreground" and "background" responses. This
suggests that the outlined controls without surface contrast did not produce, as could be expected,
salient figure-ground percepts.
Average choice response times were plotted as a function of configuration and inducer polarity, as
shown here in Figure 5. The graphs show a consistent and systematic effect of symmetry on response
times. Subjects respond markedly faster to configurations with axial symmetry, irrespective of
whether the inducers are phenomenally black (negative contrast polarity) or white (positive contrast
polarity). This effect is highlighted further by the statistics shown in Table 4 here below, which
summarizes the observations from Figure 5 in terms of results from the Two-Way Repeated
Measures ANOVA on the response times of the six subjects. The effect of symmetry is statistically
significant, the effect of inducer polarity is not. The third level of the "polarity" factor here, i.e. the
control configuration, was not included in the design plan for this ANOVA.
Figure 4. Average magnitudes of figure-ground in terms of average subjective ratings with bars
indicating +/- the standard error of the mean. Effects produced by symmetric and asymmetric
configurations are plotted as a function of the orientation of configurations in the plane (top), and as
a function of inducer polarity (bottom).
Table 3. Percentage of "foreground" responses from the choice response time task as a function
of configuration and inducer contrast polarity
Asymmetric Symmetric
White inducers
Black inducers
Control
88 %
75 %
70 %
98 %
92 %
55 %
Figure 5. Average choice response times with bars indicating +/- the standard error of the mean as a function of
configuration and inducer polarity.
Table 4. Two-Way Repeated Measures ANOVA results with the means (in milliseconds), standard
errors (SEM), and F statistics for effects of main factors and their interactions from the analysis of the
choice response times
Factor
Symmetry (S2)
Polarity (P2)
Level
asymmetric
symmetric
- - -
+++
Symmetry x Polarity
interaction
_
4. Discussion
Mean SEM
1518
900
1269
1150
73
65
64
62
_
F
F(1, 23)=36.69; p<.01
F(1, 23)=1.74; NS
F(1, 23)=1.21; NS
Although symmetry has been discussed in terms of a major grouping principle or law of good
Gestalt since Wertheimer and Metzger [1, 2], the specific effects symmetry may produce on feature
grouping, figure-ground segregation, visual discrimination, or time to respond to visual
configurations have become subject to systematic quantitative investigation in perceptual science
only recently. In the case of visual perception, symmetry may be conceived as a geometric property
that yields configurational simplicity and, therefore, represents an ecological advantage for
information processing [30]. Symmetry may also be seen as a perceptual feature that attracts
attention, enhances configural salience, and facilitates grouping [5, 31-33].
It is found that bilateral configurational symmetry, i.e. mirror symmetry within the whole
configuration, strengthens the perceptual salience of figure against ground in triangular Kanizsa
configurations. The results from the magnitude estimation (rating) experiment clearly show that the
subjective strength of the foreground at the center of the configurations is significantly higher when
the configurations have bilateral symmetry. This holds for triangular configurations when mirror
symmetric configurations are compared to asymmetric configurations with the same number of
inducers, and the same support ratio as defined by Shipley and Kellman [11]. This symmetry effect
can be exploited to further quantify critical interactions between occlusion-based surface properties,
symmetry, and figure-ground salience. Variations in symmetry could be tested against variations in
the support ratio in a first instance. Figure-ground segregation from interpolation is an early-stage
process in perceptual grouping [16, 44, 45], as is symmetry detection [39, 43]. In particular, as shown
by Erlikhman and Kellmann [44, 45], the human perceptual system uses critical spatial cues of local
position and alignment within a restricted spatiotemporal window (~165 msec) for the rapid
extraction of co-oriented edge fragments from the visual input. As predicted by the Gestalt Law of
Good Continuation, the fragments then connect by known neural interpolation processes [16, 25, 35],
producing larger surfaces that will stand out as figures against ground. The results from the
experiments here show that symmetry contributes to this early process of perceptual organization.
The results suggest no influence of the orientation (vertical versus horizontal) of the axis of
symmetry on the salience of the figure-ground percepts. Vertical and horizontal mirror symmetry
produced equally strong phenomenal salience of figure-ground. Since Mach [33], it is suggested that
symmetry around the vertical axis may be more effectively processed by the visual system than
symmetry about the horizontal, or any other, axis in the plane. Some studies have confirmed this
prediction [5]. However, more recent reviews indicate that there may be no systematic functional
advantage of vertical symmetry [34]. Effects of axis of symmetry on perception may be dependent on
what Bertamini termed "objectness" [31], i.e. whether the cognitive interpretation of the visual shape
changes with translational or rotational changes of the latter. Psychophysical data on shape
perception [43], using radial frequency patterns and other objects, indeed suggest that variations in
the location and orientation of relevant (with respect to the perceptual task) object features may
generate effects of axis of symmetry. In the two experiments here, relevant perceptual features
within and across objects (symmetric verus asymmetric) can be considered invariant, since there was
no effect of contrast polarity and no interaction between contrast polarity and symmetry. This could
explain why the axis of symmetry had no effect here either. Also, earlier psychophysical studies
have shown that bilateral symmetry is significantly more salient within objects, significantly less
between objects [39, 40]. The layout characteristics, including symmetry, of complex figure-ground
solutions are more easily processed within single perceptual objects [40]. Symmetry detection
becomes harder with complex shape configurations where other factors, such as positional
uncertainty or convexity, interact with the symmetry factor, especially when the psychophysical task
requires comparing across objects. It may be that vertical symmetry generates a measurable
advantage for perception only under such conditions.
Figure 6. An example of a configuration where effects of symmetry on visual perception cannot be tested
independently from possible effects of shape interpretation. In the square version of the Kanizsa figure,
breaking the symmetry of the classic square configuration (images on left) inevitably requires breaking the
perpendicularity of the shape borders. This inevitably results in a new and qualitatively different shape
geometry and shape interpretation. In other words, shape interpretation then becomes a confusion factor. In
this case here a "shard" with qualitatively different 3D-like shape properties emerges (images in middle and on
left). When shape orientation changes, the percept changes, again, qualitatively and a new visual object emerges
(cf. on the importance of "objectness", see Bertamini [31]). All configurations here above have roughly
equivalent, albeit not strictly identical, support ratio and central area size. Variations in inducer texture, figure
orientation, and background intensity are presented here for illustration only.
The results from the choice response time task show a consistently higher percentage of
"foreground" responses to the symmetric configurations, which is accompanied by significantly
shorter response times. Bilateral symmetry, therefore, represents a measurable functional advantage
in the perceptual processing of figure-ground. Variations in the contrast polarity of the inducing
elements had no marked effect on either the subjective strength of the figure-ground percepts or the
response times. This observation is consistent with results from earlier work with similar
configurations where symmetry was not varied [9, 12, 29], and predicted by non-linear neural
models of figure-ground based on the long-range integration of antagonistic brightness signals [12,
13, 14, 23, 24]. Interestingly, when inducers of both positive and negative contrast signs, i.e.
phenomenally white and black inducers, are present in the same configuration, the latter may be
perceived as phenomenally asymmetric with respect to brightness. The perceptual system, however,
is not influenced by the symmetry/asymmetry in contrast signals, only by geometrically defined
symmetry. Although this study here was not specifically aimed at singling out the hierarchal level of
perceptual processing at which the symmetry effect arises, it is unlikely that conscious processing
was involved here. After the experiments, subjects were asked whether they had noticed anything in
particular in the configurations or used a particular strategy to respond. Most of them stated that
"some were the same, some were different", or "some had white parts, some had black parts, some
had both", but none of them was able to specify any particular structural difference or response
strategy. We may therefore infer that subjects were not immediately aware of the systematic
variations in symmetry.
Bilateral symmetry is identified as a key prior for three-dimensional shape perception in humans
[41, 42]. The perceptual integration of symmetry in this process does not necessarily happen
consciously and, as explained above, may vary with shape complexity and shape interpretation [40]
without subjects being aware of it. Therefore, configurational complexity and shape interpretation
need to be controlled for to single out the effects of symmetry per se on any particular aspect of
perceptual organization. This is clarified further by some of the additional illustrations in Figure 6,
showing configurations where the manipulation of symmetry inevitably implies changing also the
complexity of the configuration as a whole, and the resulting shape interpretation. In the square
version of the Kanizsa figure, breaking the configurational symmetry inevitably requires changing
the shape borders. This produces a new, far more complex, qualitatively different shape geometry
leading to a radically different shape interpretation. The Kanizsa square is therefore ill-suited for
singling out effects of symmetry without ambivalence. The advantage of the triangular
configuration used in this study here is that the geometric transformations needed to manipulate
mirror symmetry affect neither the structural complexity of the configurations, nor the resulting
shape interpretation: with or without bilateral symmetry, the perceptual solution is always and only
a triangle.
The new effect found here, where symmetry strengthens the figure-ground salience of surfaces
from occlusion cues, which form through visual spatial interpolation across fragments within a
narrow temporal window of processing [44, 45], fully expresses the adaptive logic of visual
preference for symmetry. Symmetry enables perception-based decision making, and survival
relevant responses to symmetry or lack thereof can be found in animal species [37]. These highlight
the wider biological significance of symmetry as a visual signal. The early Gestalt theories
intuitively captured this fundamental importance in a large body of observations on phenomena of
human perceptual organization. Their intuitions were astute, pointing towards functional aspects of
symmetry which perception science has only just begun to quantify and predict.
Funding: This research received no external funding.
Acknowledgments: The experiments were materially supported by the CNRS. The contribution of Cyrielle
Giger-Dechavanne, who helped recruiting subjects for the experiments, is gratefully acknowledged.
Conflicts of Interest: The author declares no conflict of interest
References
1. Wertheimer, M. Perceived Motion and Figural Organization, English trans. L. Spillmann, M. Wertheimer, K.
W. Watkins, S. Lehar and V. Sarris MIT Press: Cambridge, 2012.
2. Metzger, W. Gesetze des Sehens, English trans. L. Spillmann Laws of Seeing MIT Press: Cambridge, USA
2009.
3. Deregowski, J. B. Symmetry, Gestalt and information theory. Quarterly Journal of Experimental Psychology
1991, 23, 381-385.
4. Gerbino, W.; Zhang, L. Visual orientation and symmetry detection under affine transformations. Bulletin of
the Psychonomic Society 1991, 29, 480.
5. Machilsen, B., Pauwels, M.; Wagemans, J. The role of vertical mirror symmetry in visual shape perception.
6.
Journal of Vision 2009, 9(11).
Li, Y.; Sawada, T.; Shi, Y.; Steinman, R.M.; Pizlo, Z. (2013). Symmetry is the sine qua non of shape. In: S.
Dickinson and Z. Pizlo (Eds.), Shape perception in human and computer vision. London: Springer 2013,
pp.21-40.
7. Kanizsa, G. Organization in vision: Essays on Gestalt perception; Praeger: New York, USA, 1979.
8. Victor, JD; Conte, M. Illusory contour strength does not depend on the dynamics or relative phase of the
9.
inducers. Vision Research 2000, 40, 3475 -- 3483.
Spillmann, L.; Dresp, B. Phenomena of illusory form: Can we bridge the gap between levels of
explanation? Perception 1995, 24, 1333 -- 1364.
10. Dresp, B. On "illusory" contours and their functional significance. Curr Psychol Cogn 1997, 16, 489 -- 518.
11. Shipley, T. F.; Kellman, P. J. Strength of visual interpolation depends on the ratio of physically specified-
to-total edge length. Perception & Psychophysics 1992, 52, 97-106.
12. Dresp, B; Fischer, S. Asymmetrical contrast effects induced by luminance and color configurations.
Perception & Psychophysics 2001, 63(7), 1262-1270.
13. Reid, R. C.; Shapley, R. M. (1988). Brightness induction by local contrast and the spatial dependence of
assimilation. Vision Research 1988, 28, 115-132.
14. Gerrits, H. J. M.; Vendrik, A. J. Simultaneous contrast, filling-in process, and information processing in
man's visual system. Experimental Brain Research 1970, 11, 411 -- 430.
15. Heinemann, E. G. Simultaneous brightness induction as a function of inducing and test-field luminance.
Journal of Experimental Psychology 1955, 50, 89 -- 96.
16. von der Heydt, R. Figure -- ground organization and the emergence of proto-objects in the visual cortex.
Front. Psychol. 2015, 6, 1695. Special issue on 'Perceptual Grouping-the State of the Art'.
17. Kellman, P.J.; Yin, C.; Shipley, T.F. A common mechanism for illusory and occluded object completion.
Journal of Experimental Psychology: Human Perception & Performance 1998, 24(3), 859-869.
18. Kellman, P.J. (2002). Vision - occlusion, 'illusory' contours and filling-in. In Encyclopedia of Cognitive
Science, Oxford, UK: Nature Publishing Group.
19. Yin, C.; Kellman, P.J.; Shipley, T.F. Surface completion complements boundary interpolation. Perception,
Special Issue on surface appearance 1997, 26, 1459-1479.
20. Yin, C., Kellman, P.J.; Shipley, T.F. Surface integration influences depth discrimination. Vision Research
2000, 40(15), 1969-1978.
21. Kellman, P.J. From chaos to coherence: How the visual system recovers objects. Psychological Science
Agenda of the American Psychological Association, 1997, 10(4), 8-9.
22. Dresp-Langley, B., Grossberg, S.; Reeves, A. Editorial: Perceptual Grouping -- The State of The Art. Front.
Psychol. 2017, 8, 67. Special issue on 'Perceptual Grouping-the State of the Art'.
23. Shapley, R.; Gordon, J. Nonlinearity in the perception of form. Perception and Psychophysics 1985, 37, 84 -- 88.
24. Shapley, R.; Reid, R. C. Contrast and assimilation in the perception of brightness. Proceedings of the
National Academy of Science 1985, 82, 5983 -- 5986.
25. Grossberg, S. 3D vision and figure-ground separation by visual cortex. Perception and Psychophysics 1994,
55, 48-120.
26. He, Z. J.; Ooi, T. L. Luminance contrast polarity: a constraint for occlusion and illusory surface
perception. Investigative Ophthalmology and Visual Science 1998, 39 (Suppl.), 848.
27. Nakayama, K.; Shimojo, S.; Ramachandran, V. S. Transparency, relation to depth, subjective contours,
luminance and neon color spreading. Perception 1990, 19, 497 -- 513.
28. Guibal, C. R. C.; Dresp, B. Interaction of color and geometric cues in depth perception: When does red
mean near? Psychological Research 2004, 10: 167 -- 178.
29. Dresp-Langley, B.; Grossberg, S. Neural computation of surface border ownership and relative surface
depth from ambiguous contrast inputs. Frontiers in Psychology 2016, 7, 1102. Special issue on 'Perceptual
Grouping-the State of the Art'.
30. Hu, Q.; Victor, J.D. Two-Dimensional Hermite Filters Simplify the Description of High-Order Statistics of
Natural Images. Symmetry 2016, 8, 98.
31. Bertamini, M; Griffin, L. Symmetry in Vision, MDPI: Basel, Switzerland, 2017, pp. 1-198.
32. Gillam, B. Redundant Symmetry Influences Perceptual Grouping (as Measured by Rotational Linkage).
Symmetry 2017, 9, 67.
33. Mach, E. On Symmetry, In Popular Scientific Lectures, 1893; Lasalle: Open Court Publishing.
34. Wagemans, J. Characteristics and models of human symmetry detection. Trends in Cognitive Sciences, 1997
9, 346-352.
35. Grossberg, S. Cortical dynamics of figure-ground separation in response to 2D pictures, and 3D scenes:
How V2 combines border ownership, stereoscopic cues, and Gestalt grouping rules. Front. Psychol. 2016, 6,
2054. Special issue on 'Perceptual Grouping-the State of the Art'.
36. Dresp-Langley, B. Affine Geometry, Visual Sensation, and Preference for Symmetry of Things in a Thing.
Symmetry 2016 8(11), 127.
37. Giurfa, M.; Eichmann, B.; Menzl, R. Symmetry perception in an insect. Nature 1996, 382, 458 -- 461.
38. Bertamini, M. Sensitivity to reflection and translation is modulated by objectness. Perception 2010; 39,
27 -- 40.
39. Makin, A.J.D.; Rampone, G.; Wright, A.; Martinovic, J.; Bertamini, M. Visual symmetry in objects and
gaps. Journal of Vision 2014, 14(3):12, 1 -- 12.
40. Baylis, G.C.; Driver, J. Perception of symmetry and repetition within and across visual shapes:
Part-descriptions and object-based attention. Visual Cognition 2001, 8(2), 163 -- 196.
41. Michaux, A; Kumar, V.; Jayadevan, V.; Delp, E., Pizlo, Z. Binocular 3D Object Recovery Using a
42.
Symmetry Prior. Symmetry 2017, 9, 64.
Jayadevan V., Sawada, T; Delp, E; Pizlo, Z. Perception of 3D Symmetrical and Nearly Symmetrical
Shapes. Symmetry 2018, 10, 344.
43. Poirier, F.J.A.M.; Wilson, H.R. A biologically plausible model of human shape symmetry perception.
Journal of Vision, 2010,10(1):9, 1 -- 16.
44. Erlikhman, G.; Kellman, P.J. From flashes to edges to objects: Recovery of local edge fragments initiates
spatiotemporal boundary formation. Frontiers in Psychology 2016, 7, 910. Special issue on 'Perceptual
Grouping-the State of the Art'.
45. Erlikhman, G.; Kellman, P.J. Modeling spatiotemporal boundary formation. Vision Research 2015, 126,
131-142.
|
1606.02372 | 1 | 1606 | 2016-06-08T02:01:20 | Entrainment of traveling waves to rhythmic motor acts | [
"q-bio.NC"
] | We hypothesized that rhythmic motor acts entrain neural oscillations in speech production brain regions. We tested this hypothesis in an experiment where a subject produced consonant-vowel (CV) syllables in a rhythmic fashion, while we performed ECoG recordings. Over the ventral sensorimotor cortex (vSMC) we detected significant concentration of phase across trials at the specific frequency of speech production. We also observed amplitude modulations. In addition we found significant coupling between the phase of brain oscillations at the frequency of speech production and their amplitude in the high-gamma range (i.e., phase-amplitude coupling, PAC). Furthermore, we saw that brain oscillations at the frequency of speech production organized as traveling waves (TWs), synchronized to the rhythm of speech production. It has been hypothesized that PAC is a mechanism to allow low-frequency oscillations to synchronize with high-frequency neural activity so that spiking occurs at behaviorally relevant times. If this hypothesis is true, when PAC coexists with TWs, we expect a specific organization of PAC curves. We observed this organization experimentally and verified that the peaks of high-gamma oscillations, and therefore spiking, occur at the same times across electrodes. Importantly, we observed that these spiking times were synchronized with the rhythm of speech production. To our knowledge, this is the first report of motor actions organizing (a) the phase coherence of low-frequency brain oscillations, (b) the coupling between the phase of these oscillations and the amplitude of high-frequency oscillations, and (c) TW. It is also the first demonstration that TWs induce an organization of PAC so that spiking across spatial locations is synchronized to behaviorally relevant times. | q-bio.NC | q-bio |
Entrainment of traveling waves
to rhythmic motor acts
Joaqu´ın Rapela∗
Swartz Center for Computation Neuroscience
University of California San Diego
September 25, 2018
Abstract
A remarkable early observation on brain dynamics is that when hu-
mans are exposed to rhythmic stimulation their brain oscillations entrain
to the rhythm of the stimuli (Adrian and Matthews, 1934). Currently
it is not know whether rhythmic motor acts can entrain neural oscilla-
tions. We investigated this possibility in an experiment where a subject
produced consonant-vowel (CV) syllables in a rhythmic fashion, while we
performed ECoG recordings with a dense grid covering most speech pro-
cessing brain regions across the left hemisphere. Most strongly over the
ventral sensorimotor cortex (vSMC), a cortical region that controls the
vocal articulators, we detected significant concentration of phase across
trials at the specific frequency of speech production. We also observed am-
plitude modulations. In addition we found significant coupling between
the phase of brain oscillations at the frequency of speech production and
their amplitude in the high-gamma range (i.e., phase-amplitude coupling,
PAC). Furthermore, we saw that brain oscillations at the frequency of
speech production organized as traveling waves (TWs), synchronized to
the rhythm of speech production. It has been hypothesized that PAC is a
mechanism to allow low-frequency oscillations to synchronize with high-
frequency neural activity so that spiking occurs at behaviorally relevant
times. If this hypothesis is true, when PAC coexists with TWs, we expect
a specific organization of PAC curves. We observed this organization ex-
perimentally and verified that the peaks of high-gamma oscillations, and
therefore spiking, occur at the same times across electrodes. Importantly,
we observed that these spiking times were synchronized with the rhythm
of speech production. To our knowledge, this is the first report of motor
actions organizing (a) the phase coherence of low-frequency brain oscil-
lations, (b) the coupling between the phase of these oscillations and the
amplitude of high-frequency oscillations, and (c) TWs. It is also the first
demonstration that TWs induce an organization of PAC so that spiking
across spatial locations is synchronized to behaviorally relevant times.
∗[email protected]
Glossary
CV consonant vowel.
ECoG electrocorticography. A brain recording modality that measures poten-
tials directly from the cortical surface.
EEG electroencephalography. A brain recording modality that measures elec-
trical potentials from the scalp..
ERP event-related potential. Mean of voltages recorded at a single electrode
and aligned with respect to an event of interest (e.g., the transition time
between a consonant and a vowel)..
ERSP event-related spectral perturbation. An average measure of evoked
power across trials (Section A.1.3).
ITC inter-trial coherence. A measure of phase alignment between trials (Sec-
tion A.1.2).
MEG magnetoencephalography. A brain recording modality that measures
magnetic fields from the scalp..
MI modulation index quantifying the strength of phase-amplitude coupling
(Sections 2.3 and A.1.4).
PAC phase-amplitude coupling. Coupling between the phase of low-frequency
oscillations and the amplitude of high-frequency ones (Sections 2.3 and A.1.4).
TW traveling wave (Section 2.4).
vSMC ventral sensorimotor cortex. A brain region that controls the vocal
articulators.
Contents
1 Introduction
2 Results
2.1 Phase Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
2.2 Amplitude Modulations
2.3 Phase-Amplitude Coupling (PAC)
. . . . . . . . . . . . . . . . .
2.4 Traveling Waves (TWs)
. . . . . . . . . . . . . . . . . . . . . . .
2.5 Organization of PAC in the Presence of TWs . . . . . . . . . . .
3 Discussion
4 Acknowledgments
A Supplementary Information
A.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
A.1.1 Circular statistics concepts
A.1.2 ITC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.3 ERSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.4 Phase-amplitude coupling curve and modulation index . .
A.2 Behavioral Data . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3 ITC across the vSMC . . . . . . . . . . . . . . . . . . . . . . . .
A.4 ERSPs across the vSMC . . . . . . . . . . . . . . . . . . . . . . .
A.5 ERPs across the vSMC . . . . . . . . . . . . . . . . . . . . . . .
A.6 Electrodes with strongest PAC . . . . . . . . . . . . . . . . . . .
1
2
2
4
5
7
9
10
13
2
2
2
2
2
3
3
3
16
34
34
1 Introduction
External visual (Regan, 1966) and auditory (Galambos et al., 1981) rhythmic
stimuli entrain brain oscillations (i.e., drag brain oscillations to follow the rhythm
of the stimuli), and this entrainment is modulated by attention so that the oc-
currence of attended stimuli coincides with the phase of brain oscillations of
maximal excitability (Lakatos et al., 2005, 2008, 2013; O'Connell et al., 2011;
Besle et al., 2011; Gomez-Ramirez et al., 2011; Zion Golumbic et al., 2013; Cravo et al.,
2013; Mathewson et al., 2010; Spaak et al., 2014; Gray et al., 2015). Although
speech is only quasi rhythmic (Cummins, 2012) an increasing number of studies
is showing that neural oscillations can entrain to speech sound (e.g., Zion Golumbic et al.,
2013; Gross et al., 2013; Doelling et al., 2014; Millman et al., 2015; Park et al.,
2015). For recent reviews on the entrainment of neural oscillations to speech
sound see Peelle and Davis (2012); Zion Golumbic et al. (2012); Ding and Simon
(2014). What is currently not know if whether speech production can entrain
neural oscillations. Here we present evidence showing that a speaker rhythmi-
cally producing consonant-vowel (CV) syllables at different frequencies entrains
brain oscillations in the ventral sensorimotor cortex (vSMC) at corresponding
different frequencies.
It has been suggested that a role of entrained low-frequency oscillations is to
modulate the excitability of neurons, in such a way that periods of higher ex-
citability correspond to events of interest in sensory streams (Schroeder and Lakatos,
2009). This hypothesis has been supported by reports of phase-amplitude cou-
pling (PAC), where the phase of low-frequency oscillations modulates the ampli-
tude of higher-frequency ones (e.g., Canolty et al., 2006). In speech perception
it has been proposed that theta rhythm (3-8 Hz) is related to the encoding of
slower syllabic information in the speech signal, the gamma rhythm (>30 Hz)
is involved in the linguistic coding of phonemic details, and the PAC between
the theta and gamma rhythms could modulate the excitability of neurons to
devote more processing power to the informative parts of syllabic sound pat-
terns (Giraud and Poeppel, 2012; Hyafil et al., 2015). However, PAC has not
yet been reported in speech-production brain regions. We show that over the
vSMC the low-frequency oscillations entrained to the rhythm of speech pro-
duction are strongly coupled with high-gamma oscillations.
Importantly, we
demonstrate that the phase at which the entrained oscillation couples with the
maximum amplitude of the high-gamma oscillations changes orderly from ven-
tral to dorsal electrodes over the vSMC.
Traveling waves (TWs) have been reported in animal studies (e.g., Rubino et al.,
2006), human EEG (e.g., Patten et al., 2012), and human ECoG (e.g., Bahramisharif et al.,
2013). However, these waves have not been observed in speech production brain
regions in humans. The aforementioned orderly change of PAC across the vSMC
suggested the existence of TWs. We looked for them and we found them. As
mentioned above, brain activity at single recording sites can be entrained to
external stimulation but, to our knowledge, it is unknown whether distributed
spatio-temporal activity in the form of TWs can also be entrained. We demon-
strate that TWs over the vSMC are entrained to the rhythm of speech produc-
1
tion.
PAC has been studied in the absence of TWs. When these waves are present,
in order for cells to fire at the same behaviorally-relevant time, PAC curves mea-
sured across electrodes should be different from each other and systematically
organized. If these curves were identical to each other, cells would fire in the or-
der given by the propagating TW, and not at the same time. Below we propose
an organization of PAC so that when TWs are present cells fire at the same
time, we show that this organization is present in our recordings, and provide
evidence indicating that cells along the direction of propagation of a planar TW
fire at the same behaviorally relevant time.
2 Results
2.1 Phase Alignment
We epoched the raw ECoG waveforms around the time of CV transitions (rep-
resented by the vertical line at time zero in all figures). We analyzed separately
recordings from different sessions, and present the analysis from recording ses-
sions EC2 B105 and EC2 B89, the sessions with the largest (1.62 sec) and short-
est (0.97 sec) median inter-syllable separation times, respectively (Figure A.1).
Figure 1a and 1b show the inter-trial coherence (ITC, Section A.1.2), a measure
of phase alignment in a group of trials, calculated from recordings in sessions
EC2 B105 and EC2 B89, respectively, for electrode 154 in the vSMC (Figure 2).
In agreement with our hypothesis that the rhythm of speech production entrains
low-frequency brain oscillations, in the ITC computed from data from experi-
mental session EC2 B105 with a median inter-syllable production frequency of
0.62 Hz (black horizontal line in Figure 1a), we observe a peak of ITC around
this frequency (red pixels in Figure 1a). Also, we see a peak of ITC around
a frequency of 1.03 Hz in Figure 1b, obtained from an experimental session
where the median inter-syllable separation time was 1.03 Hz. These figures il-
lustrate that low-frequency brain oscillations entrain to the rhythm of speech
production.
For each electrode, we computed the maximum ITC between 0.5 seconds
before and 0.6 seconds after the CV transition at the frequency of speech pro-
duction. Figure 2 highlights the 50 electrodes with largest ITCs. We sorted
the electrodes by their maximum ITC value, and colored them according to
their rank in this sorting. The ten electrodes with largest ITC are colored in
red and are located in the center of the ventral sensoritmotor cortex. As the
distance of an electrode to the center of the vSMC increases, its maximum ITC
decreases. Section A.3 more clearly illustrates the specificity of phase coherence
to the vSMC by showing ITC plots from the ventral to the dorsal edge of the
grid across the vSMC.
2
Figure 1: ITCs for electrode 154 computed from recordings in experimental
session EC2 B105 with an inter-syllable production frequency of 0.62 Hz (a), and
from experimental session EC2 B89 with an inter-syllable production frequency
of 1.03 Hz (b). Note that the peaks of the ITC at the lowest frequencies (red
points) align with the median frequency of speech production (black horizontal
line) both when the subject speaks slower (a) and faster (b).
3
Figure 2: Strongest entrainment occurs over the vSMC. The 50 electrodes with
largest ITC (between 0.5 seconds before and 0.6 seconds after the CV transition,
at the frequency of CV production, in the experimental session EC2 B105) are
highlighted in color.
2.2 Amplitude Modulations
The Event-Related Spectral Perturbation (ERSP, Section A.1.3) is a measure
of power modulations around an event of interest. Figure 3a shows ERSPs
around the transition time between consonants and vowels (black vertical line in
Figure 3). In electrode 133 over auditory cortex (Figure 3a) we see a significant
power increase (red blob) before the transition between consonants and vowels
in the beta range. Differently, in electrode 135 over the vSMC (Figure 3b) we
observe a significant power increase after the transition between consonants and
vowels and in the high-gamma range.
Figures A.18-A.30 show ERSPs between the ventral electrode 129 and the
dorsal electrode 141 in experimental session EC2 B105. Power in the beta range
increases from electrodes 129 to 132 (Figures A.18-A.21) in the ventral tempo-
ral cortex, and peaks over electrodes 133 and 134 around the auditory cortex
(Figures A.22 and A.23), prior to consonant-vowel transitions. A sharp transi-
tion is seen in electrode 135 over the ventral sensorimotor cortex (Figure A.24),
with strong modulations in the high-gamma range (70-200 Hz), after consonant-
vowel transitions, and almost no modulation in the beta range. This pattern
of modulations is strongest over electrode 136 (Figure A.25), and fades along
electrodes 137-141 (Figures A.26-A.30) toward the dorsal end of the vSMC.
4
(a) ERSP in Auditory Cortex
(b) ERSP in the vSMC
Figure 3: ERSPs for electrode 133 in the auditory cortex (a) and for electrode
135 in the vSMC (b). In the auditory cortex (a) we observe a peak of amplitude
(red blob) in the beta range before the CV transition (vertical black line), while
in the vSMC (b) a peak of amplitude occurs in the high-gamma range after the
CV transition.
2.3 Phase-Amplitude Coupling (PAC)
For experimental session EC2 B105, we studied the coupling between phase at
the entrainment frequency of 0.62 Hz (0.01 Hz, wavelet full width at half maxi-
mum) and amplitude at 100 Hz (2.06 Hz, wavelet full width at half maximum).
For each electrode we calculated the PAC curve (Section A.1.4), and quantified
the strength of the coupling using the modulation index (MI, Section A.1.4).
Figure 4 shows PAC curves measured at electrodes 135-138 (for visualization
purposes we show two phase cycles). The ECoG power at 100 Hz is well mod-
ulated by the phase at the entrainment frequency, with MIs 4.84e-3, 8.22e-3,
1.05e-3, and 2.82e-3 at electrodes 135-138, respectively. For electrode 136 (green
trace) larger ECoG amplitudes in the high-gamma range (100 Hz) tend to oc-
cur at the valley (phase near π) of the low-frequency oscillation (0.62 Hz). As
we move from the more ventral electrode 135 to the more dorsal electrode 138,
the phase of the low-frequency oscillation at which the amplitude of the high-
gamma oscillations is maximal shifts gradually from π (pink trace) to 0 (violet
trace). We discuss this organization in Section 2.5. Strongest PAC between
phase at the entrainment frequency and amplitude at 100 Hz was observed over
the vSMC and surrounding regions. Figure A.44 highlights the 50 electrodes
with largest MI.
For electrode 154 in the vSMC, we computed the MI for all pairs of frequency
for phase and frequency for amplitude, with the frequency for phase between
5
)
V
(
z
H
0
0
1
t
a
e
d
u
t
i
l
p
m
A
n
a
e
M
3
2
Electrode
138
137
136
135
2p
0
Phase (radians)
3p
Figure 4: Phase-amplitude coupling curves at electrodes over the vSMC. Phase
and amplitude were measured at the entrainment frequency (0.62 Hz) and at
100 Hz, respectively. All displayed electrodes showed a strong modulation of
the high-gamma amplitude by the entrained low-frequency phase. Note that
as we move from more ventral to more dorsal electrodes (i.e., from electrode
135 to electrode 138) the peak of the PAC curves occurs at earlier phases. The
significance of this organization is discussed in Section 2.5.
6
-
p
p
m
0.1 and 20 Hz, and the frequency for amplitude between 5 and 200 Hz. We only
observed large MIs for frequencies for phase between 0.1 and 2 Hz. Figure 5a
plots the obtained MIs. We found two clusters of the frequency for amplitude
with large MIs. The first cluster was on the high-gamma range (60 to 150 Hz)
and the second cluster on the beta range (20 to 40 Hz). An asterisk in Figure 5
marks a MI peak, the number next to the asterisk is the peak MI, and the
first and second numbers in the parenthesis are the frequency for phase and for
amplitude, respectively, of the peak MI. Interestingly, for the cluster on the high-
gamma range, the maximum MI occurred exactly at the bin of the entrainment
frequency (0.62 Hz for the frequency for phase). Differently, for the cluster in
beta range the maximum MI occurred below the entrainment frequency (0.5 Hz
for the frequency for phase). This figure suggests that not only low-frequency
oscillations become entrained to the rhythm of speech production, but also PAC
is optimized at this rhythm. For electrode 132 in the auditory cortex, Figure 5b,
the MIs were overall much weaker, and both the clusters in the high-gamma and
beta ranges showed a peak MI at frequencies for phase below the entrainment
frequency. This suggest that the maximization of PAC at the entrainment
frequency is specific to the vSMC.
2.4 Traveling Waves (TWs)
We filtered the ECoG recordings in experimental session EC2 B105 (Figure A.1)
with a second-order bandpass Butterworth filter around the median syllable-
production syllable of 0.62 Hz (low- and high-frequency cutoffs 0.4 Hz and
0.8 Hz, respectively). Figure 6 shows the filtered voltages from electrodes 135-
141 at the center of the vSMC in a representative time interval (390 to 400 sec-
onds after the start of the experiment). We see that, as we move from the ventral
electrode 135 to the dorsal electrode 141, the voltage traces are orderly shifted
to later times. The vertical black lines indicate the transition time between a
consonant and a vowel in the production of consonant-vowel syllables. Note
that the peaks of the voltage traces tend to occur around these consonant-vowel
transition times. Thus, this figure illustrates a wave of activity traveling from
the ventral to the dorsal edge of the ventral sensorimotor cortex and entrained
to the rhythm of speech production.
A movie showing the cosine of the phase of the bandpassed filtered voltages
across the whole grid simultaneously with the speech of the subject can be found
at https://youtu.be/dXrzj2eEuVY. The bottom/top/left/right pixels in the
movie correspond to electrodes along the ventral/dorsal/frontal/caudal sides of
the grid. We see mostly planar traveling waves moving in different directions,
but occasionally we also observe rotating traveling waves. At most times when
the subject produces a CV syllable we see a white blob over the vSMC, indicating
that filtered voltages approach their peak values. This movie demonstrates, for
the first time, the existence of traveling waves across left speech processing brain
areas synchronized to the rhythm of speech production.
7
(a)
(b)
200
150
e
d
u
t
i
l
p
m
A
100
r
o
f
y
c
n
e
u
q
e
r
F
* 0.0029 (0.70, 87.00)
* 0.0024 (0.50, 27.00)
50
0
200
150
100
50
0
MI
0.003
0.002
0.001
0.000
* 0.0004 (0.50, 63.00)
* 0.0010 (0.40, 16.00)
0.5
1.0
1.5
2.0
Frequency for Phase
0.5
1.0
1.5
Frequency for Phase
2.0
Figure 5: Modulation indices for electrode 154 over the vSMC (a) and for
electrode 132 over the auditory cortex (b) at multiple frequencies for phase and
amplitude. Over the vSMC (a) the peak MI for amplitudes in the high-gamma
range occurs at the entrained frequency for phase (0.62 Hz), while the peak MI
for amplitudes in the beta range occurs at lower frequencies. This suggests that
over the vSMC not only phase coherence, but also phase-amplitude coupling, is
matched to the speech production frequency. Over auditory cortex (b) MIs are
much smaller than over the vSMC and the peak MIs occurred at frequencies
for phase below the entrainment frequency. This suggests that PAC is largest
over speech-production brain regions and that the maximization of PAC at the
entrainment frequency is specific to the vSMC.
8
)
V
(
e
g
a
t
l
o
V
20
10
0
−10
−20
−30
Electrode
138
137
136
135
390.0
392.5
395.0
Time (sec)
397.5
400.0
Figure 6: Filtered voltage waveforms along the ventral-dorsal axis of the vSMC.
Filtering was performed between 0.4 and 0.8 Hz, around the median frequency
of CV syllable production of 0.62 Hz. Vertical black lines indicate CV transition
times. As we move from the ventral electrode 135 in red to the dorsal electrode
141 in magenta, voltage waveforms are orderly shifted to right, indicating the
existence of a traveling wave. Also, the peak of the waves occurs around the
time of CV transitions, suggesting that these traveling waves are entrained to
the rhythm of speech production.
2.5 Organization of PAC in the Presence of TWs
If we wanted to design a brain so that all neurons recorded by the four electrodes
in Figure 4 fired at the same behaviorally relevant time (e.g., the transition
between the consonant-vowel syllable at time 391.2 seconds), how would be
choose the PAC curves for these electrodes? Since the phase at electrode 135
(red trace in Figure 4) at time 391.2 is close to π/4 we would use a PAC curve
for this electrode with maximal high-gamma amplitude at this phase. Then,
because high-gamma amplitude is related to neural firing (Ray et al., 2008),
cells around electrode 135 would then tend to spike at the time of the consonant
vowel transition. Next, due to the traveling wave, the phase at electrode 136
at time 391.2 is earlier than that at electrode 135, thus we would use a PAC
curve with a peak at this earlier phase for electrode 136. Similarly, we would
use a PAC curve with an earlier peak at electrode 137 than that at electrode
136. That is, as we move along the direction of propagation of a traveling wave,
we would use PAC curves with earlier and earlier peaks. And this is what we
observed experimentally in Figure 4.
To validate that the previous organization of PAC curves allows cells to fire
at the same behaviorally relevant time in the presence of traveling waves, Fig-
ure 7 plots mean high-gamma amplitude as a function of time for electrodes
135-138 (for each time and each electrode, we extracted the phase at 0.62 Hz,
then the mean amplitude at 100 Hz for this phase from the PAC curve in Fig-
ure 4, and finally we plotted the extracted mean amplitudes as a function of the
corresponding times). As expected, we found that high-gamma amplitude, and
9
m
)
V
(m
z
H
0
0
1
t
a
e
d
u
t
i
l
p
m
A
n
a
e
M
3
2
Electrode
138
137
136
135
390.0
392.5
395.0
Time (sec)
397.5
400.0
Figure 7: Mean high-gamma amplitude as a function of time at electrodes over
the ventral sensorimotor cortex. Each time point at each electrode corresponds
to a given phase of the oscillation at the speech-production frequency of 0.62 Hz
in Figure 6, and from Figure 4 each phase corresponds to a given mean amplitude
at 100 Hz. Thus, each time point of each electrode corresponds to a mean phase
at 100 Hz. Colored traces show this correspondence across different electrodes.
We observe that mean high-gamma amplitude, and therefore spiking, is aligned
across electrodes over the ventral sensorimotor cortex and synchronized to the
rhythm of speech production. The curves are not smooth because we used coarse
bins in the calculation of the PAC curves in Figure 4.
therefore cell spiking, peaks at approximately similar times across electrodes.
In addition, we observed that the high-gamma amplitude peaks occur shortly
after the transition time between consonant and vowels, which shows that the
traveling waves are therefore entrained to the rhythm of speech production.
3 Discussion
We have shown that rhythmic speech production entrains low-frequency brain
oscillations (Figure 1) and that this entrainment is stronger over the vSMC
(Figure 2), that rhythmic speech production modulates the power of beta os-
cillations over the auditory cortex and that of high-gamma oscillations over the
vSMC (Figure 3), that the phase of the entrained low-frequency oscillations
couples with the amplitude of high-gamma oscillations (Figure 4) and that this
coupling is largest over the vSMC (Figure A.44), that over the vSMC -- but
not over the auditory cortex -- this coupling is maximal between phase at the
entrained frequency and amplitude in the high-gamma range (Figure 5), and
that the entrained oscillations organize as TWs synchronized to the produced
speech (Figure 6 and movie at https://youtu.be/dXrzj2eEuVY). Finally, we
observed that the peaks of PAC curves occur at earlier phases along the direc-
tion of propagation of the TWs and presented evidence (Figure 7) indicating
that this organization of PAC curves enables neurons along the direction of
10
propagation of the TW to fire at the same behaviorally relevant time.
These initial findings open several interesting directions for future investi-
gations. To mention a few: (1) How do entrainment, PAC, and TWs relate
to properties of the produced CV syllables? For example, does PAC occur
earlier for fricatives than for nasals? (2) How do these effects (i.e., entrain-
ment, PAC, and TWs) correspond to the specific articulators recruited in the
production of a CV syllable? For example, do planar TWs correspond to the
recruitment of the jaw but rotating TWs to the recruitment of the larynx? (3)
PAC has been reported in speech perception (e.g., Giraud and Poeppel, 2012;
Hyafil et al., 2015), but not in speech production tasks. We claimed that the
PAC we observed is related to speech production, but we have not eliminated
the possibility that it may be related to the perception of the speaker's own
speech. The detection of PAC in new experiments eliminating speech feedback
would confirm that PAC can be related to the production of speech. (4) The
observed effects appeared in the strongly rhythmic production of CV syllables.
Would similar effects be observed in colloquial speech? How do prosodic fea-
tures of speech, like intonation or stress, affect the observed effects? (5) We
have characterized phase coherence and PAC at single electrodes, but probably
phase coherence and PAC across different electrodes play an important role in
the synchronization and information transfer across different speech perception
and production brain regions, as observed in non-speech-processing tasks (e.g.,
Voytek et al., 2015).
It is possible to explain some of our results in terms of event-related poten-
tials (ERPs, the mean recorded potential around an event of interest). Previous
studies have suggested that ERPs are generated by the alignment of phases of
ongoing oscillations (Makeig et al., 2002), but this suggestion is still controver-
sial (Sauseng et al., 2007). Because the production of CV syllables occurred
at a low-frequency rhythm, ERPs should be generated at this low-frequency
and we should observe alignment of phases at this low frequency. Thus, this
ERP argument explains Figure 1. Also, since ERPs should be stronger over
the vSMC, we should expect more phase alignment over this region, explaining
Figure 2. One could argue that the production of speech is related to spiking,
and therefore to high-gamma activity, over the vSMC. Thus, the low frequency
phase at which oscillations reset for the generation of ERPs should be related
to high-gamma activity. This argument addresses Figures 4 and 5.
Section A.5 shows ERPs computed along the ventral-dorsal axis of the
recording grid.
In agreement with Figure 2, we see strongest ERPs at elec-
trodes 136-140 over the vSMC. These ERPs begin around 150 msec before the
CV transition, peak around 100 ms after this transition, and show a negative
peak around 400 ms after the CV transition. Also, in agreement with Figure 4
we see that the peak of the ERPs are shifted to later times as we move from
electrode 136 to electrode 140, also suggesting the existence of TWs.
Thus, the spectral and ERP analysis are consistent with each other. An
advantage of the former is that it decomposes the averaged effects seen in ERPs
in different frequency bands. Since several frequency-specific effects on electro-
physiological recordings are known, spectral analysis allow richer descriptions
11
of electrophysiological recordings than ERP ones. For instance by decomposing
averaged effects in different frequency bands, our spectral analysis showed that
neural spiking, as reflected by high-gamma amplitude, was coupled to relevant
times in the speech production rhythm, as indicated by the coupled phase at
the speech-production frequency (Figure 4). Also, our PAC analysis is more
quantitative than the previous argument linking ERPs and high-gamma ampli-
tude. This analysis revealed a new organization of PAC in the presence of TWs,
providing strong support to a previous hypothesis stating that, when stimuli or
behavior occurs in a rhythmic fashion, PAC is a mechanism that allows neurons
to spike at behavioral relevant times (Canolty and Knight, 2010). In summary,
the spectral methods used above to characterize rhythmic speech production
are consistent with, but superior than, ERP methods for the characterization
of rhythmic behaviors such as speech.
The effects reported above reflect brain mechanisms beyond the control of vo-
cal articulators. For example, the traveling wave movie (https://youtu.be/dXrzj2eEuVY)
shows waves moving across multiple speech processing brain regions. At sev-
eral times we see waves moving from electrodes over the auditory cortex to
electrodes over the vSMC, suggesting that these waves could be a mecha-
nism for integrating information across the brain, as suggested by others (e.g.,
Nunez and Srinivasan, 2006; Sato et al., 2012).
As speaking must be compatible with the performance limits of motor sys-
tems, several theories have linked functional properties of motor systems to
speech capability. The Frame/Content theory (MacNeilage and Davis, 2000,
2001) proposes that the features of elementary production units, i.e., sylla-
bles, are determined by mechanical properties of the speech apparatus, e.g.,
natural oscillatory rhythms (Giraud et al., 2007). Previous investigations (e.g.,
Morillon et al., 2010) have hypothesized that motor areas express low-frequency
oscillatory activity characteristic of jaw movements (4 Hz) and high-frequency
activity corresponding to movements of the tongue (e.g., trill at 35-40 Hz). How-
ever, to our knowledge this is the first study providing physiological evidence for
the relation of low- and high-frequency oscillations, as well as their interactions,
to speech production.
Natural speech production is substantially more complex than the produc-
tion of CV syllables, which suggests that the findings reported in this article
may not be relevant to natural speech. We argue on the contrary. CV sylla-
bles are the building blocks of natural speech, such as sinusoidal grating are
the building blocks of natural images. While producing CV syllables subjects
spontaneously enter a strongly rhythmic behavioral state. Natural speech is
strongly rhythmic and the production of CV syllables could be ideal to isolate
rhythmic aspects of this behavior. By characterizing responses of visual cells to
sinusoidal gratings, it has been possible to understand key neural mechanisms
of visual processing, like contrast grain control (Ohzawa et al., 1985), that are
difficult to observe in responses of visual neurons to complex natural images.
Similarly, investigating the production of CV syllables may be instrumental in
discovering key rhythmic mechanisms of speech production that are obscured
in more complex speech-production tasks.
12
The specific attributes of mesoscopic cortical activity related to speech pro-
duction are not well characterized.
In speech perception, Luo and Poeppel
(2007) have shown that phase activity (over the human auditory cortex in the
theta range (4-8 Hz) recorded with MEG) reliably tracks and discriminates
spoken sentences and that this discrimination is correlated with speech intelli-
gibility. Here we extended these results and showed that phase activity (in the
form of phase alignment, PAC, and TWs) is also a relevant mesoscopic attribute
to characterize the neurophysiology of speech production.
Having detected phase alignment, PAC and TWs in ECoG recordings, one
could use forward models to calculate how these effects propagate to scalp-
recorded potentials in the EEG or magnetic fields in the MEG. Next, one could
invert these forward models to infer the existence and properties of these oscil-
latory effects from EEG or MEG recordings (Friston et al., 2007). If successful
this approach would allow to detect phase alignment, PAC, and TWs from non-
invasive recordings, and provide novel measures to non-invasively characterize
speech production in health and disease.
4 Acknowledgments
We thank Dr. Edward Chang and Dr. Kristofer Bouchard for sharing the ECoG
recordings with us and for comments on a preliminary version of this manuscript,
and Dr. John Iversen for suggesting us the relation between PAC and TWs.
13
References
E.D. Adrian and B.H.C. Matthews. The berger rhythm : potential changes from the
occipital lobes in man. Brain, 57(4):355 -- 385, 1934.
Ali Bahramisharif, Marcel AJ van Gerven, Erik J Aarnoutse, Manuel R Mercier,
Theodore H Schwartz, John J Foxe, Nick F Ramsey, and Ole Jensen. Propagating
neocortical gamma bursts are coordinated by traveling alpha waves. The Journal
of Neuroscience, 33(48):18849 -- 18854, 2013.
J. Besle, C.A. Schevon, A.D. Mehta, P. Lakatos, R.R. Goodman, G.M. McKhann,
R.G. Emerson, and C.E. Schroeder. Tuning of the human neocortex to the temporal
dynamics of attended events. The Journal of Neuroscience, 31(9):3176 -- 3185, 2011.
R.T. Canolty and R.T. Knight. The functional role of cross-frequency coupling. Trends
in cognitive sciences, 14(11):506 -- 515, 2010.
R.T. Canolty, E. Edwards, S.S. Dalal, M. Soltani, S.S. Nagarajan, H.E. Kirsch, M.S.
Berger, N.M. Barbaro, and R.T. Knight. High gamma power is phase-locked to
theta oscillations in human neocortex. science, 313(5793):1626 -- 1628, 2006.
A.M. Cravo, G. Rohenkohl, V. Wyart, and A.C. Nobre. Temporal expectation en-
hances contrast sensitivity by phase entrainment of low-frequency oscillations in
visual cortex. Journal of Neuroscience, 33(9):4002 -- 4010, 2013.
F. Cummins. Oscillators and syllables: a cautionary note. Frontiers in psychology, 3:
364, 2012.
A. Delorme and S. Makeig. EEGLAB: an open source toolbox for analysis of single-trial
EEG dynamics including independent component analsys. Journal of Neuroscience
Methods, 134(1):9 -- 21, 2004.
N. Ding and J.Z. Simon. Cortical entrainment to continuous speech: functional roles
and interpretations. Front. Hum. Neurosci, 8(311):10 -- 3389, 2014.
Keith B Doelling, Luc H Arnal, Oded Ghitza, and David Poeppel. Acoustic land-
marks drive delta -- theta oscillations to enable speech comprehension by facilitating
perceptual parsing. Neuroimage, 85:761 -- 768, 2014.
K.J. Friston, J.T. Ashburner, S.J. Kiebel, T.E. Nichols, and W.D. Penny, editors.
Statistical Parametric Mapping. Academic Press, London, 2007.
Robert Galambos, Scott Makeig, and Peter J Talmachoff. A 40-hz auditory potential
recorded from the human scalp. Proceedings of the National Academy of Sciences,
78(4):2643 -- 2647, 1981.
A.L. Giraud, A. Kleinschmidt, D. Poeppel, T.E. Lund, R.S.J. Frackowiak, and
H. Laufs. Endogenous cortical rhythms determine cerebral specialization for speech
perception and production. Neuron, 56(6):1127 -- 1134, 2007.
Anne-Lise Giraud and David Poeppel. Cortical oscillations and speech processing:
emerging computational principles and operations. Nature neuroscience, 15(4):511 --
517, 2012.
14
M. Gomez-Ramirez, S.P. Kelly, S. Molholm, P Sehatpour, T.H. Schwartz, and J.J.
Foxe. Oscillatory sensory selection mechanism during intersensory attention to
rhythmic auditory and visual inputs: a human electrocortigraphic investigation.
The Journal of Neuroscience, 31(50):18556 -- 18567, 2011.
Michael J Gray, Hans-Peter Frey, Tommy J Wilson, and John J Foxe. Oscillatory
recruitment of bilateral visual cortex during spatial attention to competing rhythmic
inputs. The Journal of Neuroscience, 35(14):5489 -- 5503, 2015.
J. Gross, N. Hoogenboom, G. Thut, P. Schyns, S. Panzeri, P. Belin, and S. Garrod.
Speech rhythms and multiplexed oscillatory sensory coding in the human brain.
PLoS Biology, 11(12), 2013. doi: 10.1371/journal.pbio.1001752.
A. Hyafil, L. Fontolan, C. Kabdebon, B. Gutkin, and A.L. Giraud. Speech encoding
by coupled cortical theta and gamma oscillations. Elife, 4:e06213, 2015.
P. Lakatos, A.S. Shah, K.H. Knuth, I. Ulbert, G. Kamos, and C.E. Schroeder. An
oscillatory hierarchy controlling neural excitability and stimulus processing in the
auditory cortex. J. Neurophysiology, 94(3):1904 -- 1911, 2005.
P. Lakatos, G. Karmos, A.D. Mehta, I. Ulbert, and C.E. Schroeder. Entrainment of
neuronal oscillations as a mechanism of attentional selection. Science, 320:110 -- 113,
2008.
P. Lakatos, G. Musacchia, M.N. O'connel, A.Y. Falchier, D.C. Javitt, and C.E.
Schroeder. The spectrotemporal filter mechanism of auditory selective attention.
Neuron, 77(4):750 -- 761, 2013.
Huan Luo and David Poeppel. Phase patterns of neuronal responses reliably discrim-
inate speech in human auditory cortex. Neuron, 54(6):1001 -- 1010, 2007.
P.F. MacNeilage and B.L. Davis. On the origin of internal structure of word forms.
Science, 288(5465):527 -- 531, 2000.
P.F. MacNeilage and B.L. Davis. Motor mechanisms in speech ontogeny: phylogenetic,
neurobiological and linguistic implications. Current opinion in neurobiology, 11(6):
696 -- 700, 2001.
S. Makeig, M. Westerfield, T.P. Jung, S. Enghoff, J. Townsend, E. Couchesne, and
T.J. Sejnowski. Dynamic brain sources of visual evoked responses. Science, 295
(5555):690 -- 694, 2002.
K.V. Mardia. Statistics of directional data. Acadmic Press, New York, NY, 1972.
K.E. Mathewson, M. Fabiani, G. Gratton, D.M. Beck, and A. Lleras. Rescuing stimuli
from invisibility: Inducing a momentary release from visual masking with pre-target
entrainment. Cognition, 115:186 -- 191, 2010.
R.E. Millman, S.R. Johnson, and G. Prendergast. The role of phase-locking to the
temporal envelope of speech in auditory perception and speech intelligibility. Journal
of cognitive neuroscience, 2015.
15
B. Morillon, K. Lehongre, R.S.J. Frackowiak, A. Ducorps, A. Kleinschmidt, D. Poep-
pel, and A.L. Giraud. Neurophysiological origin of human brain asymmetry for
speech and language. Proceedings of the National Academy of Sciences, 107(43):
18688 -- 18693, 2010.
P.L. Nunez and R. Srinivasan. A theoretical basis for standing and traveling brain
waves measured with human eeg with implications for an integrated consciousness.
Clinical Neurophysiology, 117(11):2424 -- 2435, 2006.
M.N. O'Connell, A. Falchier, T. McGinnis, C.E. Schroeder, and P. Lakatos. Dual
mechanism of neuronal ensemble inhibition in primary auditory cortex. Neuron, 69
(4):805 -- 817, 2011.
I. Ohzawa, G. Sclar, and R.D. Freeman. Contrast gain control in the cat's visual
system. Journal of Neurophysiology, 54(3):651 -- 667, 1985.
Hyojin Park, Robin AA Ince, Philippe G Schyns, Gregor Thut, and Joachim Gross.
Frontal top-down signals increase coupling of auditory low-frequency oscillations to
continuous speech in human listeners. Current Biology, 2015.
Timothy M Patten, Christopher J Rennie, Peter A Robinson, and Pulin Gong. Human
cortical traveling waves: dynamical properties and correlations with responses. PLoS
One, 7(6):e38392, 2012.
J.E. Peelle and M.H. Davis. Neural oscillations carry speech rhythm through to com-
prehension. Front Psychol, 3(320):1 -- 17, 2012.
S. Ray, N.E. Crone, . Niebur, P.J Franaszczuk, and S.S. Hsiao. Neural correlates
of high-gamma oscillations (60 -- 200 hz) in macaque local field potentials and their
potential implications in electrocorticography. The Journal of Neuroscience, 28(45):
11526 -- 36, 2008.
D Regan. Some characteristics of average steady-state and transient responses evoked
by modulated light. Electroencephalography and clinical neurophysiology, 20(3):238 --
248, 1966.
D. Rubino, K.A. Robbins, and N.G. Hatsopoulos. Propagating waves mediate infor-
mation transfer in the motor cortex. Nature neuroscience, 9(12):1549 -- 1557, 2006.
T.K. Sato, I. Nauhaus, and M. Carandini. Traveling waves in visual cortex. Neuron,
75(2):218 -- 229, 2012.
P. Sauseng, W. Klimesch, W.R. Gruber, S. Hanslmayr, R. Freunberger, and M. Dop-
pelmayr. Are event-related potential components generated by phase resetting of
brain oscillations? a critical discussion. Neuroscience, 146(4):1435 -- 1444, 2007.
C.E. Schroeder and P. Lakatos. Low-frequency neuronal oscillations as instruments of
sensory selection. Trends in Neurosciences, 32(1):9 -- 18, 2009.
E. Spaak, F.P. de Lange, and O Jensen. Local entrainment of alpha oscillations by
visual stimuli causes cyclic modulation of perception. J Neurosci, 34(10):3536 -- 44,
2014.
16
C. Tallon Baudry, O. Bertrand, C. Delpuech, and J. Pernier. Stimulus specificity of
phase-locked and non-phase-locked 40 hz visual responses in human. Journal of
Neuroscience, 16:4240 -- 4349, 1996.
Adriano BL Tort, Robert Komorowski, Howard Eichenbaum, and Nancy Kopell. Mea-
suring phase-amplitude coupling between neuronal oscillations of different frequen-
cies. Journal of neurophysiology, 104(2):1195 -- 1210, 2010.
Bradley Voytek, Andrew S Kayser, David Badre, David Fegen, Edward F Chang,
Nathan E Crone, Josef Parvizi, Robert T Knight, and Mark D'Esposito. Oscillatory
dynamics coordinating human frontal networks in support of goal maintenance.
Nature neuroscience, 2015.
E.M. Zion Golumbic, D. Poeppel, and C.E. Schroeder. Temporal context in speech
processing and attentional stream selection: a behavioral and neural perspective.
Brain and language, 122(3):151 -- 161, 2012.
E.M. Zion Golumbic, N. Ding, S. Bickel, P. Lakatos, C.A. Schevon, G.M. McK-
hann, R.R. Goodman, R. Emerson, A.D. Mehta, J.Z. Simon, D. Poeppel, and C.E.
Schroeder. Mechanisms underlying selective neuronal tracking of attended speech
at a "cocktail party". Neuron, 77(5):980 -- 991, 2013.
E.M. Zion Golumbic, N. Ding, S. Bickel, P. Lakatos, C.A. Schevon, G.M. McK-
hann, R.R. Goodman, R. Emerson, A.D. Mehta, J.Z. Simon, D. Poeppel, and C.E.
Schroeder. Mechanisms underlying selective neuronal tracking of attended speech
at a cocktail party. Neuron, 77(5):980 -- 991, 2013.
1
A Supplementary Information
A.1 Methods
A.1.1 Circular statistics concepts
This section introduces concepts from circular statistics (Mardia, 1972) used below to
define ITC. Given a set of circular variables (e.g., phases), θ1, . . . , θN , we associate
to each circular variable a two-dimensional unit vector. Using notation from complex
numbers, the unit vector associated with variable θi is:
The resultant vector, R, is the sum of the associated unit vectors:
vec(θi) = ejθi
R(θ1, . . . , θN ) =
N
X
i=1
vec(θi)
(1)
(2)
The mean resultant length, ¯R, is the length of the resultant vector divided by the
number of variables:
¯R(θ1, . . . , θN ) =
1
N
R(θ1, . . . , θN )
The circular variance, CV , is one minus the mean resultant length:
CV (θ1, . . . , θN ) = 1 − ¯R(θ1, . . . , θN )
The mean direction, ¯θ, is the angle of the resultant vector:
¯θ(θ1, . . . , θN ) = arg(R(θ1, . . . , θN ))
(3)
(4)
(5)
Note that the mean direction is not defined when the resultant vector is zero, since
the angle of the zero vector is undefined.
A.1.2 ITC
The Inter-Trial Coherence (ITC) is a measure of phase alignment among multiple trials
at a given time and frequency (Tallon Baudry et al., 1996; Delorme and Makeig, 2004).
Mathematically, it is defined as the mean resultant length ( ¯R, Eq. 3) of the phases of
the trials at the given time and frequency.
A.1.3 ERSP
The Event-Related Spectral Perturbation (ERSP) is the mean power of a set of
baseline-normalized epochs (Delorme and Makeig, 2004).
2
A.1.4 Phase-amplitude coupling curve and modulation index
The PAC curve is a graphical means of representing the coupling between phases
at a modulating frequency and amplitudes at a modulated frequency. The MI is
a numerical quantification of the strength of this coupling. We used the methods
described in Tort et al. (2010) to compute both the PAC curve and the MI.
A.2 Behavioral Data
Subjects produced consonant-vowel syllables at different speeds in different experi-
mental sessions. This manuscript characterizes data in the experimental sessions with
largest and shortest median intersyllable separation times (Figure A.1).
A.3
ITC across the vSMC
Largest coherence values at the entrained frequency were observed over the vSMC.
Figures A.2-A.17 plot the ITC, computed from recordings in experimental session
EC2 B105, between the ventral electrode 129 and the dorsal electrode 144. For elec-
trodes over the temporal cortex there is almost no ITC (Figures A.2-A.4). ITC in-
creases over the ventral section of the vSMC (Figures A.5-A.8), peaks at its center
(Figures A.9-A.12), and tapers over its dorsal section (Figures A.13-A.17).
3
0
1
2
3
4
5
6
Inter−Syllable Separation (sec)
y
c
n
e
u
q
e
r
F
y
c
n
e
u
q
e
r
F
0
8
0
6
0
4
0
2
0
0
5
1
0
0
1
0
5
0
0
1
2
3
4
5
6
Inter−Syllable Separation (sec)
Figure A.1: Histogram of inter-syllable separations from the experimental ses-
sions with the largest (1.62 sec) and shortest (0.97 sec) intersyllable separation
time, EC2 B105 (a) and EC2 B89 (b), respectively.
4
Figure A.2: ITC for electrode 129 computed from recordings in experimental
session EC2 B105.
5
Figure A.3: ITC for electrode 130 computed from recordings in experimental
session EC2 B105.
6
Figure A.4: ITC for electrode 131 computed from recordings in experimental
session EC2 B105.
7
Figure A.5: ITC for electrode 132 computed from recordings in experimental
session EC2 B105.
8
Figure A.6: ITC for electrode 133 computed from recordings in experimental
session EC2 B105.
9
Figure A.7: ITC for electrode 134 computed from recordings in experimental
session EC2 B105.
10
Figure A.8: ITC for electrode 135 computed from recordings in experimental
session EC2 B105.
11
Figure A.9: ITC for electrode 136 computed from recordings in experimental
session EC2 B105.
12
Figure A.10: ITC for electrode 137 computed from recordings in experimental
session EC2 B105.
13
Figure A.11: ITC for electrode 138 computed from recordings in experimental
session EC2 B105.
14
Figure A.12: ITC for electrode 139 computed from recordings in experimental
session EC2 B105.
15
Figure A.13: ITC for electrode 140 computed from recordings in experimental
session EC2 B105.
A.4 ERSPs across the vSMC
Figures A.18-A.30 plot ERSPs at electrodes along the ventro-dorsal axis of the vSMC.
Note the sharp transition between the peak of large evoked beta power before the CV
transition at electrodes 132-135 over the auditory cortex to the peak of large evoked
high-gamma power after the CV transition at electrodes 135-140 over the vSMC.
16
Figure A.14: ITC for electrode 141 computed from recordings in experimental
session EC2 B105.
17
Figure A.15: ITC for electrode 142 computed from recordings in experimental
session EC2 B105.
18
Figure A.16: ITC for electrode 143 computed from recordings in experimental
session EC2 B105.
19
Figure A.17: ITC for electrode 144 computed from recordings in experimental
session EC2 B105.
20
Figure A.18: ERSP for electrode 129 computed from recordings in experimental
session EC2 B105.
21
Figure A.19: ERSP for electrode 130 computed from recordings in experimental
session EC2 B105.
22
Figure A.20: ERSP for electrode 131 computed from recordings in experimental
session EC2 B105.
23
Figure A.21: ERSP for electrode 132 computed from recordings in experimental
session EC2 B105.
24
Figure A.22: ERSP for electrode 133 computed from recordings in experimental
session EC2 B105.
25
Figure A.23: ERSP for electrode 134 computed from recordings in experimental
session EC2 B105.
26
Figure A.24: ERSP for electrode 135 computed from recordings in experimental
session EC2 B105.
27
Figure A.25: ERSP for electrode 136 computed from recordings in experimental
session EC2 B105.
28
Figure A.26: ERSP for electrode 137 computed from recordings in experimental
session EC2 B105.
29
Figure A.27: ERSP for electrode 138 computed from recordings in experimental
session EC2 B105.
30
Figure A.28: ERSP for electrode 139 computed from recordings in experimental
session EC2 B105.
31
Figure A.29: ERSP for electrode 140 computed from recordings in experimental
session EC2 B105.
32
Figure A.30: ERSP for electrode 141 computed from recordings in experimental
session EC2 B105.
33
Figure A.31: ERP for electrode 129 computed from recordings in experimental
session EC2 B105.
A.5 ERPs across the vSMC
Figures A.31-A.43 plots ERPs computed at electrodes along the ventro-dorsal axis of
the recordings grid. Note that ERPs are larger over electrodes 136-140 in the vSMC,
in agreement with Figure 2, and that the ERP peak is shifted to later times as we
move from electrode 136 to electrode 140.
A.6 Electrodes with strongest PAC
Figure A.44 highlights the 50 electrodes with largest MI.
34
Figure A.32: ERP for electrode 130 computed from recordings in experimental
session EC2 B105.
35
Figure A.33: ERP for electrode 131 computed from recordings in experimental
session EC2 B105.
36
Figure A.34: ERP for electrode 132 computed from recordings in experimental
session EC2 B105.
37
Figure A.35: ERP for electrode 133 computed from recordings in experimental
session EC2 B105.
38
Figure A.36: ERP for electrode 134 computed from recordings in experimental
session EC2 B105.
39
Figure A.37: ERP for electrode 135 computed from recordings in experimental
session EC2 B105.
40
Figure A.38: ERP for electrode 136 computed from recordings in experimental
session EC2 B105.
41
Figure A.39: ERP for electrode 137 computed from recordings in experimental
session EC2 B105.
42
Figure A.40: ERP for electrode 138 computed from recordings in experimental
session EC2 B105.
43
Figure A.41: ERP for electrode 139 computed from recordings in experimental
session EC2 B105.
44
Figure A.42: ERP for electrode 140 computed from recordings in experimental
session EC2 B105.
45
Figure A.43: ERP for electrode 141 computed from recordings in experimental
session EC2 B105.
46
Figure A.44: Strongest PAC occurs over the vSMC. The 50 electrodes with
largest MI (between 0.5 seconds before and 0.6 seconds after the CV transition,
at the frequency of CV production, in the experimental session EC2 B105) are
highlighted in color.
47
|
1612.06975 | 1 | 1612 | 2016-12-21T05:30:11 | Similarities and differences between stimulus tuning in the inferotemporal visual cortex and convolutional networks | [
"q-bio.NC"
] | Deep convolutional neural networks (CNNs) trained for object classification have a number of striking similarities with the primate ventral visual stream. In particular, activity in early, intermediate, and late layers is closely related to activity in V1, V4, and the inferotemporal cortex (IT). This study further compares activity in late layers of object-classification CNNs to activity patterns reported in the IT electrophysiology literature. There are a number of close similarities, including the distributions of population response sparseness across stimuli, and the distribution of size tuning bandwidth. Statisics of scale invariance, responses to clutter and occlusion, and orientation tuning are less similar. Statistics of object selectivity are quite different. These results agree with recent studies that highlight strong parallels between object-categorization CNNs and the ventral stream, and also highlight differences that could perhaps be reduced in future CNNs. | q-bio.NC | q-bio | Similarities and differences between stimulus tuning in the
inferotemporal visual cortex and convolutional networks.
Bryan Tripp, University of Waterloo, Canada
Submitted on December 1, 2016 to IJCNN 2017.
2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all
other uses, in any current or future media, including reprinting/republishing this material for advertising
or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
6
1
0
2
c
e
D
1
2
]
.
C
N
o
i
b
-
q
[
1
v
5
7
9
6
0
.
2
1
6
1
:
v
i
X
r
a
Similarities and differences between stimulus tuning
in the inferotemporal visual cortex and
convolutional networks.
Department of Systems Design Engineering & Centre for Theoretical Neuroscience
Bryan P. Tripp
Waterloo, Ontario, Canada
Email: [email protected]
Abstract -- Deep convolutional neural networks (CNNs) trained
for object classification have a number of striking similarities
with the primate ventral visual stream. In particular, activity in
early, intermediate, and late layers is closely related to activity
in V1, V4, and the inferotemporal cortex (IT). This study
further compares activity in late layers of object-classification
CNNs to activity patterns reported in the IT electrophysiology
literature. There are a number of close similarities, including the
distributions of population response sparseness across stimuli,
and the distribution of size tuning bandwidth. Statisics of scale
invariance, responses to clutter and occlusion, and orientation
tuning are less similar. Statistics of object selectivity are quite
different. These results agree with recent studies that highlight
strong parallels between object-categorization CNNs and the
ventral stream, and also highlight differences that could perhaps
be reduced in future CNNs.
I. INTRODUCTION
Deep convolutional neural networks (CNNs) that are trained
for object categorization have activity that resembles that of
the ventral visual stream in several respects. For example,
early layers often exhibit Gabor-like receptive fields similar to
V1 [1], and separation of early layers across multiple GPUs
has resulted in separation of color and edge encoding [2],
analogous to V1 blobs and interstripes.
Furthermore, internal activity in later layers of such CNNs
is closely related to activity later in the primate ventral visual
stream. [3] showed that much of the variance in activity of
IT neurons (up to about 30%) could be accounted for by
multilinear regression with the responses of units in the CNN's
second-last layers. They also showed that these linear predic-
tions had IT-like correlations related to object categories [3].
Models that have been specifically developed to approximate
the ventral stream, including HMAX [4] and VisNet [5], have
lacked such realistic correlations [6]. [3] also showed that
activity in intermediate layers of these CNNs predicted activity
of neurons in V4, which is a major source of input to IT [7].
to scale and
translation changes [8], and a few are also invariant to ori-
entation changes. Consistent with this, [1] found that
the
vector of population responses in the second-last later of a
CNN changed relatively little with with position, size, and
orientation, compared to responses in the first layer.
Many IT neurons are relatively invariant
More recently, [9] found that IT neurons and units of
object-classification CNNs provided comparable information
about category-orthogonal object properties, including object
position and orientation.
Importantly, CNNs perform similarly to humans [2], [1] in
"core object recognition", a rapid feedforward process that
allows primates to recognize objects in natural scenes after
seeing them very briefly [10], on similar timescales to inter-
saccade intervals.
Core object recognition is a sophisticated function, and
a major function of the ventral visual stream. It is highly
relevant to neuroscience that this this function can now be
performed by artificial systems. CNNs can not be considered
realistic models of the ventral stream, because they learn in
very different ways, consist of much simpler units, typically
lack lateral connections, etc. [11]. However,
that
they have realistic core object recognition capabilities, and
also exhibit internal representations that resemble those in
the ventral stream, constrains ideas about the roles of these
missing details in vision.
the fact
A. Purpose of this study
Linear combinations of CNN activities are currently the
best predictors of IT activity in high-variation stimulus condi-
tions, but they leave about half the variance of IT responses
unaccounted for. It would be interesting to know whether
CNNs predict certain factors of variation in IT better than
others. Also, the CNN activities themselves (as opposed to
optimal linear combinations of them) have somewhat different
correlation patterns [3], [6].
The purpose of the present study is to compare the response
properties of deep object-recognition CNNs and IT neurons,
along several dimensions of stimulus variation that have been
explored in IT. Remarkably, despite the optimization of both
IT and CNNs for natural scenes, both systems discriminate
the same simplified stimuli (e.g. line drawings) in a similar
manner (see Results). This study focuses on tuning curves of
individual units, allowing more direct comparison with IT than
previous investigations of population response vectors [1].
To summarize the results, many response properties of the
CNNs are strikingly similar to previously reported responses
of IT neurons. These including the distributions of population
sparseness across stimuli, and the distribution of size-tuning
bandwidth. Other properties, including responses to partial
occlusion and clutter, orientation tuning curves, and patterns
of size invariance, were somewhat different. Distributions of
object selectivities over neurons were quite different. Along
with differences in correlations of IT responses within and
between object categories [3], [6], these results help to clarify
the relationships between stimulus representation in object-
classification CNNs and IT.
Preliminary versions of some of these analyses appear in
[12].
II. METHODS
Some of the figures replot responses of IT neurons from
the experimental literature. The data points for these plots
were extracted from published figures using Web Plot Digitizer
(http://arohatgi.info/WebPlotDigitizer/).
A. Networks
We study responses of two well-known deep convolutional
networks. Both networks were trained to classify objects in
natural scenes for the ImageNet Large-Scale Visual Recog-
nition Challenge [13]. One was Alexnet [2]. This network
has previously been compared with IT in [14] and [6]. The
other network, VGG-16 [15], has more layers and smaller
kernels. Both of the implementations used here were taken
from https://github.com/heuritech/convnets-keras. The Alexnet
implementation was adapted from a version trained for [16].
The VGG-16 implementation was adapted from a version
from https://gist.github.com/baraldilorenzo, which was in turn
adapted from the version provided by [15].
Previous comparisons of CNNs with IT [3], [6], [9] have
focused on the second-last network layer, a fully-connected
layer that precedes the final softmax classification layer. This
layer is in a similar position to the inferotemporal cortex (IT)
in the ventral stream, the neurons of which are important
for visual object categorization, but also contain much non-
category information, e.g. [8], [9]. For comprehensiveness,
we also consider responses from the surrounding layers, i.e.
the last and third-last layers, which turn out to more closely
resemble IT in some respects.
B. Stimulus Images
Stimulus images were obtained from several sources. To
investigate stimulus selectivity and population response sparse-
ness, we used a set of 806 images developed by [17] and
used for the same purpose in IT. The images were ex-
tracted directly from their supplementary materials file. The
extracted images had a lower resolution (60 by 60 pix-
els) than the authors reported using (125 by 125 pixels).
However, the objects were easily identifiable. The extracted
images were embedded within a larger gray background. The
extracted images themselves had light borders and some-
what noisy gray backgrounds. The borders were removed
by omitting the outer two pixels, and variations were re-
duced in the backgrounds by setting to a uniform gray value
([red,green,blue]=[128, 128, 128]T) any vector with a cartesian
distance of less than 15 from this value.
A few additional higher-resolution images (banana, car,
shoe, etc.) were downloaded from various internet sources
and given uniform gray backgrounds. These images were
used to examine orientation, position, and size tuning. Images
for studying responses to rotation-in-depth were rendered in
SketchUp. 3D scooter and human head models were obtained
from the SketchUp 3D Warehouse. A wire-like object model,
as in [18] was built
in OpenSCAD. Other images were
generated using custom code in order to approximate stimuli in
[19]. Finally, some stimulus images were copied from figures
in [8] and [20].
III. RESULTS
A. Stimulus selectivity and population sparseness
Two fundamental statistical properties of IT population
responses are 1) the sparseness of responses across the pop-
ulation to various stimuli, and 2) the sparseness of individual
neurons' responses evaluated over multiple stimuli (sometimes
called stimulus selectivity). Both have been described in
terms of excess kurtosis (kurtosis in excess of that of the
Gaussian distribution). [17] compared stimulus selectivity and
population sparseness in a large dataset that included responses
of 674 IT neurons to 806 different images. They found that
distributions of selectivity and sparseness had a similar shape,
and that sparseness exceeded selectivity. They found that this
difference could be reproduced in a simple statistical model, as
a result of non-uniform mean spike rates across the population.
Datasets similar to that of [17] were collected from the
CNNs, using approximately the same set of images (see
Methods). Figure 1 plots selectivity vs. sparseness of units
in the final three layers of both CNNs (Alexnet and VGG-16).
674 units were chosen from each layer, to match the size of the
dataset in [17]. The 674 units with the highest peak activations
in response to the stimulus images were chosen from each
layer. The sparseness of the CNN activity was similar to that
of the IT neurons. However, in contrast to IT neurons, each of
the CNN layers exhibited higher selectivity than sparseness.
The sparseness of the CNN activity was disproportionately
due to very rare activation of some units, while the remaining
units were activated less sparsely than average. The sparseness
distributions had a similar shape to that of [17], but
the
selectivity distributions had longer tails (not shown).
B. Orientation Tuning
The responses of most IT neurons varies strongly with
orientation in each dimension [18], although a small fraction
(< 1◦ in [18]) are essentially orientation-invariant. Responses
are typically strongest around a single preferred orientation
unless the object is rotationally symmetric. Orientation tun-
ing curves often resemble gaussian functions. [18] reported
somewhat narrower tuning curves for rotation in depth than
rotation in the image plane. For neurons that responded to
faces, [21] reported a fairly uniform distribution of preferred
viewing angles between left and right profiles (their Figure
11).
s
s
e
n
e
s
r
a
p
S
35
30
25
20
15
10
5
0
0
IT
out-1
out
out-2
out
out-1
out-2
5
10 15 20 25 30 35
Selectivity
Fig. 1. Sparseness vs. selectivity of IT neurons (from [17]), and of CNN
units responding to a similar group of stimuli. Sparseness exceeds selectivity
in IT [17], whereas selectivity consistently exceeds sparseness, in each layer
of both CNNs. The blue circles are Alexnet layers and the green squares are
VGG-16 layers. The labels indicate the distance from the output, i.e. "out" is
the last layer, "out-1" is the second-last, and "out-2" is the third-last.
Figure 2 shows tuning curves over rotations in depth about
the vertical axis. Figure 2A shows example views of three
stimulus objects: a "wire-like" object, similar to those used
in [18]; a more complex inanimate object
that might be
encountered in life (a scooter); and a human head. Figure
2B shows examples of IT-neuron orientation-tuning curves for
a wire-like object (replotted from [18]) and a human head
(replotted from [22]). The wire (left) elicited narrow, roughly
Gaussian tuning curves from most neurons [18]. A minority
of neurons responded strongly to roughly mirror-symmetric
views separated by 180deg. Finally, a very small minority of
neurons had approximately view-invariant responses.
The left columns of Figure 2C and D show responses
of units in Alexnet and VGG-16, respectively. In contrast
with IT neurons, the CNN units typically have multimodal
tuning curves. Units in the last two layers have broad tuning.
However, the third-last layer of Alexnet has several fairly
narrow, roughly Gaussian tuning curves that resemble view-
selective neurons in [18], except that they are bimodal (perhaps
due to use of flipped images to enhance the dataset in training).
As the wire-like object is unlike objects that the networks
encountered in training, orientation tuning curves were also
plotted for an object from one of the ImageNet categories (a
scooter). As with the wire, most responses in both Alexnet and
VGG-16 were broad and multimodal. There are no examples
of narrow uni- or bimodal Gaussian tuning in this column.
The right column of 2B shows responses of several cells
in face-selective regions of IT from [22]. Some of these cells
preferred front views, and some preferred profile views, but
responded similarly to left and right profiles. Interestingly,
CNN tuning curves also showed these two patterns, although
with somewhat broader tuning.
In [18], rotations in the image plane produced responses that
were similar to rotation in depth (many neurons had roughly
Gaussian tuning) except that tuning was somewhat broader.
8
7
6
5
4
3
2
1
0
1
0 50 100150200250300350
6
5
4
3
2
1
0
1
0 50 100150200250300350
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
5
4
3
2
1
0
1
0 50 100150200250300350
6
5
4
3
2
1
0
1
0 50 100150200250300350
Angle (degrees)
Angle (degrees)
Fig. 3. Tuning curves for image-plane orientation of staple, shoe, car, and
banana images. In each case, the responses of the first ten units with peak
responses > 2 are shown. The plotted tuning curves were smoothed by
averaging over groups of four neighbouring points, so that the different curves
are easier to distinguish, according to the Gestalt law of good continuation.
Figure 3 shows examples of image-plane orientation tuning
curves from the second-last layer of Alexnet. In contrast with
typical IT responses to image-plane rotation, most of the CNN
responses are multimodal.
C. Size Tuning and Invariance
IT neurons are also tuned for size. [23] studied the distri-
bution of size tuning bandwidths, i.e. the ratio between the
larger and smaller sizes at which the response drops to half
it's peak. The mode of this distribution was approximately 1.5
octaves (their Figure 8), and bandwidths of up to 4 octaves
were reported. Figure 4B and C shows histograms of size
tuning bandwidths in Alexnet and VGG-16, respectively. They
are both quite similar to the distribution in IT. For example
the Alexnet histogram has a mode around 1 (versus 1.5 in
IT), and values up to 4.6 (vs. a few values at or above 4, the
highest value measured, in IT).
Responses of the majority of IT neurons exhibit size in-
variance [8]. Spike rates vary with stimulus size, but
the
relative spike rates in response to different stimuli are similar
regardless of sizes. A prototypical example is shown in Figure
5A. The neuron's tuning curve over six different shapes is
roughly maintained over different stimulus scales. However,
the gain of this curve depends on the scale. Figure 5D shows
responses of units in the second-last layer of Alexnet. The
stimulus shapes and relative scales are the same as in [8].
Similar to the IT neurons, many of the CNN units' tuning
curves have similar shapes when the stimuli are shown at
different scales. A 'c' in the upper-right of each plot indicates
A
B
C
D
s
n
o
r
u
e
N
T
I
60
50
40
30
20
10
e
s
n
o
p
s
e
R
0
180
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
e
s
n
o
p
s
e
R
t
e
n
x
e
A
l
6
1
-
G
G
V
180
4
3
2
1
0
1
2
5
4
3
2
1
0
1
180
"
25
20
15
10
5
0
5
180
*
(
180
4
3
2
1
0
1
24
3
1
7
6
5
4
3
2
1
0
1
180
A
?
16
14
12
10
8
6
4
2
0
2
180
J
H
Staple
Scooter
Head
0°
30°
0°
30°
0°
30°
E
18
16
14
12
10
8
6
4
2
0
180
180
60
60
Angle (degrees)
180
60
60
Angle (degrees)
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.5
1.0
1.5
180
8
7
6
5
4
3
2
1
0
1
180
%
#
25
20
15
10
5
0
5
180
-
+
180
5
4
3
2
1
0
1
29
38
7
5
10
8
6
4
2
0
2
180
D
B
20
15
10
5
0
5
180
M
K
60
60
180
60
!
60
180
60
60
180
)
Angle (degrees)
60
2
60
180
60
@
60
180
60
60
180
I
Angle (degrees)
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.5
180
8
6
4
2
0
180
&
30
25
20
15
10
5
0
5
180
0
.
180
5
4
3
2
1
0
1
2>
3=
<
:
5
4
3
2
1
0
1
180
G
E
12
10
8
6
4
2
0
2
180
P
N
60
60
180
60
$
60
180
60
60
180
,
Angle (degrees)
60
6
60
180
60
C
60
180
60
60
180
L
Angle (degrees)
60
60
180
F
60
'
60
180
60
60
180
/
Angle (degrees)
60
;
60
180
60
F
60
180
60
60
180
O
Angle (degrees)
s
l
l
e
c
f
o
n
o
i
t
c
a
r
F
s
l
l
e
c
f
o
n
o
i
t
c
a
r
F
s
l
l
e
c
f
o
n
o
i
t
c
a
r
F
s
t
i
n
u
f
o
n
o
i
t
c
a
r
F
s
t
i
n
u
f
o
n
o
i
t
c
a
r
F
s
t
i
n
u
f
o
n
o
i
t
c
a
r
F
AM
0.5
0.4
0.3
0.2
0.1
0.5 0.0
0.5
1.0
AL
0.5 0.0
0.5
1.0
ML/MF
1.0
0.0
0.5
0.4
0.3
0.2
0.1
1.0
0.0
0.5
0.4
0.3
0.2
0.1
0.0
1.0
0.5 0.0
0.5
1.0
Head orient at ion t uning dept h
output
0.5
0.4
0.3
0.2
0.1
0.5 0.0
0.5
1.0
output-1
0.5 0.0
0.5
1.0
output-2
1.0
0.0
0.5
0.4
0.3
0.2
0.1
1.0
0.0
0.5
0.4
0.3
0.2
0.1
0.0
1.0
0.5 0.0
0.5
1.0
Head orient at ion t uning dept h
Fig. 2. Tuning curves for rotation-in-depth of a wire-like object (first column), a scooter (second column), and a human head (third column). A) Example
views that were used as inputs to the CNNs. B) Tuning curves of IT neurons in response to similar (not identical) objects. Responses to a wire-left object
(left) and head (right) are redrawn from data in [18] and [22], respectively. C) Responses of units in the final three layers of Alexnet. The columns correspond
to wire, scooter, and head stimuli. The top row shows responses from the output layer, and the other rows show responses from the immediately preceeding
layers. In each case the responses of the first ten units with peak responses greater than two are shown, and the tuning curves are smoothed by averaging
over groups of three neighbouring points. D) Same format as C for VGG-16. E) Histograms of tuning depth of IT neurons in different face-responsive areas,
replotted from [22]. The tuning depth is the frontal response minus the profile response, divided by their sum. Depths are closer to zero (tuning curves are
flatter) in the anterior medial area (AM) relative to more posterior areas. F) Depth histograms of head tuning curves from the final three layers of Alexnet.
Analogous to the neural data, depth is lowest in the output layer. In contrast with the neural data, the distributions in other layers are multimodal.
A
y
c
n
e
u
q
e
r
F
D
9
8
7
6
5
4
3
2
1
0
0 1 2 3 4 5 6
Size BW (Oct aves)
B
140
120
100
80
60
40
20
0
0 1 2 3 4 5 6
y
c
n
e
u
q
e
r
F
C
160
140
120
100
80
60
40
20
0
y
c
n
e
u
q
e
r
F
0 1 2 3 4 5 6
Size BW (Oct aves)
Size BW (Oct aves)
250
200
150
100
50
0
− 50
− 100
1
A
)
s
/
s
e
k
i
p
s
(
e
s
n
o
p
s
e
R
D
largest
B
1.0
0.8
0.6
0.4
0.2
n
o
i
t
a
l
e
r
r
o
Q
n
a
e
M
Alexnet
VGG-16
smallest
4
3
2
St im ulus #
5
6
0.0
− 2
Dist an
c
− 1
0
e from out put (layers)
C
e
nT
e
r
e
f
e
r
p
h
t
i
w
n
o
i
t
aS
r
F
1.0
0.8
0.6
0.4
0.2
0.0
− 2
Dist an
R
− 1
0
e from out put (layers)
Fig. 4. A) Histogram of half-magnitude size tuning bandwidths, replotted
from [23]. The range of stimulus scales did not fully span the half-magnitude
ranges of the units plotted in red. B) Histogram of half-magnitude bandwidths
for Alexnet. Stimuli included scaled versions of banana, shoe, and car images.
For each object, the 200 units with the strongest peak responses were included
in the histogram. C) As B, but for VGG-16. D) Examples of banana images
at some of the different scales used with the CNNs.
that the mean correlation between curves is greater than 0.75.
However, only a small minority of CNN units have a clear
preferred size (marked with 'p' in the upper left) like the
canonical example of [8].
Panels 5B and C show that these results are similar for other
layers of Alexnet and VGG-16. Mean correlations are well
below those of 5A, and only a small fraction of CNN units
in any layer show a clear preference for a certain size that is
maintained across multiple shapes. Shape and size tuning may
be somewhat less separable in the CNNs than in IT.
D. Position Tuning and Invariance
Many IT neurons exhibit translation invariance as well as
size invariance [23]. Figure 6A shows examples of response
variations with horizontal stimulus position, in the second-
last layer of Alexnet. In each plot, the first 30 units with
responses > 2 are shown. As in [24], many of the tuning
curves are roughly gaussian, and most units prefer centrally
located stimuli.
The dispersion of preferences around zero (the fovea) cannot
be directly compared, because the mapping between pixels and
degrees visual angle is not well defined (it is not uniform in the
ImageNet training data). However, it is possible to compare
the ratio of this dispersion to the widths of the tuning curves.
Figure 6B plots ratios of standard deviation of tuning curve
centres of mass, divided by mean tuning curve width (distance
between points on each side of the centre of mass that fall
below half the peak). The dashed line is this ratio taken from
the IT data in [24]. The blue and green lines are from Alexnet
and VGG-16, respectively. As a fraction of tuning curve width,
the position-tuning centres in the last layers of both networks
are much more narrowly dispersed than IT position-tuning
curves. However, the second-last layer of Alexnet and third-
last layer of VGG-16 are similar to the IT neurons in this
respect.
Figure 7 shows that (in contrast with size invariance) the
CNNs have similar position invariance to IT. The last and
c
c
c
c
c
c
c
p
c
c
p
c
c
c
c
c
c
c p
c
c
c
c
c
c
c
c
Fig. 5. Size invariance in the CNNs, examined with stimuli from [8]. A)
Responses of an example IT neuron, replotted from [8]. The horizontal axis
spans different stimulus shapes, the vertical axis is the mean spike rate. The
red, green, and blue curves correspond to small, medium, and large versions
of the stimuli. B) The dashed line is the correlation between the tuning curves
in A. The blue and green are mean correlations of similar tuning curves in
Alexnet and VGG-16. The correlations are plotted versus layer, with the output
layer on the right. C) The fraction of units in each CNN layer that had the
same scale preference for all shapes (like the example neuron in A, in which
the red curve does not cross the others). D) Examples of tuning curves in the
second-last layer of Alexnet, with the same format as A. Many of the units'
responses are correlated across scales (those with mean correlations ¿0.75 are
marked 'c'). However, few have a uniform scale preference across shapes
(marked 'p'). The units in B-D were the first 64 units in each layer with peak
responses > 2 over all stimuli.
second-last CNN layers have average correlations in these
tuning curves that are similar to the example given by [8] (their
Figure 2D), which is consistent with the population averages of
[24] (their Figure 10C). A substantial fraction of the units have
a position preference that is consistent across shapes, with at
most one shape at which the preferred position is inconsistent
(as in the IT data in 7A).
E. Occlusion
[19] studied variations in the responses of IT neurons as
shape stimuli were partially occluded. Neurons maintained
their shape preferences, but responded less vigorously as the
shapes were occluded more fully.
Figure 8 shows results of a similar experiment with the
CNN. For this experiment, following [19], several shape
outlines were drawn on top of a noisy background, and these
shapes were occluded by replacing blocks of pixels with black.
As with IT neurons, shape preference order tended to be
maintained with increasing occlusion, and the amplitude of
the preferences was reduced with high occlusion. 8A shows
A
6
5
4
W
2
1
0
− 1
−
8
6_
5
4
2^
1
0
− 1
−
[
{
st aple
shoe
5
4
Z
2
1
0
5 − 25
25
U
5
V
− 1
−
X
5 − 25
25
5
Y
ar
`
5 − 25
25
ffset (pi
els)
x
e
s
n
o
p
s
e
R
6g
5
4
2f
1
0
− 1
−
5
\
anana
h
a
5 − 25
25
ffset (pi
els)
e
d
5
b
]
h
t
io
w
n
a
e
m
/
s
r
e
t
n
e
f
n
o
Alexnet
0.5
0.4
0.
j
0.2
0.1
IT
ons
pqrs
tuvwyz
0.0
− 2
− 1
0
Dist an
i
e from out put (layers)
400
50
00
250
200
150
100
50
0
− 50
1
A
)
s
/
s
e
k
i
p
s
(
e
s
n
o
p
s
e
R
«
p
2
p
p
©
1.0
n
o
i
t
a
l
e
r
r
o
n
a
e
M
0.8
0.6
0.4
0.2
¬-®¯°±
²³´µ¶·
¸¹º»¼
½¾¿ÀÁ
ÂÃÄÅ
VGG-16
Alexnet
4
5
6
t im ulus #
p
¡
0.0
− 2
Dist an
− 1
0
e from out put (layers)
p
p
p
p
¢
¥
e
ª
n~
e
r
e
f
e
r
p
h
t
i
w
n
o
i
t
a}
r
F
1.0
0.8
0.6
0.4
0.2
Alexnet
VGG-16
− 1
0
e from out put (layers)
0.0
− 2
Dist an
¦
£
§
¤
Fig. 6. A) Responses to staple, shoe, car, and banana images translated to
different horizontal positions within a gray background. The offset is the
distance from centre, in an image that was 256 pixels wide. At the largest
offsets, the stimuli nearly reached the edge of the background. B) Ratios
of standard deviation of centres of mass, over mean half-width, of position-
tuning curves. The dashed line is the ratio from data in [24]. The blue and
green lines are calculated from the responses of Alexnet and VGG-16 to the
translated staple, shoe, car, and banana images.
Fig. 7. Translation invariance in CNN units. Both the plot format and CNN
unit selection procedure are the same as in Figure 5. A) Data replotted from
an example neuron in [8]. B) Mean correlations of tuning curves in Alexnet
(blue) and VGG-16 (green) compared with that of the curves in panel A
(dashed). C) Fraction of units in each layer in which the same position was
preferred for all but at most one shape, like the IT neuron shown in panel A.
D) Example tuning curves from the second-last layer of Alexnet, in the same
format as A. The different tuning curves correspond to stimuli at the centre,
and translated up, down, left, and right from centre. The translation distance
relative to stimulus size was roughly matched to that in [8].
tuning curves over different shapes, with different occlusion
levels, for the second-last layer of Alexnet (left), the second-
last layer of VGG-16 (centre), and IT neurons from [19]. The
stimuli used to generate these plots were black outlines, and
the occluders were black squares (see example in Figure 8B).
In these examples, partial occlusion attenuated the CNN unit
responses more strongly than the IT responses. For example,
at 50% occlusion, VGG-16 units responded uniformly and
minimally regardless of the stimulus shape.
Figure 8B plots mean normalized responses to units' pre-
ferred shapes, as a function of the unoccluded percentage
of the image. Compared to the IT neurons (black line),
responses in each layer of Alexnet (left) and VGG-16 (right)
are attenuated much more by occlusion.
The CNN and IT responses cannot be compared directly,
due to temporal evolution of the IT stimulus over time that
cannot be reproduced in the CNNs. For example, in the IT
experiment, the shape appeared after the occluder, providing
an additional segmentation cue. To provide a comparable cue
for the CNNs, the simulations were repeated with the shapes
and occluders in different colors (right panel of 8C). This
reduced the attenuation of CNN responses with occlusion.
Another complication is that the IT data shown in 8A (right) is
from experiments in which the occlusion patterns moved. To
approximate such motion, the CNN experiment was repeated
with occluders that were blurred over a similar distance
relative to the stimulus sizes (right panel of 8D). This stimulus
change also reduced the attenuation of CNN responses with
occlusion. In this condition, Alexnet's occlusion responses
were similar to those of IT neurons. Notably however, the
IT neurons responded similarly to 50% occlusion regardless
of whether it was moving or stationary [19].
Relatedly, occlusion of certain parts of objects impairs the
performance of CNNs more than occlusion of other parts [1],
consistent with responses in IT [25].
F. Clutter
[26] compared the responses of IT neurons to isolated
stimulus objects with responses to multiple objects. Responses
to pairs of objects were closely clustered around the average
response (rather than the sum) of responses to each object
individually 9A.
A similar experiment was performed with the CNN, using
eights images from the dataset of [17]. Images were shown at
two locations (above and below centre) in a gray background,
as in the example of Figure 9D. Four images were used in
the upper location, and four different images were used in
k
l
m
A
e
s
n
o
p
s
e
R
n
a
e
M
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
− 0.2
1
2
Alexnet
VGG-16
IT Neurons
A
ÐÑ20
Ò50
Ó
0
Õ
Ô
e
s
n
o
p
s
e
R
n
a
e
M
4.0
.5
.0
Ù
Ø
2.5
2.0
1.5
1.0
0.5
5
4
6
Î
Shape rank
0.0
1
2
8
Ï
ÚÛ20
Ü50
Ý
0
ß
Þ
25
20
15
10
5
0
e
s
n
o
p
s
e
R
n
a
e
M
− 5
1
2
8
×
ÈÉ20
Ê50
Ë
0
Í
Ì
8
Ç
5
4
6
AE
Shape rank
5
4
6
Ö
Shape rank
B
C
D
e
s
n
o
p
s
e
R
l
eae
iå
a
m
r
o
ä
n
a
e
M
e
s
n
o
p
s
e
R
l
eò
iñ
a
m
r
o
ð
n
a
e
M
e
s
n
o
p
s
e
R
l
eþ
iý
a
m
r
o
ü
n
a
e
M
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
çèlast layer
last
2n
é
last
ë
r
ê
20
40
60
er
à
á
ent
isi
â
80
le
ã
100
óôlast layer
last
2n
õ
last
÷
r
ö
20
40
60
er
ì
í
ent
isi
î
80
le
ï
100
ÿTlast layer
last
2n
d
last
r
3
20
40
60
er
ø
ù
ent
isi
ú
80
le
û
100
e
s
n
o
p
s
e
R
l
e
iz
a
m
r
o
N
n
a
e
M
e
s
n
o
p
s
e
R
l
e
i
a
m
r
o
n
a
e
M
e
s
n
o
p
s
e
R
l
e
i
a
m
r
o
n
a
e
M
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
Ilast layer
last
2n
last
r
20
40
60
er
P
c
ent
isi
V
80
le
b
100
last layer
last
2n
last
r
20
40
60
er
ent
isi
80
le
100
last layer
last
2n
last
r
20
40
60
er
ent
isi
80
le
100
Fig. 8. Results of an occlusion experiment. A) Shape tuning curves averaged
over the 500 units with the strongest responses from the second-last layers of
Alexnet (left) and VGG-16 (center), and data from IT neurons in [19] (right).
The CNN responses are averages over ten random occlusion patterns (i.e.
different random placements of occluding blocks). An example stimulus is
shown on the right of panel B. The background, shapes, and occluders were
intended to approximate those in [19]. The lines correspond to no occulusion,
and occlusion of 20%, 50%, and 90% of the image. These responses have
baseline activity subtracted. In [19], baseline activity was taken from an inter-
stimulus delay period. In the CNNs, baseline activity is the response to a
black image. B) Responses to preferred stimuli at different occlusion levels.
The black, blue, and green lines are IT data from [19], and unit responses
from Alexnet and VGG-16, respectively. These responses were normalized
by subtracting the mean response to the least-preferred stimuli, and dividing
by the response to the unoccluded preferred stimuli. An example star-shaped
stimulus with 20% occlusion is shown on the right. C) As B, but with stimulus
shapes in red rather than black. D) As B, but with the occlusion patterns
blurred to approximate motion of the occluder in [19].
IT
ons
#$%&
B
s
r
i
a
p
t
e"
o
o
t
e
s
n
o
p
s
e
R
16
14
12
10
8
6
4
2
0
0
Ale
x'() *+
d-Last La
y,-
2
4
6
8
10 12 14 16
Sum of responses t o sin
le o
t s
e
j
g
C
Alexnet
D
output
output-1
output-2
0
pi/4
0
pi/4
0
pi/4
VGG-16
output
output-1
output-2
pi/4
0
Angle of (sum, pair) point from horizontal
pi/4
pi/4
0
0
Fig. 9. Effects of low levels of clutter. A) Responses of IT neurons to pairs of
objects (vertical axis) versus sums of their responses to each object presented
alone (reproduced with permission from [26]). Red dots indicate that each
stimulus elicited a response greater than the background rate, and blue dots
indicate that only one stimulus did so. The responses to pairs of objects are
mostly close to the average of responses to each object alone. B) Responses
of units in the second-last layer of Alexnet to pairs of objects versus sums
of responses to each object alone. Compared to IT, the CNN responses to
paired objects are more varied, and are typically greater than the mean of the
single-object responses. The plot includes responses of the 100 units with the
greatest peak responses across pair conditions. A random selection of units
produced a similar pattern but with a higher density of points in the lower left
corner. C) Angles of points (counter-clockwise from rightward) in the same
space as the scatterplots in panels A an B. Each histogram corresponds to
a layer of one of the CNNs. The red lines indicate the angle arctan(1/2).
This corresponds to points on the solid lines in A and B, i.e. responses to
paired objects that equal the average of responses to individual objects. The
average angle for the IT responses in A is approximately arctan(1/2). D)
An example of a pair of images.
of π/4 means that the unit's response to paired images equals
the sum of responses to each image alone. An angle of
arctan(1/2) (red lines) means that
the unit's response to
paired images equals the average of responses to each image
alone. In each layer of both networks, the mean and mode
of the angles is between arctan(1/2) and π/4, suggesting
that the CNN responses are not normalized as strongly as IT
responses. This is true even in the output layers, which have
softmax nonlinearities.
G. Strong Responses to Simplified Images
the lower location. Unit responses were recorded with each
of the upper and lower stimuli alone, and also with each of
the 16 possible pairs of upper and lower stimuli. As shown in
9B, responses of the second-last layer of Alexnet to pairs of
images did not cluster tightly around the average of responses
to corresponding isolated images (in contrast with [26]), but
were more broadly distributed, and typically intermediate to
the sum and mean of the isolated responses.
Figure 9B shows histograms of the angles of these points,
counter-clockwise from the positive horizontal axis. An angle
Tanaka and colleagues found that many IT neurons that
responded strongly to complex images also responded strongly
to highly simplified versions of these images [20]. After
showing a monkey many objects in a standard set, and
determining the most effective stimulus from the set (i.e. the
one that elicited the strongest response), the most effective
stimulus was progressively simplified to determine which
features were necessary for maximal activation. For example,
a cell that responded maximally to a sketch of a striped cat
also responded maximally to a sketch of a striped ball on
!
top of another striped ball. Response strength often increased
with simplification. IT neurons typically responded maximally
to moderately complex stimuli that were more complex than
primitive shapes, but less complex than natural objects.
The image-simplification process was manual and not fully
specified, so it is not clear how accurately it could be re-
produced for CNN units. Instead, the CNNs were tested with
twelve examples of complex and simplified images provided in
[20]. The CNN was first shown the twelve complex images and
56 additional images taken from ShapeNet (total 68 images,
similar to the numbers of stimuli described in [27]). For each
of the 12 complex images, units were identified that responded
maximally or nearly maximally (within 5% of the mean unit-
wise maximum) to that image among the 68 stimuli. Figure
10A shows distributions of response strengths of these units
when shown the corresponding simplified images, normalized
to the response response to the complex image. The responses
to simplified images are broadly distributed, but
in many
cases, a substantial fraction of the units responded at least
as strongly to the simplified image as to the complex image.
An exception was that very few of the units that responded
strongly to a person wearing a lab coat responded as strongly
to its simplification (a black circle over a white circle). This
exception was consistent across other Alexnet layers, and
across the VGG-16 layers.
There are infinite ways in which an image could potentially
be simplified, so it is striking that the particular simplifications
that elicited maximal responses from the IT neurons in [20]
also elicited maximal responses from some of the 4098 units
in this layer of Alexnet.
Figure 10B shows the fractions of neurons in each layer
of each network that responded maximally to the simplified
images in [20]. Interestingly, a few of the units in each layer
responded maximally to simplifications of other objects (e.g. a
unit that responded maximally to a complex bread image may
also have responded maximally to a simplified cat). However,
the fractions of maximal responses to simplification of the
most effective stimulus were uniformly greater.
IV. DISCUSSION
Past studies [1], [3], [6], [9] have shown that activity in
the next-to-last layer of object-classification CNNs is closely
related to that in the primate inferotemporal cortex (IT). The
current study examined properties of the tuning of individual
neurons to a number of stimulus variations that have been used
to understand representation in IT. The results highlight some
specific similarities between the CNN units and IT neurons,
including particularly the distributions of population response
sparseness, and size-tuning bandwidth. A number of other
tuning properties had both similarities and differences with
IT. These include scale and translation invariance, orientation
tuning, and responses to occlusion. Responses to clutter were
also systematically different. The distributions of stimulus
selectivity were quite different. Overall, in agreement with past
studies, the similarities are striking, particularly in light of the
han
:
1
2
skew er
1
2
rea
<
;
2
1
at
=
all
7
1
2
hayst a
k
8
1
2
pom
pom
9
1
2
apple
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
A
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
0.8
0.4
0.0
− 0.4
− 0.8
0
s
t
i
n
aJ
r
F
plant
1
2
hin
le
4
2
1
2
e
u
5
6
1
2
person
0.8
0.4
0.0
− 0.4
− 0.8
R
1
0
1
atio of response to simple
2
2
0
1
2
x stimulus
>?@ BCDEFH
B
1
>
o
i
t
aW
r
F
16
.G/0
0.20
0.15
0.10
0.05
le
1
A
net
0.20
0.15
0.10
0.05
0.00
− 2.0 − 1.5 − 1.0 − 0.5 0.0
0.00
− 2.0 − 1.5 − 1.0 − 0.5 0.0
Distan
^_
from output (
l`aefhi
Fig. 10. Responses of CNN units to simplified versions of most effective
stimuli. A) Responses of the second-last layer of Alexnet. Each histogram
corresponds to an example pair of complex and simplified images from Figure
1 of [20]. Neurons were found that responded maximally to each of the
complex stimuli. The blue histograms show the distribution of strength of
responses to the corresponding simplified stimuli (normalized by strength
of response to the most effective complex stimulus). As a control,
the
gray histograms show the distribution of strength of responses to other
(non-matching) simplified stimuli. B) Fractions of neurons with equal or
greater responses to simplified stimuli than complex stimuli. The blue lines
correspond to the matching simplified stimulus for each complex stimulus,
and the gray lines correspond to the other simplified stimuli.
many differences in computational mechanisms and details of
learning between the visual cortex and the CNNs.
Others [3], [14], [6] have found relationships between IT
activity and CNNs. However, they did not report individual
neuron tuning properties, so similarity with IT may have varied
differently with respect to different tuning properties in these
networks.
A. Limitations
This study has examined a small selection of tuning prop-
erties, and the responses of only two CNNs, although these
CNNs are widely used and reasonably representative. A more
comprehensive investigation would examine the effects of ad-
ditional CNN structures and training procedures. Furthermore,
although efforts were made to approximate the stimuli used
in IT studies, the stimuli could not typically be reproduced
exactly. Generally an exact analogy isn't possible. For exam-
ple, the number of pixels per degree visual angle is not well
defined in the networks. Also, the stimuli in the IT studies
K
L
M
O
Q
S
U
X
Y
Z
[
\
]
appeared and disappeared, and sometimes moved, and the IT
neurons responded dynamically. In contrast, the CNNs receive
single frames of input and produce a single corresponding
response (without dynamics). CNNs responses are probably
most closely related to the earliest (largely feedforward) IT
responses. However, IT responses change constantly, and
latencies vary across conditions, so it is not clear which time
window of IT responses would be the most closely related.
B. Toward IT-Like CNNs
The many similarities between CNNs and the ventral visual
stream suggest an opportunity to further optimize CNNs to
more closely approximate the ventral stream, perhaps in terms
of size invariance and orientation tuning. [3] trained CNNs
specifically to emulate IT neuron activity, but this approach
did not yield better IT approximations than simply training
the CNNs for object classification. However, this attempt may
have been hampered by the small size of the neural dataset.
Another approach would be to use an empirical model of
cortical activity to produce a large (but approximate) set
of labels, as in [28]. Further incorporation of physiological
mechanisms such as [29] into CNNs may also be important.
V. CONCLUSION
This work is a preliminary comparison of various stimulus
tuning properties of CNN units to corresponding properties of
IT neurons. The results show a full spectrum of close simi-
larities, subtle differences, and large differences with respect
to different tuning properties. This provides some missing
detail on the specific relationships between CNN unit and IT
neuron responses, and suggests how CNN unit activity could
be modified in future work to more closely reflect the visual
representation in IT.
ACKNOWLEDGMENT
Victor Reyes Osorio extracted the stimulus images from
[17]. Salman Khan provided comments on an earlier draft.
REFERENCES
[1] M. D. Zeiler and R. Fergus, "Visualizing and Understanding Convolu-
tional Networks arXiv:1311.2901v3 [cs.CV] 28 Nov 2013," Computer
VisionECCV 2014, vol. 8689, pp. 818 -- 833, 2014.
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet Classification
with Deep Convolutional Neural Networks," in Advances in Neural
Information Processing Systems, 2012, pp. 1 -- 9.
[3] D. L. K. 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." PNAS, may 2014.
[4] T. Serre, A. Oliva, and T. Poggio, "A feedforward architecture accounts
for rapid categorization." PNAS, vol. 104, no. 15, pp. 6424 -- 9, apr 2007.
[5] E. T. Rolls, "Invariant Visual Object and Face Recognition: Neural and
Computational Bases, and a Model, VisNet." Frontiers in Computational
Neuroscience, vol. 6, no. June, p. 35, jan 2012.
[6] S. M. Khaligh-Razavi and N. Kriegeskorte, "Deep Supervised, but Not
Unsupervised, Models May Explain IT Cortical Representation," PLoS
Computational Biology, vol. 10, no. 11, 2014.
[7] N. T. Markov, M. M. Ercsey-Ravasz, a. R. Ribeiro Gomes, C. Lamy,
L. Magrou, J. Vezoli, P. Misery, A. Falchier, R. Quilodran, M. a. Gariel,
J. Sallet, R. Gamanut, C. Huissoud, S. Clavagnier, P. Giroud, D. Sappey-
Marinier, P. Barone, C. Dehay, Z. Toroczkai, K. Knoblauch, D. C. Van
Essen, and H. Kennedy, "A weighted and directed interareal connectivity
matrix for macaque cerebral cortex." Cerebral Cortex, vol. 24, no. 1,
pp. 17 -- 36, jan 2014.
[8] E. L. Schwartz, R. Desimone, T. D. Albright, and C. G. Gross,
"Shape recognition and inferior temporal neurons," PNAS, vol. 80, no.
September, pp. 5776 -- 5778, 1983.
[9] H. Hong, D. L. K. Yamins, N. J. Majaj, and J. J. DiCarlo, "Explicit
information for category-orthogonal object properties increases along
the ventral stream," Nature Neuroscience, vol. 19, no. 4, pp. 613 -- 622,
2016.
[10] J. J. DiCarlo, D. Zoccolan, and N. C. Rust, "How does the brain solve
visual object recognition?" Neuron, vol. 73, no. 3, pp. 415 -- 434, 2012.
[11] L. Robinson and E. T. Rolls, "Invariant visual object recognition:
biologically plausible approaches," Biological Cybernetics, vol. 109, no.
4-5, pp. 505 -- 535, 2015.
[12] S. Khan and B. Tripp, "An empirical model of activity in macaque
inferior temporal cortex," Neural Networks, accepted.
[13] 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, vol. 115, no. 3, pp. 211 -- 252,
2015.
[14] 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 Computational Biology, vol. 10, no. 12, 2014.
[15] K. Simonyan and A. Zisserman, "Very deep convolutional networks for
large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.
[16] W. Ding, R. Wang, F. Mao, and G. Taylor, "Theano-based Large-Scale
Visual Recognition with Multiple GPUs," in ICLR2015, 2015. [Online].
Available: http://arxiv.org/abs/1412.2302
[17] S. R. Lehky, R. Kiani, H. Esteky, and K. Tanaka, "Statistics of visual
responses in primate inferotemporal cortex to object stimuli." Journal
of Neurophysiology, vol. 106, no. 3, pp. 1097 -- 117, sep 2011.
[18] N. K. Logothetis, J. Pauls, and T. Poggio, "Shape representation in the
inferior temporal cortex of monkeys." Current Biology, vol. 5, no. 5,
pp. 552 -- 563, 1995.
[19] G. Y. Kov´acs, R. Vogels, and G. a. Orban, "Selectivity of macaque
inferior temporal neurons for partially occluded shapes." The Journal of
Neuroscience, vol. 15, no. March 1995, pp. 1984 -- 1997, 1995.
[20] K. Tanaka, "Columns for complex visual object features in the infer-
otemporal cortex: clustering of cells with similar but slightly different
stimulus selectivities." Cerebral Cortex, vol. 13, pp. 90 -- 99, 2003.
[21] M. Hasselmo, E. Rolls, G. Baylis, and V. Nalwa, "Object-centered
encoding by face-selective neurons in the cortex in the superior temporal
sulcus of the monkey," Experimental Brain Research, vol. 75, no. 2, pp.
417 -- 429, 1989.
[22] W. Freiwald and D. Y. Tsao, "Functional compartmentalization and
viewpoint generalization within the macaque face-processing system,"
Science, vol. 330, pp. 845 -- 851, 2010.
[23] M. Ito, H. Tamura, I. Fujita, and K. Tanaka, "Size and position invariance
of neuronal responses in monkey inferotemporal cortex." Journal of
Neurophysiology, vol. 73, no. 1, pp. 218 -- 226, 1995.
[24] H. Op De Beeck and R. Vogels, "Spatial sensitivity of macaque inferior
temporal neurons," Journal of Comparative Neurology, vol. 426, pp.
505 -- 518, 2000.
[25] K. J. Nielsen, N. K. Logothetis, and G. Rainer, "Dissociation between
local field potentials and spiking activity in macaque inferior temporal
cortex reveals diagnosticity-based encoding of complex objects," The
Journal of neuroscience, vol. 26, no. 38, pp. 9639 -- 9645, 2006.
[26] D. Zoccolan, D. D. Cox, and J. J. DiCarlo, "Multiple Object Response
Normalization in Monkey Inferotemporal Cortex," Journal of Neuro-
science, vol. 25, no. 36, pp. 8150 -- 8164, 2005.
[27] E. Kobatake and K. Tanaka, "Neuronal selectivities to complex object
features in the ventral visual pathway of the macaque cerebral cortex,"
Journal of neurophysiology, vol. 71, no. 3, pp. 856 -- 867, 1994.
[28] B. Tripp, "A convolutional model of the primate middle temporal area,"
in ICANN, 2016.
[29] D. B. Rubin, S. D. V. Hooser, and K. D. Miller, "The stabilized
supralinear network: A unifying circuit motif underlying multi-input
integration in sensory cortex," Neuron, vol. 85, no. 1, pp. 1 -- 51, 2015.
[Online]. Available: http://dx.doi.org/10.1016/j.neuron.2014.12.026
|
1606.08370 | 1 | 1606 | 2016-06-27T17:16:25 | The right time to learn: mechanisms and optimization of spaced learning | [
"q-bio.NC",
"q-bio.MN"
] | For many types of learning, spaced training that involves repeated long inter-trial intervals (ITIs) leads to more robust memory formation than does massed training that involves short or no intervals. Several cognitive theories have been proposed to explain this superiority, but only recently has data begun to delineate the underlying cellular and molecular mechanisms of spaced training. We review these theories and data here. Computational models of the implicated signaling cascades have predicted that spaced training with irregular ITIs can enhance learning. This strategy of using models to predict optimal spaced training protocols, combined with pharmacotherapy, suggests novel ways to rescue impaired synaptic plasticity and learning. | q-bio.NC | q-bio | The right time to learn: mechanisms and optimization of spaced learning
Paul Smolen, Yili Zhang and John H. Byrne
Department of Neurobiology and Anatomy, W. M. Keck Center for the Neurobiology of Learning
and Memory, The University of Texas Medical School at Houston, P.O. Box 20708, Houston,
Texas 77030, USA
Abstract For many types of learning, spaced training that involves repeated long inter-trial
intervals (ITIs) leads to more robust memory formation than does massed training that involves
short or no intervals. Several cognitive theories have been proposed to explain this superiority,
but only recently has data begun to delineate the underlying cellular and molecular mechanisms
of spaced training. We review these theories and data here. Computational models of the
implicated signaling cascades have predicted that spaced training with irregular ITIs can
enhance learning. This strategy of using models to predict optimal spaced training protocols,
combined with pharmacotherapy, suggests novel ways to rescue impaired synaptic plasticity
and learning.
1
Repetitive training helps form a long-term memory. Training or learning that includes long
intervals between training sessions is termed spaced training or spaced learning. Such training
has been known since the seminal work of Ebbinghaus to be superior to training that includes
short inter-trial intervals (massed training or massed learning) in terms of its ability to promote
memory formation. . Ebbinghaus stated1: "…with any considerable number of repetitions a
suitable distribution of them over a space of time is decidedly more advantageous than the
massing of them at a single time." His studies were based on the self-testing of acquired
memory for lists of syllables, but the superiority of spaced training has now been established for
many additional forms of human learning. For example, spaced learning is more effective than
massed learning for facts, concepts and lists2, 3, 4, skill learning and motor learning5, 6, in
classroom education (including science learning and vocabulary learning)7, 8, 9, and in
generalization of conceptual knowledge in children10. Spaced training also leads to improved
memory in invertebrates, such as the mollusk Aplysia californica 11, 12, 13, 14, Drosophila
melanogaster15, 16 and bees17, and in rodents18, 19 and non-human primates20, 21. Memory
extinction is commonly considered to involve the formation of a new memory, and in rat fear
conditioning, spaced extinction trials are more effective than massed trails at establishing new
memories22.
Although it has been established that spaced training is superior to massed training in
terms of inducing memory formation, key questions remain. What are the mechanisms
underlying this superiority? Is it possible to use this mechanistic information to determine the
optimal intervals between learning trials? If so, are fixed, expanding, or irregularly spaced
intervals optimal? Another key question is whether an understanding of the mechanisms for
optimal intervals can provide insights into the design of pharmacological approaches for
memory enhancement. Computational models based on such a mechanistic understanding may
be able to predict more complex approaches to memory improvement in which the application of
multiple drugs, or combinations of drugs and training protocols, can enhance memory or treat
deficits in learning and memory.
In this Review, we describe how new insights from molecular studies may help explain
the effectiveness of spaced training, and how the molecular findings relate to the traditional
learning theories that aim to account for this effectiveness. We also review how models of
signaling pathways that are involved in synaptic plasticity can suggest, and experiments
2
empirically validate, training protocols that improve learning and that rescue plasticity impaired
by deficits of key molecular components. Finally, we discuss recent models that have suggested
combined-drug therapies that may further enhance some forms of learning and that may have
synergistic effects with optimized spaced learning on memory formation.
Traditional learning theories
We briefly summarize three of the well-known cognitive theories proposed to explain spaced
training's superiority to massed training: encoding variability theory, study-phase retrieval theory
and deficient-processing theory.
Encoding variability theory23, 24, 25 posits that repeated stimulus presentations or
learning trials are more likely to occur in multiple contexts if they are spaced further apart in
time, and that a memory trace for repeated trials therefore includes elements of each of these
contexts. Thus, spaced training would tend to bind together more contexts and hence form a
more robust memory, as a greater number of testing contexts could elicit retrieval of the
memory.
Study-phase retrieval theory26, 27, 28, 29 posits that spaced stimulus presentations or
learning trials are more effective than massed trials in reinforcing memory because each spaced
trial elicits retrieval and reactivation of a memory trace formed by the preceding trial. In contrast,
with short massed trials, the preceding memory trace is still active, so it is not retrieved or
reactivated. Therefore memory cannot be reinforced. Study-phase retrieval theory also accounts
for a decline in learning with excessively long intervals because in that case the preceding
memory trace can no longer be retrieved. A recent variant, retrieved context theory, also
incorporates elements of encoding variability theory and has succeeded in predicting the results
of subsequently performed spaced learning experiments in humans30.
Deficient processing theory posits that spaced training forms a stronger memory than
does massed training because in the latter, some processes that are necessary to form
memories are not effectively executed. The reasoning here becomes clearer by examining
variants of this theory that specify the nature of the deficient process. One variant posits that
excess habituation during massed trials prevents effective reinforcement of memory traces31,
whereas others posit that there is a failure to consolidate a memory (also known as
consolidation theory)32, 33, a lack of voluntary attention to massed presentations31, 34, or a lack
3
of cognitive rehearsals or reactivations of a memory trace within the short intervals that are
characteristic of massed training27, 35.
Consolidation occurs as a memory trace becomes more fixed and stable with time after
training2, 36. Thus, consolidation theory37, 38, 39 posits that a long-term memory trace is more
efficiently stabilized or strengthened by spaced trials. The lack of cognitive rehearsals variant of
deficient-processing theory might also be considered a more specific form of consolidation
theory, because it assumes that a minimum number of rehearsals, or autonomous reactivations,
are required to consolidate a memory trace. Variants of deficient-processing theory and relevant
experiments are further discussed in REFs28, 40, 41. Below, we focus substantially on
consolidation theory because, of all the traditional learning theories, it seems to be most closely
aligned with our current understanding of the cellular and molecular mechanisms of memory.
Landauer37 was one of the first to develop a conceptual model of the ways in which
consolidation principles could explain the effectiveness of spaced training. Although the model
was originally developed to explain the effects of short spacing intervals on memory formation, it
can readily be generalized for the effects of arbitrarily long intervals (FIG. 1). The model is
based on two assumptions. First, the state of a neural circuit following the first learning trial is
such that a second reinforcing trial soon after will not markedly increase the consolidation of the
learning trace resulting from the first trial (FIG. 1a). Thus, in massed training, overlap between
traces and, consequently, saturation of an unspecified molecular mechanism diminishes the
traces' summed impact on the consolidation of memory. Only when the effects of the first trial
decay can the effects of a second trial be fully expressed (FIG. 1b, 1c), leading to greater
potential consolidation of a memory in spaced training than in massed training (greater net gain,
FIG. 1d). The second assumption is that the probability that the second trial can successfully
reinforce the first trial declines with time (FIG. 1e). Actual consolidation is the product of these
two assumptions, yielding a prediction of an optimal interval for spaced learning (FIG. 1f).
Peterson38 described a similar model focused on the dynamics of verbal learning.
4
Figure 1 Early conceptual model of how learning trace dynamics generate an optimal interval. a–c
As described by the early model of Landauer37, spaced training is more effective at strengthening some
form of trace corresponding to memory storage in the brain, although this conceptual model does not
posit a biochemical or structural form for the trace. This model posits that memory formation becomes
more effective with longer inter-stimulus intervals between training sessions because of decreasing
temporal overlap between successive, short-lived learning traces. These learning traces do not
themselves constitute a memory. However, their net effect, denoted here as net gain, contributes to the
formation of a long-lived memory trace. In (a–c), learning traces elicited by two successive trials are
shown. The model assumes that, for each value of the inter-trial interval length, a quantity denoted "net
gain due to the reinforcing trial" is proportional to the red area. Shorter intervals are associated with more
overlap of learning traces and less net gain. Therefore, a reinforcing trial is most effective after a
refractory period following the preceding trial. For this conceptual theory, units for amplitude and time are
arbitrary. d A greater summed effect, or net gain, of reinforcing trials occurs for longer inter-stimulus
intervals. The effect reaches a plateau for long intervals as the overlap between successive learning
traces goes to zero e Over longer times, a different quantity - the probability that a reinforcing trial will
be effective at all in reactivating processes that constituted the preceding learning trace - declines. f An
optimum interval for maximizing the strength of the long-lived memory trace results when the greater net
gain of reinforcement at longer intervals (from d) is multiplied by the slowly declining probability that a
reinforcement will reactivate a previous learning trace (from e). The optimum interval for net learning is
the one that produces the peak level of the trace in f.
Wickelgren39 extended consolidation theory by positing that the resistance of a memory
trace to decay increases with the age of the trace over the total duration of a spaced learning
protocol. Thus, a trace would become not only strong but also highly resistant to decay following
spaced trials.
Molecular traces of time
Substantial progress has been made in understanding the molecular mechanisms of memory.
Given this progress, in this section we focus on potential molecular mechanisms of the spacing
effect on long-term memory formation.
5
There is now agreement that learning is implemented, at least in part, by changes in
synaptic strength (synaptic plasticity). For example, fear-conditioned memories can be
alternately erased and re-instantiated by long-term depression (LTD) and long-term potentiation
(LTP), respectively, of a defined synaptic pathway42. Thus, the molecular processes that are
essential for spaced learning might reinforce extant LTP.
Reliable correlates of LTP are the remodeling and enlargement of postsynaptic dendritic
spines, which are small protrusions that are associated with most excitatory synapses43. Thus,
studying the differential dynamics of dendritic spine remodeling following massed versus spaced
stimuli is likely to give insight into processes underlying the effectiveness of spaced training.
Studies using rat hippocampal slice found that LTP induced by multiple trains of theta-burst
stimuli was accompanied by extensive remodeling of synaptic ultrastructures44, 45 and that
subsequent spaced trains of theta-burst stimuli with intervals of 60 min or more between trains,
were needed for optimal reinforcement of LTP46. Stimulated dendritic spines were remodeled
over > 1 h, leading to enlargement of the existing functional postsynaptic density45 and the
presynaptic active zone44. The resulting increase in the numbers of AMPA-type and NMDA-type
receptors at the synapse correlated with the magnitude of LTP.
Two hypotheses that involve spine remodeling have been put forward to explain the
greater efficacy of spaced trials over massed trials in memory formation. These hypotheses
have a common theme, which is that the learning process includes a refractory period during
which the second of two closely spaced stimuli would be ineffective in enhancing the effects of
the first (FIG. 2a). One hypothesis is that spaced but not massed repetitions of a stimulus allow
the refractory period to be overcome and lead to repeated enlargement of a set of spines and a
strengthening of the synaptic connections mediated by these spines47. (FIG. 2b) A second, not
mutually exclusive, hypothesis47, 48 is that molecular processes enable later spaced stimuli to
induce LTP at spines that do not undergo initial enlargement. In this case spaced, but not
massed, inter-trial intervals would allow for a molecular process termed 'priming' to be
completed at these additional spines. After being primed, these spines would be strengthened
by subsequent stimuli and incorporated into the memory trace (FIG. 2c). Currently, the
molecular components of such a priming process are not known.
6
Figure 2 Model and hypotheses describing synaptic strengthening during spaced learning. a In
the refractory-state model, spaced stimuli (left panel, stimulus 1, red arrow 1, followed substantially later
by stimulus 2, red arrow 2) cumulatively strengthen a memory trace (as indicated by green line). By
contrast, massed stimuli (right panel, stimulus 1 followed shortly after by stimulus 2) fail to cumulatively
strengthen the memory trace. b The cumulative synaptic strengthening in spaced training may be due to
progressive enhancement of long-term potentiation, which could result from successive increases in the
in the strength of the same synaptic contacts (shown here as successive increases in the volume of the
same postsynaptic dendritic spine). Thus in one of two current hypotheses describing synaptic
strengthening during spaced learning, stimulus 1 enlarges a population of spines. If stimulus 2 follows
shortly after the first stimulus (as in massed training), it cannot further affect spines. However, if stimulus 2
comes after a refractory period (as in spaced training), it can further enlarge the same population of
spines. c Alternatively, enhancement of LTP could result from successive rounds of strengthening of new
synaptic contacts. Thus in the second current hypothesis, stimulus 1 only enlarges a subset of affected
spines, but primes additional spines. If stimulus 2 follows shortly after stimulus 1 (as in massed training), it
has no effect. If stimulus 2 comes later (as in spaced training), it does not further enlarge the first subset
of spines. Instead, stimulus 2 enlarges those spines that were primed, but not enlarged, by stimulus 1.
Through use of Schaffer–commissural projections in rat hippocampal slices, two
studies47, 48 have characterized the recruitment of additional synaptic contacts with the
application of spaced stimuli. Theta burst stimuli applied at intervals of 10 min or 40 min did not
7
cumulatively increase LTP. However, for longer intervals (60 or 90 min), a cumulative increase in
LTP was observed over three bursts of stimulation. Each theta burst stimulus (TBS) led to actin
filament polymerization in spines, which is known to be important for the stabilization of LTP49.
The second stimulus (TBS2) yielded polymerization in spines that were not apparently affected
by the first stimulus (TBS1), if TBS2 followed TBS1 by 60 or 90 min. These data do not suggest
that successive theta burst stimuli further strengthen the efficacy of the same spines. Instead,
they suggest that TBS1 initiates priming at all synaptic contacts of the stimulated afferents but
only initiates consolidation and strengthening at a subset of contacts. Spines that undergo
priming but not consolidation exhibit a refractory period of ~ 60 min, suggesting that priming
takes time to complete (FIG. 2a). If TBS2 is given after the refractory period, some or all of the
primed spines undergo consolidation. These data are consistent with the second hypothesis
presented in the preceding paragraph, because TBS1 appears to enlarge and strengthen some
spines but at others, TBS1 only initiates priming. These primed spines can then be strengthened
by TBS2.
The dynamic properties of transcription factors and their interactions could also account
for the superior efficacy of spaced training. LTP that persists for several hours or more requires
translation and transcription50, 51, which is reliant on key transcription factors such as cAMP
response element (CRE) binding protein (CREB)52. Spaced training may be more effective, in
part, because it may allow transcription factors such as CREB sufficient time to be activated,
bind to promoters and induce a round of transcription for the consolidation of LTP53 or long-term
synaptic facilitation (LTF)54. In massed training, the trials would come too close together to
initiate separate rounds of transcription. Indeed, in co-cultures of sensory and motor neurons
from Aplysia, five spaced serotonin (5-HT) applications, each of 5 min with an inter-stimulus
interval of 20 min (an analogue of spaced training) robustly elicits LTF that lasts for over 24 h14,
whereas 5-HT applied continuously over 25 min (an analogue of massed training) fails to yield
reliable LTF.
In these sensory neurons, levels of a transcription activator CREB1 are elevated for at
least 24 h after the spaced 5-HT treatments54, 55. This prolonged elevation of CREB1 owes to a
positive feedback loop in which CREB1, acting via binding to a CRE regulatory element near
creb1, increases the expression of creb154, 55 and other genes upregulated by CREB1. In
addition, in these sensory neurons, the level of the transcription repressor CREB2 shows a late
8
drop at ~12 h after treatment56. This drop in CREB2, coupled with the rise in CREB1, plausibly
corresponds to an increased potential for gene induction. Thus, an additional 5-HT pulse near
12 h after treatment might optimally reinforce LTF.
LTF at these sensorimotor synapses is associated with a simple form of learning, long-
term sensitization (LTS) of withdrawal reflexes. In vivo, four spaced electrical stimuli (30 min
intervals between the stimuli) yield LTS that lasts > 24 h, with a weak residual of LTS being
detectable at 4 d, and repetition of this spaced protocol once per day for 4 days yields much
stronger LTS that lasts for over 1 week13, 57, 58. These data suggest that in this system, the
dynamics of transcription activation and gene expression have slow components that can
summate over multiple days, yielding long-lasting memory.
Recent data also illustrates that in the hippocampus, CREB and CCAAT enhancer
binding protein (C/EBP), another transcription factor that is important for LTP, can remain active
for many hours after learning. Following inhibitory avoidance training in rats, late peaks in brain-
derived neurotrophic factor (BDNF) expression and C/EBP expression occur at ~12 h post-
training, and inhibiting BDNF action at this time blocks memory maintenance59. These BDNF
dynamics result from a positive feedback loop in which C/ebp induction leads to Bdnf
upregulation, with the resulting increase in BDNF further activating the C/EBP signaling
pathway60. Although this slow feedback loop was activated by single-trial training rather than
spaced training, it would be of interest to model these dynamics, and to examine whether an
additional spaced trial at ~ 12 h post-training, leading to a second induction of C/EBP at the time
of elevated C/EBP, might optimally reinforce learning. A second prediction would be that
massed stimuli are less effective if repeated at an interval too brief to allow the transcription
regulation and Bdnf expression necessary to activate this feedback loop. Insights that can be
obtained from computational models of learning are discussed later in the article.
time scale,
On a shorter
the dynamics of second messengers, kinases and
phosphatases may contribute to the superiority of spaced training. One study with mice61 found
marked phosphorylation and activation of CREB in the hippocampus and cortex when object
recognition trials were separated by an interval of 15 min but not by an interval of 5 min. Protein
phosphatase 1 (PP1) appeared to be necessary for this spacing effect, because PP1 inhibition
allowed the shorter interval to activate CREB. A study involving Aplysia sensory neuron–motor
neuron co-cultures62 found that protein kinase C (PKC) is activated to a greater extent during a
9
massed stimulus (continuous 5-HT application) than a spaced stimulus (15 min intervals
between applications). It is known that PKC acts to downregulate protein kinase A (PKA) and
that PKA activation is necessary for LTF, thus these data delineate crosstalk between signaling
pathways such that LTF is suppressed, in part, by stronger PKC activation during massed
training.
Another study16 characterized the dynamics of MAP kinase (MAPK) and of MAPK
phosphatase in Drosophila. In an olfactory learning protocol, each spaced training trial
generated a distinct wave of MAPK activity, whereas massed training trials occurred too close
together to generate distinct waves. The authors therefore hypothesized that effective learning
depended on the generation of distinct waves, which was only seen after the spaced trials.
Another phosphorylation-based mechanism has also been hypothesized to help explain
the efficacy of spaced intervals in Drosophila. Spaced (15 min) intervals were more effective
than massed (1 min) intervals in inducing olfactory learning, even given the same total training
time (and thus more massed presentations)63. Two isoforms of Drosophila CREB - dCREB2-a
and dCREB2-r - can activate and repress transcription, respectively. The authors proposed64
that the kinetics of the phosphorylation of these isoforms differed such that the kinase activation
generated by less frequent, spaced trials was sufficient to phosphorylate and activate dCREB2-
a, whereas dCREB2-r could only be effectively phosphorylated by massed trials. Thus, training
involving spaced intervals could maximally activate transcription and possibly induce the
formation of long-term memory by activating dCREB2-a but not the counteracting repressor
dCREB2-r.
Computational simulations have supported the plausibility of this mechanism65, but it
has not been validated empirically. However, it appears likely that a similar type of mechanism
that is based on competition between an activator of long-term memory formation and a
repressor, with the repressor only activated at short intervals, might be needed to explain any
similar data in which massed training is less effective than space training even given equal total
training times.
In experiments with Aplysia, when two electric shocks were given to induce LTS,
maximal LTS was produced when the inter-stimulus interval was 45 min. LTS was not produced
with intervals of 15 or 60 min66. The 45 min optimum was associated with activation of MAPK.
Following either a single 5-HT pulse or a single shock, MAPK activation peaked at or near 45
10
min post-trial12, 66, thus a 45 min interval might optimally reinforce the effects of MAPK. It is
known this delayed MAPK activation requires protein synthesis12, although the upstream
mechanisms underlying the dynamics of the peak in MAPK activity at ~ 45 min are not well
understood. Nevertheless, the key finding from these studies is that delayed activation of MAPK
is intimately associated with the effectiveness of spaced stimuli to induce long-term memory.
Similarly, training with intervals of 60 min, but not 20 or 120 min, enhanced object
recognition learning in wild-type mice and in a mouse model of fragile X syndrome model
(FMR1 knockout mice), at least partly by increasing synaptic activation of extracellular signal
regulated kinase (ERK1 and ERK2: isoforms of MAPK)67. This 60 min interval was predicted to
be optimal for learning because stimuli separated by 60 min had previously been found to
enhance LTP in wild-type rodents47. Thus, in Aplysia, Drosophila, and mammals, MAPK
activation appears to be a component of the molecular mechanism that underlies the spacing
effect.
Some of these molecular mechanisms appear to fit with a theory in which spaced
training sessions are effective because they reinforce the same memory trace or group of
strengthened synapses. However, spaced stimuli might also reinforce memory by recruiting new
synapses. ERK1 and ERK2 (ERK1/2) activation is needed for some forms of LTP68 and one
study69 compared ERK1/2 activation in rat hippocampal pyramidal neurons following three
spaced tetanic bursts (at 5 min intervals) with that after three massed bursts (at 20 s intervals).
About twice as many dendrites with active ERK1/2 were found following spaced bursts,
suggesting that spaced trials may recruit additional synapses on different dendrites for LTP.
Thus, a range of molecular and cellular mechanisms appears to contribute to the efficacy of
spaced training, in parallel or in series.
An extremely broad range of inter-trial intervals ranging from seconds to days has been
used for spaced training (FIG. 3). For example, in honeybee olfactory learning, efficient spaced
training can occur with intervals as short as 1 min17. Such brief intervals might allow for the
reinforcement of the activity of a short-lived second messenger such as cAMP that is produced
by preceding trials. The dynamics of kinase activation constitute a second substrate of spacing
effects. In Aplysia, Drosophila, and mammals, the data discussed above indicate that commonly
reported intervals, ranging from ~ 5 min to 1 h, may allow for the reinforcement of the activities
of key kinases essential for LTP or LTF, and consolidate structural changes in dendritic spines.
11
in
the
relatively brief
Figure 3 Different mechanisms may
underlie enhancement of learning by
intervals of widely varying
spaced
lengths. For
inter-trial
intervals (bottom trace), successive trials
may coincide with and reinforce peak
second messenger
levels generated by
preceding trials. In each trace, individual
rectangles represent individual trials, and
converging lines between traces represent
the lengthening of time scales as one
moves upwards
illustration. For
somewhat longer inter-trial intervals (several
min to ~1 h), successive trials may reinforce
the peak activities of kinases elicited by
preceding trials and elicit LTP of primed
dendritic spines. Intervals of this length may
also, in hippocampus, be needed to allow
replacement of
inactivated receptors at
stimulated spines82, enabling succeeding
stimulus repetitions
those
spines. For intervals of ~ 1 h or more,
succeeding trials may also align with peaks in transcription factor activity and gene expression owing to
preceding trials. For the longest inter-trial intervals (many h or longer) succeeding trials may reactivate
and thereby further potentiate consolidated memory traces. All these processes are likely to contribute to
the consolidation of long-term memory, in many if not all species. However, depending on the inter-trial
interval length used in a particular spaced learning protocol, the dynamics of a particular type of process
(e.g. kinase activation) may contribute in particular to the efficacy of spaced learning. In addition, trials at
one temporal domain (e.g., 1 day) may be unitary events, but also may constitute a block of spaced trials
from another temporal domain (e.g., min to h). For example, an effective protocol for long-term
sensitization training in Aplysia is the use of four trials with an intertrial interval of 30 min, with this block
repeated four times with a one-day intertrial interval13. Thus, some effective training protocols consist of a
hierarchy of temporal domains of training sessions, with briefer sessions embedded within longer ones. In
this illustration intervals are shown with regular spacing, but more effective learning may occur with
irregular spacing (see below).
to potentiate
It is plausible that the minimum inter-stimulus interval for effective learning, for a given
protocol and system, corresponds to the interval that is necessary to allow each stimulus to
contribute separately to a rate-limiting biochemical process. For example, for rapid honeybee
olfactory learning with an effective interval of 1 min, the rate-limiting process might be second
messenger accumulation or rapid activation of a kinase. For even shorter intervals, the time
scale of the rate-limiting process might be too long to permit each brief stimulus to contribute
separately to the process - a group of closely spaced stimuli would instead tend to act as just a
single stimulus. For intervals of 1 min or more, each stimulus would be able to contribute a
discrete increment to the rate-limiting process, allowing effective learning. For the spaced LTP
protocol of Gall, Lynch and colleagues47, 48, an interval of 40–60 min is needed for successive
TBS to further increase LTP. Here, the rate-limiting process would be different - plausibly
slower activation of an unspecified kinase or other intracellular signaling event, with a time
12
constant near the minimum effective interval of ~ 40 min. Stimuli at intervals much shorter than
this would not be able to generate summation of the rate-limiting process and would therefore
not cause additional LTP.
For other systems, a similar assumption may apply to the dynamics of transcription
activation. For LTF and LTS in Aplysia, transcription, as discussed above, may constitute a rate-
limiting process that helps determine the efficacy of spaced training. However, it is evident that
even for systems such as honeybee olfactory learning that involve short, spaced intervals,
effective long-term memory formation relies on the activation of transcription and translation,
downstream of the intracellular signaling pathways that are activated by these intervals17, 70.
Reactivation of memory traces may constitute an additional temporal substrate that underlies
the longest reported effective intervals, on the order of a week71. Such intervals are likely to
reactivate and reinforce consolidated patterns of strengthened synapses that correspond to
memory traces that are maintained by neuronal network activity72. Spaced learning with these
long intervals would reactivate critical components at these synapses, and in particular
reactivate NMDA receptors at these synapses. Studies using inducible and reversible NMDA
receptor knockouts have demonstrated that such NMDA receptor reactivation, which may also
in part result from spontaneous neuronal activity, is required to sustain remote memory
storage73, 74. Positive feedback loops that maintain key kinases and other molecules in
persistently active states at strengthened synapses may also contribute to such long-term
memory storage75, 76, 77, 78, 79. An important topic for future research will be to further
investigate the molecular processes that support effective spaced learning for humans that
involves inter-trial intervals of a day or more. An implication of Fig. 3 is that multiple temporal
domains of spaced training can be engaged in spaced training. For example, an effective
protocol for long-term sensitization training in Aplysia is the use of four trials with an inter-trial
interval of 30 min, repeated four times with a one-day inter-trial interval13. Thus, at least in some
cases, there appears to be a hierarchy of temporal domains of training protocols with briefer
protocols embedded within longer ones.
The above considerations, and most empirical studies, are concerned with only typical,
or minimum, inter-trial or inter-stimulus intervals for effective spaced learning or for the
summation of LTP. Only a few studies have delineated, for any specific system (that is, a given
species and stimulus protocol), both minimum and maximum effective intervals. One study80
13
found that in a hippocampal slice preparation, 5–10 min intervals between tetani were ideal for
induction of LTP and they produced similar levels of LTP, with longer or shorter intervals yielding
both less LTP and less ERK1/2 activation. In Aplysia, LTS was effectively induced by an interval
of 45 min between electrical stimuli, but not by intervals of 15 min or 60 min66. As noted above,
the authors of this study hypothesized that the coincidence of peak MAPK activation with the
second trial was necessary for effective learning. In addition, 60 min intervals were effective for
forming object location memory in mice with three trials, but intervals of 20 min or 120 min were
not67.
Owing to the small number of such studies and the lack of sufficient characterization of
the accompanying molecular processes, it is not yet possible to make detailed statements about
the ways in which intracellular signaling pathways could co-operate to generate both minimum
and maximum intervals. For maximum intervals, a reasonable qualitative assumption is that
each trial or stimulus generates a separate, relatively short-lived biochemical trace, and that for
effective spaced learning, these traces must overlap and summate, with the summed magnitude
driving the formation of long-lasting synaptic potentiation. These dynamics would be analogous
to the necessary overlap of traces in the conceptual model of Landauer (FIG. 1a-c). For
intervals longer than the maximum, the individual biochemical traces would decay and not
overlap.
Recent data and learning theories
Do the biochemical and morphological mechanisms proposed to contribute to the greater
efficacy of spaced training align with traditional cognitive theories? At this point, much of the
extant cellular data seem to be compatible with the deficient processing theory, particularly two
of its variants: the consolidation theory and the lack of cognitive rehearsals theory. In the
consolidation theory, intervals between massed trials are proposed to be too short for the
consolidation and consequent summation of memory traces that are engendered by successive
trials. In the cognitive rehearsals theory, massed trials are proposed to lead to fewer cognitive
rehearsals, or autonomous reactivations, than do spaced trials, and therefore less cumulative
consolidation and persistence of a memory.
The required refractory period of ~ 1 h between successive theta-burst stimuli in to
induce progressive increments in hippocampal LTP46, 47 may be in line with the first of these
variants, that short intervals are insufficient for consolidation and consequent summation of
14
memory traces. The refractory period appears to be necessary to complete the 'priming' of
dendritic spines that were stimulated, but not potentiated, by the first theta-burst. Priming allows
these spines to potentiate after the second burst, and thus constitutes a biochemical stimulus
trace (FIG. 1d and FIG. 2a). Kramár et al.47 noted that in hippocampal slices, additional
potentiation can be induced up to four hours after induction of the first LTP increment81. The
stimulus trace associated with priming may therefore take at least four hours to decay. Such a
long trace lifetime might allow a broad temporal window for optimal training trials.
In rat hippocampal slices, theta burst stimuli lead to proteolytic inactivation of integrin
receptors at stimulated dendritic spines82. These receptors are then replaced by vesicular
transport of new receptors over ~ 40 – 60 min, and it is hypothesized82 that subsequent theta
burst stimuli at these synaptic contacts cannot induce spine enlargement or LTP until after this
replacement has occurred, thus accounting for the refractory period of ~ 1 h in order for a
second theta burst stimulus to yield additional LTP. This receptor replacement may constitute, at
least in part, the priming of dendritic spines discussed above. These hypothesized dynamics
may be in line with deficient processing theory, with receptor replacement being the necessary
process that can only occur during spaced inter-trial intervals (FIG. 3).
Transcription factor activation also constitutes a biochemical trace, and in some
systems, training may only be effective if inter-trial intervals are long enough so that each trial
can induce a separate round of transcription and translation. Similarly, short (massed) inter-trial
intervals may not lead to sufficient levels, or a sufficient duration, of activated MAPK or other
kinases to support the consolidation of long-term memory.
The variant of deficient processing theory positing that only spaced trials can generate
sufficient cognitive rehearsals or reactivations of a memory to support long-term memory
consolidation may also correspond with the empirical finding that repeated theta-burst stimuli,
spaced by ~ 1 h, can recruit additional dendritic spines by potentiating spines that were primed
by preceding stimuli. A memory reactivation would be analogous to theta-burst stimulus in that
both events would initiate priming and potentiation. It also appears plausible that repeated
memory reactivations might induce further rounds of transcription of genes involved in LTP, such
as C/ebp and other CREB-activated genes, supporting further consolidation of long-term
memory.
15
To more strongly connect this variant of deficient-processing theory to recent cellular and
molecular data, one must also assume that reactivations of a memory reactivate some of the
same neurons and synapses that were activated in the original learning sessions. In that way,
the rehearsals and learning trials would reinforce memory in the same way. This assumption
appears plausible but requires further empirical investigation. Although finer-grain analyses are
necessary, a study using functional MRI during verbal learning supports this assumption83. In
this study, a specific brain region associated with rehearsal of verbal memory, the left frontal
operculum, was activated more during spaced learning of paired-word associations than during
massed learning. We note that these posited memory reactivations, on time scales of ~ 1 h or
longer, are distinct from voluntary rehearsals of a memory on a short time scale (seconds or ~ 1
min). Substantial behavioral evidence suggests that this latter voluntary, short-term rehearsal is
not essential for spaced learning31, 34.
The remaining variants of deficient processing theory, which focus on habituation or on a
lack of voluntary attention during massed presentations, do not appear to relate as readily to the
current single-neuron data. These variants have also been argued to not readily accommodate
certain verbal learning observations2. As regards encoding-variability theory, data on neuronal
network dynamics, rather than single-neuron data, will be needed to determine to what extent
the binding of contexts to memory occurs, which is required in this theory. Similar data will also
be needed to assess whether the binding of memories of later trials to those of earlier trials
occurs, which is required in study-phase retrieval theory. It will be important to re-assess all
these competing spaced learning theories as more information becomes available on the
dynamics of memory networks. Indeed, different theories may be more or less applicable to
different memory systems.
Irregular spacing can enhance learning
Attempts to optimize the spacing effect have generally been based on trial-and-error
approaches. Consequently, most, if not all, training protocols used in animal and human studies
are probably not optimal. For almost all learning paradigms, the training intervals are fixed,
although in one type of spaced training paradigm, the intervals between sessions progressively
lengthen2, 84. However, a meta-analysis2 and a text learning study84 found no substantial
evidence for the superiority of this approach in terms of promoting long-term memory formation.
16
It appears evident that at least part of the improvement in learning that is found with
spaced training protocols can be explained by the dynamic relationships between the training
trials and the underlying cellular and molecular mechanisms that are associated with memory
formation (FIG. 3). But is the inverse possible? Can knowledge of the dynamics of the memory
mechanisms be used to enhance memory processing by predicting optimal training protocols,
possibly with irregular training intervals? One approach is to develop models of the biochemical
cascades that underlie memory formation and use simulations to rapidly test the effectiveness
of different training protocols85. In recent years, models have described the dynamics of the
biochemical reactions that transduce stimuli into LTP86, 87, 88. These models have differential
equations that simulate and predict the dynamics of the activities of key molecular species.
Simulations have reproduced the dynamics of MAPK during LTP induction80, 86, 88. Models
have also simulated the activity time courses of PKA, CAMKII, other key enzymes and
downstream transcription factors during LTP induction88, 89, 90. Each signaling cascade in these
models displays a characteristic activity time course, thus it is likely some irregular sequence of
intervals would be predicted to maximize the induction of LTP. For example, subsequent trials
that are delivered at times that coincide with kinase activity peaks might optimally reinforce
learning.
One study from our laboratory developed a model describing the 5-HT-induced PKA and
ERK signaling pathways essential for LTF in Aplysia85. In the model (FIG. 4a), the necessity of
PKA and ERK activation for LTF was simply represented with a variable termed 'inducer'. The
value of inducer was proportional to the product of PKA and MAPK activities. The amount of LTF
and LTS was predicted to increase with an increase in the peak value of inducer. Ten thousand
different protocols consisting of five trials that were separated by intervals of 0 to 45 min were
simulated (FIG. 4b). The ability to simulate and predict the effects of so many protocols in a
relatively short space of time represents a distinct advantage of computational studies over
empirical studies. The simulations determined that of these protocols, a massed protocol (FIG.
4b) produced the lowest peak value of inducer, consistent with data that massed 5-HT
application fails to produce LTF14. The 'best' protocol yielding the highest peak value for
inducer, termed the 'enhanced' protocol (FIG. 4b), had irregular intervals. The protocol termed
the 'standard' protocol (five 5-min pulses of 5-HT, with uniform inter-stimulus intervals of 20 min)
(FIG. 4b), has been commonly used to induce LTF in empirical studies for ~30 years91. This
standard protocol yielded an intermediate peak value of inducer and was predicted to have an
17
intermediate effectiveness (FIG. 4c). These predictions were empirically validated. LTF and LTS
produced by the enhanced protocol exceeded that from the standard protocol (FIG. 4d)85. An
explanation for this enhancement of LTF, consistent with data11, 12, 92, is as follows. In response
to each 5-HT pulse, PKA activates rapidly and decays rapidly (FIG. 4c). MAPK activity rises and
decays more slowly, not peaking until ~45 min after a pulse. The four initial pulses initiate a
surge of MAPK activity, which peaks near the time of the last pulse. This last pulse activates
PKA, so that a PKA peak is approximately coincident with peak MAPK activation, maximizing
inducer and the predicted LTF.
Figure 4 Dynamics of a model which has successfully predicted greater efficacy for a learning
protocol with irregular spaced intervals. a A simplified mathematical model85 describes the activation
and effects of two key kinases necessary for LTF, a cellular correlate of a simple form of learning, long-
term sensitization. Brief applications of 5-HT activate protein kinase A (PKA) via elevation of the
secondary messenger cAMP, and activate the ERK isoform of MAP kinase via a Ras-Raf-MEK cascade.
PKA and ERK interact, at least in part via phosphorylation of transcription factors, to induce LTF. In the
model, the variable 'inducer' represents the PKA–ERK interaction. A higher peak value of inducer was
assumed to predict a greater amplitude of LTF. b Six samples of the 10,000 5-HT protocols that were
simulated with the model. All protocols consist of five 5-min pulses of 5-HT, diagrammed as square
18
waves, with inter-pulse intervals chosen as multiples of 5 min, in the range from 5–50 min. The standard
protocol (green trace) is the protocol most commonly used in studies of LTF in vitro. The enhanced
protocol (red trace) produced the largest peak value of inducer, whereas the massed protocol (blue trace)
produced the smallest peak value of inducer. The standard protocol has uniform inter-pulse intervals of 20
min, whereas the enhanced protocol has non-uniform intervals of 10, 10, 5 and 30 min. The massed
protocol has no gaps between the 5-HT pulses. c Simulated time courses of activated PKA, activated
ERK and inducer in response to the standard protocol (green traces), enhanced protocol (red traces), and
massed protocol (blue traces). The unit of concentration (y axes) is µM. d In an empirical validation of
the model's prediction, the LTF induced by the enhanced protocol, as determined by the percentage
increase in the amplitude of excitatory postsynaptic potentials (EPSPs), was greater than the LTF
produced by the standard protocol. Figure adapted from REF. 85.
for several
factors
transcription
If irregularly spaced protocols can enhance normal learning, might modeling also predict
protocols capable of restoring learning that is impaired by a genetic mutation or other
physiological insults? A recent study90 tested this hypothesis. CREB binding protein (CBP) is an
acetyltransferase and essential co-activator
including
phosphorylated CREB (pCREB). CBP is also required for the consolidation of long-term
memory93. Mutations that decrease CBP activity cause a human genetic disorder termed
Rubinstein–Taybi syndrome (RTS)94 that is associated with intellectual disability and learning
deficits, and Cbp+/− mice show impaired LTP and LTM95. The recent study90 used small
interfering (siRNA) knockdown of CBP in Aplysia sensory neurons to impair LTF. In this study,
the model previously used85 to predict optimal, irregularly spaced protocols that would enhance
LTF (FIG. 4a) was extended to represent induction of c/ebp, a transcription factor known to be
essential for LTF96. In simulations of the effects of different spaced protocols, greater peak
levels of phosphorylated C/EBP (pC/EBP) were taken to predict greater LTF. Simulations
showed a substantial decrease in pC/EBP when CBP was reduced by a decrement that
corresponded to the siRNA effect. A 'rescue' protocol with irregularly spaced intervals was
predicted to restore peak pC/EBP and, correspondingly, LTF. This rescue protocol was
empirically validated to restore normal LTF in Aplysia. A similar predicted rescue protocol of
irregularly spaced intervals rescued a deficit in LTF that was produced by a siRNA knockdown
of CREB197.
Although these empirical studies were conducted in Aplysia, it should be noted that key
molecular mechanisms of memory are substantially conserved from simple model organisms
such as Aplysia to mammals52, 96. For example, LTF and LTP both rely on PKA and ERK
activation12, 92, 98 and both rely on co-operative gene induction by pCREB and CBP53, 96, 99,
100. LTF relies on deactivation of CREB2, a transcriptional repressor101. Similarly, relief of
19
transcriptional repression due to ATF4, a mammalian analogue of CREB2, appears to be
important for the maintenance of hippocampal LTP102, 103. Thus, the results with Aplysia
suggest it may be possible, in complex organisms including mammals, to computationally
predict the efficacies of numerous learning or training protocols, a process that is impractical
using empirical studies alone.
Given that knowledge of the underlying biochemical cascades can help develop models
to predict optimal training protocols, can models also be used to predict pharmacological targets
to improve memory? The time may also be right for such an approach. For example, if
simulated LTP deficits were rescued by combined parameter changes corresponding to known
drug effects, these 'best' parameter combinations might prioritize drug combinations for testing
in animal models. A recent study104 took a first step by modelling LTP induction and
transcriptional regulation by CREB, and simulating the effects of drugs on LTP by altering the
model's parameters. In this model, the magnitude of LTP induction was represented by an
increase in a synaptic weight variable. LTP impairment seen in a mouse model of RTS95 was
first simulated. Then, starting from this simulation, the parameters were altered in ways
corresponding to plausible single-drug effects. However, no single-drug effect completely
rescued LTP. Therefore, pairs of parameter changes were considered, corresponding to
plausible paired-drug effects. Two pairs were identified that restored LTP. In the first case, an
increased rate constant
to application of an
acetyltransferase activator, was paired with an increased duration of stimulus-induced cAMP
elevation, corresponding to application of a cAMP phosphodiesterase (PDE) inhibitor. The
second pair corresponded to a PDE inhibitor paired with a deacetylase inhibitor. For both pairs,
additive synergism, defined as a combined-drug effect that exceeds the summed effects of the
individual drugs, was also evident, as quantified by a simple additive measure (FIG. 5). A
subsequent empirical study by another group did find that pairing a PDE inhibitor with a
deacetylase inhibitor was effective in rescuing a deficit of LTP in a mouse model of Alzheimer
disease105. A further extension of these strategies might similarly predict, and empirically test,
enhancement of synaptic plasticity when pharmacotherapy is combined with computationally
designed spaced protocols.
for histone acetylation, corresponding
Figure 5 A model predicts that a pair of drugs
can act synergistically to enhance LTP. A CREB
binding protein (Cbp) mutation impairs hippocampal
LTP and impairs learning in mice, and Cbp+/- mice
are considered a model for aspects of Rubinstein-
20
Taybi syndrome in humans104. We developed a model to examine whether drugs could be used to
overcome this impairment in LTP. This figure was generated from a series of simulations of the effects of
two drugs on the induction of LTP. LTP was modeled as the percent increase in a synaptic weight
variable. In the absence of drugs, simulated LTP induced by a high-frequency tetanic stimulus was
strongly impaired. Only a 50% increase in synaptic weight for Cbp+/- occurred, compared to an increase in
synaptic weight of 148% with non-mutated Cbp. The effect of each drug was simply modeled as a change
in the value of a kinetic parameter. In this series of simulations, the doses of two drugs - drug 1, a cAMP
phosphodiesterase inhibitor, and drug 2, an acetyltransferase activator - were concurrently varied. The
effect of drug 1 was simulated by decreasing a rate constant for cAMP degradation, and the effect of drug
2 was simulated by increasing a rate constant for histone acetylation. The 'dose' of drug 1 - the
amplitude of the rate constant change - was increased and simultaneously the dose of drug 2 was
decreased. 80 pairs of drug doses were simulated. Both drugs substantially enhanced LTP. For drug 2
alone (left endpoint of graph) LTP was 155%, and for drug 1 alone (right endpoint), LTP was 116%. For
both drugs together, with smaller doses of each drug, intermediate LTP amplitudes were observed
(combined effect curve). This series of simulations further shows that additive synergism persists over a
substantial range of drug doses. Additive synergism is quantified as the difference (red arrow) between
the LTP simulated when both drugs are applied together (combined effect curve), and the LTP simulated
by adding together the effect of the drugs applied individually in separate simulations (summed effects
curve).
Future directions
There is reason for optimism that more predictive models for determining optimal intervals
between learning trials will be available in the near future, because the molecular data that are
necessary for the development of such models, delineating dynamics of signaling pathways that
are important for LTP and long-term memory, continue to accumulate rapidly. However, despite
the progress being made in understanding the molecular mechanisms of the spacing effect,
some aspects of this effect cannot be explained by current models and constitute important
directions for future research. For example, in human verbal learning, an interesting positive
correlation exists between the length of inter-trial intervals for effective spaced learning and the
retention interval (that is, the interval between the final training trial and the test of memory
retention). With relatively short retention intervals (~1 min–2 h), training intervals in the broad
range of ~1 min to 3 h yield greater verbal learning than do training intervals of 2 d or more2.
With a longer retention interval of 1 d, a 1 d training interval yielded greater learning than did a
very short (< 30 s) interval. For verbal learning with a retention interval of 6 months, a training
interval of 7 d was superior to an interval of 3 d71. This correlation between longer training and
retention intervals suggests that longer training intervals preferentially form a memory trace with
a very long lifetime. For the temporal range of minutes versus hours, it is plausible that a longer
trace lifetime corresponds, at least in part, to increased activation of transcription by the longer
training intervals. However, this explanation may not suffice when comparing training intervals
of ~ 1 d versus many days. It would be of interest to determine whether reactivation of stored
21
memory representations at the network level, or transfer of these representations between brain
regions, contribute to this correlation.
Another challenge will be to use innovative strategies to test the predictions of the
cognitive theories for the spacing effect. For example, consider the variant of deficient-
processing theory positing that repeated cognitive rehearsals of a memory are needed for
consolidation. A neuronal correlate of rehearsals is, plausibly, repeated activation of a specific
neuron assembly that serves as a locus of storage of a long-term memory trace. Empirically, is
such repeated activation necessary for persistence of memory for days or longer? Repeated
spontaneous activation of neuron assemblies does occur106, 107 as does repeated replay or
rehearsal of assemblies that encode recent experiences108, 109. One study supporting the
necessity of such replay found that the post-training suppression of activity of neurons that were
engineered to overexpress CREB in the amygdala for several h after conditioning blocked the
consolidation of a memory of association between cocaine and a location110. Similar blocking
effects were obtained by the indiscriminate activation of neurons that overexpressed CREB.
Although encouraging, these manipulations lack the cellular precision that is necessary to
demonstrate conclusively that reactivation of a particular assembly of neurons is essential for
the persistence of long-term memory. Future studies using optogenetic techniques could
provide that precision. Similarly, innovative strategies will be needed to address whether
effective spaced learning requires the binding of contextual and episodic memories at the
neuronal network level, such as posited by encoding variability theory, or increased binding due
to greater retrieval effort, as posited by study-phase retrieval theory.
The successful prediction of the interval structure of behavioral training protocols that
may overcome some human learning deficits (when applied alone or in combination with
pharmacotherapy) will require improved knowledge of the signaling pathways that underlie LTP
and long-term memory and of the ways in which the deficits affect those pathways. Future
models are still likely to be incomplete owing to gaps in knowledge. For example, data will be
incomplete and associated with unavoidable uncertainties in the values of biochemical
parameters in models such as enzyme activities or protein concentrations. In model
development, data from several preparation types (for example, cell cultures and slices) and
species (for example, primates and rodents) commonly need to be used to estimate different
parameters86, 88. However, although these limitations are important, the potential benefits of
combining modeling with experiments in the ways discussed herein are extensive, such that this
22
strategy may have promise for improving the clinical and educational outcomes for patients with
learning and memory deficits. In addition, it is possible that education and learning in individuals
without such deficits could benefit from such a strategy. Indeed, enhancing normal learning by
judicious pharmacotherapy has recently received attention111, and combining drugs with
optimized spaced learning protocols might yield even better outcomes.
Glossary
Reinforcement: A broad term used here to describe a stimulus or item that enhances the
strength or lifetime of a memory. Reinforcement stimuli activate biochemical and molecular
processes and thereby regulate changes in synaptic strength associated with memory.
Memory Extinction: The decline of a learned behavioral response to a conditioned stimulus
following the withdrawal of reinforcement stimuli that were previously paired with repetitions of
the conditioned stimulus. This term is distinct from forgetting, which is the decline of a learned
response during the prolonged absence of stimulus repetitions.
Habituation: A decrease of the behavioral response to a stimulus following frequent repetitions
of that stimulus. This term is distinct from extinction, because habituation can denote a
decrease of response to a stimulus that was never paired with a reinforcing stimulus.
Memory Reactivation: Re-instantiation of a conditioned behavioral response or of neural activity
associated with a specific response. Reactivation can be elicited by presentation of a
conditioning stimulus or of the context in which learning previously occurred, or it can be
spontaneous, occurring as a part of normal ongoing neural activity.
Drug Synergism: In combined-drug treatment, a synergistic effect of the combination is an
effect greater than what would be predicted by considering the individual drugs as independent
and not interacting. Several methods for assessing drug synergism exist. The method referred
to here (REF. 104) is one of the simplest.
References
23
1. Ebbinghaus, H. Memory: a Contribution to Experimental Psychology, Chapter 8, p. 89
(Teachers College, Columbia University, New York, 1885/1913) (reprinted Thoemmes
Press, Bristol, 1999).
2. Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T. & Rohrer, D. Distributed practice in verbal
recall tasks: A review and quantitative synthesis. Psychol. Bull. 132, 354-380 (2006).
This extensive review and meta-analysis delineates the comprehensive body of
knowledge describing human spaced and massed verbal learning, as well as
theories posited to explain spaced training's superiority to massed training in terms
of inducing long-term memory formation.
3. Godbole, N. R., Delaney, P. F. & Verkoeijen, P. P. The spacing effect in immediate and
delayed free recall. Memory 22, 462-469 (2014).
4. Raman, M., McLaughlin, K., Violato, C., Rostom, A., Allard, J. P. & Coderre, S. Teaching in
small portions dispersed over time enhances long-term knowledge retention. Medical
Teacher 32, 250-255 (2010).
5. Donovan, J. J. & Radosevich, D. J. A meta-analytic review of the distribution of practice
effect. J. Appl. Psychol. 84, 795-805 (1999).
6. Shea, C. H., Lai, Q., Black, C. & Park, J. H. Spacing practice sessions across days benefits
the learning of motor skills. Human Movement Sci. 19, 737-760 (2000).
7. Gluckman, M., Vlach, H. A. & Sandhofer, C. M. Spacing simultaneously promotes multiple
forms of learning in children's science curriculum. Appl. Cognit. Psychol. 28, 266-273
(2014).
8. McDaniel, M. A. & Callender, A. A. Cognition, memory, and education. In: Learning and
Memory: A Comprehensive Reference, Vol. 2 (ed. Roediger, H. L.) 819-844 (Academic
Press, Oxford, 2008).
9. Sobel, H. S., Cepeda, N. J. & Kapler, I. V. Spacing effects in real-world classroom vocabulary
learning. Appl. Cognit. Psychol. 25, 763-767 (2011).
10. Vlach, H. A. The spacing effect in children's generalization of knowledge: allowing children
time to forget promotes their ability to learn. Child Devel. Persp. 8, 163-168 (2014).
This study found that, in children's education, spacing of learning sessions promoted
not only better fact retention but also generalization of currently learned items to
bind with and extend previously established conceptual categories.
24
11. Philips, G. T., Kopec, A. M, & Carew, T. J. Pattern and predictability in memory formation:
from molecular mechanisms to clinical relevance. Neurobiol. Learn. Mem. 105, 117-124
(2013a).
12. Philips, G. T., Ye, X., Kopec, A. M. & Carew, T. J. MAPK establishes a molecular context that
defines effective training patterns for long-term memory formation. J. Neurosci. 33, 7565-
7573 (2013b).
This work characterized the dynamics of MAPK and its activation of C/EBP, a
transcription factor that is important for LTM, identifying the response of MAPK
within a specific time window as essential for generating an optimal inter-trial
interval.
13. Wainwright, M. L., Zhang, H., Byrne, J. H. & Cleary, L. J. Localized neuronal outgrowth
induced by long-term sensitization training in Aplysia. J. Neurosci. 22, 4132-4141 (2002).
14. Mauelshagen, J., Sherff, C. M. & Carew, T. J. Differential induction of long-term synaptic
facilitation by spaced and massed applications of serotonin at sensory neuron synapses of
Aplysia californica. Learn. Mem. 5, 246–256 (1998).
15. Beck, C. D., Schroeder, B. & Davis, R. L. Learning performance of normal and mutant
Drosophila after repeated conditioning trials with discrete stimuli. J. Neurosci. 20, 2944-
2953 (2000).
16. Pagani, M. R., Oishi, K., Gelb, B. D. & Zhong, Y. The phosphatase SHP2 regulates the
spacing effect for long-term memory induction. Cell 139, 186-198 (2009).
17. Menzel, R., Manz, G., Menzel, R. & Greggers, U. Massed and spaced learning in
honeybees: the role of CS, US, the intertrial interval, and the test interval. Learn. Mem. 8,
198-208 (2001).
18. Anderson, M. J., Jablonski, S. A. & Klimas, D. B. Spaced initial stimulus familiarization
enhances novelty preference in Long-Evans rats. Behav. Processes 78, 481-486 (2008).
19. Bello-Medina, P. C., Sanchez-Carrasco, L., Gonzalez-Ornelas, NR., Jeffery, K. J. &
Ramirez-Amaya, V. Differential effects of spaced vs. massed training in long-term object-
identity and object-location recognition memory. Behav. Brain Res. 250, 102-113 (2013).
20. Medin, D. L. The comparative study of memory. J. Human Evol. 3, 455-463 (1974).
21. Robbins, O. & Bush, C. T. Memory in great apes. J. Exp. Psychol. 97, 344-348 (1973).
25
22. Urcelay, G. P., Wheeler, D. S. & Miller, R. R. Spacing extinction trials alleviates renewal and
spontaneous recovery. Learn. Behav. 37, 60-73 (2009).
23. Benjamin, A. S. & Tullis, J. What makes distributed practice effective? Cogn. Psychol. 61,
228-247 (2010).
24. Madigan, S. A. Intraserial repetition and coding processes in free recall. J. Verb. Learn.
Verb. Behav. 8, 828-835 (1969).
25. Melton, A. W. The situation with respect to the spacing of repetitions and memory. J. Verb.
Learn. Verb. Behav. 9, 596-606 (1970).
26. Braun, K. & Rubin, D. C. The spacing effect depends on an encoding deficit, retrieval, and
time in working memory: evidence from once-presented words. Memory 6, 37-65 (1998).
27. Greene, R. L. Spacing effects in memory: Evidence for a two-process account. J. Exp.
Psychol. Learn. Mem. Cogn. 15, 371-377 (1989).
28. Hintzman, D. L., Summers, J. J. & Block, R. A. Spacing judgments as an index of study-
phase retrieval. J. Exp. Psychol. Hum. Learn. 1, 31-40 (1975).
29. Tzeng, O. J. & Cotton, B. A study-phase retrieval model of temporal coding. J. Exp. Psychol.
Hum. Learn. 6, 705-716 (1980).
30. Siegel, L. L. & Kahana, M. J. A retrieved context account of spacing and repetition effects in
free recall. J. Exp. Psychol. Learn. Mem. Cogn. 40, 755-764 (2014).
31. Hintzman, D. L. Theoretical implications of the spacing effect. In: Theories in Cognitive
Psychology: The Loyola Symposium (ed. Solso, R. L.) 77-90 (Erlbaum, Potomac, 1974).
32. Dempster, F. N. Spacing effects and their implications for theory and practice. Educ.
Psychol. Rev. 1, 309-330 (1989).
33. Toppino, T. C. The spacing effect in young children's free recall: support for automatic-
process explanations. Mem. Cogn. 19, 159-167 (1991).
34. Hintzman, D. L. Repetition and memory. In: The Psychology of Learning and Motivation, Vol.
10 (ed. Bower, G. H.) 47-91 (Academic Press, New York, 1976).
35. Rundus, D. Analysis of rehearsal processes in free recall. J. Exp. Psychol. 89, 63-77 (1971).
36. Lechner, H. A., Squire, L. R. & Byrne, J. H. 100 years of consolidation-remembering Müller
and Pilzecker. Learn. Mem. 6, 77-87 (1999).
26
37. Landauer, T. K. Reinforcement as consolidation. Psychol. Rev. 76, 82-96 (1969).
This early paper posited and described elements of the dynamics of spaced learning,
such as the saturation of memory formation by closely spaced trials and the failure
of reinforcement for long trial spacing, and how these elements generate an optimal
inter-trial interval.
38. Peterson, L. R. Short-term verbal memory and learning. Psychol. Rev. 73, 193-207 (1966).
39. Wickelgren, W. A. Trace resistance and the decay of long-term memory. J. Math. Psychol. 9,
418-455 (1972).
40. Cuddy, L. J. & Jacoby, L. J. When forgetting helps memory: an analysis of repetition effects.
J. Verb. Learn. Verb. Behav. 21, 451-467 (1982).
41. Toppino, T. C. & Bloom, L. C. The spacing effect, free recall, and two-process theory: a
closer look. J. Exp. Psychol. Learn. Mem. Cogn. 28, 437-444 (2002).
42. Nabavi, S., Fox, R., Proulx, C. D., Lin, J.Y., Tsien, R. Y. & Malinow, R. Engineering a
memory with LTD and LTP. Nature 511, 348-352 (2014).
43. Bosch, M. & Hayashi, Y. Structural plasticity of dendritic spines. Curr. Opin. Neurobiol. 22,
383-388 (2012).
44. Bell, M. E., Bourne, J. N., Chirillo, M. A., Mendenhell, J. M., Kuwajima, M. & Harris K. M.
Dynamics of nascent and active zone ultrastructure as synapses enlarge during LTP in
mature hippocampus. J. Comp. Neurol. 522, 3861-3864 (2014).
This work, one of a group of related studies by the Harris laboratory, characterized in
detail pre- and postsynaptic structural changes that are necessary for the
consolidation of LTP.
45. Bourne, J. N. & Harris, K. M. Coordination of size and number of excitatory and inhibitory
synapses results in a balanced structural plasticity along mature hippocampal CA1
dendrites during LTP. Hippocampus 21, 354-373 (2011).
46. Cao, G. & Harris K. M. Augmenting saturated LTP by broadly spaced episodes of theta-burst
stimulation in hippocampal area CA1 of adult rats and mice. J. Neurophysiol. 112, 1916-
1924 (2014).
27
47. Kramár, E. A., Babayan, A. H., Gavin, C. F., Cox, C. D., Jafari, M., Gall, C. M., Rumbaugh,
G. & Lynch, G. Synaptic evidence for the efficacy of spaced learning. Proc. Natl. Acad. Sci.
USA 109, 5121-5126 (2012).
This study, together with REF. 48, substantiates the key hypothesis that spaced stimuli
first 'prime' and then strengthen upon repetition successive groups of dendritic
spines; creating a morphological correlate of the memory trace that spaced, but not
massed, protocols effectively consolidate.
48. Lynch, G., Kramár, E. A., Babayan, A. H., Rumbaugh, G. & Gall, C. M. Differences between
synaptic plasticity thresholds result in new timing rules for maximizing long-term
potentiation. Neuropharmacology 64, 27-38 (2013).
49. Rex, C. S., Chen, L. Y., Sharma, A., Liu, J., Babayan, A. H., Gall, C. M. & Lynch, G. Different
Rho GTPase-dependent signaling pathways initiate sequential steps in the consolidation of
long-term potentiation. J. Cell Biol. 186, 85-97 (2009).
50. Nguyen, P. V., Abel, T. & Kandel, E. R. Requirement of a critical period of transcription for
induction of a late phase of LTP. Science 265, 1104-1107 (1994).
51. Scharf, M. T., Woo, N. H., Lattal, K. M., Young, J. Z., Nguyen, P. V. & Abel, T. Protein
synthesis is required for the enhancement of long-term potentiation and long-term memory
by spaced training. J. Neurophysiol. 87, 2770-2777 (2002).
52. Kandel, E. R. The molecular biology of memory storage: a dialogue between genes and
synapses. Science 294, 1249-1254 (2001).
53. Silva, A. J., Kogan, J. H., Frankland, P. W. & Kida, S. CREB and memory. Annu. Rev.
Neurosci. 21, 127-148 (1998).
54. Liu, R. Y., Fioravante, D., Shah, S. & Byrne, J. H. cAMP response element-binding protein 1
feedback loop is necessary for consolidation of long-term synaptic facilitation in Aplysia. J.
Neurosci. 28, 1970-1976 (2008).
55. Liu, R. Y., Cleary, L. J. & Byrne, J. H. The requirement for enhanced CREB1 expression in
consolidation of long-term synaptic facilitation and long-term excitability in sensory neurons
of Aplysia. J. Neurosci. 31, 6871-6879 (2011a).
56. Liu, R. Y., Shah, S., Cleary, L. J. & Byrne, J. H. Serotonin- and training-induced dynamic
regulation of CREB2 in Aplysia. Learn. Mem. 18, 245-249 (2011b).
28
57. Frost, W. N., Castellucci, V., Hawkins, R. D., & Kandel, E. R. Monosynaptic connections
made by the sensory neurons of the gill- and siphon-withdrawal reflex in Aplysia participate
in the storage of long-term memory for sensitization. Proc. Natl. Acad. Sci. USA 82, 8266-
8269 (1985).
58. Pinsker, H. M., Hening, W. A., Carew, T. J., & Kandel, E. R. Long-term sensitization of a
defensive withdrawal reflex in Aplysia. Science 182, 1039-1042 (1973).
59. Beckinschtein, P., Cammarota, M., Igaz, L. M., Bevilaqua, L. R., Izquierdo, I. & Medina, J. H.
Persistence of long-term memory storage requires a late protein synthesis- and BDNF-
dependent phase in the hippocampus. Neuron 53, 261-277 (2007).
60. Bambah-Mukku, D., Travaglia, A., Chen, D. Y., Pollonini, G. & Alberini, C M. A positive
loop via C/EBPβ mediates hippocampal memory
autoregulatory BDNF
consolidation. J. Neurosci. 34, 12547-12559 (2014).
feedback
This study characterizes the dynamics of a molecular positive feedback
loop
demonstrated to be important for the consolidation of long-term (24 h or more)
memory, as well as the late repression of Bdnf that terminates the feedback.
61. Genoux, D., Haditsch, U., Knobloch, M., Michalon, A., Storm, D. & Mansuy, I. M. Protein
phosphatase 1 is a molecular constraint on learning and memory. Nature 418, 970-974
(2002).
62. Farah, C. A., Weatherill, D., Dunn, T. W., & Sossin, W. S. PKC differentially translocates
during spaced and massed training in Aplysia. J. Neurosci. 29, 10281-10286 (2009).
63. Tully, T., Preat, T., Boynton, S. & Del Vecchio, M. (1994) Genetic dissection of consolidated
memory in Drosophila melanogaster. Cell 79, 35-47.
64. Yin, J. C., Del Vecchio, M., Zhou, H. & Tully, T. CREB as a memory modulator: induced
expression of a dCREB2 activator isoform enhances long-term memory in Drosophila. Cell
81, 107-115 (1995).
65. Smolen, P., Baxter, D. A. & Byrne, J. H. Frequency selectivity, multistability, and oscillations
emerge from models of genetic regulatory systems. Am. J. Physiol. 274, C531-542 (1998).
66. Philips, G. T., Tzvetkova, E. I, & Carew, T. J. Transient mitogen-activated protein kinase
activation is confined to a narrow temporal window required for the induction of two-trial
long term memory in Aplysia. J. Neurosci. 27, 13701-13705 (2007).
29
67. Seese, R. R., Wang, K., Yao, Y. Q., Lynch, G. & Gall, C. M. Spaced training rescues memory
and ERK1/2 signaling in fragile X syndrome model mice. Proc. Natl. Acad. Sci. USA 111,
16907-16912 (2014).
68. Rosenblum, K., Futter, M., Voss, K., Erent, M., Skehel, P. A., French, P., Obosi, L., Jones, M.
W. & Bliss, T. V. The role of extracellular regulated kinases I/II in late-phase long-term
potentiation. J. Neurosci. 22, 5432-5441 (2002).
69. Ajay, S. M., & Bhalla, U. S. A propagating ERKII switch forms zones of elevated dendritic
activation correlated with plasticity. HFSP J. 1, 49-66 (2007).
70. Chen, C., Wu, J., Lin, H., Pai, T., Fu, F., Wu, C., Tully, T. & Chiang, A. Visualizing long-term
memory formation in two neurons of the Drosophila brain. Science 335, 678-685 (2012).
71. Cepeda, N. J., Coburn, N., Rohrer, D., Wixted, J. T., Moser, M. C. & Pashler, H. Optimizing
distributed practice: theoretical analysis and practical implications. Exp. Psychol. 56, 236-
246 (2009).
72. Wei, Y. & Koulakov, A. A. Long-term memory stabilized by noise-induced rehearsal. J.
Neurosci. 34, 15804-15815 (2014).
73. Cui, Z., Wang, H., Tan, Y., Zaia, K. A., Zhang, S. & Tsien, J. Z. Inducible and reversible NR1
knockout reveals a crucial role of the NMDA receptor in preserving remote memories in the
brain. Neuron 41, 781-793 (2004).
This study demonstrated the key point that ongoing neural activity and NMDA receptor
activation are essential for long-lasting preservation of stored memory.
74. Tetzlaff, C., Kolodziejski, C., Timme, M., Tsodyks, M. & Worgotter, F. Synaptic scaling
enables dynamically distinct short- and long-term memory formation. PLoS Comput. Biol. 9,
e1003307 (2013).
75. Miller, P., Zhabotinsky, A. M., Lisman, J. E. & Wang, X. J. The stability of a stochastic
CaMKII switch: dependence on the number of enzyme molecules and protein turnover.
PLoS Biol. 3, e107 (2005).
76. Pettigrew, D. B., Smolen, P., Baxter, D. A. & Byrne, J. H. Dynamic properties of regulatory
motifs associated with induction of three temporal domains of memory in Aplysia. J.
Comput. Neurosci. 18, 163-181 (2005).
30
77. Shema, R., Sacktor, T. C. & Dudai, Y. Rapid erasure of long-term memory associations in
the cortex by an inhibitor of PKM ζ. Science 317: 951-953 (2007).
78. Volk, L. J., Bachman, J. L., Johnson, R., Yu, Y. & Huganir, R. L. PKM ζ is not required for
hippocampal synaptic plasticity, learning and memory. Nature 493, 420-423 (2013).
79. Fioriti, L., Myers, C., Huang, Y. Y., Li, X., Stephan, J. S., Trifillief, P., Colnaghi, L., Kosmidis,
S., Drisaldi, B., Pavlopoulos, E., & Kandel, E. R. The persistence of hippocampal-based
memory requires protein synthesis mediated by the prion-like protein CPEB3. Neuron 86,
1433-1448 (2015).
80. Ajay, S. M. & Bhalla, U. S. A role for ERKII in synaptic pattern selectivity on the time scale of
minutes. Eur. J. Neurosci. 20, 2671-2680 (2004).
81. Frey, U. & Morris R. G. Synaptic tagging: implications for late maintenance of hippocampal
long-term potentiation. Trends Neurosci. 21, 181-188 (1998).
82. Babayan, A. H., Krámar, E. A., Barrett, R. M., Jafari, M., Häettig, J., Chen, L. Y., Rex, C. S.,
Lauterborn, J. C., Wood, M. A., Gall, C. M. & Lynch, G. Integrin dynamics produce a
delayed stage of long-term potentiation and memory consolidation. J. Neurosci. 32, 12854-
12861 (2012).
83. Callan, D. E. & Schweighofer, N. Neural correlates of the spacing effect in explicit verbal
semantic encoding support the deficient-processing theory. Hum. Brain Mapp. 31, 645-659
(2010).
84. Karpicke, J. D. & Roediger, H. L. Is expanding retrieval a superior method for learning text
materials? Mem. Cogn. 38, 116-124 (2010).
85. Zhang, Y., Liu, R. Y., Heberton, G. A., Smolen, P., Baxter, D. A., Cleary, L. J. & Byrne, J. H.
Computational design of enhanced learning protocols. Nat. Neurosci. 15, 294-297 (2012).
This study found that computational models describing
long-term synaptic
strengthening can successfully predict optimal spaced training protocols, with
irregular inter-trial intervals, that enhance learning.
86. Bhalla, U. S. & Iyengar, R. Emergent properties of networks of biological signaling
pathways. Science 283, 381-387 (1999).
87. Kim, M., Huang, T., Abel, T. & Blackwell, K. T. Temporal sensitivity of protein kinase A
activation in late-phase long term potentiation. PLoS Comput. Biol. 6, e1000691 (2010).
31
88. Smolen, P., Baxter, D. A. & Byrne, J. H. A model of the roles of essential kinases in the
induction and expression of late long-term potentiation. Biophys. J. 90, 2760-2775 (2006).
89. Aslam, N. & Shouval, H. Z. Regulation of cytoplasmic polyadenylation can generate a
bistable switch. BMC Syst. Biol. 6, 12-20 (2012).
90. Liu, R. Y., Zhang, Y., Baxter, D. A., Smolen, P., Cleary, L. J. & Byrne, J. H. Deficit in long-
term synaptic plasticity is rescued by a computationally predicted stimulus protocol. J.
Neurosci. 33, 6944-6949 (2013).
91. Montarolo, P. G., Goelet, P., Castellucci, V. F., Morgan, J., Kandel, E. R. & Schacher, S. A
critical period for macromolecular synthesis in long-term heterosynaptic facilitation in
Aplysia. Science 234, 1249-1254 (1986).
92. Müller, U. & Carew, T. J. Serotonin induces temporally and mechanistically distinct phases of
persistent PKA activity in Aplysia sensory neurons. Neuron 21, 1423-1434 (1998).
93. Korzus, E., Rosenfeld, M. G & Mayford M. CBP histone acetyltransferase activity is a critical
component of memory consolidation. Neuron 42, 961-972 (2004).
94. Park, E., Kin, Y., Ryu, H., Kowall, N. W., Lee, J. & Ryu, H. Epigenetic mechanisms of
Rubinstein-Taybi syndrome. Neuromol. Med. 16, 16-24 (2014).
95. Alarcon, J. M., Malleret, G., Touzani, K., Vronskaya, S., Ishii, S., Kandel, E. R. & Barco, A.
Chromatin acetylation, memory, and LTP are impaired in CBP+/- mice: a model for the
cognitive deficit in Rubinstein-Taybi syndrome and its amelioration. Neuron 42, 947-959
(2004).
96. Alberini, C. M. Transcription factors in long-term memory and synaptic plasticity. Physiol.
Rev. 89, 121–145 (2009).
97. Zhou, L., Zhang, Y., Liu, R. Y., Smolen, P., Cleary, L. J. & Byrne, J. H. Rescue of impaired
long-term facilitation at sensorimotor synapses of Aplysia following siRNA knockdown of
CREB1. J. Neurosci. 35, 1617-1626 (2015).
98. Abel, T. & Nguyen, P. V. Regulation of hippocampus-dependent memory by cyclic AMP-
dependent protein kinase. Prog. Brain Res. 169, 97-115 (2008).
99. Bartsch, D., Casadio, A., Karl, K. A., Serodio, P. & Kandel, E. R. CREB1 encodes a nuclear
activator, a repressor, and a cytoplasmic modulator that form a regulatory unit critical for
long-term facilitation. Cell 95, 211-223 (1998).
32
100. Kida, S. A functional role for CREB as a positive regulator of memory formation and LTP.
Exp. Neurobiol. 21, 136-140 (2012).
101. Bartsch, D., Ghirardi, M., Skehel, P. A., Karl, K. A., Herder, S. P., Chen, M., Bailey, C. H. &
Kandel, E. R. Aplysia CREB2 represses long-term facilitation: relief of repression converts
transient facilitation into long-term functional and structural change. Cell 83, 979-992
(1995).
102. Dong, C., Upadhya, S. C., Ding, L., Smith, T. K. & Hegde, A. N. Proteosome inhibition
enhances the induction and impairs the maintenance of late-phase long-term potentiation.
Learn. Mem. 15, 335-347 (2008).
103. Pasini, S., Corona, C., Liu, J., Greene, L. A., & Shelanski, M. L. Specific downregulation of
hippocampal ATF4 reveals a necessary role in synaptic plasticity and memory. Cell
Reports 11, 183-191 (2015).
104. Smolen, P., Baxter, D. A. & Byrne, J. H. Simulations suggest pharmacological methods for
rescuing long-term potentiation. J. Theoret. Biol. 360, 243-250 (2014).
105. Cuadrado-Tejedor, M., Garcia-Barroso, C., Sanzhez-Arias, J., Mederos, S., Rabal, O.,
Ugarte, A., Franco, R., Pascual-Lucas, M., Segura, V., Perea, G., Oyarzabal, J., & Garcia-
Osta, A. Concomitant histone deacetylase and phosphodiesterase 5
inhibition
synergistically prevents the disruption in synaptic plasticity and it reverses cognitive
impairment in a mouse model of Alzheimer's disease. Clin. Epigenetics 7, 108 (2015).
106. Ikegaya, Y., Aaron, G., Cossart, R., Aronov, D., Lampl, I., Ferster, D. & Yuste, R. Synfire
chains and cortical songs: temporal modules of cortical activity. Science 304, 559-564
(2004).
107. Miller, J. E., Ayzenshtat, I., Carrillo-Reid, L. & Yuste, R. Visual stimuli recruit intrinsically
generated cortical ensembles. Proc. Natl. Acad. Sci. USA 111, E4053-4061 (2014).
108. Karlsson, M. P. & Frank, L. M. Awake replay of remote experiences in the hippocampus.
Nat. Neurosci. 12, 913-919 (2009).
109. Wu, X. & Foster, D. J. Hippocampal replay captures the unique topological structure of a
novel environment. J. Neurosci. 34, 6459-6469 (2014).
110. Hsiang, H. L., Epp, J. R., van den Oever, M. C., Yan, C., Rashid, A. J., Insel, N., Ye, L.,
Niibori, Y., Deisseroth, K., Frankland, P. W. & Josselyn, S. A. Manipulating a "cocaine
engram" in mice. J. Neurosci. 34, 14115-14127 (2014).
33
111. Lynch, G., Cox, C. D. & Gall, C. M. Pharmacological enhancement of memory or cognition
in normal subjects. Front. Syst. Neurosci. 8, doi: 10.3389/fnsys.2014.00090 (2014).
Acknowledgements
This work was supported by National Institutes of Health grants NS073974 and NS019895.
34
|
1210.6317 | 1 | 1210 | 2012-10-23T18:21:16 | On the geometric structure of fMRI searchlight-based information maps | [
"q-bio.NC",
"q-bio.QM",
"stat.AP",
"stat.ML"
] | Information mapping is a popular application of Multivoxel Pattern Analysis (MVPA) to fMRI. Information maps are constructed using the so called searchlight method, where the spherical multivoxel neighborhood of every voxel (i.e., a searchlight) in the brain is evaluated for the presence of task-relevant response patterns. Despite their widespread use, information maps present several challenges for interpretation. One such challenge has to do with inferring the size and shape of a multivoxel pattern from its signature on the information map. To address this issue, we formally examined the geometric basis of this mapping relationship. Based on geometric considerations, we show how and why small patterns (i.e., having smaller spatial extents) can produce a larger signature on the information map as compared to large patterns, independent of the size of the searchlight radius. Furthermore, we show that the number of informative searchlights over the brain increase as a function of searchlight radius, even in the complete absence of any multivariate response patterns. These properties are unrelated to the statistical capabilities of the pattern-analysis algorithms used but are obligatory geometric properties arising from using the searchlight procedure. | q-bio.NC | q-bio | On the geometric structure of fMRI searchlight-based information maps
Shivakumar Viswanathan1, Matthew Cieslak, Scott T. Grafton
Department of Psychological and Brain Sciences,
University of California, Santa Barbara, California, USA
Abstract
2
1
0
2
t
c
O
3
2
]
.
C
N
o
i
b
-
q
[
1
v
7
1
3
6
.
0
1
2
1
:
v
i
X
r
a
Information mapping is a popular application of Multivoxel Pattern Analysis (MVPA) to fMRI. Information
maps are constructed using the so called searchlight method, where the spherical multivoxel neighborhood
of every voxel (i.e., a searchlight) in the brain is evaluated for the presence of task-relevant response patterns.
Despite their widespread use, information maps present several challenges for interpretation. One such chal-
lenge has to do with inferring the size and shape of a multivoxel pattern from its signature on the information
map. To address this issue, we formally examined the geometric basis of this mapping relationship. Based
on geometric considerations, we show how and why small patterns (i.e., having smaller spatial extents) can
produce a larger signature on the information map as compared to large patterns, independent of the size
of the searchlight radius. Furthermore, we show that the number of informative searchlights over the brain
increase as a function of searchlight radius, even in the complete absence of any multivariate response pat-
terns. These properties are unrelated to the statistical capabilities of the pattern-analysis algorithms used but
are obligatory geometric properties arising from using the searchlight procedure.
Keywords: MVPA, searchlight, FMRI, pattern-classification
1. Introduction
In FMRI, a functional map is an important represen-
tation of how cognitive function is related to neu-
roanatomy. Such maps provide a topographic rep-
resentation of the brain regions that are (and are
not) systematically responsive to differing values of
a cognitive variable. The size, shape, number and
location of the “blobs” (i.e., voxel-clusters meeting
some statistical relevance criterion) on the functional
maps are the basis for inferences about the neural
substrates of the cognitive process. Given the impor-
tance of functional maps, there is a continuing need
to scrutinize the sensitivity, precision and technical
assumptions of the mapping procedure itself. The
topic of the current technical note is the mapping
procedure used to generate the widely used informa-
tion maps (Kriegeskorte et al., 2006).
The motivation for information mapping is the sta-
tistical concern that a region’s responses to the cog-
nitive variable under study might take a complex
1Corresponding author: [email protected]
Preprint submitted to arXiV
multivariate form. For example, a group of mul-
tiple voxels might conjointly respond in a task-
relevant manner even though individual voxels may
not detectably do so (Haxby et al., 2001; Cox and
Savoy, 2003; Haynes, 2006; Norman et al., 2006;
Mur et al., 2008; Tong, 2010; Wagner and Rissman,
2010; Formisano and Kriegeskorte, 2012; Serences
and Saproo, 2012). Such distributed response pat-
terns might be effectively undetectable with conven-
tional univariate statistical tests restricted to indi-
vidual voxel responses, but detectable with an ex-
plicit multivariate test for multivoxel response pat-
terns, i.e., using some Multivoxel Pattern Analysis
(MVPA) technique. To address this concern in the
context of functional mapping, Kriegeskorte et al.
(2006) proposed a simple procedure to enable so-
phisticated MVPA methods to be readily applied to
detect and map brain regions that contain informa-
tion about the experimental conditions, irrespective
of whether the informative responses are univariate
or multivariate.
In the proposed procedure, the unit of evaluation is
not the single voxel but a “searchlight” – the group of
May 3, 2014
voxels contained in a spherical neighborhood of ra-
dius r around a single voxel. The searchlight statis-
tic is a measure of whether the conjoint responses
of this group of voxels contain information about
the experimental conditions being tested. Based
on these abstractions, an information map is gen-
erated as follows: the searchlight statistic is evalu-
ated for searchlights centered at every voxel in the
brain; and the statistic’s value for each searchlight
is mapped to the central voxel of that searchlight.
The resulting topographic representation generated
by the searchlight-procedure has been referred to as
the information map. Such searchlight-based infor-
mation maps are now routinely reported in studies
that employ MVPA methods (for example, Haynes
et al., 2007; Soon et al., 2008; Johnson et al., 2009;
Poldrack et al., 2009; Chadwick et al., 2010; Ooster-
hof et al., 2010; Nestor et al., 2011; Alink et al., 2011;
Golomb and Kanwisher, 2011; Peelen and Kastner,
2011; Stokes et al., 2011; Woolgar et al., 2011; Mor-
gan et al., 2011; Oosterhof et al., 2012; Connolly et al.,
2012; Kaplan and Meyer, 2012).
Notwithstanding their popularity,
interpreting a
searchlight-based information map presents a va-
riety of challenges (Kriegeskorte et al., 2006; Pol-
drack et al., 2009; Pereira and Botvinick, 2011; Jimura
and Poldrack, 2012). One such challenge is posed
by the topographic ambiguity of the information
map. Recall that a searchlight statistic computed on
the responses of an entire multivoxel searchlight is
mapped to a single voxel on the information map,
namely, that searchlight’s central voxel. This map-
ping protocol is applied to searchlights across the
brain irrespective of the number or the spatial loca-
tions of the information-carrying voxels within each
searchlight. Consequently, the spatial position of
an informative voxel on the information maps is a
coarse index to the actual location of the informative
“pattern” within that voxel’s searchlight. Further-
more, since a searchlight has a unique central voxel
v, an informative voxel on the information map is
not indicative of the actual number of voxels con-
stituting the informative pattern within that voxel’s
searchlight neighborhood. Given these properties of
the information map, we asked: what, if anything,
can be reliably inferred about the size and shape of a
multivoxel pattern from its corresponding signature
on the information map?
Previous studies have treated this question as a
qualitative concern requiring cautious interpreta-
tion. Nonetheless, here we show that information
maps are in fact subject to several crisply quantita-
tive geometric constraints that strongly govern how
such maps can be interpreted.
Our analytical results are based on a simple geomet-
ric intuition. Since a multivoxel searchlight is de-
fined at every voxel across the brain, searchlights
centered at different voxels systematically overlap
each other, i.e., have voxels in common. Using over-
lapping searchlights is crucial to obtain a continuous
topographic coverage especially when the locations
and spatial extents of voxel-neighborhoods that are
task-responsive are unknown a priori. We observed
that due to these overlaps, multiple searchlights
would be deemed informative merely by virtue of
sharing the same task-relevant multivoxel response
patterns. Thus we reasoned that the size and shape
of a multivoxel group G’s signature on the informa-
tion map should be defined by exactly those vox-
els which have searchlight-neighborhoods that con-
tain G. Using this observation and simple geometric
reasoning, we formally deduce some key properties
of the relationship between an informative pattern
and its corresponding signature on the information
map.
Based on our formal analysis, we prove here that,
for any searchlight radius, a single task-responsive
voxel produces a larger signature on the informa-
tion map as compared to a distributed multivoxel
response pattern. Furthermore, the number of in-
formative searchlights over the brain can increase as
a function of searchlight radius, without necessar-
ily revealing any new information and even in the
complete absence of any multivariate response pat-
terns. Importantly, these properties are largely inde-
pendent of the type of machine-learning algorithm
or the testing protocol used to compute the search-
light statistic.
2. Model
2.1. Definition: The searchlight decomposition
The basis of the searchlight analysis is the geomet-
ric structure of the voxel-space in which the brain
images are defined. The voxel-space V is defined
here as the set of all voxels V augmented with a ge-
ometric structure defining the relative spatial posi-
tion of the voxels in V , and a distance measure be-
tween these voxels. For analytical convenience, we
treat the voxel-space as being uniform and connected
as described below.
A d-dimensional voxel-space V is deemed to be uni-
form if every voxel has a neighboring voxel in all d
principal directions. Additionally, we assume that
the voxel-space V is connected. Specifically, there
is a path connecting every pair of voxels vi and vj
in V with a path defined here to be an ordered se-
quence of voxels (cid:104)vi,··· , vk, vk+1, ....vj(cid:105) where voxel
vk+1 is a neighbor of vk along one of the d princi-
pal directions. These simplifying assumptions are
intended to emphasize the general geometric princi-
ples entailed by the searchlight method while delib-
erately ignoring the special cases associated with (i)
the boundaries of V where a searchlight may be trun-
cated; and (ii) distinctions between gray-matter and
white-matter voxels and any masking of the latter
from the searchlights. Although we refer to search-
lights as being volumes in a voxel-space having di-
mensionality d = 3, the properties derived here are
agnostic to the specific value of d and apply to sur-
faces (d = 2) where the searchlights are discs (as in,
Oosterhof et al., 2011; Chen et al., 2011).
The key abstraction defined by the searchlight
method is a decomposition of V into subsets of
voxels based on a geometric criterion.
Given
a voxel space V, we define a searchlight voxel-
decomposition using the following indexing func-
tion
S : V × R → P(V )
(1)
where P(V) is the powerset of V, namely, the set of
all subsets of V. This indexing function S takes two
inputs – the identity of a voxel v in the voxel-space
V, and a real-value r ∈ R specifying the search-
light’s radius. The searchlight indexing function
uses these parameters in conjunction with the geo-
metric structure of V to extract and output a set of
voxels S(r, v) ∈ P(V ). A voxel v(cid:48) ∈ V is a mem-
ber of S(r, v) if and only if the distance between v(cid:48)
and v is less than or equal to r. For convenience,
we henceforth write Sr(v) to denote the searchlight
S(r, v). The resulting searchlight voxel-decomposition
of V for a given radius r is defined as
Sr(V) = {Sr(v) for all v ∈ V}
(2)
For clarity, we restrict our usage of the term “search-
light” to the cases when the value of the radius of
a searchlight r is such that each Sr(V ) is a multi-
voxel entity that is not identical with V, that is, 1 <
Sr(v) (cid:28) V, for any Sr(v) ∈ Sr(V). We refer to
the univariate case where Sr(v) = 1 as the univoxel
decomposition.
A schematic of the searchlight indexing scheme is
shown in Figure 1.
2.2. Definition: Informativeness function
The searchlight statistic is a measure of whether the
voxels in the searchlight, as a unit, exhibit differ-
ences in their conjoint responses to the experimen-
tal conditions. More generally, it is a measure of
whether the searchlight contains information about
the experimental condition, i.e., whether the search-
light is informative. As with the radius of the search-
light, the specific statistical procedure used to com-
pute the searchlight statistic is a discretionary choice
made by the researcher (for example, see Pereira and
Botvinick, 2011).
To describe the searchlight statistic in a procedure-
independent manner, we use a binary indicator
function, which we refer to as the informativeness
function, I : P(V ) → {0, 1}. Given a subset of voxels
G ∈ P(V ), the function I returns a value of 1 if the
responses of G are deemed to be informative; or 0 if
they are not, based on some appropriately specified
statistical criterion.
Evaluating the informativeness function on the re-
sponses corresponding to each searchlight Sr(v) in
Sr(V) defines the overall information set for a par-
ticular radius
I(Sr(V)) = {I(Sr(v)) for all Sr(v) in Sr(V)}
(3)
The information map is the object obtained when the
information set defined above is augmented with the
geometric structure of the voxel-space V by map-
ping the informativeness value of each searchlight
I(Sr(v)) to its corresponding central voxel v.
The performance measure of interest here is the total
number of informative searchlights for a particular
Figure 1: Cartoon of searchlight indexing scheme: Left panel shows two example voxels va and vb that are mapped to searchlights
Sr(va) (orange) and Sr(vb) (light gray) respectively, having some radius r, as shown in the right panel. Two searchlights can overlap
to varying degrees depending on the distance between their center voxels and the radius. Here the overlap is indicated in dark-
gray. The searchlight statistic computed on each searchlight is mapped back to the corresponding central voxels to generate the
information map (see text).
searchlight decomposition
V(cid:88)
Fr =
I(Sr(vi))
(4)
i=1
2.3. Linking assumptions
Two simple properties link the structure of the
searchlight decomposition Sr(V) to the structure of
the information set I(Sr(V)).
The first property is that, by virtue of the regularity
of their shape and relative positioning, searchlights
in Sr(V) can overlap. Consequently, the same voxels
in V can be included or sampled by multiple search-
lights. The second property is that since a sphere
is an arbitrarily chosen and regular shape, it is un-
likely that every voxel is necessarily task-relevant in
every informative searchlight. Consequently, the in-
formativeness of the responses in some particular
searchlight volume Sr(v) can be alternatively and
accurately interpreted as indicating that some group
of voxels in that searchlight volume exhibits task-
dependent responses. These two properties can be
combined as follows. Let G be a group of task-
relevant voxels in a searchlight Sr(v), where G ⊆
Sr(v). Since searchlights share voxels, if some other
searchlight Sr(v(cid:48)) also contains G, that is G ⊂ Sr(v(cid:48)),
then it implies that Sr(v(cid:48)) should also contain task-
relevant information as it includes the task-relevant
voxels G.
Based on this observation, we make two linking
assumptions about the behavior of the procedures
used to compute the searchlight statistic and hence
the informativeness function I. The first is that we
restrict the focus of our analysis to the common mul-
tivariate procedure that does not include geomet-
ric information about the relative spatial positions
of the voxels in a searchlight while computing that
searchlight’s informativeness. The second is the Su-
perset informativeness (SIN) assumption which postu-
lates that:
Superset informativeness assumption: If a group of
voxels G is informative then every searchlight that con-
tains G is also informative.
That is, according to the SIN assumption, if I(G) = 1
then I(Sr(v)) = 1 for all v ∈ V where G > 0 and
G ⊆ Sr(v). Unless otherwise stated, we will over-
load the symbol I to denote an informativeness func-
tion that explicitly satisfies these two model require-
ments.
Although the SIN assumption is based is on a sound
deduction, the empirical requirement that it poses
may not necessarily be satisfied in practice. Specif-
ically, even if it is known that I(G) = 1, the sta-
tistical procedure used to evaluate informativeness
might fail to detect that a searchlight Sr(v) is in-
formative even if G ⊂ Sr(v). Such a Type II er-
ror (i.e., failing to reject a false null hypothesis that
H0 : I(Sr(v)) = 0) might occur for any of a va-
!!!"!"#$%&’()*$+$),*-%./-0’"##!!$"##!$$ riety of reasons, for example, the use of an inap-
propriate machine-learning algorithm (Pereira and
Botvinick, 2011), insufficient power due to a limited
number of samples, and so on. In this regard, the
SIN assumption treats the multivoxel pattern anal-
ysis techniques as being more sensitive and reliable
than might actually be the case in practice. That is,
the SIN assumption allows us to establish the infor-
mation map’s properties in the best-case indepen-
dent of the performance idiosyncrasies of the spe-
cific multivariate method being used.
Proof. Consider a searchlight Sr(va) centered at
voxel va. Since a searchlight is defined at every voxel
in V (Equation 2), it follows that there is a search-
light defined at every voxel in Sr(va). By definition,
a voxel vb ∈ V is a member of Sr(va) if and only if the
distance between va and vb is less than or equal to the
radius r. Since there is a searchlight Sr(vb) centered
at vb ∈ Sr(va), and the distance between va and vb is
less than or equal to r, it follows that va is a member
of searchlight Sr(vb). Therefore, if vb is a member of
Sr(va) then va is a member of Sr(vb).
3. Analytical results
Our focus of the current section is to establish how
the structure of the sampling bias arises from the
searchlight decomposition. We first prove that due
to the geometric regularities of a searchlight decom-
position, single-voxels and multivoxel groups are
sampled with different frequencies, i.e., included in
a different number of searchlights. Specifically, sin-
gle voxels are included in more searchlights than
multivoxel groups. This sampling difference is in-
dependent of the searchlight radius. We then ex-
tend these results to prove that the frequency with
which voxel-groups are sampled increases with the
radius of the searchlights, irrespective of the number
of voxels in the group. Finally, we prove that the in-
formation map mirrors these sampling biases in an
optimistic manner, i.e., in a manner that is not neces-
sarily warranted by the data.
3.1. Single-voxels and multivoxel-groups are sampled
with different frequencies
The regularity in the shape of the searchlights and
their relative positions the voxel-space define a sys-
tematic relationship between each voxel v ∈ V and
the searchlights in Sr(V) that contain that voxel v.
Firstly, if a voxel va is a member of the searchlight
Sr(vb), then by symmetry, the voxel vb is a mem-
ber of the searchlight Sr(va) (Lemma 1). Secondly,
two distinct voxels va and vb are not simultaneously
included in every searchlight that contains either of
these voxels (Lemma 2).
Lemma 1. If a voxel vb is a member of Sr(va) then the
voxel va is a member of Sr(vb), where va, vb ∈ V and
Sr(va), Sr(vb) ∈ Sr(V.
Lemma 2. For any two non-identical voxels va and vb,
where va, vb ∈ V and Sr(va) (cid:54)= Sr(vb), there necessar-
ily exists a searchlight that contains va but not vb and a
different searchlight that contains vb but not va.
Proof. This claim can be proved in two steps based
on the distance between va and vb.
First consider the case where the distance between
va and vb is greater than 2r, that is, the diameter
of a searchlight. By definition, a searchlight con-
tains voxels that have a distance less than or equal
to r from that searchlight’s central voxel. Due to
the spherical shape of the searchlight, the maximum
distance between any two voxels in a searchlight is
equal to 2r.
If the distance between va and vb is
greater than 2r, there does not exist any searchlight
of radius r that contains both va and vb as members.
Thus, it follows that there exists some searchlight
that contains va but not vb; and some other search-
light that contains vb but not va.
Now consider the second case where the distance be-
tween va and vb is less than or equal to 2r. Since the
distance between these two voxels is less than the
maximum distance between some two voxels in a
searchlight, in a uniform voxel-space there necessar-
ily exists some searchlight Sr(v) that contains both
va and vb as members. Contrary to the proposition,
let us assume that both va and vb are contained in ev-
ery searchlight that contains either va or vb. That is,
if a searchlight Sr(v) contains va, then it necessarily
contains vb, and vice versa. Recall that, from Lemma
1, a voxel va is contained in every searchlight Sr(v)
where v ∈ Sr(va). Now, based on the contradictory
assumption, it implies that vb is also contained in
every such searchlight Sr(v) where v ∈ Sr(va). By
the same reasoning, va should be contained in every
searchlight Sr(v) where v ∈ Sr(vb). If these condi-
tions hold true, then it implies that every voxel in
Sr(va) is also contained in Sr(vb); and every voxel in
Sr(vb) is also contained in Sr(va).
If this the case,
then the searchlights S(va) and S(vb) are identical
as they contain exactly the same voxels. This rela-
tionship, however, contradicts the requirement that
Sr(va) (cid:54)= Sr(vb). Thus, the assumption that va and vb
are both contained in every searchlight that contains
either va or vb cannot be true.
Therefore, there necessarily exists a searchlight that
contains va that does not contain vb, and some other
searchlight that contains vb but not va.
Armed with the properties described by Lemmas 1
and 2, we can now numerically estimate the num-
ber of searchlights that include a given individual
voxel.
Theorem 3. A voxel v is contained in exactly Nr(v) dif-
ferent searchlights, where Nr(v) is the number of voxels
contained in the searchlight Sr(v).
Proof. From Lemma 1, a voxel v is contained in each
searchlight Sr(v(cid:48)), if and only if v(cid:48) is a voxel in
Sr(v). Let Nr(v) be the number of voxels in Sr(v).
Therefore, v is present in each of these Nr(v) search-
lights.
For simplicity, we treat Nr(v) as being the same
for every searchlight, and write Nr to indicate the
canonical number of voxels contained in a spheri-
cal volume of radius r, for a given resolution of the
voxel-space.
From Theorem 3, we see that the radius, a parameter
chosen by the researcher, directly specifies how often
information in a particular voxel is sampled by mul-
tiple searchlights. For voxels of size 3mm × 3mm ×
3mm, the number of voxels contained in searchlights
of different radii are shown in Figure 2. As can be
seen, the number of voxels in a searchlight, that is
Nr, grows rapidly with the radius of the searchlight
r, and consequently so do the number of searchlights
that include a particular voxel.
Figure 2: Number of voxels in a searchlight volume (Nr) as a
function of searchlight radius (r) in a 3×3×3 mm3 voxel-space.
on the relative spatial locations of the voxels in the
group.
Theorem 4. A group of voxels G containing more than
one voxel is contained in strictly less than Nr search-
lights.
Proof. A voxel is contained in Nr searchlights, from
Theorem 3. Consequently, every voxel in G is each
contained in Nr searchlights. From Lemma 2, for any
two voxels va and vb, there is necessarily a search-
light that contains va and not vb, and vice versa.
Therefore, of the Nr searchlights containing va, there
necessarily exists at least one searchlight that con-
tains va but not vb. Thus, the number of searchlights
that simultaneously contain both va and vb must be
less than Nr. Since G contains multiple voxels, any
pair of voxels in G must be simultaneously con-
tained in less than Nr searchlights. Therefore, all the
voxels in G cannot be simultaneously contained in
Nr searchlights, and G must be contained in strictly
less than Nr searchlights.
3.2. The sampling frequency of voxel(s) increases with
searchlight radius
Since searchlights are intended to identify multi-
voxel response patterns, we extend the single-voxel
property in Theorem 3 to quantify the membership
of a group of multiple voxels placing no constraint
Although single-voxels and multivoxel groups are
included in different numbers of searchlights for any
radius r, we now show that the absolute number
of searchlights that include either a single-voxel or
456789101112050100150200250Radius (mm)Number of voxels in searchlight ( Nr ) a multivoxel group increases with the radius of the
searchlight.
Lemma 5. A searchlight of radius ra is fully contained
in more than one searchlight of radius rb, where rb > ra
and Sra(v) ⊂ Srb(v) for all v ∈ V.
Proof. Consider two searchlights centered at the
same voxel v – one that has a radius ra, and the
other having radius rb. By definition, since Sra(v) ⊂
Srb(v), all the voxels in Sra(v) are members of Srb(v),
and there exists at least one voxel in Srb(v) that is not
in Sra(v).
Now, consider the searchlight Sz(v), having radius
z = rb − ra. Due to the spherical shape of search-
lights, the maximum distance between a voxel v(cid:48)(cid:48) in
Sz(v) and some voxel v(cid:48) in Sra(v) is equal to ra + z =
rb. Therefore, all other voxels in Sra(v) must have
distances less than or equal to rb.
Since the distance between these two maximally dis-
tant voxels v(cid:48) and v(cid:48)(cid:48) is equal to rb, the voxel v(cid:48) must
be contained in a searchlight of radius rb that is cen-
tered at v(cid:48)(cid:48), namely, Srb(v(cid:48)(cid:48)). Since all other vox-
els in Sra(v) have a distance less than or equal to
rb from v(cid:48)(cid:48), it follows that every voxel in Sra(v) is
also contained in the searchlight Srb(v(cid:48)(cid:48)). Thus every
voxel in Sra(v) is contained in at least two search-
lights having radius rb, namely, Srb(v) and Srb(v(cid:48)(cid:48)).
Therefore, a searchlight Sra(v) is contained in more
than one searchlight of radius rb, where rb > ra and
Sra(v) ⊂ Srb(v).
Using Lemma 5, we can now prove a general scaling
property. Irrespective of the size of a voxel group,
the frequency with which it is sampled by different
searchlights increases with the radius of the search-
light - a property that we prove next.
Theorem 6. A group of voxels G is contained in more
searchlights of radius rb than searchlights of radius ra,
where G ⊆ Sra, rb > ra and Sra(v) ⊂ Srb(v), for all
v ∈ V.
Proof. Let Kra and Krb be the number of searchlights
of radius ra and rb that contain G.
Since Sra(v) ⊂ Srb(v) for every v ∈ V, it follows, by
transitivity, that if G ⊆ Sra(vi) for some voxel vi ∈ V,
then G ⊆ Srb(vi). Therefore, the number of search-
≮ Kra.
lights of radius rb that contain G cannot be strictly
less than that for ra, that is, Krb
By the transitivity of the subset relation, if G ⊆
Sra(vi) and Sra(vi) ⊂ Srb(vj) for some vi, vj ∈ V,
then it follows that G ⊂ Srb(vj). From Lemma 5,
a searchlight of radius ra is contained in multiple
searchlights of radius rb where rb > ra and Sra(v) ⊂
Srb(v) (for all v ∈ V). Since there is more than one
searchlight of radius rb containing Sra(vi), for every
searchlight for which G ⊆ Sra(vi) holds true, it im-
plies that Krb ≥ Kra.
From Theorems 3 and 4, the number of searchlights
of radius ra that can contain G is less than or equal
to Nra. Consequently, in a uniform and connected
voxel-space, it follows that there exist two adjacent
voxels vi and vj in V such that G is a subset of
Sra(vi) but is not a subset of Sra(vj). From Lemma
5, searchlights of radius rb centered at voxels within
rb − ra from vi fully contain all voxels in Sra(vi).
Since Sra(v) ⊂ Srb(v), it implies that the distance of
vj to vi is less than or equal to rb − ra. Therefore,
Sra(vi) ⊂ Srb(vj) and consequently G ⊂ Srb(vj).
Since G (cid:54)⊂ Sra(vj) and G ⊂ Srb(vj), it implies that
there exists at least one voxel at which a searchlight
of radius rb contains G, but where a searchlight of ra-
dius ra does not contain G. Consequently, Krb must
be strictly greater than Kra, that is, the group of vox-
els G is contained in more searchlights of radius rb
than ra.
Theorem 6 above establishes that the number of
searchlights that include either a voxel or group of
voxels increases monotonically with the radius of
the searchlight. How then does this scaling of the
sampling bias influence the properties of the infor-
mation map?
3.3. An optimistic bias in the information map
Recall that Fr (Equation 4) is an index of the sen-
sitivity of the searchlight method in detecting mul-
tivoxel response patterns, and is equal to the to-
tal number of informative searchlights with a par-
ticular search decomposition. We now prove that
as a direct consequence of how the sampling bias
scales with the searchlight radius, the value of Fr
also increases strictly monotonically with increasing
searchlight radius.
Theorem 7. For two searchlight radii, ra and rb, where
rb > ra and Sra(v) ⊂ Srb(v) for every v ∈ V, if 0 <
Fra < V then Frb > Fra.
Proof. Since Sra(v) ⊂ Srb(v) for every v ∈ V, by the
SIN assumption, it follows that if I(Sra(v)) = 1 then
I(Srb(v)) = 1, for any voxel v ∈ V. Therefore, the
number of informative searchlights of radius rb can-
≮ Fra,
not be strictly less than that for ra, that is, Frb
for any value of Fra.
From Lemma 5, a searchlight of radius ra is con-
tained in multiple searchlights of radius rb where
rb > ra and Sra(v) ⊂ Srb(v) (for all v ∈ V). For every
searchlight for which I(Sra(v)) = 1, there is more
than one searchlight of radius rb containing Sra(v).
Since each informative searchlight of radius ra is a
subset of multiple searchlights of radius rb, by the
SIN assumption, it implies that Frb ≥ Fra.
Let 0 < Fra < V . Since Fra < V , there neces-
sarily exist two adjacent voxels vi and vj such that
I(Sra(vi)) = 1 and I(Sra(vj)) = 0. By the same
logic used to prove Theorem 6, searchlights of ra-
dius rb centered at voxels within rb − ra from vi fully
contain all voxels in Sra(vi). Consequently, by the
SIN assumption, I(Srb(vj)) = 1. This implies that
a searchlight centered at voxel vj is informative if it
has a radius rb but not if it has a radius ra. Therefore
Frb > Fra.
What does Theorem 7 have to do with optimism?
The monotonic increases in the number of informa-
tive searchlights is due to increases in the sampling
bias, which in turn is due to the use of a multivoxel
searchlight. Specifically, it is possible to obtain an
increased “sensitivity” of the information map sim-
ply by increasing the radius of the multivoxel search-
lights, with no reference to the statistical properties
of the voxel-responses, i.e., whether they in fact ex-
hibit multivariate response differences.
4. An illustration
In this section, we present simulations to provide
a concrete intuition for the analytical results above,
and their implications. For ease of demonstration,
the voxel-space V for all simulations consisted of a
single axial slice having two principal directions. All
the voxels in this voxel-space were populated with
simulated response information from two fictitious
experimental conditions A and B. These simulated
data were subjected to the searchlight-procedure to
produce information maps. The radius r of the
searchlights used for the searchlight decomposition
Sr(V) was varied systematically to produce a cor-
responding information map for each radius value.
The radius took the values: 4 mm, 6 mm, 8 mm, 10
mm and 12 mm, corresponding to searchlights con-
taining 5 voxels, 13 voxels, 21 voxels, 37 voxels and
49 voxels respectively.
The simulated response-data differed in the num-
ber and relative spatial location of the voxels that
were responsive to the experimental conditions. In
the first of these simulations discussed next, a single
voxel contained task-relevant information while all
the remaining voxels did not.
4.1. The needle-in-the-haystack effect
Suppose there exists some voxel in V, say v0, that ex-
hibits a response difference to the experimental con-
ditions such that the informativeness function iden-
tifies v0 as being task-relevant, that is, I(v0) = 1.
Since I(v0) = 1, by the SIN assumption it follows
that each of the searchlights that contain v0 should
also be deemed to be informative as well. Recall
that, according to Theorem 3, each voxel v in V is
contained in exactly Nr searchlights where Nr =
Sr(v). It then follows that the signal-carrying voxel
v0 should be contained in Nr searchlights, each be-
ing centered at a voxel in Sr(v0). Thus, a single
signal-carrying voxel (a “needle”) should produce a
cluster having Nr voxels on the information map (a
“haystack”).
To simulate this “needle-in-the-haystack” effect, the
task-relevant responses of v0 in conditions A and
B took the form illustrated in Figure 3(a). The re-
sponses to both conditions were drawn randomly
from a normal distribution with standard deviation
σ = 1. The voxel v0’s mean response to condition
A was µA = +2; and µB = −2 for condition B.
The responses of all other (non task-relevant) voxels
were drawn from normal distributions having σ = 1
where µA = µB = 0. To maximize the sensitivity
of the searchlight statistic and emulate the require-
ments of the SIN assumption, a total of 300 samples
were drawn for each condition. The spatial position
of voxel v0 is shown in blue in Figure 3(b). The voxel
(a)
(b)
Figure 3: Single-voxel response: (a) Scatter plot of simulated responses of voxel v0 to conditions A and B. A total of 300 samples
were drawn per condition (see text). (b) Spatial location of voxel v0 (indicated by blue square) in the simulated single-slice voxel
space.
was placed far from the boundaries of the slice to
avoid truncations of the searchlights and to emulate
a uniform voxel-space in the vicinity of v0.
With this setup,
the searchlight decomposition
and testing procedure was implemented using the
PyMVPA toolbox (Hanke et al., 2009). Each search-
light’s informativeness was determined by evaluat-
ing the decodability of its responses, i.e., testing for
the existence of a model that accurately classifies a
sample’s membership in each condition based on
the searchlight’s responses (Pereira and Botvinick,
2011; Pereira et al., 2009). Decodability was tested
using a linear Support Vector Machine (SVM) with
a soft-margin regularization parameter, C = 1. The
searchlight statistic was the mean classification accu-
racy obtained using a Leave-One-Out (LOO) cross-
validation procedure.
Figure 4(a) shows the information maps obtained
(thresholded at 60%). In the upper-panel, going from
the left to the right in order of increasing radius, we
see that there is a single high accuracy cluster (red-
colored voxels) centered at the signal-carrying voxel
v0, and this cluster grows in size with increasing
radius. The lower-panel shows an expanded view
of this high accuracy cluster, thresholded at 80%.
Consistent with the predictions described above, for
each radius, the size and shape of these clusters on
the information map correspond exactly to the size,
shape and location of the searchlight Sr(v0) centered
at voxel v0. Furthermore, consistent with Theorem
7, the number of informative searchlights identified
(Fr) increases in a monotonic manner with the ra-
dius of the searchlight, even though there is no dif-
ference in the actual information present or even any
multivoxel response patterns to speak of.
Figure 4(b) shows the values on the information
map from a single 1D segment running horizontally
through the voxel v0 through the diameter of the
searchlights centered at v0. The voxel v0 is assigned
a value 0. Consistent with the SIN assumption, the
accuracies on the information map do not exhibit
a smooth degradation as a function of the distance
from v0. Critically, this pedestal-like profile is unlike
the profile that would be expected if the searchlights
were the equivalent of a “spatial smoothing” kernel
on the information map.
What is the comparable effect on the information
map when the task-relevant signal is distributed
over multiple voxels? We next consider this sce-
nario.
4.2. The haystack-in-the-needle effect
Suppose there are two voxels, v1 and v2 in V, that
conjointly exhibit a response difference to the ex-
perimental conditions. However, neither voxel by
itself shows a task-relevant difference.
That is,
I({v1, v2}) = 1 and I(v1) = I(v2) = 0. By the SIN
assumption, every searchlight that contains both v1
and v2 should be informative, but searchlights that
AB−5−4−3−2−1012345ConditionResponse of voxel v0!! (a)
(b)
Figure 4: Information maps for single-voxel response: (a) Upper panel shows the information-maps across the entire slice as a
function of increasing searchlight radius going from left to right (threshold = 60% accuracy). Lower panel shows the corresponding
expanded view of the high-accuracy clusters (threshold = 80% accuracy) for each searchlight radius value. Horizontal lines are
to provide a common reference to compare relative sizes of clusters. (b) Cross-sectional profile of the information map showing
classification accuracies for searchlights centered at voxels on a horizontal 1D slice through the voxel v0. The dotted horizontal line
indicates the thresholding value of 80%. Only the profiles for radii 4mm, 8mm and 12mm are shown.
contain either v1 or v2 alone would not necessarily be
informative. Recall that, according to Theorem 4, a
group of multiple voxels (i.e., having more than one
voxel) is contained in strictly less than Nr search-
lights. It then follows that the signal-carrying voxel
group {v1, v2} should produce a cluster having less
than Nr voxels on the information map, i.e., a mul-
tivoxel “haystack” should produce a “needle”-like
cluster, unlike the needle-in-the-haystack scenario in
Section 4.1 above.
To simulate this “haystack-in-the-needle” effect, the
task-relevant responses in the two voxels v1 and v2
took the form shown in Figure 5(a). The responses to
each condition were drawn randomly from a normal
distribution having standard deviation σ = 1. Each
voxel’s mean response to conditions A and B are
shown as dotted lines. The voxel v1 had an identical
mean response to both conditions A and B, specif-
ically, µA = µB = 0 (the horizontal dotted line);
while voxel v2’s mean response to condition A was
µA = +0.5 and to condition B was µB = −0.5 (in-
dicated by each of the dotted vertical lines). Impor-
tantly, the responses of voxel v1 and v2 to both con-
ditions were correlated negatively. The response of
voxel v1 on condition A, denoted as X1,A was equal
to −X2,A, the response of voxel v2 to condition A.
Similarly, for condition B, X1,B = −X2,B. The simu-
lated responses of all other voxels were drawn from
distributions having σ = 1 and µA = µB = 0, and
were uncorrelated with the responses in either voxel
v1 or v2. As with the previous simulation above, a
total of 300 samples were drawn for each condition.
With signals of this form, the conjoint responses of
voxels v1 and v2 to conditions A and B are linearly
separable (see Figure 5(a)). However, A and B can-
not be distinguished from the responses in v1, but
should be weakly discriminable from the responses
in v2.
The relative spatial positions of v1 and v2, indicated
as blue squares, are shown in Figure 5(b). We con-
sidered two cases, where v1 and v2 were separated
by 2 voxels in one case; and by 3 voxels in the other.
When v1 and v2 have a separation of 2 voxels, there
is no one searchlight of radius 4mm that can con-
tain both of these voxels. With a separation of 3
voxels, there are no searchlights of radius 4 mm, 6
mm, or 8 mm that can contain both v1 and v2. With
this setup, the searchlight decomposition and testing
procedure was simulated in the same manner as in
Section 4.1.
Figure 6 shows the portions of the information maps
INFORMATION MAPSr = 4mmr = 8mmr = 10mmr = 12mmr = 6mm5 voxels21 voxels37 voxels13 voxels49 voxelsAccuracy > 80%Accuracy > 60%60%100%−10−6 −2 2 6 10 405060708090100Voxel indexClassification accuracy (%) 4mm8mm12mm (a)
(b)
Figure 5: Multivoxel response: (a) Scatter plot of simulated responses of voxels v1 and v2 to conditions A (gray squares) and
B (black dots). Dotted lines indicate the mean response of each voxel alone to A and B. (b) Spatial location of voxels v1 and v2
(indicated by blue squares) in the simulated single-slice voxel space. The separation between v1 and v2 was either 2 or 3 voxels (see
text). The responses of voxel v2 show a weak response difference to the conditions, indicated by the dotted circle.
in the vicinity of voxels v1 and v2 (thresholded
at 60%).
In all the information maps, the above-
threshold cluster takes the size and shape of the cor-
responding searchlight and is centered at voxel v2,
namely, the voxel exhibiting a weak response dif-
ference to conditions A and B. This “needle-in-the-
haystack” organization is consistent with the simu-
lations in Section 4.1, and is invariant to the number
of voxels separating v1 and v2.
Now, observe that the clusters in several, but not all,
of the information maps contain sub-clusters con-
sisting of voxels having high classification accuracies
(indicated in red). These voxels on the information
map correspond to the centers of searchlights that
contain both v1 and v2. As required by Theorem 4,
for each radius, the number of high-accuracy vox-
els in the cluster are less than Nr. Due to the geo-
metric constraint defined by the separation between
v1 and v2, the presence of any high-accuracy voxels
at all in an information map depends on the radius
of the searchlights used. For example, information
maps obtained with searchlights of radius 4 mm do
not contain any high-accuracy voxels for both sep-
arations (top row), while the information maps for
searchlights of radius 8mm contain high-accuracy
voxels for the 2 voxel separation but not for the 3
voxel separation.
Figures 6(a) and (b) show the 1D cross-section of
the information map through the horizontal diame-
ter of the clusters for the 2 voxel and 3 voxel separa-
tions respectively. As evident, there is a “smearing”,
rather than smoothing, of the accuracies with grow-
ing radius values, as in Figure 4(b). Furthermore,
when a searchlight is large enough to include both
v1 and v2, there is a large increase in the classifica-
tion accuracy.
The above simulations confirm the basic statisti-
cal premise motivating the searchlight-procedure,
namely, the ability of a multivoxel pattern anal-
ysis method to detect distributed response pat-
terns. However, for any radius, the size of the
clusters produced by multivoxel response patterns
are smaller than those produced by single voxel
response-differences. Consistent with Theorem 7,
the number of informative searchlights identified in-
creases in a monotonic manner with the radius of the
searchlight.
4.3. Whole-brain inflation maps
The previous two simulations demonstrated signal-
dependent effects caused by the sampling bias inher-
ent in the searchlight decomposition. However, ac-
cording to Theorem 7, there should be a monotonic
increase in the number of informative searchlights as
a function of radius, irrespective of the actual distri-
bution of task-relevant voxels/voxel-groups across
the brain. This monotonic scaling of the size of the
“blobs” on the information map makes plausible a
rather unusual scenario – an information map where
−3−2−10123−3−2−10123Response of voxel v2Response of voxel v1 Condition ACondition B!!!"!"#$%$&’() (a)
(b)
(c)
Figure 6: Information maps for multivoxel response: (a) Left and right panels show portions of information map (threshold =
60% accuracy) centered around voxels v1 and v2 as a function of increasing searchlight radius (top to bottom); and the separation
between v1 and v2 – 2 voxel separation (left panel), and 3 voxel separation (right panel). Plots (b) and (c): Cross-sectional profile of
the information map showing classification accuracies for searchlights centered at voxels on a horizontal 1D slice through the voxels
v1 and v2 for separation of 2 voxels (b), and 3 voxels (c). The lower dotted horizontal line indicates the thresholding value of 60%,
and the upper dotted line indicates the regime of the high-accuracy searchlights (80%).
!"#$""#!!"!#$%%!!"!#&%%!!"!’%%!!"!(%%!!"!)%%&!*+,-./%&’()(*+,-0!*+,-./%&’()(*+,-12345678142!679:−8 −4 0 4 8 12 405060708090100Voxel indexClassification accuracy (%) 4mm8mm12mm−8 −4 0 4 8 12 405060708090100Voxel indexClassification accuracy (%) 4mm8mm12mm every searchlight in the brain is deemed to be infor-
mative.
This scenario was motivated by results recently re-
ported by Poldrack et al. (2009).
In that study, in-
formation maps were generated using searchlights
of radius 4 mm and 8 mm. Rather remarkably, with
a radius of 8 mm, only one region in the informa-
tion map (the bilateral dorsolateral prefrontal cor-
tex) was found to be uninformative while every other
searchlight was informative. This whole-brain cov-
erage was, however, not the case with the 4 mm
searchlights. Given the inflationary relationship be-
tween Fr (the number of informative searchlights)
and searchlight radius r that established in the pre-
vious sections, curiosity asked: could an informative
whole-brain arise (i.e., Fr = V) by random chance
with a suitably chosen searchlight radius?
This question can be formulated as a covering prob-
lem. Consider a finite 3D voxel space correspond-
ing to one containing the brain, approximated as a
cubic volume of size NX × NY × NZ, where Ni is
the number of voxels along the principal direction i.
Suppose there is a minimum covering set of search-
lights Cr ⊂ Sr(V) such that every voxel in V is con-
tained in some searchlight in Cr. Recall that a single-
voxel signal can produce a cluster having Nr voxels
on the information map, due to Theorem 3. If the
central voxel of each of the searchlights in Cr was in-
formative, it would follow that searchlights centered
at every voxel in every one of the searchlights in Cr
would also be informative. Since every voxel in V
is present in some searchlight in Cr, it implies that
a rather sparse distribution of informative single-
voxels specified by Cr could produce an information
map where every searchlight in Sr(V) would be in-
formative (with the proviso that the SIN assumption
holds true.)
The sparsity of these informative single-voxels can
be readily approximated if we use cubical volumes
as a proxy for the spherical shape of the searchlight
volumes. A cube of side w voxels would fully con-
√
tain a sphere having radius r = w/2, and would be
fully contained in a sphere of radius w
3/2. With
this simplification, the minimal number of search-
light cubes required to cover the voxel space V is
readily approximated as the volume of the voxel
space divided by the volume of each searchlight
cube, that is, ≈ (cid:100)(NX NY NZ)/(w3)(cid:101).
Figure 7: Number of non-intersecting cubes required to fill a
volume of dimensions Nx = 52, Ny = 63, Nz = 45 (Ni = num-
ber of voxels in principal direction i) as a function of cube-size.
Values next to each data-point indicate the actual number of fill-
ing cubes for each cube-size.
For voxels of size 3 × 3 × 3 mm3, we approximate
the size of the voxel space with the following val-
ues NX = 52 voxels, NY = 63 voxels and NZ = 45
voxels. Figure 7 shows the minimum number of cu-
bical volumes required to cover V as a function of w,
where w took values 1, 3, 5, 7, 9, 11.
A searchlight cube of side w = 1 is equivalent to a
single voxel so the size of the covering set Cr is equal
to the total number of voxels in V, namely, 147420.
However, increasing values of w produce a rapid
decrease in the size of the covering set. For w = 3
voxels, a cubical volume that would fully contain a
spherical searchlight of radius 4 mm, a total of 5460
equally spaced signal-carrying voxels can produce
an information map where every searchlight is in-
formative. However, for a cubical volume with side
w = 7 voxels, corresponding to spherical volumes
of radius 8mm, a mere 430 voxels are required for
such a fully informative map. Stated differently, an
information map with a single task-relevant cluster
made up of every voxel in V can be generated from a
mere 430 regularly spaced voxels of the 147, 420 vox-
els in V, that is, 430 voxels that enable the conditions
to be distinguished whether due to the presence of
true signal or by random chance. This potential
for a small number of single voxels (i.e., ≈ 0.003%
1357911102103104105Cube width w (voxels)Number of cubes covering V (log10)147,4204301112031,1805,460 of V) to drive the structure of the entire informa-
tion map simply by the choice of the searchlight ra-
dius presents an important consideration for draw-
ing neurobiological interpretation.
it is maximal when explicitly assuming a highly sen-
sitive and robust MVPA technique, namely one sat-
isfying the superset informativeness (SIN) assump-
tion.
5. Discussion
6. Acknowledgments
Knowledge of the actual information-carrying vox-
els in each informative searchlight would make the
information map irrelevant. These actual informa-
tive voxels could be directly reported, hence resolv-
ing the over-counting that arises from their inclusion
in multiple searchlights. One possible implementa-
tion would be to identify task-relevant voxels in each
searchlight, and then combine these identified vox-
els across searchlights. However, requiring the iden-
tification of the actual informative voxels in each
searchlight could reduce the generality of the search-
light method. When pattern classifiers are used to
compute the searchlight statistic, each voxel (or fea-
ture) in a searchlight is typically assigned a weight,
and the weighted combination of the multivoxel re-
sponses is used to make a classification decision.
However, the specific basis for assigning weights to
individual features is highly dependent on the spe-
cific machine learning algorithm and its inductive
assumptions (Mitchell, 1980; Wolpert, 1996; Guyon
et al., 2002; Pereira et al., 2009). Consequently, ap-
propriate techniques would be required to allow re-
sults to be compared across studies that use different
MVPA-techniques.
Until such advances are made, the analytical frame-
work described above provides several constraints
on alternate interpretations of the information map.
Our results present a strong argument against mea-
suring the sensitivity of information mapping by a
count of the number of informative searchlights. The
seemingly high sensitivity of the searchlight method
as judged by such a performance measure in part has
a rather trivial explanation. Specifically, an expla-
nation in the obligatory geometric properties of the
searchlight-method as discussed above rather than
an explanation related to underlying neural orga-
nization, or the sophisticated machine-learning al-
gorithms used to analyze multivoxel response pat-
terns, or the widely discussed merits of multivariate
statistical evaluations. Indeed, the upshot of the op-
timistic scaling of this performance measure is that
This work was supported by the U.S. Army Re-
search Office through the Institute for Collaborative
Biotechnologies under Contract No. W911NF-09-D-
0001.
7. References
References
Alink, A., Euler, F., Kriegeskorte, N., Singer, W., Kohler, A., 2011.
Auditory motion direction encoding in auditory cortex and
high-level visual cortex. Human Brain Mapping.
Chadwick, M. J., Hassabis, D., Weiskopf, N., Maguire, E. A.,
2010. Decoding individual episodic memory traces in the hu-
man hippocampus. Current Biology 20 (6), 544–7.
Chen, Y., Namburi, P., Elliott, L. T., Heinzle, J., Soon, C. S., Chee,
M. W. L., Haynes, J.-D., 2011. Cortical surface-based search-
light decoding. Neuroimage 56 (2), 582–92.
J. S., Gors,
J., Hanke, M.,
Halchenko, Y. O., Wu, Y.-C., Abdi, H., Haxby, J. V., 2012. The
representation of biological classes in the human brain. Jour-
nal of Neuroscience 32 (8), 2608–18.
Connolly, A. C., Guntupalli,
Cox, D. D., Savoy, R. L., Jun 2003. Functional magnetic reso-
nance imaging (fmri) ”brain reading”: detecting and classi-
fying distributed patterns of fmri activity in human visual
cortex. Neuroimage 19 (2 Pt 1), 261–70.
Formisano, E., Kriegeskorte, N., 2012. Seeing patterns through
the hemodynamic veil - the future of pattern-information
fMRI. Neuroimage 62 (2), 1249–56.
Golomb, J. D., Kanwisher, N., 2011. Higher level visual cortex
represents retinotopic, not spatiotopic, object location. Cere-
bral Cortex Epub.
Guyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene selec-
tion for cancer classification using support vector machines.
Machine learning 46 (1), 389–422.
Hanke, M., Halchenko, Y. O., Sederberg, P. B., Hanson, S. J.,
Haxby, J. V., Pollmann, S., 2009. PyMVPA: A python toolbox
for multivariate pattern analysis of fMRI data. Neuroinfor-
matics 7 (1), 37–53.
Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten,
J. L., Pietrini, P., 2001. Distributed and overlapping represen-
tations of faces and objects in ventral temporal cortex. Science
293 (5539), 2425–30.
Haynes, J., 2006. Decoding mental states from brain activity in
humans. Nature Reviews Neuroscience 7 (7), 523–534.
Haynes, J.-D., Sakai, K., Rees, G., Gilbert, S., Frith, C., Passing-
ham, R. E., Feb. 2007. Reading Hidden Intentions in the Hu-
man Brain. Current Biology 17 (4), 323–328.
tions in human visual cortex. Neuroimage 56 (3), 1540–5.
Tong, F., Dec. 2010. Pattern Classification Analysis. Annual Re-
view of Psychology 63 (1), 110301102248092.
Wagner, A. D., Rissman, J., Dec. 2010. Distributed representa-
tions in memory: insights from functional brain imaging. An-
nual Review of Psychology 63 (1), 110301102248092.
Wolpert, D., 1996. The lack of a priori distinctions between
learning algorithms. Neural computation 8 (7), 1341–1390.
Woolgar, A., Thompson, R., Bor, D., Duncan, J., May 2011. Multi-
voxel coding of stimuli, rules, and responses in human fron-
toparietal cortex. Neuroimage 56 (2), 744–52.
Jimura, K., Poldrack, R. A., 2012. Analyses of regional-average
activation and multivoxel pattern information tell comple-
mentary stories. Neuropsychologia 50 (4), 544–52.
Johnson, J. D., McDuff, S. G. R., Rugg, M. D., Norman, K. A.,
2009. Recollection, familiarity, and cortical reinstatement: a
multivoxel pattern analysis. Neuron 63 (5), 697–708.
Kaplan, J. T., Meyer, K., 2012. Multivariate pattern analysis
reveals common neural patterns across individuals during
touch observation. Neuroimage 60 (1), 204–12.
Kriegeskorte, N., Goebel, R., Bandettini, P., 2006. Information-
based functional brain mapping. Proceedings of the National
Academy of Sciences 103 (10), 3863–8.
Mitchell, T., 1980. The need for biases in learning generaliza-
tions. Tech. Rep. CBM-TR-5-110, Department of Computer
Science, Rutgers University.
Morgan, L. K., Macevoy, S. P., Aguirre, G. K., Epstein, R. A.,
2011. Distances between real-world locations are represented
in the human hippocampus. Journal of Neuroscience 31 (4),
1238–45.
Mur, M., Bandettini, P. A., Kriegeskorte, N., 2008. Revealing
representational content with pattern-information fMRI–an
introductory guide. Social Cognitive and Affective Neuro-
science 4 (1), 101–109.
Nestor, A., Plaut, D. C., Behrmann, M., 2011. Unraveling the
distributed neural code of facial identity through spatiotem-
poral pattern analysis. Proceedings of the National Academy
of Sciences 108 (24), 9998–10003.
Norman, K. A., Polyn, S. M., Detre, G. J., Haxby, J. V., Sep. 2006.
Beyond mind-reading: multi-voxel pattern analysis of fMRI
data. Trends in Cognitive Sciences 10 (9), 424–430.
Oosterhof, N. N., Tipper, S. P., Downing, P. E., Jan 2012. View-
point (in)dependence of action representations: An MVPA
study. J Cogn Neurosci.
Oosterhof, N. N., Wiestler, T., Downing, P. E., Diedrichsen, J.,
May 2011. A comparison of volume-based and surface-based
multi-voxel pattern analysis. Neuroimage 56 (2), 593–600.
Oosterhof, N. N., Wiggett, A. J., Diedrichsen, J., Tipper, S. P.,
Downing, P. E., Aug 2010. Surface-based information map-
ping reveals crossmodal vision-action representations in hu-
man parietal and occipitotemporal cortex. Journal of Neuro-
physiology 104 (2), 1077–89.
Peelen, M. V., Kastner, S., Jul 2011. A neural basis for real-world
visual search in human occipitotemporal cortex. Proceedings
of the National Academy of Sciences 108 (29), 12125–30.
Pereira, F., Botvinick, M., May 2011. Information mapping with
pattern classifiers: A comparative study. NeuroImage 56 (2),
476–496.
Pereira, F., Mitchell, T., Botvinick, M., Mar. 2009. Machine learn-
ing classifiers and fMRI: A tutorial overview. NeuroImage
45 (1), S199–S209.
Poldrack, R. A., Halchenko, Y. O., Hanson, S. J., Nov. 2009. De-
coding the Large-Scale Structure of Brain Function by Clas-
sifying Mental States Across Individuals. Psychological Sci-
ence 20 (11), 1364–1372.
Serences, J. T., Saproo, S., Mar 2012. Computational advances
towards linking bold and behavior. Neuropsychologia 50 (4),
435–46.
Soon, C. S., Brass, M., Heinze, H.-J., Haynes, J.-D., Apr. 2008.
Unconscious determinants of free decisions in the human
brain. Nature Neuroscience 11 (5), 543–545.
Stokes, M., Saraiva, A., Rohenkohl, G., Nobre, A. C., Jun 2011.
Imagery for shapes activates position-invariant representa-
|
1512.08339 | 1 | 1512 | 2015-12-28T08:00:27 | Detecting local processing unit in drosophila brain by using network theory | [
"q-bio.NC",
"physics.bio-ph"
] | Community detection method in network theory was applied to the neuron network constructed from the image overlapping between neuron pairs to detect the Local Processing Unit (LPU) automatically in Drosophila brain. 26 communities consistent with the known LPUs, and 13 subdivisions were found. Besides, 45 tracts were detected and could be discriminated from the LPUs by analyzing the distribution of participation coefficient P. Furthermore, layer structures in fan-shaped body (FB) were observed which coincided with the images shot by the optical devices, and a total of 13 communities were proven closely related to FB. The method proposed in this work was proven effective to identify the LPU structure in Drosophila brain irrespectively of any subjective aspect, and could be applied to the relevant areas extensively. | q-bio.NC | q-bio | Detecting local processing unit in drosophila brain by using network theory
Dongmei Shi1,2, Chitin Shih3,4, Chungchuan Lo5 , Yenjen Lin1, and Annshyn Chiang1,6,7
Brain Research Center, National Tsing Hua University, Hsinchun 30013 Taiwan,
Department of Physics, Bohai University, Jinzhou Lianning 121000, P.R.C,
Department of Physics, Tunghai University, Taichung 40704 Taiwan,
Physics Division, National Center for Theoretical Sciences, Hsinchu 30043, Taiwan,
Institute of Systems Neuroscience, National Tsing Hua University, Hsinchu 30013, Taiwan,
Institute of Biotechnology, National Tsing Hua University, Hsinchu 30013, Taiwan
1.
2.
3.
4.
5.
6.
7. Genomics Research Center, Academia Sinica, Taipei 11529, Taiwan
Community detection method in network theory was applied to the neuron network constructed
from the image overlapping between neuron pairs to detect the Local Processing Unit (LPU)
automatically in Drosophila brain. 26 communities consistent with the known LPUs, and 13
subdivisions were found. Besides, 45 tracts were detected and could be discriminated from the
LPUs by analyzing the distribution of participation coefficient P. Furthermore, layer structures in
fan-shaped body (FB) were observed which coincided with the images shot by the optical devices,
and a total of 13 communities were proven closely related to FB. The method proposed in this
work was proven effective to identify the LPU structure in Drosophila brain irrespectively of any
subjective aspect, and could be applied to the relevant areas extensively.
Keywords: Drosophila Local processing unit Complex networks Modularity
Instruction
Some neuron circuits in Drosophila are believed worth studying in understanding the
fundamental features of neurons by most neuroscientists [1-4]. Genetic screens in Drosophila
have identified many genes involved in neural development and function [5-9]. One reason is the
power of the Drosophila genetic toolbox [10-12], another basic reason is that fruit flies have
about 1000-fold fewer neurons, while exhibit extensive cognitive abilities.
Local processing unit (LPU) [13-14] was proposed in constructing a mesoscopic map of the brain
circuit in Drosophila. A LPU is defined as a brain region consisting of its own local interneurons
(LN) whose nerve fibers are completely restricted to this region. Further, each LPU is contacted by
at least one nerve tract. LPU is a functional unit dealing with the signals, which plays a critical
part in information transmission in the brain. Connectomics-Based analysis of information flow in
the drosophila brain based on the LPUs has been studied recently, and it was shown that the
network showed hierarchical structure, small world, and rich-club organization [14]. Normally,
LPU and its boundary are determined by observation based on the optical devices, which exists
some subjective factors.
In this work, complex network theory [15-18] was introduced to detect the LPU structure in
Drosophila brain based on LPU's definition. It was proven that LNs and neuron tracts could be
both identified successfully, and so that LPU would be determined by LNs and the corresponding
neuron tracts, irrespectively of any subjective aspect.
Methods
Community structure (Figure 1) [19]: A network is realized to have the community structure if the
nodes of network can be easily grouped into sets of nodes such that each set of nodes is densely
connected internally.
Modularity Q: Modularity is one measure of the structure of networks or graphs. It was designed
to measure the strength of division of a network into modules (also called groups, or
communities).
where
, and
is the module which the node i belongs to.
is the strength of node
i, and
is the element in adjacency matrix.
is determined by maximizing the modularity Q,
and larger Q indicates more distinct community structure in the networks.
Participation coefficient P:
where
is the strength of node i connected to the nodes in module
, and
is the total
strength of node i in the network.
is close to 0 if most of the edges node i linked to are in its
own module
. On the other hand,
is close to 1 if the links are uniformly distributed among
all the modules of the network.
Simulations
Neuron network construction: Neuron network was based on the 23380 female fruit fly data. The
nodes in the network indicated the neurons of brain, and the links represented the interactions
among the neurons. The corresponding adjacency matrix was constructed as follows: the matrix
dimension was the total number of neurons, and the matrix element
was the overlapping
strength in space of neuron i and neuron j. The neuron network was an undirected-weighted
complex network.
Relations between LPU in the brain neuron system and community structure in a network (Figure
2): Neurons in the fruit fly brain could be categorized into two functional distinct populations:
local
interneurons (LNs) whose processes were restricted within a single brain region, and
projection neurons (PNs) whose dendrites and axons connected two or more brain regions. Local
Processing Unit (LPU) was defined as a brain region consisting of its own LNs population whose
nerve fibers were completely restricted to the region, and meanwhile, each LPU was connected
by at least one neural tract. So LPU boundary could be determined by LNs and the neuron tract
fibers restrained to this region. On other hand, according to the definition of community
structure in a network, two types of nodes could be classified: nodes whose links were all in a
single community (P=0), and nodes whose links were distributed into several communities (P>0).
Accordingly, a correspondence could be built up between LPU in the brain and community
structure in a network: the LNs corresponded to the nodes with P=0, and PNs corresponded to
the nodes with P>0. Figure 2 showed the representation of this relation.
Community division: Based on the analysis above, detecting LPU structure was tantamount to
find or form the communities in which participation coefficient P for all the nodes met P=0. Firstly,
modularity calculation was used to divide the neuron networks into communities. In order to
detect the basic LPU structure in which no smaller LPUs existed, modularity calculation was
performed three times sequentially based on the original neuron network. Precisely (Fig. 3), the
original network was divided into 8 communities by calculating Q (Cal. Q), and these 8
communities were called level 1 communities (L1-Comm). Then, L1-Comm would be further
divided into smaller communities which were called level 2 communities (L2-Comm), and further,
L2-Comm would be done by the same analogy into level 3 communities (L3-Comm). Taking the
5th community on level 1 for example, it was seen in Fig. 3 that the 5th community on L1-Comm
was divided into 3 small communities on level 2, 5-1, 5-2, 5-3. And the second community 5-2 on
level 2 could be further divided into 3 smaller communities on level 3, 5-2-1, 5-2-2, 5-2-3. And so
on, for each of the 8 communities on level 1.
Community detection method: Representation of community detection method was shown in Fig.
4, one can see that three steps were needed to detect the LPU structure. On the first step, the
original brain network was divided into L3-Comm communities by Cal. Q, among which LPU
communities were included. On step 2, neuron tract communities would be discriminated. It had
been proven according to the statistical results in this study that a community was the tract
community if
, or if there existed a distinct segmentation point
or P>0.5
(f(P)~0) in the distribution of P, then the neurons corresponding to the last segmentation
belonged to a neuron tract.
is the minimum of participation coefficient in this community,
and f(P) is the distribution function of P. See an example of tract distribution in Fig. 5.
LPU communities were left by manipulations in step 2, but there still existed PNs in these
communities. In order to extract LNs from LPU communities, LPU community would be optimized
in step 3. Based on correspondence between LPU and community structure (Fig. 2), our solution
was to remove the nodes with large P piece by piece from the LPU community by modularity
calculating successively, until
of this community. The community could be realized
isolated when
in our study.
Before our further study, there were two points needed to be discussed. According to the
theory in community detection method,
the nodes with participation coefficient P=0
corresponded to the LNs whose fibers were all restricted to one region. However, considering the
existence of warping area differences, when the neurons were put into the standard brain [20-21],
the critical value of P could not reach 0 accurately. Actually P should be a function of warping
area, and was different in different brain regions.
On other hand, considering that some LPUs might be destroyed by the successively divisions in
step 1, operations on step 3 would be applied directly to the communities in level 1 to
compensate the destroyed LPUs.
Results
Community detection method was applied to the Drosophila brain neuron system by means of
the above manipulations. Figure 6 showed the LNs of different brain regions in the standard brain
from forward, middle and backward levels of the brain, and it was proven that these LNs
belonged to 26 known LPUs: DLP (dlp), MB (mb), AL (al), VLP (vlp), MED (med), LOB (lob), nod, FB,
EB, LH (lh), LOP (lop), PB, CMP (cmp). Two subdivisions in VLP (vlp), three subdivisions in FB, and
three subdivisions in MB (mb) were also illustrated. Generally, 49 LPUs and hubs had been found,
and the lost LPUs in this work were mainly due to the fewer LNs supplied in these LPUs.
45 neuron tracts were discriminated by the community detection method, and were shown in Fig.
7. 9 types of neuron tracts were proven connected AL, which were sighed 1-9 in Fig. 8.
Subdivisions in FB and MB were shown in Fig. 8 (1) and Fig. 8 (2). Three subdivisions in FB were
detected which had been not observed in experiment, but were proven to be consistent with the
results only based on FB data by applying the community detection method. Three subdivisions
(
,
,
)
in MB (mb) shown in Fig. 8 (2) coincided with the observations in experiment by
optical devices.
Four layers in FB were detected, and illustrated in Fig. 9. Only neurons in Fig. 9 (3) were the LNs,
and neurons in Fig. 9 (1), (2), (4) were all PNs according to the community detection method.
Figure 10 shows the communities related to FB, including three LPU communities and other 10
tract communities. Apparently FB played an important part in information transmission in the
brain.
Summary
We applied the community detection method to detect the LPUs in Drosophila brain system, and
had proven that 26 known LPUs were detected, and subdivisions were also found. Precisely, three
subdivisions in MB (mb), and two subdivisions in VLP (vlp) were consistent with the observations
in experiments, and unknown three subdivisions in FB would give us important information to
understand FB's internal structure. It was further proven that the same subdivisions in FB were
also found if the community method was only applied to FB neuron system. Besides, layer
structures in FB were detected, and one layer was a subdivision in FB, and other three layers
were tract communities. A total of 13 communities were closely related to FB,
including
subdivisions of FB, neuron tract communities inside FB, and neuron tract communities on FB's
boundary. Apparently, FB played an important part in information transmissions in Drosophila
brain. Moreover, it had been proven that neuron tracts could be discriminated by analyzing the
distributions of participation coefficient P.
Community detection method was proven effective to detect LPU in Drosophila brain system in
this study, and its boundary was determined by LNs and the neuron tracts totally, irrespectively
of any subjective factor. This method would be more perfect if the participation coefficient P as a
function of warping area was known, and would be applied to the relative areas extensively.
Appendix [Detecting subdivisions in FB and subdivisions in neuron tract communities]
Manipulations in step 3 of community detection method (Fig. 4) were directly applied to FB
neuron system in which 544 neurons in FB had been obtained by experiment. Precisely, nodes
with large P would be removed piece by piece from the FB neuron network successively until
of this community. It had been proven that three subdivisions were found in FB
coinciding with the results in the above simulations. Moreover, this method was also applied to
neuron tract communities discriminated in step 2 in Fig. 4 in order to find the basic neuron tracts
in these communities.
Acknowledgments
This work was supported by the MOE 5-Year-50-Billion Project and NSC Nano National Project in
Taiwan, Specialized Foundation for Theoretical Physics of China (Grant No. 11247239), National
Natural Science Foundation of China (Grants No. 11305017).
[1]Shawn R. Olsen and Rachel L. Wilson, Cracking neural circuits in a tiny brain: new approaches
for understanding the neural circuitry of Drosophila, Cell: Trends in Neurosciences, 31 (2008).
[2]T.C. Holmes, etal, Circuit-breaking and behavioral analysis by molecular genetic manipulation
of neural activity in Drosophila. In Invertebrate Neurobiology (North, G. and Greenspan, R.j., eds)
Cold Spring Harbor Laboratory, (2007), pp. 19--52.
[3]L. Luo etal, Genetic dissection of neural circuits, Neuron, 57 (2008), pp. 634--660.
[4]S. Tang, etal, Visual pattern recognition in Drosophila is invariant for retinal position, Science,
305 (2004), pp. 1020--1022.
[5]G. Liu, etal, Distinct memory traces for two visual features in the Drosophila brain, Nature, 439
(2006), pp. 551--556.
[6]K. Neuser, etal, Analysis of a spatial orientation memory in Drosophila, Nature, 453 (2008), pp.
1244--1247.
[7]C. Faucher, etal, Behavioral responses of Drosophila to biogenic levels of carbon dioxide
depend on lif-stage, sex and olfactory context, J. Exp. Biol. 209 (2006), pp. 2739-2748.
[8]A. Yurkovic, etal, Learning and memory associated with aggression in Drosophila melanogaster,
Proc. Natl. Acad. Sci. U. S. A., 103 (2006), pp. 17519--17524.
[9]J.W. Bohland, C. Wu, H. Barbas, H. Bokil, M. Bota, H.C. Breiter, H.T. Cline, J.C. Doyle, P.J. Freed,
R.J. Greenspan, etal, A proposal for a coordinated effort for the determination of brain-wide
neuroanatomical connectivity in model organisms at a mesoscopic scale, PLOS Comput. Biol. 5
(2009), e1000334.
[10]S.R. Olsen, and R.I. Wilson, Cracking neural circuits in a tiny brain: New approaches for
understanding the neural circuitry of Drosophila, Trends Neurosci. 31 (2008), pp. 512--520.
[11] A. Claridge-Chang, R.D. Roorda, E. Vrontou, L. Sjulson, H. Li, J. Hirsh, and G. Miesenböck,
Writing memories with light-addressable reinforcement circuitry. Cell 139 (2009), pp. 405--415.
[12]M.J. Krashes, S. DasGupta, A. Vreede, B. White, J.D. Armstrong, and S. Waddell, A neural
circuit mechanism integrating motivational state with memory expression in Drosophila. Cell 139
(2009), pp. 416--427.
[13] A.S. Chiang, C.Y. Lin, C.C. Chuang, H.M. Chang, C.H. Hsieh, C.W. Yeh, C.T. Shih, J.J. Wu, etal,
Three-Dimensional Reconstruction of Brain-wide Wiring Networks in Drosophila at Single-Cell
Resolution}, Current Biology 21 (2011), pp. 1--11.
[14]C.T. Shih, O. Sporns, S.L. Yuan, T.S. Su, Y.J. Lin, C.C. Chuang, T.Y. Wang, C.C. Lo, R.J. Greenspan,
A.S.Chiang, Connectomics-Based Analysis of Information Flow in the Drosophila Brain, Current
Biology 25 (2015), pp. 1249--1258.
[15] N. Ganguly, A. Deutsch and A. Mukherjee,
Applications
Science,
978-0-8176-4750-6.
[16] A. Barrat, M. Barthelemy and A. Vespignani, Dynamical processes on complex networks,
Cambridge University Press, (2008), ISBN 978-0-521-87950-7.
[17]G. Caldarelli, Scale-Free Networks, Oxford University Press, (2007), ISBN 978-0-19-921151-7.
[18]E. Estrada, The Structure of Complex Networks: Theory and Applications, Oxford University
Press, (2011), ISBN 978-0-199-59175-6.
[19] M.E.J. Newman, Modularity and community structure in networks, Proc. Natl. Acad. Sci. USA
103 (2006), pp. 8577--8582.
[20]A.W. Toga, Neuroimage databases: The good, the bad and the ugly, Nat. Rev. Neurosci. 3
(2002), pp. 302--309.
[21]K. Rein, M. Zöckler, M.T. Mader, C. Grübel, and M. Heisenberg, The Drosophila standard brain,
Curr. Biol. 12 (2002), pp. 227--231.
Dynamics On and Of Complex Networks,
ISBN
to Biology, Computer
and the
Social
Sciences,
(2009)
Figure 10. 13 communities closely related to FB.
Figure 9. Layer structures in FB.
Figure 8. (1) Subdivisions in FB; (2) Subdivisions in MB (mb).
Figure7. 45 neuron tracts in the standard brain, and neuron tracts signed 1-9 were the tracts
connecting the AL (al).
Figure6. LNs in different regions in the standard brain from forward, middle and backward level,
respectively.
Figure5. Tract community discrimination.
Figure 4. Graphic representation of community detection method. Step 1, community division;
step 2, neuron tract discrimination; step 3, LPU community optimization.
Figure3. Community division by calculating modularity ($Calc. Q$). The original brain neuron
system was divided into 8 communities called level 1 community ($L1-Comm$), and these 8
communities could be further divided into sub-communities called level 2 community
($L2-Comm$). Based on $L2-Comm$, they would be further divided to communities called level 3
community ($L3-Comm$). Taking the 5th community on level 1 for example, it was divided into 3
smaller communities 5-1,5-2,5-3 on level 2, and community 5-2 was further divided into 3
sub-communities on level 3.
Figure 2. Relations between LPU and community. (1) Black-solid curves represented the local
interneurons (LNs), and other color-solid curves indicated the projection neurons (PNs). Green,
blue, and red-solid curves represented neuron tracts; (2) Nodes in the same shadow formed a
community, and two types of nodes were included, type of node A whose links were all in a single
community ($P_{A}=0$), and type of B whose links were distributed into one more communities
($P_{B}>0$).
Figure 1. Representation of community structure in a graph. Nodes in the same shadow form a
community.
|
1309.1654 | 2 | 1309 | 2013-09-09T20:52:59 | A Generic Approach to Solving Jump Diffusion Equations with Applications to Neural Populations | [
"q-bio.NC"
] | Diffusion processes have been applied with great success to model the dynamics of large populations throughout science, in particular biology. One advantage is that they bridge two different scales: the microscopic and the macroscopic one. Diffusion is a mathematical idealisation, however: it assumes vanishingly small state changes at the microscopic level. In real biological systems this is often not the case. The differential Chapman-Kolmogorov equation is more appropriate to model population dynamics that is not well described by drift and diffusion alone.
Here, the method of characteristics is used to transform deterministic dynamics away and find a coordinate frame where this equation reduces to a Master equation. There is no longer a drift term, and solution methods there are insensitive to density gradients, making the method suitable for jump processes with arbitrary jump sizes. Moreover, its solution is universal: it no longer depends explicitly on the deterministic system.
We demonstrate the technique on simple models of neuronal populations. Surprisingly, it is suitable for fast neural dynamics, even though in the new coordinate frame state space may expand rapidly towards infinity. We demonstrate universality: the method is applicable to any one dimensional neural model and show this on populations of leaky- and quadratic-integrate-and-fire neurons. In the diffusive limit, excellent approximations of Fokker-Planck equations are achieved. Nothing in the approach is particular to neuroscience, and the neural models are simple enough to serve as an example of dynamical systems that demonstrates the method. | q-bio.NC | q-bio |
1
A Generic Approach to Solving Jump Diffusion Equations with
Applications to Neural Populations
Marc de Kamps1
1 School of Computing, University of Leeds, Leeds, West Yorkshire, United Kingdom
∗ E-mail: Corresponding [email protected]
Abstract
Diffusion processes have been applied with great success to model the dynamics of large populations
throughout science, in particular biology. One advantage is that they bridge two different scales: the
microscopic and the macroscopic one. Diffusion is a mathematical idealisation, however:
it assumes
vanishingly small state changes at the microscopic level. In real biological systems this is often not the
case. The differential Chapman-Kolmogorov equation is more appropriate to model population dynamics
that is not well described by drift and diffusion alone.
Here, the method of characteristics is used to transform deterministic dynamics away and find a
coordinate frame where this equation reduces to a Master equation. There is no longer a drift term,
and solution methods there are insensitive to density gradients, making the method suitable for jump
processes with arbitrary jump sizes. Moreover, its solution is universal: it no longer depends explicitly
on the deterministic system.
We demonstrate the technique on simple models of neuronal populations. Surprisingly, it is suitable
for fast neural dynamics, even though in the new coordinate frame state space may expand rapidly
towards infinity. We demonstrate universality: the method is applicable to any one dimensional neural
model and show this on populations of leaky- and quadratic-integrate-and-fire neurons. In the diffusive
limit, excellent approximations of Fokker-Planck equations are achieved. Nothing in the approach is
particular to neuroscience, and the neural models are simple enough to serve as an example of dynamical
systems that demonstrates the method.
Author Summary
Biological systems stand out by their complexity and multi-scale interactions. Modeling techniques that
can explain the behavior of systems - the macroscopic level - in terms of their individual components
- the microscopic level - are very valuable. Diffusion models do that, so, unsurprisingly, they pervade
biology. One key assumption underlying the validity of the diffusion approach is that state changes
of individuals are minuscule.
In real biological systems this can often not be justified. The method
presented here dispenses with the assumption and broadens the application range of population modelling
considerably. Moreover, the method reduces the mathematically difficult problem of solving a partial
differential equation to a simpler one: understanding the noise process that gave rise to it.
We demonstrate the approach on neuronal populations. We envisage that the study of networks of such
populations will elucidate brain function, as large-scale networks can be simulated with unprecedented
realistic neural dynamics. This may help explain imaging signals, which after all are the observable
collective signal of large groups of neurons, or it may help researchers to develop their ideas on neural
coding. We also expect that the technique will find broad application outside of neuroscience and biology,
in particular in finance.
Introduction
Diffusion models are of crucial importance to biological modeling; literally, text books have been filled
on the subject, e.g. [1, 2]. The application domain is enormous: fluid dynamics, chemotaxis [3], animal
2
population movement [4, 5], neural populations e.g. [6 -- 17] are just a few examples (see [18] for a recent
review and further references). Diffusion models allow an understanding of group behaviour in terms
of that of its constituents, i.e. they link macroscopic and microscopic behaviour. In general diffusion
models emerge as follows: microscopic deterministic laws describing individual behavior in the absence
of interactions are determined. Interactions among individuals and between individuals and the outside
world are described by a stochastic process ('noise'). Diffusion arises if one assumes that changes in
the microscopic state as a consequence of an interaction are vanishingly small. Sometimes, this assump-
tion is well justified:
for example, in Brownian motion these interactions are due to the collisions of
individual molecules, interactions that are indeed minuscule compared to the macroscopic state. Often,
this assumption is not appropriate, but made in order to use the familiar mathematical machinery of
diffusion equations. For example, in computational neuroscience the diffusion approximation hinges on
the assumption that synaptic efficacies - a measure for how strongly neurons influence each other - are
small. This is simply not the case in many brain areas, e.g. [19], and leads to sizable corrections to diffu-
sion results [20]. Nevertheless, the diffusion approximation predominates in computational neuroscience.
Given the widespread distribution of diffusion models in biology, it stands to reason that this situation
is common.
Experience from computational neuroscience [20] and finance - a field that has put considerable effort
into understanding jump diffusion equations e.g. [21 -- 23] - has shown that results valid beyond the diffusion
limit are very hard to obtain, limited to a specific study and often require a perturbation approach [20].
Due to their nature, jump diffusion processes may lead to very jagged, locally discontinuous probability
density profiles [24]. For this reason numerical solution schemes that are well established for diffusion
processes may be unsuitable for finite size jumps. Some solution schemes are applicable for finite jumps if
it can be assumed that the underlying density profile is smooth. Brennan and Schwartz [25] demonstrated
that an implicit scheme for solving the Black-Scholes equation is equivalent to a generalized jump process.
The assumption must be justified, however. In contrast, the method presented here makes no assumptions
about the structure of the underlying structure of the density profile at all. Below, we will present an
example that can be easily modeled by the method presented here, but where we are not aware of more
conventional discretization schemes that would apply.
Standard text books such as [24, 26] demonstrate how deterministic laws of motion for individuals
and a Master equation describing the noise process under consideration can be combined. This results
in the so-called differential Chapman-Kolmogorov equation, a partial integro-differential equation in
probability density (it is derived in the context of neuronal populations in "Methods: Derivation of
the population density equation"). When the assumption is made that the transitions due to the
stochastic processes induce very small state transitions, this equation reduces to a partial differential
equation of the Fokker-Planck type [24, 26]. Diffusion models ultimately derive from this equation.
Clearly, a solution method for the Chapman-Kolmogorov equation is highly valuable: it would still allow
the study of diffusion processes, but also processes with large jumps. Below, we will give an example of
the gradual breakdown of diffusion. This opens up an important application area: it will be possible to
re-examine results already obtained in the diffusion limit and to investigate the consequences of finite
jump sizes.
Many expositions immediately proceed to the diffusion regime. Here we follow a radically different
approach: we use the method of characteristics to define a coordinate frame that co-moves with the flow
of the deterministic system. In this frame the differential Chapman-Kolmogorov equation reduces to the
Master equation of the stochastic process under consideration. This approach neatly sidesteps the need
for solving a partial (integro-) differential equation as the Master equation is a set of ordinary differential
equations, and can be solved by relatively simple numerical methods.
This simplicity comes at a price, however: the probability density is represented in an interval that is
moving with respect to the new coordinate frame. How this interval behaves becomes dependent on the
topology of the flow of the deterministic system. We will illustrate the problem on two neuron models
3
that represent two ends of the dynamic spectrum. Leaky-integrate-and-fire neurons have slow underlying
deterministic dynamics. Quadratic-integrate-and-fire neurons model the rapid onset of a spike. The
dynamics is so fast that the original interval representing the probability density profile moves to infinity
in finite time in the new coordinate frame.
We find that these neural models are a demonstration of the broad applicability of the technique. Not
only are they simple examples of dynamical systems that can easily be understood without a background
in computational neuroscience. They also provide boundary conditions that any good solution method
must be able to handle: absorbing boundaries. In general neurons spike and then return to pre-spike
conditions. In one dimensional neural models this is simulated by resetting the neuronal state as soon as
an absorbing boundary (the 'threshold') is reached. The reintroduction of a neuron once it has hit the
threshold (a 'reset') is particular to neural systems. It is a marked difference from financial derivatives -
as options become worthless after their expiration date - and one reason why results from finance do not
carry over to neuroscience despite similarities.
The method is manifestly stable, and the jump size is immaterial. The method can be considered a
generalization of the diffusion approach: for a stochastic process characterized by small jumps diffusion
results are recovered, as one would expect. However, it is not always practical to study diffusion by
a small finite jump process as the convergence to diffusion results is not always uniform and a direct
application of the Fokker-Planck equation may then give better results. We believe the main application
of the method is for processes with truly finite jumps, or to study deviations from the diffusion limit.
Results
We obtained the following results.
1. The method of characteristics was employed to define a new coordinate frame that moves along
the flow of deterministic dynamical system. In the new coordinate frame the differential Chapman-
Kolmogorov equations transforms to the time dependent Master equation of the noise process. This
result was used earlier in the limited setting of leaky-integrate-and-fire neurons [27,28]. It is restated
here for convenience.
2. The characteristics themselves were used to define a representation of the probability density. This
is a key new insight as it allows an accurate probability representation even when the deterministic
process is fast. This is explained in some detail below.
3. The density representation is dependent on the topology of the deterministic flow. Although the
number of topologies is typically limited, potentially this could require a novel implementation of
the algorithm. Somewhat surprisingly, we found that the simplest implementation of the density is
for a deterministic process that corresponds to periodically spiking neurons. Importantly, we found
that under broad conditions it is possible to 'switch topology' by a process we dubbed 'current
compensation'.
It is possible to modify the topology of the deterministic neural dynamics by
adding a DC current to the neurons. We compensate by renormalizing the noise input such that its
mean is lower by an amount equal to the DC current. Therefore, often a single topology suffices.
4. The algorithm that results from the application of the technique is independent of the neural model.
Exploring different models does not require recoding of the algorithm, but merely the recalculation
of a grid representation. The algorithm can be considered to be a universal population density solver
for one dimensional neural models whenever current compensation is appropriate. We delivered an
implementation of the algorithm in C++ that is publicly available (http://miind.sf.net). We used
this implementation in the examples that illustrate the use of this algorithm.
4
Transforming away neural dynamics
Assume that a neuron is characterised by a vector ~v ∈ M summarising its state, where M is a open
subset of Rn. Further, assume that a smooth vector field ~F (~v) exists everywhere on M and that a
density function ρ(~v, t) is defined for every ~v ∈ M . Now consider a large population of neurons. ρ(~v)d~v
is the probability for a neuron to have its state vector in d~v. It is also assumed that the population is
homogeneous in the sense that interactions between neurons and the outside world can be accounted for
by the same stochastic process for all neurons (although individual neurons each see different realisations
of this process). In the absence of noise, neurons follow trajectories through state space determined by
the flow of vector field ~F (~v):
τ
d~v
dt
= ~F (~v),
(1)
where τ is the neuron's time constant. We must allow for the possibility of v being driven across ∂M ,
the edge of M , at which time it must be reset, immediately or after a refractive period at v = Vreset.
As explained in "Methods: Derivation of the Population Density Equation", conventional
balance arguments lead to an advection equation for the density. When a noise process is also taken into
consideration, this equation becomes:
∂ρ
∂t
+
∂
∂~v · (
~F ρ
τ
) =ZM
d ~w {W (~v ~w)ρ( ~w) − W ( ~w ~v)ρ(~v)}
(2)
W ( ~w ~v) is the probability per unit time for a noise event that will cause a state transition from d~v to
d ~w. This equation is known as the differential Chapman-Kolmogorov equation [24].
Under the usual assumptions that guarantee the existence and uniqueness of solutions of Eq. 1 on
M , one can find integral curves ~v′(t, ~v0(t0)). If ~v0 and t0 are suitably chosen on ∂M , these curves cover
the entire manifold M and every point of M is uniquely defined by a coordinate pair (~v′, t).
These curves are the characteristics of Eq. 2. Applying the method of characteristics one finds that
the total time derivative of the density in the (~v′, t) system is given by:
dρ′
dt
=ZM′
where ρ′(~v′, t) = eR t ∂F (~v′ )
τ ∂~v′
dt′ ρ(~v′, t).
d ~w′nW (~v′ ~w′)ρ′( ~w′) − W ( ~w′ ~v′)ρ′(~v′)o
(3)
For simplicity, the remainder of the paper will consider one dimensional neuronal models subject to
external Poisson distributed spike trains, each spike causing an instantaneous jump in the membrane
potential (the case of more than one input, or step sizes h that originate from a distribution of synaptic
efficacies can be handled easily):
leading to:
W (v′ v) = νδ(v′ − h − v),
dρ
dt
= ν {ρ(v′ − v′h, t) − ρ(v′, t)} + r(t)δ(v′ − v′reset),
(4)
(5)
Here r(t) =R v′th
ρ(w, t)dw; the δ peak reflects the reset of neurons that spiked to their reset potential.
Equation 5 is the Master equation of an inhomogeneous Poisson process. When considered over a
sufficiently short time scale, the solution of this process can be approximated by a homogeneous one, for
which efficient numerical and analytic methods are available [28]. The main technical problem associated
with solving Eq. 5 is to find a grid that adequately represents the density, not only at t = 0, but also
at later times. Note that in v′-space all evolution is due to synaptic input, as in the absence of noise
dρ(v′,t)
v′th,h
dt = 0.
5
10
)
s
m
(
t
8
6
4
2
10
)
s
m
(
t
8
6
4
2
0
10
5
5
0
10
V(t=0) (mV)
15
20
0
10
5
10 15 20 25 30
5 0
V(t=0) (mV)
Figure 1. Solutions to Eq. 1 for LIF neurons (Vth = 20 mV, τ = 10 ms) for I = 0 mV (left)
or I = 30 mV (right).
This idea works well for leaky-integrate-and-fire neurons [27]:
τ
dV
dt
= −V + I,
(6)
where V is the membrane potential and I an external (non-stochastic) current. Here v′-space slowly
expands, which can be accommodated for by adding points to the grid representing the density profile
during simulation. Occasionally one must rebin to curb the resulting growth of the grid (see [27] for
details).
In general neuronal dynamics is not slow compared to synaptic input, e.g. consider the quadratic-
integrate-and-fire model given by:
τ
dV
dt
= V 2 + I,
(7)
where V is a rescaled dimensionless variable. The characteristics can be calculated analytically (Table
1), and are shown in Fig. 2. Their topology depends on the value of I. It is clear that some curves run
away to infinity in finite time: these neurons are intrinsically spiking. Upon reaching infinity, or some
threshold potential Vth, these neurons are reintroduced in the system at V = Vreset.
At first sight, it seems hard to find an adequate density representation without being forced to rapid
and expensive rebinning operations that must be applied in brief succession. However, by adopting a grid
whose bin limits are defined by the characteristics rebinning operations can be avoided altogether. Fast
dynamics will result in large bins, slow dynamics in small ones. Large bins, however, do not introduce
10
)
s
m
(
t
8
6
4
2
0
10
0
5
15
V(t=0) (rescaled)
10
5
10
)
s
m
(
t
8
6
4
2
20
0
10
0
5
15
V(t=0) (rescaled)
10
5
10
)
s
m
(
t
8
6
4
2
20
0
10
0
5
15
V(t=0) (rescaled)
10
5
6
20
Figure 2. Characteristics for the values of I: I = −10 (top), I = 0 (middle) and I = 10
(bottom), on an interval D = [−10, 10].
I > 0
I = 0
I < 0
t = τ
t = τ
√I narctan V
V0 − 1
lnn( V −
t = τ
2√I
√I
V +√I
V
√I − arctan V0√Io V (t) = √I tann√I t
V (t) = V0
1−V0
V (t) = √I (√I+V0)e−2√I t
(√I+V0)e−2√I t
)( V0+√I
√I
t
τ
V0−
)o
τ + arctan( V0√I
)o
τ −(√I−V0)
τ +(√I−V0)
The solutions to Eq. 1, for different values of I.
Table 1. QIF Characteristics
inaccuracy, because all density within the same bin will share a common fate. This is true even after
reset. This observation is crucial and applies quite generally. For this reason we will explain it in some
detail in the next section. 1
7
Figure 3. Modelling advection requires only a pointer update. Top: At time t each
probability bin is associated with a density interval. Bottom: Update at t + 1.
Representing fast dynamics
Starting with a model of the form of Eq. 1, we assume that dV
more general cases later on.
dt > 0 everywhere on (Vmin, Vth), considering
Consider now the initial value problem, posed by Equation 1 with boundary condition V (t = 0) = V0.
Assuming V0 = Vmin, the neuron will reach threshold Vth in finite time tperiod. The neuron will then be
reintroduced at Vreset. For simplicity Vreset = Vmin will be assumed, for a practical implementation of
the algorithm this is immaterial. Consider now a division of time interval tperiod into N time steps:
tstep ≡
tperiod
N
(8)
The density profile ρ(v) for given time t will be represented in a grid consisting of N bins. We define
the bin limits as follows: each bin i (i = 0,··· , N − 1) corresponds to a potential interval [vi, vi+1) in the
following way: v0 = Vmin, and when V (t = 0) = Vmin, vi is given by:
vi ≡ V (itstep), i = 0,··· , N,
(9)
where V (t) is the solution of Eq. 1. Note that there are N + 1 bin limits and that by definition VN ≡ Vth.
In general, the binning is non-equidistant. When a neuron spikes, for example, it traverses a considerable
potential difference in a short period of time and the bins covering this traversal will be very large.
The evolution of a population density profile defined on (Vmin, Vth) can now be modeled as follows. A
probability grid P is created of size N . An array V contains the bin limits of P. We define element i
of V, Vi ≡ vi. Assume an initial density profile ρ(v, 0) is defined at t = 0. We represent this profile by
setting Pi to:
(10)
Pi ≡ ρ(vi, 0)(vi+1 − vi),
1The reset is a crucial difference between neuroscience and finance. It is one reason why analytic results for financial
derivatives do not carry over to neuroscience and why analytic results there are almost unobtainable. The method must be
able to handle reset, but will also work for systems where individuals are not reintroduced.
8
so that Pi approximates the total probability between V = vi and V = vi+1. Remarkably, having defined
the contents of V and P, they can remain constant, yet describe advection: the evolution of the density
in the absence of synaptic input. To see this, consider the evolution over a period of time tstep. All
neurons that previously had a potential between V = vi and V = vi+1, will now have a potential between
V = vi+1 and V = vi+2, except for those that had a potential between V = vN−1 and Vth who now have
a potential between V = Vmin and V = v1. So, the relationship between V and P changes, but not the
actual array contents themselves. Specifically, the density profile at simulation time tsim = jtstep is given
by:
ρ(vi, tsim) =
P(i−j) mod N
vi+1 − vi
This is all that is required to represent the density profile in v′-space; the computational overhead is
negligible as an implementation of this idea only requires a pointer update without any need for shuffling
data around (see Fig. 3).
Equation 5 can be solved by assuming that during tstep neurons only leave their bin due synaptic
input. It is also assumed that, although the magnitude of h is dependent on v′, it is constant during
the small time tstep. The solution method is not really different from that described in [27 -- 29], but is
complicated by the non-equidistance of the probability grid P. Where in the original implementation only
the magnitude of the synaptic efficacy had to be recalculated in v′-space to find the transition matrix of
the Poisson process, here a search is needed to locate the density bin that will receive probability from a
given density bin, as the jump size h′ becomes dependent on v′ in v′-space. This issue is rather technical
and is be explained in detail in "Methods: Synaptic Input: Solving the Master Equation".
Switching Topologies: Current Compensation
The idea is most easily illustrated for a Poisson process. In one dimension, for a single synaptic input
with postsynaptic efficacy h receiving a single Poisson distributed input rate ν, the population density
equation is given by [8]:
∂ρ
∂t
+
∂
∂v
(
F (v)ρ(v)
τ
) = ν {ρ(v − h) − ρ(v)}
(12)
Consider the limit h → 0, such that νh = constant. A Taylor expansion of the right hand side up to
order h2 then gives:
(11)
(13)
(14)
(15)
Now define:
∂ρ
∂t
+
∂
∂v (cid:26) F (v)ρ
τ
+ ρhν −
νh2
2
∂ρ
∂v(cid:27) = 0
and Eq. 13 becomes a Fokker-Planck equation:
µ ≡ νhτ
σ2 ≡ νh2τ
∂ρ
∂t
+
∂
∂v (cid:26)ρ
F (v) + µ
τ
σ2
2τ
−
∂ρ
∂v(cid:27) = 0
This equation describes the evolution of the density due to the deterministic dynamics, determined by
F (v) and an additive Gaussian white noise with parameters µ and σ.
It is clear that Eq. 15 is invariant under the transformation:
(cid:26) F (v) → F (v) + Ic
µ → µ − Ic
(16)
9
1
-110
-210
-10
1
-110
-210
-10
1
-110
-210
-10
0
10
V
0
10
V
0
10
V
Figure 4. Density profile of QIF population decorrelated by large synapses. Line:
population density; markers: Monte Carlo simulation.
for any constant current value Ic, i.e. one add a DC component to a neuron when a similar mean
contribution is subtracted from its input. This result is valid beyond the diffusion approximation, as it
still holds when higher than second order derivatives are considered in the procedure used to derive Eq.
15.
Table 2. Default simulation parameters for QIF neurons.
N
τ
τref
Vmin
Vmax
Vreset
Ic
σc
hdif f
300
10−2 s
0 s
-10
10
-10
0.2
0.01
0.03
r
r
r
10
Examples
We will give a number of examples. For quadratic-integrate-and-fire neurons the default simulation
parameters are given Tab. 2. In the remainder only deviations will be listed.
Example 1: quadratic-integrate-and-fire neurons with extremely large synapses
Figure 4 demonstrates the validity of the method beyond the diffusion limit. A periodically firing group
of neurons, firing in synchrony at the start of the simulation, is decorrelated by a low frequency (5 Hz)
high impact (h = 5, this is half of the entire interval!) input. The density profile is a slowly collapsing
delta-peak, travelling along the characteristics of a periodically firing quadratic-integrate-and-fire neuron
for several seconds, before reaching its steady state profile. We show the density profile in Fig. 4 at
t = 0.02, 0.12 and 9.9 s.
At no moment the population density is well described by a diffusion process, and the steady state
firing rate deviates considerably from that predicted by numerical calculations based on the diffusion
approximation. At t = 0.12 s, the density profile shows a fine structure that is indeed borne out by
large-scale Monte Carlo simulations. We are not aware of a method sensitive to the density gradient that
could have modeled this accurately.
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
-10
6
5
4
3
2
1
0
-10
= -0.1
= 0.1
t
0.019 s
0.029 s
steady state
analytic
-8
-6
-4
-2
0
2
4
6
8
= 0.3
= 0.1
t
0.008 s
0.028 s
0.051 s
steady state
analytic
-8
-6
-4
-2
0
2
4
6
8
10
V
10
V
1.4
1.2
1
0.8
0.6
0.4
0.2
0
-10
1
0.8
0.6
0.4
0.2
0
-10
= -0.1
= 0.5
t
0.019 s
0.029 s
steady state
analytic
-8
-6
-4
-2
0
2
4
6
8
10
V
= 0.3
= 0.5
t
0.008 s
0.028 s
0.051 s
steady state
analytic
0
10
V
Figure 5. Examples of the evolution of the density of a population of QIF neurons under
the influence of Gaussian white noise input.
Example 2: quadratic-integrate-and-fire neurons in the diffusion limit
A single Poisson input, or a combination of one excitatory and one inhibitory Poisson input, can emulate
r
m
s
r
m
s
r
m
s
r
m
s
11
Figure 6. Steady state density profiles in response to Gaussian white noise.
a Gaussian white noise (see "Methods: Emulation of Gaussian white noise"). We will study both
the density transient density profiles and the resulting firing rate response of the population (a neuron
reaching threshold will emit a spike and reset to reset potential. The population firing rate is fraction of
neurons that in a time interval, divided by that time interval).
In Fig. 5 the transient density profiles are shown for four different (µ, σ) combinations. The negative
input, small noise distribution (µ = −0.1, σ = 0.1) peaks at the stable equilibrium point. Almost no
neurons are pushed across the second, instable equilibrium point and this population does not fire in
response to its input (see Fig. 7). For higher noise values, the peak is smeared and some neurons are
pushed across the unstable, leading to an appreciable firing rate; a deterministic current with similar
mean would evoke no response (see Fig. 7).
A larger mean µ = 0.3 evokes a clear response, and the peak of the density is clearly driven across
V = 0. Unlike in Example 1, this peak will soon collapse. It will approach its steady state distribution
but demonstrate small oscillations that are clearly visible in the firing rate for a long time. At higher
noise, the steady state is reached faster and the firing rates transients die out much faster (Fig. 7).
Figures 6 and 7 provide a summary of the diffusion results for µ ∈ [−0.5, 0.5] and σ = 0.1, 0.2, 0.5, 0.7.
Figure 7 shows the transient firing rates are shown and compared to Monte Carlo simulations (red
markers). Figure 6 shows the density profiles for large time t = 9.9 s, when profiles should assume their
steady state (grey line, but barely visible since Monte Carlo and analytic results agree and are overlaid).
Markers show Monte Carlo results, red lines are calculated directly from the diffusion approximation,
using the integration scheme from [15]. Figure 8 shows the so-called gain curves, the steady state firing
rates that can be read of Fig. 7 in black markers. Monte Carlo results are indicated by red markers, lines
12
0.5
-0.1
-0.2
-0.4
=-0.5
=0.7
=0.5
=0.2
=0.1
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
-0.3
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
0.3
0.2
0.1
0
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
0.4
50
45
40
35
30
25
20
15
10
50
45
40
35
30
25
20
15
10
50
45
40
35
30
25
20
15
10
50
45
40
35
30
25
20
15
10
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Figure 7. Transient firing rates inresponse to Gaussian white noise.
50
45
40
35
30
25
20
15
10
5
0
50
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
50
0
45
40
35
30
25
20
15
10
5
0
0
s
m
s
s
s
13
)
z
H
(
f
25
20
15
10
5
0
=
0.1
0.2
0.5
0.7
-0.4
-0.2
0
0.2
0.4
0.6
Figure 8. Gain curves. Lines: analytic steady state firing rates. Red error bars: Monte Carlo
simulation, black markers: population density results.
are calculated analytically from the diffusion approximation [30]. All results are in excellent agreement
with each other.
Example 3: The gain curve is changed by large synapses.
A single Poisson input can only approximate Gaussian white noise as long as the variability is much
smaller than the mean of the signal. Figure 9 (top) shows the efficacy h as a function of µ for different
σ values, and hence indicate where deviations can be expected (namely for large h).
Figure 9 shows a population receiving input from a single Poisson input with firing rate and efficacy
calculated using Eq. 30. µ must not be interpreted as the DC contribution of a Gaussian white noise
input.
A value of I = −1(Ic = 1.1) was chosen and simulated both in Monte Carlo and using the population
density approach. Here, one expects the population to fire when the neurons' negative I value is overcome,
i.e. for µ > 1.
The steady state values of the population density approach are in excellent agreement with the Monte
Carlo, but both deviate from the gain curves that are calculated in the diffusion approximation (Fig. 9
(bottom)), precisely where this would have been expected according to Fig. 9 (top). Note, however, that
for the first time it was necessary to smear the synaptic efficacies slightly (σ = 0.01). In the absence of
smearing, there would have been a disagreement between Monte Carlo and data for low σ. This is due
to the fact that the variability of the compensation current Ic cannot be chosen arbitrarily low as this
would entail artificially high firing rates that would render the solution of the Master equation inefficient.
A reasonable minimum value is σ = 0.01.
As can be seen from Fig. 9, the jumps are really large for σ = 0.7 and low µ. Indeed, there the
deviation from the diffusion approximation is substantial, but there is no visible disagreement between
Monte Carlo and population density approach. This is all the more impressive because to model negative
I values, current compensation with Ic = 1.1 had to be used, and demonstrates the viability of the
current compensation approach.
When the steady state density profile obtained with the population density method is compared
to that obtained from the Monte Carlo, again good agreement is found, while a small but significant
deviation from the density profile calculated from the steady state, of the Fokker-Planck equation Eq. 15
can be seen for large σ and low µ (Fig. 9 (top)).
Example 4: Universality: quadratic- and leaky-integrate-and-fire neurons are handled
by the same algorithm.
m
s
14
=
=
=
=
0.1
0.1
0.1
0.1
0.2
0.2
0.2
0.2
0.5
0.5
0.5
0.5
0.7
0.7
0.7
0.7
h
1
0.8
0.6
0.4
0.2
0
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
)
z
H
(
f
25
20
15
10
5
0
=
0.1
0.2
0.5
0.7
0.6
0.8
1
1.2
1.4
Figure 9. For large synapses the gain curve changes. Top: Jump size for a single Poisson input
emulating Gaussian white noise. Bottom: Monte Carlo (red markers), population density results (black
markers) and diffusion prediction (lines).
We use current compensation to turn leaky-integrate-and-fire neurons into spiking neurons. We add a
small constant current Ic = 1.1 as in Eq. 6. Without this current the characteristics move away from
threshold, with this current they cross threshold (compare Fig. 1 left and right).
A Gaussian white noise current emulated by Poisson input with mean µ = −1.1 and small variability
(σ = 0.05) is presented as input. This is combined with a single Poisson input of 800 Hz with synaptic
efficacy h = 0.03. This example was used in earlier studies as a benchmark [8, 27].
In Fig. 10 the
steady state density profile and the firing rate of the population are shown. They are compared to the
simulation without current compensation, described earlier [27], using a density representation tailored
to leaky-integrate-and-fire characteristics (red line).
The results are almost identical. In the original simulation the steep peaks are the consequence of a
single synaptic efficacy. Sirovich [29] has shown analytically that for the first peak the density profile is
discontinuous, for the second peak the first derivative is discontinuous, and so on. This is not necessarily
realistic in terms of neuroscience, but it demonstrates once more that the method can handle the jagged
density profiles that result from finite size jump processes.
In the current compensation version of the simulation, the small variability of the compensation
current smears these peaks. The noise also slightly reduce the extremes of the firing rate (Fig. 10),
but otherwise the agreement is excellent. The effect of the smearing due to the compensation current
is equivalent to effects caused by spread in the synaptic efficacies [30]. Moreover, many brain systems
receive an unspecific background that must be modeled as noise anyway, so current compensation can be
m
s
s
s
s
m
s
15
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
20
18
16
14
12
10
8
6
4
2
0
0
i
)
s
/
s
e
k
p
s
(
e
t
a
r
g
n
i
r
i
f
0
0.2
0.4
0.6
0.8
1
V (dimensionless)
0.05
0.1
0.15
0.2
0.25
0.3
t (s)
Figure 10. Algorithm works for both QIF and LIF neurons. Top: steady state density profile
for current compensation (black), dedicated LIF neuron implementation (red). Bottom: output firing
rates.
applied quite generally in neuroscience.
Importantly, the only change that had to made to change over from quadratic- to leaky-integrate-
and-fire neurons is in the calculation of array V, which is done during the initialization of the algorithm.
Discussion
We presented a method for numerically solving the differential Chapman-Kolmogorov equation for large
jumps. We demonstrated the validity of the approach by comparing population density results for neu-
ronal populations receiving input from large synapses to Monte Carlo simulations, and neurons receiving
input from small synapses to analytic or numerical results obtained in a diffusion approach. We presented
two examples (Example 1 and 4) that demonstrate that the method can handle arbitrary jump sizes and
the extremely jagged density profiles that result.
We also demonstrated that the algorithm is universal:
it can handle any one dimensional neural,
subject to the condition that a (pseudo) diffusive background of small variability is allowed. This point is
obvious mathematically, the transformation from the differential Chapman-Kolmogorov equation to the
Master equation of the noise process can always be performed [27]. The challenge is to find a suitable
density representation. Periodically spiking neurons yield a particularly simple density representation,
In practice the
and using current compensation a non spiking neural model can be made to spike.
r
16
Figure 11. High level overview of the approach. Neural model (red) only enters explicitly is in V.
compensation current will always introduce some variability. Where variability is present already, for
example because the input is noise or because synaptic efficacies are variable, this extra variability can
be absorbed and has no bearing on the simulation results.
The universality of the algorithm is reflected in the implementation. Although the neuronal model
defines the grid boundaries of the density representation - the content of grid V, it does so before simulation
starts. This means that neuronal models can be exchanged with great simplicity. To provide a neural
model is to provide the contents of grid V. This is the only place were the neuronal model enters the
method explicitly. This important point is emphasized in Fig. 11, which provides a high level overview
of the method.
We demonstrated the method using a Poisson process and showed that if the jump size becomes
small, diffusion results are recovered. The use of Poisson statistics is appropriate in neural simulations.
An individual neuron will receive at most of the order of 105 spikes per seconds. At such rates, a Poisson
Master equation can solved efficiently. We found that 1 simulation second could be handled in between a
0.5 and 2.0 s real time, which is an order of magnitude faster than Monte Carlo - based on comparisons to
leaky-integrate-and-fire 2 simulations in NEST [31] where a 10000 neuron population achieves real time
ratios of approximately 30. We found an excellent recovery of diffusion results whenever the jump size is
5 % of the interval for low noise and of the order of 10 % for high noise conditions (Fig. 9).
2Quadratic-integrate-and-fire neurons were not implemented at the time of writing. We performed Monte Carlo simula-
tions by writing an event-driven simulator based on the analytic solutions of Table 1.
17
It is conceivable that in other applications jump sizes are smaller. Artificially small jump sizes would
lead to very high firing rates for a Poisson master equation. We find that adopting a jump size of less
than 1 % of the interval leads to inconveniently high firing rates. The method is clearly best used for
finite jump sizes.
To study diffusion, it would be interesting to consider a finite difference scheme for solving Eq.
15 with F (v) = 0. This would allow a direct comparison to diffusion results, while still benefitting
from the universality of this approach. For example, it would still be possible to compare diffusion in
different neural models without the need to derive a separate Fokker-Planck equation for each case. Cox
and Ross [32] have shown that explicit finite difference schemes for solving equations of the type 15
correspond to a jump process with three possibilities: a move left, right or no move in the grid. Brennan
and Schwartz [25] have shown that implicit finite difference scheme corresponds to a jump process where
jumps to every other position are allowed. For suitable parameters and in the limit of an infinitely fine
grid both schemes converge to a diffusion process, where the implicit scheme is unconditionally stable.
These results are interesting because they suggest suitable Master equations to arrive at good diffusion
approximations.
In the context of neural systems this discussion is to some extent academic as true
diffusion conditions are not realized in the brain, but most analytic results on population densities have
been derived in the diffusion approximation, and a comparison is valuable.
The ideas presented here constitute an approach rather than an algorithm. There is an endless
variation of noise processes and dynamical systems that can be combined. Colored noise can be handled
by a two-step approach: the Master equation of a colored noise process can be simulated by leaky-
integrate-and-fire population receiving white noise.
The technique was demonstrated on one dimensional neural populations. Such models are simple
enough to demonstrate the applicability of the technique to a wider audience. Whether it can be success-
fully employed to more complex neural models is an open question. Fortunately, there is a tendency to
move from complex conductance-based models to simpler low dimensional effective model neurons, see
e.g. [33]. The population density approach still is competitive in two dimensions [14].
We hope that the technique will find widespread application. Depending on the application area,
the methods described here may need adaptation. The current compensation technique may not work
for all applications as the extra variability introduced by the compensation current might not always be
acceptable. However, this is no fundamental objection to using this approach. Applying a combination of
techniques described here and earlier [27, 28] useful density representations may still be found, although
their implementation may be limited to a specific topology of the deterministic flow.
Methods
Solutions in absence of synaptic input
Consider a neuron in the absence of synaptic input, but receiving a constant input current I. In general
Equation 1 can be used to evolve the state of a neuron that has potential V0 at t = 0. For some cases
analytic solutions exist. For leaky-integrate-and-fire neurons:
V (t) = I − (I − V0)e−
t
τ
(17)
For quadratic-integrate-and-fire neurons, the solutions of Eq. 7 depend on the sign of I, and there are
clear topological differences between solutions for different I. They are given in Tab. 1 and shown in
Fig. 2.
18
Derivation of the population density equation
The population density approach can be motivated as follows and is not restricted to one dimensional
neuronal models. To reflect that, this section will consider the more general model:
τ
d~v
dt
= ~F (~v)
(18)
Assume that M is an open subset of Rn, where n is the dimensionality of the neuronal point model
defined by Eq. 18. Assume that the neuronal state is given by point ~v in M. Further, assume that a
density function ρ(~v, t) is defined on M and that ρ(~v, t)d~v represents the probability for a given neuron
to be in the state space element d~v at time t. Consider the evolution of neuronal states for neurons in
d~v during a short period of time dt. If external input is ignored, then under the influence of neuronal
dynamics all neurons, and no others, that were in d~v at time t are in volume element d~v′ at time t′, with
~v′ and t′ defined by:
Conservation of probability requires:
(cid:26) t → t′ = t + dt
~v → ~v′ = ~v + ~dv = ~v +
~F
τ dt
J =
so that, collecting all terms up to O(dt):
dt +O(dt2),
∂~v
( 1
τ
∂ ~F (~v)
∂~v
ρ +
∂ρ
∂t
+
~F (~v)
τ
∂ρ
∂~v) d~v = 0.
·
Since this equation must hold for infinitesimal but arbitrary volumes d~v, it follows that:
∂ρ
∂t
+
1
τ
∂
∂~v · ( ~F ρ) = 0
Equation 26 is an advection equation for probability flow under the influence of neuronal dynamics. When
noise is also considered the equation becomes:
∂ρ
∂t
+
1
τ
∂
∂~v · ( ~F ρ) =ZM
d ~w {W (~v ~w)ρ( ~w) − W ( ~w ~v)ρ(~v)} .
(27)
This is an example of what Gardiner calls the differential Chapman-Kolmogorov equation [24]. In the
remainder only Poisson processes will be considered, but the method is applicable to any noise process
that can be modeled by a Master equation. For a single Poisson input W (~v′ ~v) is given by:
W (~v′ ~v) = νδ(v′0 − v0 − h)
(28)
Using:
Equation 20 becomes:
Since
with:
ρ(~v, t)d~v = ρ(~v′, t′)d~v′.
ρ(~v′, t′) = ρ(~v, t) + dt
∂ρ(~v, t)
∂t
+ d~v ·
∂ρ(~v, t)
∂~v
,
(cid:26)ρ(~v, t) + d~v ·
∂ρ(~v, t)
∂~v
+ dt
∂ρ(~v, t)
∂t (cid:27) d~v′ = ρ(~v, t)d~v
d~v′ = J d~~v,
∂ ~F (~v)
∂~v′
∂~v = I +
1
τ
(19)
(20)
(21)
(22)
(23)
(24)
(25)
(26)
19
Here ν is the rate of the Poisson process, it is assumed that one of the neuronal state variables, v0
instantaneously transitions from v0 to v0 + h, where h is the synaptic efficacy.
In a one dimensional
neuronal model v0 is V , the membrane potential. Clearly, not all synapses in a population have a single
value, and often one assumes:
W (~v′ ~v) = νZ dhp(h)δ(v′0 − v0 − h),
(29)
where p(h) is often assumed to be Gaussian [8, 30]. In a network where many populations connect to
any given population, multiple contributions of the form Eq. 29 must be included - one for each separate
input. In principle, each input must be modeled separately, but often many inputs can be subsumed
into a single one using the central limit theorem, leading to a Gaussian white noise input contribution.
Sometimes kernels p(h) of separate inputs overlap and can be integrated into one input contribution.
Figure 12. Emulation of a Gaussian white noise with Poisson inputs. Left: isoline of
h = 0.01Vth, h = 0.1Vth. Right: synaptic efficacies are fixed at 3 % of Vth. Blue (red) lines are isolines of
frequency of excitatory (inhibitory) inputs.
20
Figure 13. The Master equation follows from translating the contents of grid V the
synaptic efficacy.
Emulation of Gaussian white noise and the breakdown of the diffusion approx-
imation
It is evident that by choosing a small efficacy h and large input rate ν, the Poisson input approximates
additive Gaussian white noise. These parameters must be chosen such that they provide the correct µ
and σ. Inverting Eq. 14 gives:
h =
ν =
σ2
µ
µ2
τ σ2
(30)
A single Poisson input can only emulate a Gaussian white noise when h is small compared to the relevant
potential difference, e.g. for leaky-integrate-and-fire neurons when h ≪ Vth, or h ≪ I for quadratic-
integrate-and-fire neurons, i.e. not for all values of (µ,σ) can values for h and ν be found so that the
diffusion regime is reliably reproduced. This is borne out by Fig. 12 (left). Here a neuron is considered
with τ = 10 ms, Vth(≡ θ) = 20 mV. For a (µ, σ) plane isolines of h are plotted: one where h is precisely
at 1% of the threshold potential, one where it is at 10%.
It is clear that a single Poisson input can emulate a broad range of µ values, but only at low σ. Using
two inputs, one excitatory and one inhibitory, one can fix the synaptic efficacies at a low value compared
to Vth. If this value is J one finds:
µ = τ J ν(νe − νi)
σ2 = τ νJ 2(νe + νi)
(31)
One then has two frequencies νe and νi as free parameters. This setup covers most of the (µ, σ) plane,
except for values at low σ, where the input frequency of the inhibitory input falls below 0. In Fig. 12 only
frequency isolines for positive input frequencies are shown. The boundary where the inhibitory input rate
falls below 0 is forbidden for two inputs, but this is precisely the area covered by a single input (clearly
if νi = 0, we have a single excitatory input). The arguments presented here cover a positive range of µ,
similar arguments can be made for a net inhibitory input.
In summary, one needs either one or two Poisson inputs to emulate a Gaussian white noise.
In
particular, this implies that we can use a single excitatory input as a model for the gradual breakdown
of the diffusion limit:
if we increase σ and calculate the synaptic efficacy using Eq. 30, we should see
deviations from the diffusion limit as h can no longer be considered to be small.
This is shown in Fig. 9 which shows the step size for a single Poisson input given µ, σ. For σ
comparable to µ this step size is not necessarily small compared to the potential scale on which the
neuronal model is defined and deviations of the diffusion approximation can be expected.
21
Whenever a Gaussian white noise with very small variability is required, this must be delivered by
a single Poisson input. The firing rate is then inversely proportional to the jump size. For very low
variability a numerical solution of the Poisson process may become inefficient.
Synaptic input: solving the master equation
Now consider a single Poisson input. The evolution of the density profile is straightforward in principle
but complicated by the bins being non-equidistant. Formally the noise process can be represented as a
transition matrix. Figure 13 how this matrix can be determined: the contents of grid V are shifted by an
amount h, where h is the synaptic efficacy. Probability moves from potential V − h to V as consequence
of an input spike. In large bins this means that neurons stay mostly within the same bin. Neurons that
are in small bins may end up several bins higher. Below we will examine this process in detail.
10
5
0
5
10
0
10
20
30
40
0
10
20
30
40
0
10
20
30
40
0.0
0.2
0.4
0.6
0.8
0
10
20
30
40
Figure 14. Transition matrices for QIF (top) and LIF neurons (bottom). Loss of probability
is indicated by black, gain by white.
To determine the Poisson Master equation for excitatory input (h > 0), consider that every neuron
arriving in the interval Di = [vi, vi+1) must have come from Di,h = [max(vi−h, Vmin), vi+1−h). Consider,
the N + 1 integers defined by:
pi ≡( vi − h < v0:
vi − h ≥ v0:
-1
arg min
k
vk ≤ vi − h < vk+1
(32)
Now consider i = 0 ··· N − 1; if pi = pi+1, αi,pi is given by:
vi+1 − vi
vpi+1 − vpi
αi,pi ≡
For pi = −1, αi,l = 0. If pi < pi+1, αi,l is given by:
;
αil ≡
with:
Finally,
l = pi:
pi < l < pi+1:
l = pi+1:
vpi +1−(vi−h)
vpi +1−vpi
θl
1
vi+1−h−vpi+1
vpi+1+1−vpi+1
θl ≡(cid:26) l < 0:
l ≥ 0:
0
1
From this the Master equation at t = ktstep follows:
0
vpN +1−vN +h
vpN +1−vpN
1.
i < pN :
i = pN :
i > pN :
βi ≡
= − Pi−k mod N +
dPi−k mod N
dt
22
(33)
(34)
(35)
(36)
(37)
(38)
(39)
αi,lPl−k mod N
pi+1
Xl=pi
βlPl−k mod N ),
N−1
+δiR(
Xl=0
vk ≤ Vreset ≤ vk+1
R ≡ arg min
k
where R is defined by:
and δ is the Kronecker δ. The term including it reflects the re-entry of neurons at the reset potential
once they have spiked.
For inhibitory input (h < 0), the formulae are almost identical as now v−h refers to a higher potential.
Again, define
pi ≡( vi − h > vN : N
vi − h ≤ vN :
arg min
k
vk ≤ vi − h < vk+1
Again consider i = 0 ··· N − 1; if pi = pi+1, αi, pi is given by:
vi+1 − vi
v pi+1 − v pi
αi,pi ≡
;
if pi < pi+1, αi,l is given by:
αil ≡
Finally, γi is defined as follows:
l = pi:
pi < l < pi+1:
l = pi+1:
v pi +1−(vi+h)
v pi+1−v pi
1
θN−1− pi+1
vi+1+h−v pi+1
v pi+1+1−v pi+1
i < p0:
i = p0:
i > p0:
1
vp0+1−(v0+h)
0
vp0+1−vp0
γi ≡
(40)
(41)
(42)
(43)
The Master equation for an inhibitory input is:
dPi−k mod N
dt
= −(1 − γi)Pi−k mod N +
pi+1
Xl= pi
αi,lPl−k mod N
Note that Eqs. 38 and 44 are of the form:
d ~P
dt
= M(k) ~P ,
23
(44)
(45)
where M(e) is the transition matrix resulting from adding individual input contributions of the form 38
and 44.
Solving the master equation
Equation 38 and 44 were solved numerically using Runge-Kutta-Fehlberg integration. This is not optimal,
but is very straightforward to implement, and yields a performance that is faster than Monte Carlo
simulation by at least one order of magnitude. Note that in Eq. 38 one simply ignores the zero elements
of M(e). The calculation of the time derivatives then scales as O(N )O(Ninput), where Ninput is the
number of inputs that need to considered. In many neural systems, such as cortex it can be assumed
that most inputs deliver a background that can be treated as a single noisy input [14, 30]
Different neuronal models: different transition matrices
Although the noise process and the neuronal dynamics are independent, the neuronal dynamics determines
the representation of the density profile. The bin boundaries of the density profile in v-space, V are given
by the evolution of the neuronal state, and therefore the transition matrix is directly dependent on the
neuronal model. This is illustrated in Fig. 14, which shows the transition matrix for identical h, ν for
quadratic-integrate-and-fire and leaky-integrate-and-fire neurons. First, it should be observed that to
accommodate for spiking behaviour, the state space of the quadratic-integrate-and-fire neuron is much
larger, for the leaky-integrate-and-fire neuron it is limited by the threshold, Vth. Second, the quadratic-
integrate-and-fire neuron is a truly spiking model: once it spikes, its state runs away to infinity in finite
time. The leaky-integrate-and-fire neuron runs over threshold towards the externally applied current in
infinite time.
It is clear that probability lost from a bin will end up in a higher bin, i.e. a bin corresponding to
a higher potential, as expected for an excitatory event. For leaky-integrate-and-fire neurons the same
is done. The differences between a spiking neuron model and a non-spiking model are clearly visible.
The outer bins of the quadratic-integrate-and-fire grid are huge, because the potential difference bridged
during a simulation step tstep is large during a spike. Relative to their size the outer bins are barely shifted
with respect to each other, meaning that most probability originating from this bin during a synaptic
event will end up there again. This makes intuitive sense: during a spike the influence of synaptic input
is negligible. Indeed the transition matrix shows that no probability transfer takes place in the outer bins
for quadratic-integrate-and-fire neurons. For leaky-integrate-and-fire neurons, the picture is different:
neurons will mainly pass threshold due to synaptic input. After crossing threshold, their potential will
be reset. The influence of the reset is clearly visible in the leaky-integrate-and-fire transition matrix as a
horizontal line in the row corresponding to the reset bin.
24
Acknowledgments
References
1. Ricciardi LM (1977) Diffusion Processes and Related Topics in Biology (Lecture Notes in Biomath-
ematics). Springer-Verlag.
2. Murray JD (2007) Mathematical Biology: I. An Introduction (Interdisciplinary Applied Mathe-
matics) (Pt. 1). Interdisciplinary applied mathematics. Springer, 3rd edition.
3. Alt W (1980) Biased random walk models for chemotaxis and related diffusion approximations.
Journal of Mathematical Biology 9: 147 -- 177.
4. Skellam JG (1951) Random dispersal in theoretical populations. Biometrika 38: 147 -- 218.
5. Skellam J (1973) The Formulation and Interpretation of Mathematical Models of Diffusionary
Processes in Population Biology. In: Bartlett MS, Hiorns RW, editors, The Mathematical Theory
of the Dynamics of Biological Populations, New York: Academic Press.
6. Stein RB (1965) A Theoretical Analysis of Neuronal Variability. Biophysical Journal 5: 173 -- 194.
7. Knight BW (1972) The Relationship between the Firing Rate of a Single Neuron and the Level of
Activity in a Population of Neurons. The Journal of General Physiology 59: 767 -- 778.
8. Omurtag A, Knight BW, Sirovich L (2000) On the simulation of large populations of neurons.
Journal of Computational Neuroscience 8: 51 -- 63.
9. Brunel N, Chance F, Fourcaud N, Abbott L (2001) Effects of synaptic noise and filtering on the
frequency response of spiking neurons. Physical Review Letters 86: 2186 -- 2189.
10. Haskell E, Nykamp DQ, Tranchina D (2001) Population density methods for large-scale modelling
of neuronal networks with realistic synaptic kinetics: cutting the dimension down to size. Network
12: 141 -- 174.
11. Casti ARR, Omurtag A, Sornborger A, Kaplan E, Knight B, et al. (2002) A Population Study of
Integrate-and-Fire-or-Burst Neurons. Neural Computation 14: 957 -- 986.
12. Fourcaud-Trocm´e N, Hansel D, van Vreeswijk C, Brunel N (2003) How spike generation mechanisms
determine the neuronal response to fluctuating inputs. The Journal of Neuroscience 23: 11628 --
11640.
13. Mattia M, Del Giudice P (2004) Finite-size dynamics of inhibitory and excitatory interacting
spiking neurons. Physical Review E 70: 052903+.
14. Apfaltrer F, Ly C, Tranchina D (2006) Population density methods for stochastic neurons with
realistic synaptic kinetics: Firing rate dynamics and fast computational methods. Network 17:
373 -- 418.
15. Richardson MJE (2007) Firing-rate response of linear and nonlinear integrate-and-fire neurons to
modulated current-based and conductance-based synaptic drive. Physical Review E 76: 021919+.
16. Helias M, Deger M, Rotter S, Diesmann M (2011) Finite Post Synaptic Potentials Cause a Fast
Neuronal Response. Frontiers in Neuroscience 5.
17. Dumont G, Henry J (2013) Population density models of integrate-and-fire neurons with jumps:
well-posedness. Journal of Mathematical Biology 67: 453 -- 481.
25
18. Codling EA, Plank MJ, Benhamou S (2008) Random walk models in biology. Journal of The Royal
Society Interface 5: 813 -- 834.
19. Sirovich L (2007) Populations of Tightly Coupled Neurons: The RGC/LGN System. Neural Com-
putation 20: 1179 -- 1210.
20. Helias M, Deger M, Rotter S, Diesmann M (2010) Instantaneous Non-Linear Processing by Pulse-
Coupled Threshold Units. PLoS Comput Biol 6: e1000929+.
21. Merton R (1976) Option pricing when underlying stock returns are discontinuous. Journal of
Financial Economics 3: 125 -- 144.
22. Kou SG (2002) A Jump-Diffusion Model for Option Pricing. Management Science 48: 1086 -- 1101.
23. Cont R, Voltchkova E (2005) A Finite Difference Scheme for Option Pricing in Jump Diffusion and
Exponential L´evy Models. SIAM Journal on Numerical Analysis 43: 1596 -- 1626.
24. Gardiner CW (1997) Handbook of Stochastic Methods. Springer, 2 edition.
25. Brennan MJ, Schwartz ES (1978) Finite Difference Methods and Jump Processes Arising in the
Pricing of Contingent Claims: A Synthesis. The Journal of Financial and Quantitative Analysis
13: 461+.
26. Van Kampen NG (1992) Stochastic Processes in Physics and Chemistry, Volume 1, Second Edition
(North-Holland Personal Library). North Holland, 2 edition.
27. de Kamps M (2003) A simple and stable numerical solution for the population density equation.
Neural Computation 15: 2129 -- 2146.
28. de Kamps M (2006) An analytic solution of the reentrant Poisson master equation and its appli-
cation in the simulation of large groups of spiking neurons. In: International Joint Conference on
Neural Networks, 2006. pp. 102 -- 109.
29. Sirovich L (2003) Dynamics of neuronal populations: eigenfunction theory; some solvable cases.
Network (Bristol, England) 14: 249 -- 272.
30. Amit DJ, Brunel N (1997) Model of global spontaneous activity and local structured activity during
delay periods in the cerebral cortex. Cerebral cortex (New York, NY : 1991) 7: 237 -- 252.
31. Gewaltig MO, Diesmann M (2007) Nest (neural simulation tool). Scholarpedia 2: 1430.
32. Cox JC, Ross SA (1976) The valuation of options for alternative stochastic processes. Journal of
Financial Economics 3: 145 -- 166.
33. Brette R, Gerstner W (2005) Adaptive exponential integrate-and-fire model as an effective descrip-
tion of neuronal activity. Journal of Neurophysiology 94: 3637 -- 3642.
|
1706.05702 | 1 | 1706 | 2017-06-18T19:13:26 | Perfect spike detection via time reversal | [
"q-bio.NC",
"math.DG",
"physics.bio-ph",
"q-bio.QM"
] | Spiking neuronal networks are usually simulated with three main simulation schemes: the classical time-driven and event-driven schemes, and the more recent hybrid scheme. All three schemes evolve the state of a neuron through a series of checkpoints: equally spaced in the first scheme and determined neuron-wise by spike events in the latter two. The time-driven and the hybrid scheme determine whether the membrane potential of a neuron crosses a threshold at the end of of the time interval between consecutive checkpoints. Threshold crossing can, however, occur within the interval even if this test is negative. Spikes can therefore be missed. The present work derives, implements, and benchmarks a method for perfect retrospective spike detection. This method can be applied to neuron models with affine or linear subthreshold dynamics. The idea behind the method is to propagate the threshold with a time-inverted dynamics, testing whether the threshold crosses the neuron state to be evolved, rather than vice versa. Algebraically this translates into a set of inequalities necessary and sufficient for threshold crossing. This test is slower than the imperfect one, but faster than an alternative perfect tests based on bisection or root-finding methods. Comparison confirms earlier results that the imperfect test rarely misses spikes (less than a fraction $1/10^8$ of missed spikes) in biologically relevant settings. This study offers an alternative geometric point of view on neuronal dynamics. | q-bio.NC | q-bio |
Perfect spike detection via time reversal
J. Krishnan1 2 4, P.G.L. Porta Mana1, M. Helias1 2 3, M. Diesmann1 2, E. Di Napoli2 4
1Institute of Neuroscience and Medicine (INM-6), Jülich Research Centre, Germany
2Institute for Advanced Simulation (IAS), Jülich Research Centre, Germany
3Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany
4Aachen Institute for advanced study in Computational Engineering Science
(AICES), RWTH Aachen University, Germany
Correspondence*:
Jeyashree Krishnan
[email protected]
ABSTRACT
Spiking neuronal networks are usually simulated with three main simulation schemes: the
classical time-driven and event-driven schemes, and the more recent hybrid scheme. All three
schemes evolve the state of a neuron through a series of checkpoints: equally spaced in the first
scheme and determined neuron-wise by spike events in the latter two. The time-driven and the
hybrid scheme determine whether the membrane potential of a neuron crosses a threshold at the
end of of the time interval between consecutive checkpoints. Threshold crossing can, however,
occur within the interval even if this test is negative. Spikes can therefore be missed.
The present work derives, implements, and benchmarks a method for perfect retrospective
spike detection. This method can be applied to neuron models with affine or linear subthreshold
dynamics. The idea behind the method is to propagate the threshold with a time-inverted
dynamics, testing whether the threshold crosses the neuron state to be evolved, rather than vice
versa. Algebraically this translates into a set of inequalities necessary and sufficient for threshold
crossing. This test is slower than the imperfect one, but faster than an alternative perfect tests
based on bisection or root-finding methods. Comparison confirms earlier results that the imperfect
test rarely misses spikes (less than a fraction 1/108 of missed spikes) in biologically relevant
settings.
This study offers an alternative geometric point of view on neuronal dynamics.
Keywords: State-space analysis, NEST, time-driven, event-driven, simulation, LIF neuron, differential geometry
1 INTRODUCTION
In the last decade, considerable work has been devoted to improve the accuracy of simulators that are
capable of efficiently simulating large networks of spiking neurons (Hansel et al., 1998; Mattia and
Del Giudice, 2000; Shelley and Tao, 2001; Dehaene and Changeux, 2005; Morrison et al., 2007; Brette,
2007; D'Haene et al., 2009; van Elburg and van Ooyen, 2009; Zheng et al., 2009; Hanuschkin et al., 2010).
The field is driven by the ideal of combining the capability to cope with the high-frequency of synaptic
1
J. Krishnan et al.
Perfect spike detection via time reversal
events arriving at a neuron in nature with a mathematically accurate implementation of the threshold
process a wide class of neuron models is based on.
Two classical schemes to simulate neuronal networks are the time-driven and the event-driven schemes
(Fujimoto, 2000; Zeigler et al., 2000; Ferscha, 1996). Both schemes describe the state of the neurons by a
set of variables and the action potentials as events that mediate the interaction between them.
In a time-driven scheme, the state of a neuron is updated on a time grid defined by the simulation
step (for a review see Morrison and Diesmann, 2008). After all neurons are updated, their membrane
potential is checked for threshold crossings. If the membrane potential of a neuron is above the threshold at
this checkpoint, a spike is delivered to all neurons it is connected to. Subsequently, a new iteration step
begins. The step size stipulates how frequently occurrences of threshold crossings are inspected during the
simulation. The choice of the step size a trade-off between spike accuracy and the speed of the simulation
Morrison et al. (2007). Such grid-constrained simulations force each spike event to a position on the
equidistant temporal grid spanned by the step size and therefore induce artificial synchronization of the
network dynamics (Hansel et al., 1998; Shelley and Tao, 2001; Morrison et al., 2007; Brette, 2007; van
Elburg and van Ooyen, 2009; Hanuschkin et al., 2010).
In an event-driven scheme, the state of a neuron is updated only when it receives a spike. A central
queue of events is maintained and each spike is inserted into this queue with its own time stamp. Upon
update a neuron predicts when its next spike will occur in the absence of further input. This preliminary
event is inserted into the queue and confirmed if it becomes due or removed when invalidated by further
input. Efficient and elegant predicition methods have been developed for classes of neuron models without
invertable dynamics (Brette, 2007; van Elburg and van Ooyen, 2009; D'Haene et al., 2009; D'Haene and
Schrauwen, 2010; Ferscha, 1996). However, maintaining a central queue in a distributed simulation is
challenging and may compromise the time performance of the simulator (Hanuschkin et al., 2010, for a
detailed review see).
A hybrid scheme circumvents the shortcomings of both these schemes by embedding a locally event-
driven algorithm for each neuron into a globally time-driven scheme (Morrison et al., 2007). The arrival
of a spike at a neuron introduces one additional update and check point. The dynamics of a given neuron
is then propagated from incoming spike to incoming spike and eventually to the end point of the global
timestep. If the membrane potential of a neuron is above the threshold at a local or global check point,
the precise point of threshold crossing is determined in continuous time, and a spike is emitted. Next to
their location on the time grid, in this scheme spike events carry a floating point offset. Thus, in contrast
to an event-driven scheme, the hybrid scheme does not predict future spike times but identifies threshold
crossings only retrospectively. Hanuschkin et al. (2010) demonstrate that the latter scheme is equally
accurate as the former at lower computational costs.
The hybrid scheme still has a loophole, however: spikes can be missed. The reason is that, as in the
time-driven scheme, crossing of the threshold voltage is tested by inspecting whether the membrane
potential of the neuron is above threshold at the end of a checkpoint (see Figure 1 for an illustration of the
scenario). Nevertheless, the membrane potential V , evolved by the subthreshold dynamics, can be below
threshold θ at two consecutive checkpoints t and t + h while a double, quadruple, etc. threshold-crossing
occurred in between. The first crossing constitutes a missed spike. Symbolically,
V (t + h) ⩾ θ =⇒ threshold crossing
This is a provisional file, not the final typeset article
(1)
2
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 1. Illustration of undetected threshold crossing between two consecutive checkpoints t and t + h
on the time grid. The short black vertical bar represents an incoming spike which causes an increase in
the membrane voltage of the neuron, leading to a threshold crossing at tθ. The subthreshold dynamics,
however, brings the voltage under threshold again at the next checkpoint. Since the test V (t + h) ⩾ θ yields
false, the outgoing spike at tθ is missed. The red dots indicate points where the values of state variables are
known.
so the test V (t + h) ⩾ θ is a sufficient but not a necessary condition for the occurrence of a threshold
crossing during the time interval ]t, t + h]. For future reference we call this test "standard test".
The conceptual question therefore remains whether a globally time-driven scheme can formulated such
that it detects every threshold crossing. Although Hanuschkin et al. (2010) argue that the loss of spikes of
the standard test is not of practical relevance in natural parameter regimes, the availability of a method
perfect by construction would free the researcher from inquietude and costly controls when faced with
previously unexplored neuron models or network architectures.
In this work we propose a new spike-detection method, which we term "lossless method" or "lossless test"
because it is a necessary and sufficient condition for threshold crossing to occur in a given time interval. The
method is based on state-space analysis and works with any neuron model with affine or linear subthreshold
dynamics. It consists of a system of inequalities – some linear, some non-linear in the state-space variables
– that together determine whether the initial state of a neuron will or will not reach threshold within the
time interval until the next checkpoint. The lossless method replaces the standard test (1) in the time-driven
and the hybrid scheme. Alone, it does not solve the problem of artificial synchronization the time-driven
scheme suffers from. Hence the method is most meaningful within the hybrid scheme. Thanks to its perfect
spike detection the lossless method can in fact be used to benchmark the hybrid scheme based on the
standard test; Hanuschkin et al. (2010) use the method of D'Haene et al. (2009) for this purpose.
In Section 2 we present the idea behind the lossless method for a general neuron model with an affine or
linear dynamics, and develop its mathematical construction. Parts of this construction must be addressed on
a case-by-case basis; therefore in Section 3 we provide a concrete implementation of the lossless method
for the leaky integrate-and-fire model with exponential synaptic currents (Fourcaud and Brunel, 2002a),
within the hybrid scheme. The method can be algorithmically expressed in different ways. We explore
two alternative cascades of inequalities and assess their costs in terms of time-to-completion relative to
each other and to the hybrid scheme based on the standard test (1). For the latter scheme we also assess
the number of missed spikes in commonly considered network regimes. The hybrid scheme based on the
lossless method delivers the desired exact implementation of the mathematical definition of the neuron
model without any further approximation up to floating point precision.
Frontiers
3
J. Krishnan et al.
Perfect spike detection via time reversal
Preliminary results have been published in abstract form (Kunkel et al., 2011; Krishnan et al., 2016). The
technology described in the present article will be made available with one of next major releases of the
open-source simulation software NEST. The conceptual and algorithmic work described here is a module
in the long-term collaborative project to provide the technology for neural systems simulations (Gewaltig
and Diesmann, 2007).
2 A TIME-REVERSED STATE-SPACE ANALYSIS
2.1 Idea: moving a surface backwards instead of a point forward
Let us summarize the problem mentioned in the previous section. We assume that a neuron's state evolves
according to three different dynamics: (a) an integrable subthreshold dynamic as long as the neuron's
membrane potential is below threshold and there are no changes in input currents; (b) discrete jumps in the
subthreshold dynamics at predetermined times, corresponding to incoming spikes or to sudden changes in
external currents; these can be formally incorporated into the subthreshold dynamics (a) via delta functions;
and (c) a "spike", i.e. an instantaneous jump of the membrane potential from threshold to a reset value,
as soon as the potential reaches the threshold value. The jump may be followed by a refractory period in
which the membrane potential remains constant at the reset value. Then the integrable dynamics takes
place again.
The advantage of the integrable dynamics is that the state of the neuron at a time t + h can be analytically
determined by that at time t; here h can be negative or positive. The evolution can thus be calculated in
discrete time steps, in particular in between times at which jumps (b) occur. The spike part of the dynamics,
however, forces us to check whether the membrane potential V reached a threshold value θ within the
timestep interval ]t, t + h]. We call this event threshold crossing (by "crossing" we also mean tangency). A
sufficient condition for threshold crossing is that the membrane potential be above threshold at the end
t + h of the time step: by continuity, it must have assumed the threshold value at some time in the interval
]t, t + h]. But this condition is not necessary: during the time step the potential may touch or surpass the
threshold value and then go below it again, as in Figure 1, an even number of times. Its value is then
below threshold at both ends of the time step, t and t + h. A test that only relies on the sufficient condition
V (t + h) ⩾ θ – the standard test (1) – can therefore miss some spikes, leading to an incorrect dynamics.
We need a test based on a necessary and sufficient condition.
A necessary and sufficient condition for threshold crossing is that the trajectory of the state during ]t, t+h]
intersect the hypersurface "membrane potential = threshold value". Translated into analytic geometry
this means finding the solutions of a system of parametric equations – one representing the threshold
hypersurface, the other the trajectory – and to test whether its solution set is empty (no threshold crossing) or
not (threshold crossing). This idea is illustrated in Figure 2A for a two-dimensional state space. This system
is usually transcendental and its solutions have to be found numerically. Unfortunately, numerical solutions
typically rely on bisection algorithms (Press et al., 2007, ch. 9), involving an increasingly finer timestepping
of the dynamics. This nullifies the advantage of having an integrable dynamics with coarse-grained time
steps.
The problem is that the test of intersection between trajectory and threshold tells us not only whether
a threshold crossing occurs, but also the time at which it does. The test's high computational cost partly
comes from delivering this additional information. This problem is avoided if we formulate a different
geometric test that tells us whether a threshold crossing occurs but does not deliver the crossing time.
This is a provisional file, not the final typeset article
4
J. Krishnan et al.
Perfect spike detection via time reversal
Mathematics often offer non-constructive proofs: "there exists a solution to the problem, but we do not
know what the solution is". It turns out that this ignorance is exactly what we need in our problem.
Instead of evolving the state of the neuron forwards in time, tracing a trajectory in state-space, and
checking if and when it crosses the threshold hypersurface, we can evolve the threshold hypersurface
backwards in time, sweeping a hypervolume, and check if and when it "crosses" the initial neuron state,
which is a point. In other words we are testing whether a point belongs to a particular state-space region.
The test for the intersection of a 1-dimensional curve with an (N − 1)-dimensional surface is replaced by
the test for the membership of a point in an N-dimensional volume. This idea is illustrated in Figure 2B for
a two-dimensional state-space.
Mathematically the latter test translates into a system of inequalities that the initial state at t must satisfy
if it does not cross threshold within ]t, t + h]. Even if transcendental functions appear in this system, we do
not need to find their roots: we only need to test whether the inequalities are satisfied by simply inserting
the value of the state-space variables in the functions.
The equations corresponding to these inequalities represent the piecewise-differentiable boundary between
the set of states that will cross the threshold within the timestep h – which we call spike region – and the
set of those that won't – which we call no-spike region. Finding these inequalities in explicit form is the
most important point of this method, and can be achieved with this heuristic procedure:
I. Find the hypervolume swept by the threshold, in parametric form. This is done by representing the
backpropagation of the threshold in state-space as a map between two manifolds: the product manifold
threshold × time, and the state-space manifold.
II. Find the boundary of the hypervolume, in parametric form. This is done by determining the placement of
the images of the boundaries of the product manifold and, most important, of the images of the critical
points of the map. The latter are the points at which the map becomes singular, thus mapping hypervolume
elements into hypersurface elements (Spivak 1999, ch. 2; Choquet-Bruhat et al. 1996, § II.A.1).
III. Transform the equations of the boundary above from parametric to implicit form. If this is not possible,
the boundary can still be approximated by a triangularization, as finely as we please.
IV. Finally, exclude the parts of the boundary that lie in the interior of the hypervolume (so technically not
a boundary). Some guidelines to achieve this can be given based on the the convex properties of the
hypervolume.
In the following analysis we mathematically develop the first two steps in a general way for any neuron
model with an affine dynamic. Let us point out that they can be generalized to other kinds of dynamics.
The procedure for the last two steps depends on the particular neuron model, so we can only give general
guidelines. Section 3 provides a concrete example of all steps. An advantage of this procedure is that it can
be geometrically explained and needs to be carried out only once for any given neuron model.
The advantage of this novel test is that its computational cost is distributed differently from the approach
where the trajectory of the initial point is followed. In hand-waving terms, in the trajectory-based test
the crossing time is immediately available from the parametrization of the orbit; the threshold-crossing
condition is represented by computationally costly inequalities, because the crossing time, when it exists,
can be read from them immediately. In the novel test the crossing time is a more complicated function of the
hypervolume parametrization; the threshold-crossing condition in this case is represented by inequalities
that are computationally less costly, because they do not explicitly deliver the crossing time. Knowledge of
the latter requires an additional computational cost. The novel test thus allows us to avoid this extra cost
when we check for the existence of a threshold-crossing; we only need to pay it if the crossing does occur.
Frontiers
5
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 2. Illustration of two exact methods to check whether an initial state (red dot) crosses the threshold
(blue horizontal line) during the evolution from t to t + h. Upper panel: the idea of a root-finding method is
to evolve the state forwards in time by a step ∆t = +h, and to check whether its trajectory (light-red curve)
intersects the threshold. Such a method informs us of where the intersection occurs on the threshold, and at
which time. Lower panel: the idea of the lossless method is to evolve the whole threshold "backwards
in time" from t = h to t = 0 by a step ∆t = −h and to check whether its trajectory, which is a volume
in state-space (light-blue region), contains the initial state, which is kept fixed. The threshold shifts and
rotates as it evolves, and the trajectories of its individual points are unknown: this method does not inform
us of when and where on the crossing it occurs, and is therefore computationally faster.
From a mathematical point of view the two approaches are equivalent. The state-space technique is
just a useful reformulation of the conventional trajectory view that is obtained by suitable mathematical
manipulations. The conceptual device of propagating the threshold backwards in time is useful because it
performs those manipulations in a transparent manner and gives them an intuitive dynamical meaning.
This is a provisional file, not the final typeset article
6
θVt=0∆t=+ht=hIθVt=0∆t=−ht=h J. Krishnan et al.
Perfect spike detection via time reversal
2.2 Mathematical preliminaries
The final equations to be obtained, eq. (23), can be derived by concepts from vector analysis, Cartesian
geometry, and functional analysis; but the derivation is lengthy. To shorten it we use concepts and
terminology from affine spaces (Coxeter, 1969; Rockafellar, 1972; Nomizu and Sasaki, 1994; Artin, 1955;
Porta Mana, 2011), and differential manifolds (Choquet-Bruhat et al., 1996; Burke, 1987, 1995; Simon
et al., 1992; Nomizu and Sasaki, 1994; Marsden and Ratiu, 2007; Bossavit, 1991, 2002; Schouten, 1989;
de Rham, 1984; Dodson and Poston, 1991; Ramanan, 2005).
The state-space S of a neuron has a natural vector-space structure (an affine-space structure would also
suffice), inherited from the physical quantities that define it: the membrane potential V and other N − 1
physical quantities I whose exact number and definition depend on the specific neuron model (e.g., I could
represent currents or additional voltages, for example of different compartments). For our purpose it is
useful to consider S as an N-dimensional differential manifold: its points {s} are the neuron states, and the
quantities (I, V ) are coordinates V : S → R and I : S → RN−1 e.g., the membrane potential of a state s
is V (s). These coordinates respect the vector structure of the state-space, i.e. V (s1 + s2) = V (s1) + V (s2)
and likewise for I.
⊺ being
Every hyperplane in the state-space is defined by an affine equation k
the normal to the hyperplane, and −κ being the affine term. The inequality k
s > −κ defines one of the
two half-spaces delimited by the hyperplane. The threshold hyperplane is especially important: it is the set
of states s whose membrane potential has the threshold value:
s = −κ, the covector k
⊺
⊺
Its equation k
⊺
s = κ in coordinates (I, V ) has coefficients
V (s) = θ.
⊺
⊺
= (0
s = κ on threshold,
k
, 1),
⊺
k
κ = θ,
s < κ below threshold.
⊺
k
An affine transformation of the state-space onto itself,
s (cid:55)→ Ms + m,
where M is a linear transformation and m a state, maps each hypersurface and half-space k
hypersurface and half-space k′⊺
s ⩾ κ′ with
k′⊺
⊺M−1,
κ′ = κ + k
⊺M−1m
= k
⊺ shows why it is a covector rather than a vector).
(the transformation of the normal k
(2)
(3)
(4)
s ⩾ κ to a
⊺
(5)
We now show that the integrable part of the neuron dyamics within a finite time step h is an affine
transform. Similarly, the integrable part of of the neuron dynamics we consider an affine evolution
determined by the equation
(6)
(7)
7
In a time interval h, this dynamics propagates an initial state s0 at time t into the final state
s(t) = As(t) + q.
s(t + h) = ehAs(t) +(cid:0)ehA − 1(cid:1)A−1q,
s(t) = s0.
Frontiers
J. Krishnan et al.
Perfect spike detection via time reversal
(cid:18)B d
(cid:19)
This, for each h, is an affine transformation of the form (4). In coordinates (I, V ) the linear operator A and
vector q have the block form
A =
(8)
where B is an (N − 1, N − 1) matrix, α and β numbers, and the dimensionalities of the rectangular matrices
⊺, d follow accordingly.
c
2.3 Derivation of the threshold-crossing condition
q =
α
β
c
⊺
,
(cid:18)r
(cid:19)
,
Let us mathematically summarize the first threshold-crossing test discussed in Section 2.1. We said that
the state evolution (7) can be efficiently used in a time-step scheme in numerical simulations, but we
need to test whether a threshold crossing occurred at some time t ∈ ]t, t + h]. A necessary and sufficient
condition would be the existence of solutions of the trascendental equation in t
t ∈ ]0, h],
V(cid:2)etAs(t) +(cid:0)etA − 1(cid:1)A−1q(cid:3) = θ,
(9)
which corresponds to the intersection of the trajectory (7) and the threshold hyperplane (2); but it is a costly
condition to test.
We now develop the second kind of threshold-crossing test discussed in Section 2.1, according to the
steps I–IV. Steps I and II are performed in full generality for an affine dynamics. Steps III and IV have to
be solved on a case-by-case basis, so their analysis below is only a guideline; a concrete example on how
to perform them is given in Section 3.3.3 and Section 3.3.4.
2.3.1 Hypervolume in parametric form
Consider the threshold hyperplane V (s) = θ as an (N − 1)-dimensional manifold with coordinates x. It
is embedded in state-space via the map
x (cid:55)→ (I, V ) = (x, θ).
(10)
Each state (x, θ) on the threshold, when propagated backwards in time for an interval h, traces a curve in
state-space (the yellow lines of Figure 3).
The union of these curves is an N-dimensional product manifold, called the extrusion (Bossavit, 2003) of
the threshold hyperplane. We can use coordinates (x, t) ∈ (RN−1 × [0, h]) on this manifold. Its mapping
into state-space is given, with the help of eq. (7), by
E : (x, t) (cid:55)→ e−tA
(cid:19)
(cid:18)x
θ
+(cid:0)e−tA − 1(cid:1)A−1
(cid:18)r
(cid:19)
β
,
(11)
where t > 0 is the direction of the past. This map is analytic, but generally not an embedding because it
can have self-intersections. We will see the significance of this in Section 2.4. It is not an immersion either
because it can have singular points; these will be especially important for us because they constitute part of
the boundary between spike and no-spike regions. See Figure 3 for a two-dimensional example.
For fixed t, the map x (cid:55)→ E(x, t) is affine, and its image is a hyperplane representing the states on the
threshold propagated backwards in time for an interval t. Combining it with eqs (3)–(5) we find that the
This is a provisional file, not the final typeset article
8
V = θ,
(cid:0)0
⊺
1(cid:1) ehA
(cid:18)I
(cid:19)
V
= θ +(cid:0)0
⊺
TE(x, t) = (∂xE, ∂tE) = −e−tA
1(cid:1)(cid:0)1 − ehA(cid:1)A−1
(cid:18) I
Bx + dθ + r
⊺
x + αθ + µ
c
⊺
0
(cid:18)r
(cid:19)
β
.
(cid:19)
.
(13)
(14)
(15)
J. Krishnan et al.
Perfect spike detection via time reversal
backpropagated threshold has t-dependent normal and affine terms
⊺
t =(cid:0)0
κt = θ +(cid:0)0
1(cid:1) etA,
1(cid:1)(cid:0)1 − etA(cid:1)A−1
⊺
⊺
k
(cid:19)
.
(12)
(cid:18)r
β
⊺
t s < κt determines the backpropagated half-space, which is below threshold.
The inequality k
2.3.2 Hypervolume boundary in parametric form
We must now find the boundary of the image of the map E. The latter is a closed set, being the image
of a closed set under a continuous map; its boundary must therefore be the image of some points of the
domain. Such points must either lie on the boundaries of the domain, RN−1 × {0} and RN−1 × {h}, or
be critical points of E, or both, because E is differentiable. See the example of Figure 3.
The images of the boundary are easily found from (11): one (image of t = 0) is the threshold hyperplane,
⊺
hs = κh, with coefficients given by (12). Explicitly, in coordinates
the other (t = h) is the hyperplane k
(I, V ),
Let us find the image of the critical points of E. The tangent map of E at a point (x, t) is
The tangent map is also denoted E∗ in some differential-geometry texts.
Its determinant is the inverse ratio between a volume element at that point and its image in state-space
(the sign determines their relative orientation). Hence this ratio vanishes at points where volume elements
are mapped onto area elements (in other words, N linearly independent vectors in the domain are mapped
onto N linearly dependent vectors), which is a feature of the boundary. See the two-dimensional example
of Figure 3.
Let us look for points where det TE(x, t) = 0. In eq. (15), the determinant of the exponential never
vanishes, so we only have to consider the determinant of the matrix on the right. This is easily calculated
by Laplace expansion along the last row, whose elements all vanish except the last. The cofactor of the last
element is det I (modulo a sign). Hence
det TE(x, t) = 0 ⇐⇒ c
⊺
x + αθ + β = 0
and the coordinates (x, t) of critical points satisfy
⊺
c
x + αθ + β = 0,
0 ⩽ t ⩽ h.
Frontiers
(16)
(17)
9
J. Krishnan et al.
Perfect spike detection via time reversal
This equation says that one of the coordinates x has an affine dependence on the remaining ones; let us call
these y. For example, if cN−1 = 0, the equation above has the parametric solution
x1 = y1,
xN−2 = yN−2,
xN−1 = −c1y1 + ··· + cN−2yN−2 + β + αθ
. . . ,
cN−1
.
(18)
Denote this affine dependence by x(y). By taking the derivative of eq. (17) with respect to y we have
⊺
c
∂yx = 0
⊺
,
(19)
a property we will use later.
The locus of critical points in state-space is then given parametrically by a map Γ , found by substitution
of eq. (18) in (11):
Γ : (y, t) (cid:55)→ e−tA
(cid:18)x(y)
(cid:19)
θ
+(cid:0)e−tA − 1(cid:1)A−1
(cid:18)r
(cid:19)
β
,
0 < t < h.
(20)
This locus has four important interrelated features:
First: from the form of eqs (17) and (20), the locus of critical points is flat along N − 2 dimensions
– corresponding to a fixed value of the coordinate t – and is curved normally to the direction t: it is an
(N − 2)-ruled surface.
Second: the points on the locus corresponding to fixed t belong to a hyperplane with coefficients (12),
as is easily checked by substitution. In other words, the locus of critical points is the envelope of the
⊺
backpropagated threshold hyperplanes k
t s = κt, eq. (12), at different times t, and its tangent hyperplanes
have normals k
⊺
t given by (12).
Third: considering the feature above for t = 0, the locus of critical points is tangent to the threshold
hyperplane.
Fourth: comparing the dynamics (6), the map for the threshold hyperplane (10), and the critical-point
condition (17), we notice that the latter is also the condition for the trajectory (7) of a state s0 = (x, θ) on
the threshold hyperplane to have an extremum in the membrane potential V . Hence, the locus of critical
points is the trajectory of the intersection between the threshold and the V -nullcline.
In view of the second property above, let us call the locus of critical points envelope, for brevity.
2.3.3 Hypervolume boundary in implicit form
The next step is the elimination of the parameters (y, t) to express the envelope Γ as one or more implicit
equations χ(s) = 0 defined on domains Dχ for the state-space coordinates s. This step can involve a
transcendental equation, therefore we cannot give a general solution for it. In this regard, a criticism might
be raised, which we immediately address: it looks like we have already taken great care in avoiding to solve
eq. (9), only to face another difficult transcendental equation? The current problem is more manageable
for three reasons. First, we observe that the equations needed to re-express the curved surface in explicit
form are generally easier to handle than (9). Section 3.3.3 gives a concrete example. Second, the problem
of finding the explicit form of the curved surface must be solved only once, whereas the solution of the
original equation (9) has to be found for each initial state s0. Third, the equation for the curved surface
This is a provisional file, not the final typeset article
10
J. Krishnan et al.
Perfect spike detection via time reversal
is easier to approximate numerically than the original one (9), again because the approximations do not
depend on an initial state.
Suppose we have found a function χ such that χ(s) = 0 is the envelope:
{s χ(s) = 0, s ∈ Dχ} = {Γ (y, t) y ∈ RN−2, 0 ⩽ t ⩽ h}.
(21)
The condition above leaves χ completely undetermined (apart from smoothness requirements) outside of
the envelope. Therefore, if possible, it is useful to extend the condition as follows. Since χ(s) = 0 is the
⊺
envelope of the backpropagated threshold hyperplanes k
t s = κt, eq. (12), at different times t, it must
⊺
t s − κt < 0, for each t, correspond to the backpropagated
be tangent to each of them. The inequalities k
below-threshold half-plane, and it is useful to choose χ in such a way that the inequality χ(s) < 0,
s ∈ Dχ, determines the side corresponding to the intersection of these half-spaces. An example is given
⊺
in Section 3.3.3. The differential dχ then is a positive multiple of k
t are collinear
because of the tangency condition).
⊺
t (note that dχ and k
For the following, let us assume that χ has been chosen this way.
2.3.4 Boundary intersections and final system
If a state does not cross the threshold hyperplane at any time t ∈ [0, h], then it must, by definition, belong
for every t to the image of the half-space that lies below threshold, when this half-space is backpropagated
by the time t. Mathematically, this is simply the statement of the equivalence
⊺
(Ms + m) < κ ⇔ (k
⊺M)s < (κ − k
⊺
k
m)
(22)
established in Section 2.2, where M and m are the coefficients of the affine evolution by time t.
By construction in the previous subsection, the inequality χ(I, V ) < 0 defines the region of intersection
of all such backpropagated half-spaces, bounded by the envelope χ(I, V ) = 0. We just need to join to it
⊺
hs < κh, discussed
the condition for the boundaries corresponding to the times t = 0 and t = h, V < θ, k
in Section 2.3.1.
The states that do not cross the threshold during the time interval [0, h] belong therefore to the no-spike
region defined, in coordinates, by
s0 = (I, V ) ∈ no-spike region ⇔
V < θ
(cid:0)0
⊺
1(cid:1) ehA
(cid:18)I
(cid:19)
V
< θ +(cid:0)0
⊺
1(cid:1)(cid:0)1 − ehA(cid:1)A−1
(cid:18)r
(cid:19)
β
and
Frontiers
(23a)
(23b)
11
J. Krishnan et al.
Perfect spike detection via time reversal
and
χ(I, V ) < 0,
(I, V ) ∈ Dχ.
(23c)
Some inequalities in this system may turn out to be redundant, i.e. automatically satisfied if the remaining
ones are, and can thus be dropped. Section 3.3.3 illustrates such a redundancy.
2.4 The region of missed spikes
In the previous section we mentioned that the N-dimensional product manifold formed by the (N − 1)-
dimensional threshold hyperplane and the time interval [0, h] presents self-intersections when mapped to
the state-space. Two-dimensional examples of such a region are shown in Figure 3, lower panel, and in
Figure 5, the region called S2.
A state s in the self-intersection region corresponds to two or more different coordinates (x, t):
s = E(x, t) = E(x′, t′),
t = t′.
(24)
Note that x = x′, t = t′ is impossible for an affine transformation, since the first column of (15) can never
vanish. The condition above simply means that s crosses the threshold at x after an interval t and at x′
after an interval t′. By continuity of the dynamics (6), one of the two must be a crossing from above, and
one from below: this is exactly the scenario of double threshold crossing illustrated in Figure 1.
Suppose we have a probability distribution for the initial states, for example one that is invariant under
the dynamics (6). The probability of the self-intersection region P (X) is the probability that the initial
state will lead to a double-crossing of the threshold, and therefore be missed. This fact will be used in
Section 3.5.2 to estimate the number of spikes missed.
2.5 Convexity, approximations, optimization
The affine dynamics (7) preserves affine combinations of solutions – and therefore convex combinations
as well. If s1, s2 are two arbitrary initial states in the no-spike region, then their propagated states also
satisfy V [s1(t)] < θ and V [s2(t)] < θ when 0 ⩽ t ⩽ h; and also the propagation of their convex
combination λs1 + (1 − λ)s2, 0 < λ < 1 satisfies
V [λs1(t) + (1 − λ)s2(t)] < θ,
0 < λ < 1,
(25)
i.e. it lies in the no-spike region. This proves that the no-spike region is convex.
Convexity is important for approximations. If transforming the parametric equation for the envelope (20)
into an implicit form (21) turns out to be analytically impossible, we can numerically find some points
on the envelope and then approximate the latter by simplices constructed on these points, as finely as
needed. In other words we can triangularize the envelope. Owing to convexity, the triangularization can be
done completely on the side of the no-spike region (the corners of the simplices touch the envelope), or
completely from that of the spike region (the barycentres of the simplices touch the envelope). This way
we can formulate a test with no false positives, or one with no false negatives, or both.
This is a provisional file, not the final typeset article
12
J. Krishnan et al.
Perfect spike detection via time reversal
The envelope is moreover a ruled surface, as shown in Section 2.3.2. The triangulation on the side of the
no-spike region can therefore be conveniently chosen in such a way that one face of each simplex fully lies
on the envelope.
3 IMPLEMENTATION EXAMPLE: LEAKY INTEGRATE-AND-FIRE NEURON WITH
EXPONENTIALLY DECAYING POST-SYNAPTIC CURRENTS
3.1 The example model
In the previous section we mathematically developed the idea of propagating the threshold backward in
time in order to check whether a threshold-crossing occurs in a time-stepped dynamics. The derivation,
valid for an affine subthreshold dynamics, is general and therefore also quite abstract; moreover, it involves
a couple of mathematical steps (III and IV in Section 2.1) for which no general formulae can be given.
To explain the idea in more concreteness and to give an example of how to face all its steps, we now apply
the scheme to a simple but relevant model with a 2-dimensional state-space: the leaky integrate-and-fire
neuron with exponentially decaying post-synaptic currents. This model has a homogeneous linear dynamics
on a 3-dimensional state-space (Rotter and Diesmann, 1999), where the third coordinate is the input current.
If this current is constant, the dynamics can be rewritten as a 2-dimensional affine one.
Leaky integrate-and-fire models, despite their simplicity, approximate the behavior of real neurons with
high accuracy (Rauch et al., 2003). The model with exponential synaptic currents captures important
properties of real neurons: The postsynaptic potential has a finite rise and decay time and the membrane
potential is a continuous function of time. Continuity avoids artificial synchronization, present in simpler
models. Moreover, the model is to some extent analytically tractable. For short synaptic time constants, the
mean firing rate (Fourcaud and Brunel, 2002a) as well as the linear response to small inputs (Schücker
et al., 2015) can be obtained analytically.
3.2 Mathematical preliminaries: terms in block form
The example model has a 2-dimensional state-space for a single neuron, defined by the post-synaptic
current I and the membrane potential V , which are also our coordinates. Its subthreshold interspike
dynamics (a) in Section 2 is affine:
I(t),
I(t) = − 1
τs
I(t) − 1
τ
V (t) =
1
C
(cid:18)− 1
(cid:19)
(cid:18) I
V
=
V (t) +
Ie,
1
C
(cid:19)
(cid:19)(cid:18) I
V
+
(cid:19)
,
(cid:18) 0
1
C Ie
0
τs
1
C − 1
τ
or in matrix form
d
dt
(26a)
(26b)
where C is the membrane capacitance, τ is the membrane time constant, I is the synaptic input current and
Ie is the external input current. The membrane potential V is subject to dissipation with time constant τ
and integrates the post-synaptic current I. The latter decays exponentially with time constant τs. Typical
values of the parameters are τ = 10 ms, C = 250 pF, τs = 2 ms, and the threshold θ = 20 mV.
Frontiers
13
J. Krishnan et al.
Perfect spike detection via time reversal
Incoming spikes are incorporated in the equation for the current as Dirac deltas, and the external current Ie
has jump discontinuities in time. In the timestepped evolution, such discontinuous events are implemented
as instantaneous changes in the initial state s0 at each timestep. Hence we do not need to consider them
explicitly in the equations above (Rotter and Diesmann, 1999).
In terms of the block form of Section 2.2 we have
The exponential of A is
exp(−tA) =
B = ( − 1
⊺
= ( 1
C ),
c
τs ), d = ( 0 ),
α = −1/τ,
r = ( 0 ),
Ie.
1
C
,
β =
0
t
τ
e
(27)
(28)
(cid:0)e
τs(cid:1)τ τs
t
e
τs
τ −e
t
C(τ−τs)
t
which determines the evolution of the neuron state s0 by an affine map as in eq. (7).
Since the state space is 2-dimensional, the "hyperplanes" and "hypersurfaces" of Section 2 are straight
lines and curves. In particular, the threshold hyperplane is a line; when propagated by a time t it maps onto
a line with covector and affine term given by eq. (12), explicitly
(cid:19)
⊺
t =
k
(cid:18)(cid:0)e− t
τs(cid:1)τ τs
κt = θ − Ieτ(cid:0)1 − e− t
τ(cid:1)
− t
τ −e
C(τ−τs)
C
e− t
τ
.
,
(29)
e− t
τ V +
Ieτ
C
0 ⩽ t ⩽ h,
(cid:18)
For this model, testing for threshold-crossing by checking the intersection of a propagated state and the
threshold line means solving the following transcendental system in t, given the initial state s0 = (I, V ):
(1 − e− t
τ ) + I
τ τs
C
e− t
τ − e− t
τ − τs
τs
= θ,
(30)
for which we cannot find a solution in analytic form; we would have to resort to bisection algorithms.
3.3 The threshold-crossing condition
3.3.1 Hypervolume in parametric form
The product manifold "threshold line × time interval" is in this case 2-dimensional, with coordinates
(x, t). Here x is the value of the current at threshold crossing and t the corresponding time point. Its
mapping E, eq. (11), to the state-space is
E : (x, t) (cid:55)→ (I, V ) =
e− t
τs x,
e− t
τ θ +
Ieτ
C
(1 − e− t
τ ) + x
τ τs
C
e− t
τ − e− t
τ − τs
τs
.
(31)
(cid:19)
This map is shown in Figure 3: the x isocurves are yellow and in the lower panel they represent the
trajectories of states terminating on the threshold; the t isolines are blue and represent the threshold line
This is a provisional file, not the final typeset article
14
J. Krishnan et al.
Perfect spike detection via time reversal
propagated at different times. Each area element dx ∧ dt in the domain – the small rectangles in the upper
plane – is mapped into an area element dI ∧ dV in the image. Note how these area elements are rotated
and sheared. The thicker green curve is the set of singular points where the determinant of the tangent map
vanishes: det(TE) = 0. Such points are singular because around them the images of the area elements get
flattened to one dimension. In 3.4 we discuss the region of self-intersection bounded by the thick light cyan
line, the thick dark blue line, and the thick green curved line.
3.3.2 Hypervolume boundary in parametric form
The boundaries of the image of the map E must be, as explained in Section 2.3.3, a subset of the images
of the boundaries of the domain, R × {0} and R × {h}, and of the envelope.
The images of the boundaries, with general equations (13) and (14), in terms of (I, V ) follow
The set of critical points of the map E is in this case a 1-dimensional curve, given in parametric form by
V = θ,
V = e
h
τ θ +
Ieτ
C
(cid:18)
(cid:16) θC
τ
Γ : t (cid:55)→
t
τs
e
− Ie
,
τ(cid:1) +
h
Iτ
C
(cid:0)1 − e
(cid:17)
τ e
τs
e
τse− h
(cid:16)θC
h
τs
h
τ − e
τ − τs
(cid:17)
t
τs
t
τ − τse
τ − τs
τ
C
− Ie
τ
+
τ
C
Ie
;
(cid:19)
.
(32)
(33)
the coordinate y of the general form (20) do not exist in this case, because the threshold is 1-dimensional.
Figure 3 visualizes the set by the green curve.
3.3.3 Hypervolume boundary in implicit form
The next step in our procedure is to convert the parametric equation (34) of the envelope into an explicit
or implicit equation for the coordinates (I, V ). As Section 2.3.3 does not provide a general algorithm,
below we illustrate the process using our example model.
By equating the first component of eq. (34) to the coordinate I and solving for t, we find
t = −τs ln
− Ie
(cid:104)(cid:16) θC
(cid:16)θC
τ
,
/I
(cid:105)
(cid:17)
(cid:17) ⩽ 1,
subject to the condition for I
e− h
required for 0 ⩽ t ⩽ h and a real logarithm.
Substituting eq. (35a) into the V coordinate of eq. (34) we find
τs ⩽ 1
I
− Ie
τ
τ(cid:2)(cid:0) θC
τ − Ie
(cid:1)/I(cid:3)1− τs
τ − τs
τ − τs
V =
τ Ie
C
+
τ I
C
Frontiers
(34)
(35a)
(35b)
(36)
15
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 3. Example of map E : (x, t) (cid:55)→ (I, V ), eq. (11), in two dimensions. Upper panel: the abstract
manifold with coordinates (x, t) corresponding to the values of current and time at threshold crossing.
Lower panel: the image of the map in state space. Thick green curve corresponds to the set of singular
points where det(T E) = 0, eq. (9). Yellow lines, constant x, are trajectories of states ending on the
threshold. Violet lines, constant t, are snapshots of the threshold moving "backwards in time". The thicker
violet and blue lines correspond to the boundaries t = 0 and t = −h. For t = 0 we have I = x, V = θ.
subject to the condition (35b). Let us analyse this equation, in view of its extension to an inequality of the
form χ(I, V ) < 0 as required in Section 2.3.3. First, we observe that
τ e
t
τs
t
τ − τse
τ − τs
⩽ 1
for
0 ⩽ t ⩽ h.
This is a provisional file, not the final typeset article
(37)
16
x−h0tIθV J. Krishnan et al.
Perfect spike detection via time reversal
The inequality can be proven by studying the derivative of the fraction with respect to t. The derivative is
always negative in the range above and the only maximum of the fraction is the value unity assumed at
t = 0.
Inspection of equation (36) and of its parametric form (34) shows that we must consider three cases:
Ie ⋚ θC/τ, i.e. whether the external current is smaller or larger than the rheobase current
this is the current necessary to reach threshold in an infinite time starting from any state with I ⩽ 0.
• If Ie < Iθ, then I is restricted to
Iθ = θC/τ ;
0 < Iθ − Ie ⩽ I ⩽ e
h
τs (Iθ − Ie).
(38)
(39)
In this case, using the inequality (37), the V component of the envelope (34) is always smaller than the
threshold:
τ(cid:2)(Iθ − Ie)/I(cid:3)1− τs
τ − τs
τ − τs
• If Ie > Iθ, then I must be negative and restricted to
+
V ≡ τ Ie
C
τ I
C
⩽ θ when Ie < Iθ.
h
τs (Iθ − Ie) ⩽ I ⩽ Iθ − Ie < 0.
e
(40)
(41)
τ(cid:2)(Iθ − Ie)/I(cid:3)1− τs
τ − τs
In this case, using the inequality (37), the V component of the envelope (34) is always larger than the
threshold:
V ≡ τ Ie
C
τ I
C
+
(42)
• If Ie = Iθ, the envelope degenerates to a point: Γ (t) = (0, θ), which is the limit point reached in infinite
time from any initial point in the no-spike region; this is the geometric interpretation of the equality of
external and rheobase currents.
τ − τs
⩾ θ when Ie > Iθ.
The function representing the envelope, eq. (21) of Section 2.3.3, has in this case the explicit form
χ(I, V ) := V − τ Ie
C
− τ I
C
τ(cid:2)(Iθ − Ie)/I(cid:3)1− τs
τ − τs
τ − τs
,
with (I, V ) ∈ Dχ :=
(cid:104)
Iθ − Ie, e
(cid:105) × R.
h
τs (Iθ − Ie)
(43)
If we calculate its differential, as discussed in Section 2.3.3, we find that the latter is a positive multiple of
the differential of the backpropagated threshold for each t. This is true in each of the three cases above.
Consequently, the inequality χ(I, V ) < 0, (I, V ) ∈ Dχ always includes the no-spike region.
In summary, we arrive at an analytic, implicit equation for the curved boundary (43). The expression
is a transcendental function in I, owing to the generally irrational exponent 1 − τs/τ, but it is used in an
inequality, hence we do not need to find its roots. This is in contrast to the original transcendental equation
for the threshold-crossing condition (30), which requires a bisection algorithm to find its solution.
Frontiers
17
J. Krishnan et al.
Perfect spike detection via time reversal
3.3.4 Boundary intersections and final system
We can now assemble the system of inequalities defining the no-spike region consisting of the
boundaries (32) and (33), and the envelope (43).
With some rearrangements, simplifications, and the introduction of two new functions f, b, the condition
reads:
s0 = (I, V ) ∈ no-spike region ⇔
V < θ
and
and
V < fh,Ie(I) := e
h
τ θ +
τ Ie
C
h
(cid:0)1 − e
τ(cid:1) + τse− h
τ(cid:2)(Iθ − Ie)/I(cid:3)1− τs
τs
h
τs
h
τ − e
τ − τs
e
τ I
C
V < bIe(I) :=
τ Ie
C
+
τ I
C
τ − τs
τ − τs
if I ∈(cid:104)
Iθ − Ie, e
(cid:105)
h
τs (Iθ − Ie)
(44a)
(44b)
(44c)
Figure 4 illustrates the system for the two cases Ie < Iθ and Ie > Iθ. In the former case all three
inequalities are necessary; in the latter, as well as for Ie = Iθ, the last inequality is automatically enforced
by the first because its right-hand side is larger than the threshold θ (see eq. (42)).
3.4 The region of missed spikes
In Figure 3, lower panel, the trajectories of several states during a timestep h and ending on the threshold
line are represented by yellow curves. In that figure we can identify a region were such trajectories self-
intersect: it is bounded by a segment of the thick light blue line, a segment of the thick dark blue line, and a
portion of the thick green curved line. Trajectories with initial states in this region must therefore cross
the threshold twice during the interval ]t, t + h]. As explained in Section 2.4, their threshold-crossing is
not detected by the sole condition V [s(t + h)] ⩾ θ. All states in this region thus generate spikes that are
missed by the standard test (1).
This region is crucial for the comparison of the performances of schemes implementing the present
lossless method and schemes relying on the standard test (1). This comparison is quantitatively made in
Section 3.5.3.
This is a provisional file, not the final typeset article
18
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 4. System of inequalities, eq. (44), determining the no-spike region. The colored areas represent
the complementary inequalities of the system (44), so the solution of that system is the white area. The
red region delimited by the horizontal line corresponds to the first equation of the system, the blue region
delimited by the inclined line to the second, and the yellow region delimited by the curved line to the third.
Upper panel: case Ie < Iθ, all inequalities necessary. Lower panel: case Ie > Iθ, the third inequality is
redundant.
Frontiers
19
0Iθ−Ieeh/τs(Iθ−Ie)θVIe<IθIθ−Ieeh/τs(Iθ−Ie)0IθVIe>Iθ J. Krishnan et al.
Perfect spike detection via time reversal
3.5 Optimization, time performance, and accuracy
3.5.1 Numerical implementation and optimization
In the last section we arrived at the system of three inequalities (44) that determines whether or not
the current state (V, I) will cross the threshold within the timestep h. The state will cross the threshold
if the system is not satisfied, and it will not cross the threshold if the system is satisfied. This system of
inequalities constitutes the lossless method in the present model. The system requires the current timestep
h, state (V, I), and external electric current Ie as inputs (cf. Section 2). The timestep h is the minimum
between the global timestep and the time interval up to the next input coming from other neurons, hence it
can differ every time the test is called. The external electric current Ie may also vary, stepwise, during the
simulation, hence it may also be distinct at each call of the test.
If the system of inequalities is satisfied, thus predicting the absence of a spike within a timestep h, then
the state of the neuron is evolved by applying the propagator (7) with (27)–(28), leading to a new state, and
the procedure starts again. If the system is not satisfied, thus predicting the occurrence of a spike within h,
it is then necessary to compute the time tθ at which the threshold is crossed. This calculation, explained in
Appendix A, is made by interpolation between the current state and time, and the state and time tmax at
which the membrane potential would reach its maximum if allowed to increase above threshold. Once tθ is
calculated, a spike is emitted, communicated to the postsynaptic neurons, and the membrane potential is
reset to Vreset for a refractory period. When this refractory period is over the procedure starts again.
In the evolution loop just described, the membrane potential is reset to a value below threshold as soon
as it crosses the latter. Thus no initial state can have V ⩾ θ. This means that the first inequality (44a) in
the system is always satisfied and can be dropped. Only inequalities (44b) and (44c) have to be assessed,
leading to the reduced system
(45)
(cid:40) V < fh,Ie(I),
V < bIe(I).
We first discuss the case Ie < Iθ ≡ θC/τ.
The geometric meaning of the inequalities above is illustrated in Figure 5. The figure shows four regions:
NS1, NS2, S1, S2. The no-spike region is the union of NS1 and NS2, the spike region the union of S1 and S2.
In the figure they are separated by a thick line, partly curved and black, partly straight and blue. Subregion
S2 is particularly important: it is the region of missed spikes discussed in Section 3.4, corresponding to
the self-intersection region of Figure 3, lower panel. It contains those states that lead to spikes missed by
schemes that rely on the standard test (1).
Region S1 is separated from S2 by a dashed blue line, and from NS1 by a continuous blue line, the
continuation of the dashed one. This partly dashed, partly continuous blue line corresponds to the equation
V = fh,Ie(I). Hence if the inequality V < fh,Ie(I) is not satisfied then the initial state is in region S2 or on
its blue boundary, and there will be a spike. If the inequality is satisfied the state could be in S2 – spike – or
NS1 ∪ NS2 – no spike; an undetermined case. This inequality requires modest computational costs because
it is linear in I and Ie and involves exponentials of h. Regions S2 and NS2 are separated by a black curve:
. Hence if the
inequality V < bIe(I) is not satisfied the initial state is in S2 or on its boundary, and there will be a spike.
If the inequality is satisfied the initial state is either in NS2, or in NS1 with Iθ − Ie < I < e
τs (Iθ − Ie), and
no spike will occur.
this is the envelope, corresponding to the equation V = bIe(I) for I ∈(cid:104)
Iθ − Ie, e
h
τs (Iθ − Ie)
(cid:105)
h
This is a provisional file, not the final typeset article
20
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 5. State-space subregions formed by the intersections of the reduced inequalities (45): V < fh,Ie(I)
(straight blue line, partly dashed partly continuous) and V < bh(I) (black curve), and by the auxiliary
inequality (47): V < gh,Ie(I) (dot-dashed straight purple line). The spike region is S1 ∪ S2, the no-spike
region is NS1 ∪ NS2. The subregion S2 contains all states that emit spikes undetected by the standard
test (1); they are detected by the lossless method.
The computationally most expensive inequality is V < bIe(I) because it involves irrational powers of I
and Ie. It is advisable to avoid its direct computation as often as possible by pre-testing a linear inequality.
In Section 2.5 we discussed how such a pre-test is indeed possible thanks to the convexity of the no-spike
region. There, we argued that the curved envelope can be approximated by triangular hypersurfaces, which
simply reduce to one straight segment in the present two-dimensional case: this is the dot-dashed red line
in Figure 5, separating NS1 and NS2. This line has equation V = gh,Ie(I) with
gh,Ie(I) := θ +
h
τs
τse
τ − τs
τ
C
I + e
h
τ
(Iθ − Ie),
τ
C
and the corresponding inequality
has the same computational costs as V < fh,Ie(I).
V < gh,Ie(I)
(46)
(47)
If the auxiliary inequality V < gh,Ie(I) is satisfied, the initial state is in NS1 and V < bIe(I) is also
satisfied. It is therefore convenient to test the auxiliary inequality before the computationally costly one,
which can be discarded if the test is positive. Figure 5 suggests that this test might be positive for the
Frontiers
21
Iθ−Ieeh/τs(Iθ−Ie)IθVNS1S1NS2S2fh,Ie(I)gh,Ie(I)b(I) J. Krishnan et al.
Perfect spike detection via time reversal
majority of initial states because region NS1 is much wider than NS2. This possibility would be very
advantageous, but we now argue that it should be verified by a dynamical analysis.
The average time cost of the algorithm in a long simulation is given by(cid:80)
The system (45) can be translated into a computational algorithm in several different ways, depending
on the order of evaluation of its two inequalities and of the auxiliary inequality (47). In simplified terms,
such an algorithm consists in a sequence of tests – variously implemented as if, and, or constructs –
for finding the initial state in space-time regions R1, R2, and so on. The order of these tests is important.
i pici, where pi is the frequency
with which states are found in region Ri, which we call "occupation frequency", and ci is the cumulative
time cost of the test for region Ri. This time cost ci is cumulative in the sense that all tests up to the
(i − 1)th must have been performed, with false outcomes, to arrive at the test for Ri. The efficiency of
an algorithm therefore depends on the mathematical form of the inequalities defining a region and on the
occupation frequencies of the regions, determined by the dynamics. These two factors can be extrapolated
by a theoretical analysis, or more practically measured by running long test simulations with typical
network setups corresponding to the cases one is interested in.
We now try to determine the most efficient algorithm for the present case. Region S1 is the least costly,
because bounded by one line and therefore involving one inequality linear in I; then region NS1, bounded
by two lines involving two linear inequalities; and finally regions NS2 and S2, bounded by the curve that
involves rational exponentiation. For this example model, we tried different orderings but show here only
two possible extreme cases to illustrate that there is no significant difference in the computational cost.
Algorithm 1
Algorithm 2
bool is_spike(h):
pre-compute gh,Ie(I)
if V ⩽ fh,Ie(I) and V < gh,Ie(I) then
else if V ⩾ gh,Ie(I) then
else if V ⩾ bIe(I) then
return false
return true
return true
else
return false
bool is_spike(h):
if V ⩾ gh,Ie(I) or
[V ⩾ fh,Ie(I) and V ⩾ bIe(I)] then
return true
else
return false
Algorithm 1 is based on the assumption that the occupation frequency of a subregion is proportional
to that subregion's relative size. If we check the two largest first, in the order NS1, S1, S2, NS2, we are
therefore more likely to exit the test in its first if branches. The value of gh,Ie is used in two if branches,
so it is computed just once and saved before the if sequence in order to save some computations.
Algorithm 2 uses a composite or-and condition rather than several ifs. Assuming left-to-right evalu-
ation, the algorithm corresponds to testing first S1 (left side of or), then S2 (right side of or), either case
leading to a spike. If neither is true, no further tests are necessary because the state must necessarily be in
the no-spike region. The test for S2 is made less costly on average by using the auxiliary inequality (47).
This algorithm uses one test less overall than the previous one, but it may require one more test on average,
if NS1 is the region with highest occupation frequency. The left-to-right evaluation assumption does not
always hold in modern processors, which build their own statistics to optimize the test order of logical
constructs.
This is a provisional file, not the final typeset article
22
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 6. (A) Schematic of the simulation setup used to calculate occupation frequencies and to compare
the hybrid scheme with lossless method Algorithm 1, with lossless method 2, and with the standard test.
The neuron model (empty red circle), implementing one of the three schemes, receives input from an
external current and from one excitatory (E) and one inhibitory (I) Poisson generator. The total input has
mean µ and variance σ2. (B) Sample of membrane-potential dynamics for µ = 15 mV and σ2 = 25 mV2.
The analysis assumed Ie < Iθ ≡ θC/τ. The reduced system (45) and the Algorithm 1 and 2 are, however,
also valid in the case that Ie ⩾ Iθ, corresponding to the lower panel of Figure 4. In this case subregions
NS2 and S2 do not exist below the threshold, and the inequalities V < bIe(I) and V < gh,Ie(I) are always
satisfied when V < θ; they always evaluate to true in both algorithms. Both algorithms therefore correctly
distinguish spiking from non-spiking states in this case, although they become inefficient owing to the
additional superfluous evaluations of bIe(I) and gh,Ie(I) for spiking states. We decide not to modify them
in the present work because the case Ie ⩾ Iθ is unusual in real applications. More efficient algorithms for
this case can be designed by interested readers following the guidelines just given in this section.
3.5.2 Occupation frequencies
We want to measure the occupation frequencies in the typical case of a neuron embedded in a recurrent
network, receiving fluctuating synaptic input. The setup for this simulation is illustrated in Figure 6A and
its formulae explained in Appendix B. One neuron is coupled with strengths J and −J to an excitatory
and an inhibitory Poisson generator and also receives a constant external current. The Poisson generators
each mimic an excitatory and an inhibitory population. The neuron thus receives a fluctuating input current
having average µ and variance σ2. In this instance the code of the simulation includes a subroutine that
informs us of the current state every time the lossless method test is called, without altering the test or the
dynamics.
Figure 7A gives a visual idea of the occupation frequencies for µ = 15 mV, σ2 = 25 mV2, and
J = 0.1 mV (all expressed in volts through multiplication by a resistance of τ /C = 40 MΩ), which
correspond to the case Ie < Iθ. These values correspond to a composite average input of 250 000 spikes/s,
and a total average input current of 400 pA. This presynaptic input makes the neuron fire at an average
rate of 7 spikes/s. A sample of its membrane dynamics is shown in Figure 6B. Subregion NS1 has
the overwhelmingly largest occupation frequency. The other three subregions have actually very small
Frontiers
23
IAF neuron J. Krishnan et al.
Perfect spike detection via time reversal
Figure 7. (A) Frequency density of states over state space at each call of the threshold-crossing test,
for network parameters µ = 15 mV, σ2 = 25 mV2, J = 0.1 mV. The colourbar is in units of 6 ×
109 mV−1 pA−1 (obtained from total number of events × area element). The dotted purple line is the
threshold θ. The only visible region in this plot is NS1. and the horizontal blue line is the boundary between
no-spike and spike regions. The spike region S1∪ S2 and subregion NS2 are not visible on this scale because
of the exceedingly small average timestep h = 4 × 10−3 ms. To discern them we need to zoom in, as done
in upper panel (B): the curved envelope and the two straight lines that separate S1, S2, and NS2 extend
horizontally and vertically for just about 1 pA and 10−5 mV. (B) lower panel. In contrast, for a much
larger timestep h = 5 ms the three boundaries would have a larger extension, about 5 500 pA and 20 mV,
and be discernible in plot (A).
areas, as clear from the axis ranges of Figure 7B, owing to the very small value of the average timestep,
h = 4 × 10−3 ms, given by the inverse of the input rate in this hybrid scheme.
A more precise comparison of the occupation frequencies of the four regions N S1, N S2, S1, S2 is
shown in Figure 8 for several combinations of three network parameters, producing different dynamic
regimes. The parameters are the average µ, the variance σ2 of the input current, and the presynaptic
coupling strength J (all expressed in volts through multiplication by a resistance of τ /C = 40 MΩ).
The values of the parameters (µ, σ2, J) include typical realistic cases as well as some extreme cases, like
unusually high coupling strengths. Each panel of Figure 8 shows the occupation frequencies for a set of
dynamical regimes with constant J, β and several σ2. The panels in the last row correspond to the case
Ie ≡ µC/τ ⩾ Iθ ≡ θC/τ, or µ ⩾ θ, in which subregions NS2 and S2 do not exist below threshold.
It is important to remember that the boundary and size of the regions of Figure 5 vary with the timestep
h, which is a parameter of the simulation scheme, not of the dynamics per se. In an event-driven or
hybrid scheme, this step varies inversely with the event input rate, which for Poisson input generators is
proportional to σ2/J 2. As a consequence, the frequencies displayed in Figure 8 are not determined by the
Liouville distribution of the dynamics (26) alone, but also by the details of the numerical-implementation
scheme. The dependence of the boundaries on h is illustrated in Figure 7B. As h decreases, the line
V = fh,Ie(I) and the auxiliary line V = gh,Ie(I) gets closer to the threshold, and the subregions S1, S2,
NS2 disappear. This is plausible since in the limit of h = 0 we are not evolving the initial state at all. As
h increases the point of tangency between the envelope V = bIe(I) and the line V = fh,Ie(I) moves to
This is a provisional file, not the final typeset article
24
p J. Krishnan et al.
Perfect spike detection via time reversal
increasingly lower voltages and higher currents; subregion S2 takes over S1 and subregion NS1 becomes
wider. For typical timestep values of several milliseconds, though, subregions S1, S2, NS2 are still very
small.
A rough estimate of the dependence of the areas of the bounded regions S2 and NS2 on the parameters
(µ, σ2, J), for µ < θ, can be obtained by looking at Figure 8 and considering that these areas together form
a triangle with vertices
(Iθ − Ie, θ),
with If such that fh,Ie(If ) = θ. (48)
This triangle has base If − (Iθ − Ie) and height θ − bIe[eh/τs (Iθ − Ie)]. Expressing h and Ie in terms of
µ and σ2 using eqs (53), where h is inversely proportional to the input rate rI + rE, we find
(cid:0)eh/τs (Iθ − Ie), bIe[eh/τs (Iθ − Ie)](cid:1),
(If , θ),
(cid:20)
1− τ eh/τ − τseh/τs
(cid:21)(cid:20)eh/τs (τ − τs) (1 − eh/τ )
(cid:21)
(θ−µ)2
areas of S2 and NS2 ∝ C
τ
τ J 2
σ2 .
(49)
When τs ≲ τ and J 2 ≲ σ2 a Taylor expansion in J 2/σ2 to fourth order gives a good approximation, with a
relative error below 10 %:
τs (eh/τ − eh/τs)
τ − τs
with h =
−1
(cid:34)
(cid:18) J 2
(cid:19)3
σ2
τ 2
4τ 2
s
(cid:18) J 2
(cid:19)4
τ 2 (τ + τs)
+
8τs3
σ2
(cid:19)5(cid:35)
(cid:18) J 2
σ2
+ O
.
(50)
areas of S2 and NS2 ∝ C
τ
(θ − µ)2
This approximate formula shows that subregions NS2 and S2 grow with the square of the input mean µ
and with the third or fourth power of the ratio J 2/σ2. Recall that these regions do not exist for µ ⩾ θ. The
occupation frequencies do not depend on the areas alone, however, but also on the dynamics, as explained
in the previous section. We can identify several other dynamical mechanisms for their dependence on the
parameters (µ, σ2, J):
• an increase in mean input µ leads to more frequent threshold crossings, thus frequently bringing the
voltage to its reset value, underneath subregions NS2 and S2. The occupation frequencies of these
subregions may therefore decrease with µ even though their areas grow with µ;
• for low mean input µ, an increase in variance σ2 means a higher chance of high-V regions, and thus an
increase in the occupation frequencies of NS2 and S2, even though their areas shrink with σ2;
• for mean input µ close to the threshold, an increase in the variance σ2 leads to more frequent threshold
crossings, and may thus increase occupation frequency of S2 with σ2, even though its area shrinks with
σ2.
The occupation frequency of subregion NS1 (green circles) dominates all others, varying from 90 % to
100 % depending on the network parameters. Subregion S1 (yellow crosses) follows in order of frequency
and is the most frequently visited between the two spike subregions. Subregions NS2 (blue squares) and S2
(red stars) are scarcely visited for lower synaptic amplitudes, with frequencies from 0 to 10−6; and slightly
more often at higher synaptic amplitudes (frequencies from 0 to 10−2).
3.5.3 Time performance and accuracy of a hybrid scheme based on the lossless method
The average time costs of Algorithm 1 and Algorithm 2 within a hybrid scheme can be assessed by
real-time simulation measurements. The average time cost of a hybrid scheme based on the standard test
(1) can also be assessed in the same way for comparison. We therefore compare the two algorithms of the
lossless method and the sufficiency test in this section.
Frontiers
25
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 8. Occupation frequencies of the four subregions NS1 (green circles), S1 (yellow crosses), NS2
(blue squares), S2 (red stars) of Figure 5, for various sets (µ, σ2, J) of input-current mean and variance,
and synaptic strength. Each panel shows the frequencies vs current variance for fixed current mean and
synaptic strength. The columns have the same J, ranging from 0.1 mV (leftmost) to 5 mV (rightmost).
The rows have the same µ, ranging from 10 mV (top) to 22 mV (bottom). The frequencies were measured
from N samples, depending on (µ, σ2, J). The dotted lines in each plot show the inverse number of
samples 1/N for that network regime. The thickness of the segments connecting the data points equals
one standard deviation. The shaded regions show where the limiting frequencies (N → ∞) are expected
to lie with 87 % probability, using a Johnson-Dirichlet model with parameter k = 0.05 determined by
posterior maximization (Johnson, 1932; Zabell, 1982; Good, 1966; Bernardo and Smith, 2000, § 3.2.5).
The frequencies of region S2 (red stars) are particularly important: they are the frequencies of spike-misses
of the standard test (1).
This is a provisional file, not the final typeset article
26
10-810-710-610-510-410-310-210-1100frequencyJ=0.1mVJ=0.5mVNS1S1NS2S2J=1.0mVJ=5.0mVµ=10mV10-810-710-610-510-410-310-210-1100frequencyµ=14mV10-810-710-610-510-410-310-210-1100frequencyµ=18mV101418222630σ2/mV210-810-710-610-510-410-310-210-1100frequency101418222630σ2/mV2101418222630σ2/mV2101418222630σ2/mV2µ=22mV J. Krishnan et al.
Perfect spike detection via time reversal
The basic setup is the same as for the frequency analysis of 3.5.1, explained in Appendix B, with network
parameters (µ, σ2, J). The only difference is that in the present case the code does not include the subroutine
that informs us of the frequencies, which would otherwise increase and bias the real-time durations of the
simulations. Three instances of the basic setup are prepared: in the first the neuron is modelled by a hybrid
scheme with lossless method Algorithm 1, in the second the neuron is modelled by a hybrid scheme with
lossless method Algorithm 2, and in the third the neuron is modelled by a hybrid scheme with the standard
threshold-crossing test (1). The various random-number-generator seeds of the three instances are exactly
the same, so that the three neurons receive exactly the same input, spike-for-spike; this is essential for a
fair comparison between the three schemes. The three instances are run for a long time (2 × 106 ms) and
repeated (in parallel) for several times (10), enough to collect reliable statistics. The statistics are collected
for the same sets of (µ, σ2, J) values as in the frequency analysis. The total real-time length of a simulation
depends, in all three schemes, on how often the checkpoints and threshold-crossing tests occurs, and this in
turn depends on the presynaptic input frequency, as already discussed.
Rather than showing the results for all sets of parameters (µ, σ2, J), which in this case are not very
informative, we show in Figure 9 those with the (J, µ) values that yield the slowest and fastest performances.
The average computation costs of Algorithm 1 and Algorithm 2, which embody the lossless method, turn
out to be very similar – within each other's standard deviations – and basically identical in comparison
with the cost of the scheme based on the standard test (1). Both are slower than the hybrid scheme with
the standard test: from around 33 % slower in the case of high-activity regime with frequent incoming
spikes (µ = 18 mV, J = 0.1 mV), to around 8 % slower in the case of low activity regime and infrequent
incoming spikes (µ = 10 mV, J = 5 mV). These are the extremes shown in Figure 9. In most other sets
of network parameters the hybrid scheme with lossless method was around 20 %–25 % slower than the
standard hybrid scheme with threshold-crossing test (1).
As mentioned in 3.4, subregion S2 contains all states for which the standard threshold-crossing test
misses a spike. The occupation frequencies of this subregion are therefore a direct measure of the number
of spike missed, per neuron, by the standard hybrid scheme. They are shown as red stars in Figure 8
for the various sets of network parameters (µ, σ2, J). The frequency of missed spikes does not have a
simple monotonic dependence on the three parameters, owing to the interaction of several mechanisms,
discussed in Section 3.5.2. An increase in synaptic coupling J or input current µ generally leads to more
frequently missed spikes because the neuron spikes more often overall. For very high – suprathreshold –
input currents, however, the frequency decreases again until no spikes are missed anymore; this change
in trend happens because the checkpoints become more frequent. Regimes of low J are diffusion-like
processes, where frequent arrival of synaptic events does not lead to missed spikes. Regimes of high J
are shot-noise processes, where sudden and infrequent arrival of synaptic events leads to suprathreshold
excursions and missed spikes. Separate simulations show that the hybrid scheme based on the lossless
method, with either algorithm, reproduce the analytic solution of the neuron model within floating point
precision.
For biologically realistic synaptic couplings, J < 1 mV, the standard hybrid scheme is 33 % faster than
the scheme with the lossless method, and misses less than 1 spike every 106 test calls, per neuron; for
very low couplings J < 1 mV this figure even becomes less than 1 spike every 108 test calls. For synaptic
couplings J > 1 mV the standard hybrid scheme starts to miss more spikes, reaching even 1 missed spike
every 500 test calls for subthreshold average input currents; and it is only 8 % faster than the hybrid scheme
with the lossless method. From these figures a user can decide to use the hybrid scheme with the standard
or the lossless test, depending on the desired balance of accuracy and speed.
Frontiers
27
J. Krishnan et al.
Perfect spike detection via time reversal
Figure 9. Computational costs of the hybrid scheme with lossless method, Algorithm 1 (red), lossless
method, Algorithm 2 (green), and standard threshold-crossing test (1) (blue). (A) µ = 18 mV, J = 0.1 mV
describes regimes where the lossless method has the highest increase in computational cost, around 33 %.
These are regimes of high activity and frequently incoming spikes. (B) µ = 10 mV, J = 5 mV describe
regimes where the lossless method has the lowest increase in computational cost, around 8 %. These are
regimes of low activity and infrequently incoming spikes. The data come from 10 simulations of 2× 106 ms
simulation-time each.
This is a provisional file, not the final typeset article
28
(mV) J. Krishnan et al.
Perfect spike detection via time reversal
4 SUMMARY AND DISCUSSION
We here present a general method to solve the threshold-crossing detection problem for an integrable,
affine or linear neuronal dynamics. The method is based on the geometric idea of propagating the threshold
plane backwards in time and to determine whether the swept volume contains the initial state, rather than
propagating the initial state forward in time to check whether it crosses the threshold. These two procedures
are obviously mathematically equivalent, but they distribute the computational load in different ways. The
forward-propagation of the state looks for the value of the crossing time; if no such value exists, it means
there is no threshold crossing. The backward-propagation of the threshold first tests whether a crossing
time exists at all, without yielding its value; the latter is calculated afterwards.
The different distribution of computational load in the two procedures can be explained geometrically and
algebraically. The first procedure geometrically checks for the intersection of a curve (the state trajectory)
with a hypersurface (the threshold). Algebraically, this corresponds to finding the roots of a system of
equations, often transcendental. The second procedure geometrically checks for the "intersection" of a
point (the state) with a hypervolume (the threshold trajectory), i.e. the inclusion of the former in the latter.
Algebraically, this corresponds to testing a set of inequalities. The latter procedure is more efficient for
numerical computations, because it only relies on inequality tests – in which the presence of transcendental
functions is much less costly than in an equation that needs to be solved for a particular variable. The
determination of the exact crossing-time, which involves bisection algorithms and is the costlier part, is
done only when the existence of this value is certain. No root search is unnecessarily performed.
We have calculated the system of inequalities expressing the threshold-crossing condition, for a generic
affine or linear neuronal dynamics in any dimension. The result is the conjunction of inequalities (23).
It consists of two affine linear inequalities in the state-space variables, voltage and currents, and a non-
linear one. The numerical implementation of this system of inequalities can be further optimized, on a
case-by-case basis. In order to give a concrete implementation example of the generic inequalities (23), to
show their geometrical meaning, and to give an example of optimization,in the present work we apply our
procedure, step-by-step, to the 2-dimensional case of a leaky integrate-and-fire neuron with exponentially
decaying post-synaptic currents (Rotter and Diesmann, 1999). The generic inequalities (23) take in this
case the concrete form (44).
The quantitative data in the present work are obtained by integrating these inequalities into a combined
event-and-time-driven simulation framework (Morrison et al., 2007) for large-scale spiking neuronal
network models as released by Bos et al. (2015). Implementation and comparison to earlier work show that:
• the system of inequalities, the non-linear one in particular, can be expressed analytically in terms of
the state-space variables even when the original threshold-crossing condition involves a transcendental
equation – and would therefore require bisection algorithms. Compare (44) with (30);
• the computationally expensive non-linear function in the system can be conveniently triangularized,
speeding up the algorithm even further by testing a linear inequality first, ruling out the majority of initial
states;
• the new method reproduces the analytic solution of the neuron model within floating point precision.
It detects all threshold crossings, in particular those that the approximate test (1) of Hanuschkin et al.
(2010), misses in some ranges of mean activity, fluctuations, and synaptic-coupling strength (Figure 8);
• at the default spike accuracy of 0.1 ms of the reference simulator the new method is 8 %–33 % slower
than the fastest available solver with spike loss (Hanuschkin et al., 2010). It is therefore of comparable
speed as embedded event driven methods (see Hanuschkin et al. (2010) Fig 5, inset).
Frontiers
29
J. Krishnan et al.
Perfect spike detection via time reversal
In practice the method of Hanuschkin et al. (2010) rarely misses spikes for biologically realistic synaptic
amplitudes. At low frequencies of afferent synaptic events, however, such missed threshold-crossings can
happen. Our new scheme therefore offers an alternative for users who need guaranteed spike detection and
are willing to pay a price in terms of slightly longer computation time.
In state-spaces of higher dimensions it is be more difficult to derive the non-linear inequality of the
system (23) in implicit form, but its set of approximating flat surfaces can still be easily calculated. In the
worst-case scenario of an inequality not expressible analytically, it is still possible to construct a nested
sequence of approximating flat surfaces to be tested hierarchically. Such construction only needs to be
done once for any given model and detects spikes with any desired precision. Being linear, such nested
inequalities likely are computationally less expensive than a bisection algorithm.
The problem of detecting some sort of threshold crossing in a system of coupled first-order linear
differential equations appears in many other applications and phenomena like switching, friction, and
saturation (Hiebert and Shampine, 1980). For example, in an air-conditioning unit a thermostat controls the
on-off state based on a certain threshold value of the room temperature (Shampine, 1994). The dynamics
of this system is similar to that described in Sections 3.1–3.2. Another example is the problem of ejecting a
pilot such that collision with the aircraft stabilizer is avoided.
The present work use concepts from differential geometry, in particular extrusions (Bossavit, 2003) and
critical points of maps between manifolds, and shows that these concepts have a readily understandable
geometrical and visual meaning. The notion of extrusion has recently found applications in numerical and
discretization techniques for partial and integral differential equations (Desbrun et al., 2005). The notion of
critical points of a manifold mapping is ubiquitous in science: from the caustics of propagating seismic
fronts, at which the seismic wave changes its phase (Romanowicz and Dziewonski, 2007, § 1.04), to the
singularities between two coordinate charts in general relativity (Misner et al., 2003), which affect the
accuracy of global navigation satellite systems (Coll et al., 2012; Sáez and Puchades, 2013).
Indeed, the neuron model analysed in Section 3 exhibits a similarity with the dynamics of a point mass
near a black hole. If the simulation timestep h is very large the curved surface separating the states that
lead to a spike from those that do not acts like an event horizon in general relativity: a state evolved from
the spike region can enter the no-spike region, but once there it cannot escape and will always remain a
"no-spike" state. This is only true for the dynamics (6), though, with constant affine term and no resets
at threshold. Inputs from other neurons lead to discontinuous changes in the affine term of the dynamics,
causing a "transport" of initial states out of the event horizon, from the no-spike to the spike region.
Nevertheless, maybe such similarities are more than mere coincidences. For example, the trajectory of the
threshold surface of a leaky integrate-and-fire model with α-shaped post-synaptic currents (Bernard et al.,
1994) can be implicitly expressed in terms of the Lambert-W function (Corless et al., 1996), as an analysis
along the lines of Section 2.2 shows. This function also appears in the implicit expression of point-mass
trajectories in (1 + 1)-dimensional general relativity (Mann and Ohta, 1997). It is surely worthwhile to
bring the nascent field of neuronal dynamics closer to ideas and techniques from differential geometry and
general relativity.
This is a provisional file, not the final typeset article
30
J. Krishnan et al.
Perfect spike detection via time reversal
A INTERPOLATION FOR THRESHOLD-CROSSING TIME
If we know that the neuron voltage V is below threshold at times t and t + h but above threshold
somewhere in the interval ]t, t + h], then by continuity it must reach a maximum above threshold at a time
tmax ∈ ]t, t + h]. This time can be obtained solving the equation V (tmax) = 0, with V (t) given by (26).
The solution (Hanuschkin et al., 2010) is
(cid:20) τ
τs
(cid:18) Ie
I
− τ − τs
τ
− CV
τ I
(cid:19)(cid:21)
,
tmax = − τ τs
τ − τs
ln
(51)
from which the potential V (tmax) > θ can also be easily calculated.
The time tθ ∈ ]t, tmax] at which the first threshold crossing occurs, V (tθ) = θ, must lie between t and
tmax, with V (t) < V (tθ) < V (tmax), and can thus be interpolated using a bisection algorithm (Press et al.,
2007).
B NETWORK SETUP
The dynamics of the membrane potential V and synaptic current I can be described, if the input is treated
stochastically and for weak synaptic couplings, by a diffusion process with equations (Fourcaud and Brunel,
2002b)
τ
dV
dt
d(RI)
dt
τs
= −V (t) + (RI)(t),
= −(RI)(t) + µ + σ
√
τ ξ(t).
(52)
{Ji} and firing rates {ri} of the input neurons via µ = τ(cid:80)
with ξ a zero-mean Gaussian process. The parameters µ and σ2 characterize the stochastic input and are
the mean and variance of the total incoming synaptic current. They are related to the synaptic couplings
2ri. If we have one
excitatory and one inhibitory population of input neurons, mimicked by two Poisson generators with rates
rE and rI coupled to the neuron with strengths JE and JI (where J = τ
C w and w is the synaptic weight of
the current), and by an input current Ie, then
i Jiri and σ2 = τ(cid:80)
i Ji
µ =
Ieτ
C
σ2 = τ (JI
+ τ (JI rI + JErE)
2rE),
(cid:90) θ−µ
2rI + JE
√
σ +
ζ(1/2)√
2
√ τs
√ τs
τ
(53)
1
r
where r is the output firing rate of the neuron, approximated to linear order in(cid:112)τs/τ, τr is the refractory
ζ(1/2)√
Vr−µ
σ +
2
τ
= τr + τ
π
ey2
[1 + erf(y)] dy,
time, Vr the reset voltage, and ζ the zeta function (Abramowitz and Stegun, 1972).
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.
Frontiers
31
J. Krishnan et al.
Perfect spike detection via time reversal
AUTHOR CONTRIBUTIONS
All authors have made significant, direct, and intellectual contribution to the work.
FUNDING
This work was supported by DFG Grant: GSC11; Helmholtz association: VH-NG-1028 and SMHB; EU
Grant 604102 (HBP); Juelich Aachen Research Alliance (JARA).
ACKNOWLEDGMENTS
PGLPM thanks the Forschungszentrum librarians for their always prompt and kind help, Mari & Miri
for continuous encouragement and affection, Buster Keaton for filling life with awe and inspiration, and
the developers and maintainers of LATEX, Emacs, AUCTEX, MiKTEX, arXiv, biorXiv, PhilSci, Python,
Inkscape, Sci-Hub for making a free and unfiltered scientific exchange possible.
REFERENCES
Abramowitz, M. and Stegun, I. A. (eds.) (1972). Handbook of Mathematical Functions: With Formulas,
Graphs, and Mathematical Tables, vol. 55 of National Bureau of Standards Applied mathematics series
(Washington, D.C.: U.S. Department of Commerce), tenth printing, with corrections edn. First publ.
1964
Artin, E. (1955). Geometric Algebra, vol. 3 of Interscience tracts in pure and applied mathematics (New
York: Interscience)
Bernard, C., Ge, Y. C., Stockley, E., Willis, J. B., and Wheal, H. V. (1994). Synaptic integration of NMDA
and non-NMDA receptors in large neuronal network models solved by means of differential equations.
Biol. Cybern. 70, 267–273
Bernardo, J.-M. and Smith, A. F. (2000). Bayesian Theory. Wiley series in probability and mathematical
statistics (New York: Wiley), reprint edn. First publ. 1994
Bos, H., Morrison, A., Peyser, A., Hahne, J., Helias, M., Kunkel, S., et al. (2015). Nest 2.10.0. http:
//dx.doi.org/10.5281/zenodo.44222, http://www.nest-simulator.org/
Bossavit, A. (1991). Differential Geometry: for the Student of Numerical Methods in Electromagnetism.
http://butler.cc.tut.fi/~bossavit/
Bossavit, A. (2002). Applied Differential Geometry (A Compendium). http://butler.cc.tut.
fi/~bossavit/BackupICM/Compendium.html. First publ. 1994
Bossavit, A. (2003). Extrusion, contraction: their discretization via Whitney forms. COMPEL 22, 470–480.
http://butler.cc.tut.fi/~bossavit/Papers.html
Brette, R. (2007). Exact simulation of integrate-and-fire models with exponential currents. Neural Comput.
19, 2604–2609
Burke, W. L. (1987). Applied Differential Geometry (Cambridge: Cambridge University Press), reprint edn.
First publ. 1985
Burke, W. L. (1995). Div, Grad, Curl Are Dead. http://count.ucsc.edu/~rmont/papers/
Burke_DivGradCurl.pdf; 'preliminary draft II'. See also http://www.ucolick.org/
~burke/
Choquet-Bruhat, Y., DeWitt-Morette, C., and Dillard-Bleick, M. (1996). Analysis, Manifolds and Physics.
Part I: Basics (Amsterdam: Elsevier), rev. ed. edn. First publ. 1977
This is a provisional file, not the final typeset article
32
J. Krishnan et al.
Perfect spike detection via time reversal
Coll, B., Ferrando, J. J., and Morales-Lladosa, J. A. (2012). Positioning systems in Minkowski space-time:
Bifurcation problem and observational data. Phys. Rev. D 86, 084036. arXiv:1204.2241
Corless, R. M., Gonnet, G. H., Hare, D. E. G., Jeffrey, D. J., and Knuth, D. E. (1996). On the Lam-
bert W function. Adv. Comput. Math. 5, 329–359. https://cs.uwaterloo.ca/research/
tr/1993/03/W.pdf
Coxeter, H. S. M. (1969). Introduction to Geometry (New York: Wiley), 2 edn. First publ. 1961
de Rham, G. (1984). Differentiable Manifolds: Forms, Currents, Harmonic Forms, vol. 266 of Grundlehren
der mathematischen Wissenschaften (Berlin: Springer). Transl. by F. R. Smith. First publ. in French
1955
Dehaene, S. and Changeux, J.-P. (2005). Ongoing spontaneous activity controls access to consciousness: A
neuronal model for inattentional blindness. public library of science 3, 0910–0927
Desbrun, M., Hirani, A. N., Leok, M., and Marsden, J. E. (2005). Discrete exterior calculus. arXiv:
math/0508341
D'Haene, M. and Schrauwen, B. (2010). Fast and exact simulation methods applied on a broad range of
neuron models. Neural Comput. 22, 1468–1472
D'Haene, M., Schrauwen, B., Van Campenhout, J., and Stroobandt, D. (2009). Accelerating event-driven
simulation of spiking neurons with multiple synaptic time constants. Neural Comput. 21, 1068–1099.
doi:10.1162/neco.2008.02-08-707
Dodson, C. T. J. and Poston, T. (1991). Tensor Geometry: The Geometric Viewpoint and its Uses, vol. 130
of Graduate texts in mathematics (Springer), 2 edn. First publ. 1977
Ferscha, A. (1996). Parallel and distributed simulation of discrete event systems. In Parallel and Distributed
Computing Handbook, ed. A. Y. Zomaya (McGraw-Hill), chap. 35. 1003–1041
Fourcaud, N. and Brunel, N. (2002a). Dynamics of the firing probability of noisy integrate-and-fire neurons.
Neural Comput. 14, 2057–2110
Fourcaud, N. and Brunel, N. (2002b). Dynamics of the firing probability of noisy integrate-and-fire
neurons. Neural Comp. 14, 2057–2110. http://galton.uchicago.edu/~nbrunel/pdfs/
fourcaud02.pdf
Fujimoto, R. M. (2000). Parallel and distributed simulation systems (New York: Wiley)
Gewaltig, M.-O. and Diesmann, M. (2007). NEST (NEural Simulation Tool). Scholarpedia 2, 1430
Good, I. J. (1966). How to estimate probabilities. J. Inst. Maths. Applics 2, 364–383
Hansel, D., Mato, G., Meunier, C., and Neltner, L. (1998). On numerical simulations of integrate-and-fire
neural networks. Neural Comput. 10, 467–483
Hanuschkin, A., Kunkel, S., Helias, M., Morrison, A., and Diesmann, M. (2010). A general and efficient
method for incorporating precise spike times in globally time-driven simulations. Front. Neuroinform. 4,
113
Hiebert, K. L. and Shampine, L. F. (1980).
(Albuquerque, USA: Sandia Laboratories)
Implicitly Defined Output Points for Solutions of ODEs
Johnson, W. E. (1932). Probability: The deductive and inductive problems. Mind 41, 409–423. With some
notes and an appendix by R. B. Braithwaite
Krishnan, J., Porta Mana, P., Helias, M., Diesmann, M., and Di Napoli, E. (2016). Perfect detection of
spikes via time-reversal. In Proceedings of the Bernstein Conference. doi:http://dx.doi.org/10.12751/
nncn.bc2016.0152
Kunkel, S., Helias, M., Diesmann, M., and Morrison, A. (2011). Fail-safe detection of threshold crossings
of linear integrate-and-fire neuron models in time-driven simulations. BMC Neuroscience 12, P229
Frontiers
33
J. Krishnan et al.
Perfect spike detection via time reversal
Mann, R. B. and Ohta, T. (1997). Exact solution for the metric and the motion of two bodies in (1 +
1)-dimensional gravity. Phys. Rev. D 55, 4723–4747. arXiv:gr-qc/9611008
Marsden, J. E. and Ratiu, T. (2007). Manifolds, Tensor Analysis, and Applications, vol. 75 of Applied
mathematical sciences (New York: Springer), 3 edn. http://www.esm.vt.edu/~sdross/pub/
books/. Written with the collaboration of Ralph Abraham. First publ. 1983
Mattia, M. and Del Giudice, P. (2000). Efficient event-driven simulation of large networks of spiking
neurons and dynamical synapses. Neural Comput. 12, 2305–2329
Misner, C. W., Thorne, K. S., and Wheeler, J. A. (2003). Gravitation (New York: W. H. Freeman and
Company), 25th printing edn. First publ. 1970
Morrison, A. and Diesmann, M. (2008). Maintaining causality in discrete time neuronal network simu-
lations. In Lectures in Supercomputational Neuroscience: Dynamics in Complex Brain Networks, eds.
P. beim Graben, C. Zhou, M. Thiel, and J. Kurths (Springer), Understanding Complex Systems. 267–278
Morrison, A., Straube, S., Plesser, H. E., and Diesmann, M. (2007). Exact subthreshold integration with
continuous spike times in discrete time neural network simulations. Neural Comput. 19, 47–79
Nomizu, K. and Sasaki, T. (1994). Affine Differential Geometry: Geometry of Affine Immersions, vol. 111
of Cambridge tracts in mathematics (Cambridge: Cambridge University Press)
Porta Mana, P. G. L. (2011). Notes on affine and convex spaces. arXiv:1104.0032
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. (2007). Numerical Recipes: The Art of
Scientific Computing (Cambridge: Cambridge University Press), 3 edn. First publ. 1988
Ramanan, S. (2005). Global Calculus, vol. 65 of Graduate studies in mathematics (Providence, USA:
American Mathematical Society)
Rauch, A., La Camera, G., Lüscher, H., Senn, W., and Fusi, S. (2003). Neocortical pyramidal cells respond
as integrate-and-fire neurons to in vivo like input currents. Journal of Neurophysiology 90, 1598–1612
Rockafellar, R. T. (1972). Convex Analysis, vol. 28 of Princeton mathematical series (Princeton: Princeton
University Press). First publ. 1970
Romanowicz, B. and Dziewonski, A. (eds.) (2007). Seismology and the Structure of the Earth, vol. 1 of
Treatise on Geophysics (Amsterdam: Elsevier), 1 edn.
Rotter, S. and Diesmann, M. (1999). Exact digital simulation of time-invariant linear systems with
applications to neuronal modeling. Biol. Cybern. 81, 381–402
Sáez, D. and Puchades, N. (2013). Relativistic positioning systems: Numerical simulations. Acta Futura 7,
103–110. arXiv:1404.1000
Schouten, J. A. (1989). Tensor Analysis for Physicists. Dover books on physics and chemistry (New York:
Dover), corr. second ed. edn. First publ. 1951
Schücker, J., Diesmann, M., and Helias, M. (2015). Modulated escape from a metastable state driven by
colored noise. Phys. Rev. E 92, 052119. arXiv:1411.0432
Shampine, L. F. (1994). Numerical Solution of Ordinary Differential Equations (New York: Chapman &
Hall/CRC)
Shelley, M. J. and Tao, L. (2001). Efficient and accurate time-stepping schemes for integrate-and-fire
neuronal networks. Journal of Computational Neuroscience 11, 111–119
Simon, U., Schwenk-Schellschmidt, A., and Viesel, H. (1992). Introduction to the Affine Differential
Geometry of Hypersurfaces. Lecture notes of the Science University of Tokyo (Tokyo: Science University
of Tokyo)
Spivak, M. (1999). A Comprehensive Introduction to Differential Geometry. Vol. One (Houston, USA:
Publish or Perish), 3 edn. First publ. 1970
This is a provisional file, not the final typeset article
34
J. Krishnan et al.
Perfect spike detection via time reversal
van Elburg, R. A. J. and van Ooyen, A. (2009). Generalization of the event-based Carnevale-Hines
integration scheme for integrate-and-fire models. Neural Computation 21, 1913–1930. doi:10.1162/
neco.2009.07-08-815
Zabell, S. L. (1982). W. E. Johnson's "sufficientness" postulate. Ann. Stat. 10, 1090–1099. Repr. in (Zabell,
2005, pp. 84–95)
Zabell, S. L. (2005). Symmetry and Its Discontents: Essays on the History of Inductive Probability.
Cambridge studies in probability, induction, and decision theory (Cambridge: Cambridge University
Press)
Zeigler, B. P., Praehofer, H., and Kim, T. G. (2000). Theory of Modeling and Simulation: Integrating
Discrete Event and Continuous Complex Dynamic Systems (Amsterdam: Academic Press), 2 edn.
Zheng, G., Tonnelier, A., and Martinez, D. (2009). Voltage-stepping schemes for the simulation of spiking
neural networks. Journal of Computational Neuroscience 26, 409–23
Frontiers
35
|
1704.02342 | 1 | 1704 | 2017-04-07T18:33:59 | Sleep Paralysis: phenomenology, neurophysiology and treatment | [
"q-bio.NC"
] | Sleep paralysis is an experience of being temporarily unable to move or talk during the transitional periods between sleep and wakefulness: at sleep onset or upon awakening. Feeling of paralysis may be accompanied by a variety of vivid and intense sensory experiences, including mentation in visual, auditory, and tactile modalities, as well as a distinct feeling of presence. This chapter discusses a variety of sleep paralysis experiences from the perspective of enactive cognition and cultural neurophenomenology. Current knowledge of neurophysiology and associated conditions is presented, and some techniques for coping with sleep paralysis are proposed. As an experience characterized by a hybrid state of dreaming and waking, sleep paralysis offers a unique window into phenomenology of spontaneous thought in sleep. | q-bio.NC | q-bio | Elizaveta Solomonova1,2
Abstract
Sleep Paralysis: phenomenology, neurophysiology and treatment
Clinical Conditions. Fox, K & Christoff, K. Eds.
1Université de Montréal, Individualized program (Cognitive Neuroscience & Philosophy).
2Center for Advanced Research in Sleep Medicine, Dream and Nightmare Laboratory, Montreal, Canada
To appear in: The Oxford Handbook of Spontaneous Thought: Mind-Wandering, Creativity, Dreaming, and
Sleep paralysis is an experience of being temporarily unable to move or talk during the transitional
periods between sleep and wakefulness: at sleep onset or upon awakening. Feeling of paralysis may be
accompanied by a variety of vivid and intense sensory experiences, including mentation in visual, auditory,
and tactile modalities, as well as a distinct feeling of presence. This chapter discusses a variety of sleep
paralysis experiences from the perspective of enactive cognition and cultural neurophenomenology. Current
knowledge of neurophysiology and associated conditions is presented, and some techniques for coping with
sleep paralysis are proposed. As an experience characterized by a hybrid state of dreaming and waking, sleep
paralysis offers a unique window into phenomenology of spontaneous thought in sleep.
Introduction
"I had a few terrifying experiences a few years ago. I awoke in the middle of the night. I was sleeping
on my back, and couldn't move, but I had the sensation I could see around my room. There was a
terrifying figure looming over me. Almost pressing on me. The best way I could describe it was that it
was made of shadows. A deep rumbling or buzzing sound was present. It felt like I was in the presence
of evil... Which sounds so strange to say!" (31 year old man, USA)
Sleep paralysis (SP) is a transient and generally benign phenomenon occurring at sleep onset or upon
awakening. Classified as a rapid-eye-movement (REM) sleep-related parasomnia, SP represents a
psychophysiological state characterized simultaneously by qualities of both sleep and wakefulness, wherein
the experiencer can open her eyes (Hishikawa & Kaneko, 1965), can be aware of her physical environment
but is unable to move and may start seeing, hearing, feeling or sensing something.
While documented instances of SP seem to be very consistent across cultures , SP's lived qualities,
phenomenology, and interpretation as a meaningful experience varies depending on the cultural and religious
background. Rooting SP in a particular belief system may either help the experiencer recognize that SP is
common and transient, or amplify negative qualities of SP by giving more concrete shape to an already
terrifying experience of a supernatural assault.
The emotional experience of SP is often one of fear, terror and panic. Threatening presences, the
vulnerability of being in a paralyzed state, uncontrollable visions - all these elements contribute to intense,
predominantly dysphoric, negative affect. Some qualities of spontaneous thought associated with felt
presence during SP can be seen as paranoid (Cheyne & Girard, 2007), spatial, or interpersonal/social
(Nielsen, 2007, Solomonova, 2008). The vast majority of SP experiences are associated with intense feelings
of realism, and are most often characterized by fear and distress, which may carry-over into wakefulness and
create a vicious cycle of negative emotional association with sleep, including aversion to going to bed, and
even, in extreme cases, can result in symptoms reminiscent of post-traumatic stress disorder (McNally and
Clancy, 2005). Yet, some SP experiences are described in positive terms, especially vestibulo-motor
phenomena that include out-of-body experiences (OBE), or sensations of flying or floating. While the intuitive
and immediate reaction to SP is typically negative, as will be discussed in the section on practical
considerations, there are numerous reports of neutral/positive SP. Furthermore, there is a possibility of
harnessing the power and potential of the dissociative/overlapping state in order to take active charge of
1
one's experience and to use the opportunity presented by the simultaneity of waking and sleeping cognition -
a potential for entering into a lucid dream state or for contemplative self-observation.
Neurophysiologically, SP is currently understood as a state dissociation or a state overlap between
REM sleep and wakefulness (American Academy of Sleep Medicine, 2014). During SP one can open her eyes,
look around the room, become aware of her environment and simultaneously experience REM sleep-related
paralysis (muscle atonia) as well as intense and realistic imagery1 of all sensory modalities - a nightmare
spilling into the real world. Normally, during REM sleep, skeletal muscle atonia blocks most motor output,
effectively preventing the sleeper from acting out her dreams (Peever, Luppi, & Montplaisir, 2014). SP can
also occur in the context of narcolepsy (Sharpless & Barber, 2011; Terzaghi, et al, 2012), but the majority of
those who experience SP report it in its isolated form (often referred to as Isolated Sleep Paralysis), without
known medical or neurological association.
In current medical and neuroscientific literature, SP is discussed in terms of its presentation and
negative factors: SP-associated mentation is generally seen as a non-desirable effect of REM sleep intrusion
into waking. In this chapter I propose that situating SP experiences as a dream phenomenon within the
framework of embodied mind and enactive cognitive science, including contemplative approaches to
consciousness, is an alternative that accounts for the phenomenology of SP as a lived experience, allows for
rich and detailed cultural framing of the experience, and offers avenues for a cross-cultural social
neurophenomenology of SP.
First, I will discuss the phenomenology and neurophysiology of SP experiences. I will present SP in
general, without distinguishing between its isolated and narcolepsy-related form, unless such a separation is
warranted. I will start by presenting the current state of knowledge of SP prevalence, as well as the varieties
of imagery accompanying SP and their cultural significance. I will then discuss SP in terms of a REM sleep
parasomnia and outline its precipitating and enabling factors, as well as sleep and dream characteristics of
those who experience SP. Finally, I will examine SP in light of various cultural and shared practices, including
preventative measures, and practices aimed at interrupting and transforming SP experiences.
The experiential examples of SP used in the present chapter are all derived from an Internet-based
study of SP (Solomonova et al, 2008). Our research group collected 193 responses from people with recurrent
SP experiences, using a modified version of the Waterloo Unusual Sleep Experiences Questionnaire (Cheyne,
Newby-Clark, & Rueffer, 1999). Our participants were recruited online, using word of mouth, and via
advertising on SP-related forums, information and support.
Idiopathic SP (SP not associated with narcolepsy, and without known cause) is a benign and transient
parasomnia (Howell, 2012), occurring during transitions between wake and sleep: at sleep onset or upon
awakening. The Diagnostic and Statistical Manual of Mental Disorders (DSM-5) classifies isolated SP
accompanied by fearful mentation as an instance of a nightmare disorder (American Psychiatric Association,
2013). During an episode of SP, characteristics of REM sleep intrude upon seemingly awake consciousness:
thus the person experiencing SP, while having an impression of being awake and aware of her environment,
is unable to initiate voluntary movements (i.e., experiences REM sleep muscle atonia/paralysis), and may also
experience intense and realistic sensations in any sensory modality - REM sleep-related mentation (American
Academy of Sleep Medicine, 2014). SP should be distinguished from night terrors – early night awakenings
with feelings of panic/terror, typically associated with somnambulism-spectrum arousal from slow wave
(Stages 3 and 4) non-REM (NREM) sleep. Night terrors are characterized by sudden awakening in an agitated
state, anxiety, body motility and general amnesia with regards to underlying cognitive experience
(Szelenberger, Niemcewicz & Dabrowska, 2009).
Prevalence estimates of SP range widely, and may depend on geographic and cultural factors. A
systematic review of SP prevalence has revealed that studies report SP lifetime prevalence from as low as
1.5% to possibly 100% in the general population (Sharpless & Barber, 2011). The authors indicated that
1 In this chapter I will use the terms 'imagery' and 'mentation' interchangeably to refer to visual, auditory, somatosensory
and even social experiences during SP. The term 'imagery' here is therefore not restricted to the visual domain. I prefer
'imagery' and 'mentation' to 'hallucination' in order to emphasize the dream-like process of spontaneous imagination that
takes place during SP, and to de-emphasize the association with delusional thought and pathologies, associated with the
term 'hallucination'.
Definitions and prevalence
2
The 4EA cognition and oneiric mentation
about one in five individuals may have experienced SP at least once in their lifetime (of 36,533 persons in
their review). Prevalence estimation of SP is difficult due to numerous factors such as ethnicity and cultural
background, including variable familiarity with the phenomenon and in the wording of questions (Fukuda,
1993). For example, in a cross-cultural study, Fukuda and colleagues (2000) reported that while it is unclear
whether SP is equally prevalent in Canada and Japan, the lack of familiarity and of a normative cultural
framework for SP in Canada may contribute to the fact that many Canadian, but not Japanese, respondents
qualified SP as a kind of a dream, and would not have, therefore, readily recognized SP in a prevalence study.
An additional reason for under-diagnosis of SP in the West may be the fact that those who experience SP may
be misdiagnosed as having psychiatric disturbances (Hufford, 2005). The developmental trajectory of SP is
traditionally associated with an onset during adolescence, which may indicate a process associated with sleep
architecture maturation (Wing, Lee & Chen, 1994). However, in one study of older adults, a bimodal onset
pattern was reported, with a second pattern of onset of SP episodes after the age of 60 years old (Wing et al,
1999), suggesting a possibility that SP may have a variety of onset conditions.
Neurologically and phenomenologically SP is situated on the REM-sleep based dream/nightmare
continuum. In this chapter, I will approach SP experiences as a variant of intensified or disturbed dreaming,
and will situate them within a framework of embodiment and enactivism.
Recent years have seen the development of a paradigm shift from a strictly neurocentric view of the
mind, a position that can be stated as "embrained" (Morris, 2010), to a diverse family of approaches that
consider the mind embodied (Varela, Thompson & Rosch, 1991; Gallagher, 2005), enactive (Noe, 2004;
Thompson, 2005; Stewart, Gapenne & Di Paolo, 2010), extended into and embedded in the physical and social
world (Clark & Chalmers, 1998; Menary, 2010), and affective (Colombetti, 2013; Pessoa, 2013). While these
approaches are in many respects quite different, they have sometimes been labeled as 4EA (embodied,
embedded, extended, enactive, affective) cognition, with a common theme of offering a robust alternative to
computational, connectionist, and neuro-reductionist views of the mind (Wheeler, 2005; Protevi, 2012).
These theories attempt to situate cognition, brain activity, and psychophysiology within the larger contexts of
lived subjective experience, by emphasizing the roles of developmental sensorimotor attunement to the
world, as well as of the active and motivated processes of perception and sense-making, the importance of the
social and cultural milieu, and the role of emotion and affect.
Sleep and dreaming phenomena have been only rarely addressed by 4EA theorists (with the exception of
Thompson 2014, 2015a,b), and the prevailing view of the sleeping mind today situates sleep mentation as
being firmly constrained within the brain (Rechtschaffen, 1978; Hobson, Pace-Schott & Stickgold, 2000;
Revonsuo, 2006). As Revonsuo states: "The conscious experiences we have during dreaming are isolated from
behavioral and perceptual interactions with the environment, which refutes any theory that states that
organism-environment interaction or other external relationships are constitutive of the existence of
consciousness" (Revonsuo et al, 2015: 3). Alternatively, situating dream mentation within a framework of
4EA approaches implies that the dreaming subject is not entirely isolated or disconnected from
environmental and somatic stimuli, and that her experiential self retains affective, social, sensorimotor and
sense-making qualities. Dreaming then is not passively lived as a purely mental simulation (Revonsuo et al,
2015), but can be seen as a process of active imagination (Thompson, 2014) rooted in the dreamer's physical,
social and affective world (Solomonova & Sha, 2016). I propose that SP experiences, by virtue of their special
kind of overlap between and simultaneous presence of both waking and dreaming cognition, are perfect
candidates for neurophenomenological research on spontaneous thought in sleep, which would help
illuminate particular qualities of dreaming cognition that may otherwise be inaccessible to reflective
consciousness upon awakening from a dream.
3
Phenomenology of sleep paralysis experiences
Figure 1. A representation of sleep paralysis experience2. The sleeper is awakened suddenly and sees a menacing
shadowy creature on top of him. He experiences the sensation of being pushed into the bed, while the bed itself is swirling
in a sort of a tornado. The two faces of the dreamer represent the "double" consciousness during sleep paralysis: he is
simultaneously terrified of the supernatural attacker and also knows that if he does not resist the experience and allows
himself to drift back into sleep he may have a lucid dream (this lucid consciousness is represented as a sleeping face with
a colorful brain, denoting vibrant possibilities of lucidity. Artist: Benjamen Samaha, Montreal, Canada 2016. Reproduced
with artist's permission.
In addition to transient experience of muscle paralysis, the most dramatic quality of SP is its sensory
content, characterized by vivid, intrusive audio-visual and somatosensory imagery. The experience of SP can
be extremely realistic, have a quasi-perceptual and wake-like quality, and may be accompanied by tactile and
kinesthetic sensations. Reflective thought processes, self-awareness and metacognitive abilities seem to be
relatively preserved during SP experiences, and people who have had multiple SP experiences may develop a
"feel" for recognizing SP imagery.
SP-associated experiences are typically referred to as hallucinations (hypnagogic, when occurring at
sleep onset, or hypnopompic, when happening upon awakening), since these occur during otherwise
seemingly awake consciousness (Liddon, 1967; American Academy of Sleep Medicine, 2014). This entails that
a person who experiences SP sees something that is not there, something that is distorted or false. Such a
view presupposes that during SP one is effectively awake and is misinterpreting her experience. Another way
of looking at SP is to situate it within the spectrum of dream mentation and dreaming imagination. While
dreaming too has been seen as a delusional/hallucinatory activity, an alternative view, in line with embodied
mind theories and enactive approach, has also been proposed: "When you hallucinate, you seem to perceive
2 Excerpt from the dreamer's account: "The transition between wake and sleep is a crucial moment to enter into the world of dreams.
During this transition on countless occasions I would awaken suddenly not being able to move. During this experience it seems that the
very essence of fear permeates my consciousness. Eeriness goes through my soul, freezes my blood and interrupts all substantial notion
of my being. No words can describe that visceral sensation, and in parallel, no words can come out of my mouth. Aware of the lack of
muscle tonus, I try to escape this inevitable Machiavellian black beast, that materializes in my head before my eyes and on my chest,
slowing down my breathing. Moreover, my senses are grabbed by an impression of fighting a hurricane that may drag me out of my body.
This cyclone, that has a black hole in lieu of an eye, forces me to fight it, and this fight seems crucial to my survival (…) In addition, there
are auditory experiences, an amalgam of petrifying words and vibrations that feel like sudden gusts of wind in the eardrums. All this
happens when by body feels like a statue, without a possibility of screaming.
(…) On occasion, with determination and lucidity, I can have power over this swirl of stillness. (…) I take back some control of my
imaginary hands, and then I hold them out to Morpheus for a dazzling and colorful dance in a deep and enlightened night". (Male SP
sufferer, also diagnosed with narcolepsy. Montreal, Canada. Account translated from French).
4
Kinds of sleep paralysis experience
what is not there. (…) When you imagine, you evoke something absent and make it mentally present to your
attention" (Thompson, 2014: 179). In this chapter I adopt this latter view and will refer to SP experiences as
variant of spontaneous thought/mentation or mental imagery, rather than hallucinations or delusions.
A factor analysis by Cheyne and colleagues (1999) showed that SP mentation typically falls into three
general categories. The first category is Intruder, and it is characterized by a felt presence, fear, as well as
auditory and visual imagery. The person who experiences SP feels that someone is in the house or in their
room. This experience is sometimes accompanied by seeing or hearing someone or something sentient move
around the house. The second category of SP experiences is known as Incubus, in which the felt presence is
interpreted as a supernatural assault and is often accompanied by a sensation of shallow breathing, a feeling
of being smothered, pressure on the chest, or pain. In this case, the sleeper often sees and feels the maleficent
being on top of her. The third category, Unusual Bodily Experiences, appears to be a separate, less well-known,
and a qualitatively different kind of SP experience: these are often described as positive events, such as
sensations of floating, out-of-body experiences (OBEs), and feelings of bliss. Both Intruder and Incubus
categories typically include the experience of felt presence – a distinct sensation that someone sentient is in
the immediate vicinity of the sleeper (Cheyne, 2005).
Most literature on SP focuses almost exclusively on the first two kinds, Intruder and Incubus, possibly
due to their particularly intensified felt presence imagery, which contributes to distressing SP experiences
(Solomonova et al, 2008; Cheyne, 2013). However, neutral and positive instances of SP have also been
described, and the third category, Unusual Body Experiences, or vestibulo-motor experiences, is often
characterized by pleasant sensations and a spirit of exploration, accompanying sensations of flying, out-of-
body experiences, or autoscopy (observation of one's own body from an unusual/novel point of view)
(Brugger, Regard & Landis, 1997).
figure is standing over the bed staring down at me, or pacing back and forth." (22 year old, gender not reported,
USA). Among all SP experiences, felt presence, the distinct sensation that another sentient being, human or
not, is present in the extracorporeal space of the experiencer, is arguably the most salient, terrifying, and
rich.. Felt presence is consistently reported as the most common SP-associated experience – about 80% of
episodes (Cheyne et al, 1999), which produces most fear and SP-related state of distress (Solomonova et al,
2008). One salient feature of felt presence experiences during SP is the fact that it is a distinct sensation, and
may occur in the absence of visual, auditory, or tactile imagery. Felt presence experiences during SP have
been classified as a paranoid delusion (Cheyne & Girard, 2007), an expression of spatial social imagery
(Nielsen, 2007), and as a variant of basic intersubjective experience of the world (Solomonova, Frantova, &
Nielsen, 2010).
Felt presence experiences are often interpreted within the cultural framework available to the
experiencer (see the following section on the cultural neurophenomenology of SP), but some basic
characteristics seem to be common across cultures and ages (Cheyne, 2001): 1) felt presence often manifests
from ambiguous stimuli: it is often described as "shadowy", and its physical characteristics are often unclear;
2) the experiencer may report a distinct sensation of being watched, and that the presence has some
intentions towards the dreamer; these range from some vague interest to full-blown assault; 3) felt presence
is usually accompanied by intense emotions (often fear when the presence is interpreted as threatening),
sometimes to the point of a distinct feeling of dread, imminent death, or being in the presence of evil. Positive
emotions, however, are also possible, especially when the experience is understood as visitations by deceased
relatives or visions of the divine.
Consider the following examples of felt presence experiences of the Intruder type: A 26-year-old man
from the United States reports: "It felt as if someone was watching me but silently standing behind me". In
this example the presence is felt in a distinct and clear way, but not seen or heard, yet the experiencer knows
where in space the presence is located. Similarly, a 29-year-old woman from USA regularly experiences the
"Just before going to sleep or if awoken suddenly I feel as though a presence, usually a dark shadow
Felt Presence
Intruder
5
Incubus
malevolent presence without ever seeing it: "…feeling of evil that is watching or monitoring; never able to
actually see this "evil entity". Even in the absence of direct visual, auditory or tactile imagery she feels that
she is observed and that the presence is "evil". The ambiguous qualities of the physical attributes of SP
visitors can be illustrated by the following two examples. A 39-year-old man from USA writes: "The
"presence" is a tall black/darkest grey shadow of a human form without any features. It stands in the
doorway to my bedroom waiting to be "noticed". Likewise, a 30-year-old woman experienced various ways in
which the presence was manifesting during her SP attacks: "Once it seemed a shadow was leaving the room.
One other time the shadow seemed to have "wild" hair or if it doesn't have hair at all, it looked as some sort of
black something".
The Incubus experience happens when the Intruder physically oppresses the sleeper, sometimes in a
rather dramatic way. In words of a 52-year-old man from the United States: "My worst experience was being
choked by a man who burst into my bedroom. The experience was so real and frightening that I was very
afraid of my SP for many months after." The Incubus takes many forms, including human, supernatural and
more rarely, animal: "I often hallucinate creatures like large cats - lions or tigers, … wrapping themselves
firmly around me and crushing my body" writes a 20-year-old woman from England.
Some of the most dramatic and potentially traumatic SP Incubus experiences are instances that are
lived as sexual assault or alien abduction. Consider the following example, reported by a 40-year-old man
from the United States: "When it is a "Dark Man" episode, he most likely touches me. Either by laying across
my body, in a sexual way or in the beginning, he would grab me and drag me. I always felt that if I let go, he
would pull me out of my body". Similarly, a 29-year-old woman from Spain describes her distressing SP
experience: "…extreme terror, the feeling that air is dense and darker, that shadows boil and take shape… I
hear some low tone noises, voices, tactile feeling of grabbing, of naked cold skin, and, very rarely, a presence.
Very dark with round eyes, spider-like fingers, that laughs, messes up the bed, and makes me feel terror, with
some sexual approaches…". In a study linking reports of space alien abduction to SP episodes McNally and
Clancy present this case: "…female abductee… was completely paralyzed, and felt electrical vibrations
throughout her body. She was sweating, struggling to breathe, and felt her heart pounding in terror. When
she opened her eyes, she saw an insect-like alien being on top of her bed" (McNally & Clancy, 2005: 116).
While most easily recognizable and most commonly documented cases of felt presence during SP
have to do with a threatening and ominous "visitor", some evidence suggests, however, that the presence is
not always understood as hostile. Such experiences include perception of friends and family; visitation from
deceased relatives or benevolent spirits; and erotic encounters where the sense of presence is comforting. A
20-year-old SP sufferer from the United States writes: "Once or twice I have thought that my friend or
roommate was standing over me. I was confused but not afraid." Similarly, encountering deceased family
members in visions or in dreams can be experienced as a positive spiritual event, and possibly play a healing
role in processes of bereavement (Garfield, 1996; Belicki et al, 2003).
Finally, another rare kind of SP-related felt presence episode involves first an experience of
"someone there", and then a doubling of the dreamer's own body, a self-projection into the extracorporeal
space. Some individuals report that the felt presence entities are becoming an externalized view of
themselves: "Sometimes I feel that the presence is myself, that I can watch myself", reports a 21-year-old man
from Jamaica; "I switch to another world and I myself become a presence", writes a 19-year-old man from
Russia.
Most (if not all) SP episodes are defined by an altered experience of the body. These include simple
experience of muscle paralysis; sensations associated with supernatural assault, including touch, pressure on
the chest, or even choking; feelings of unusual vibrations or falling into a vortex; and out-of-body experiences,
including flying, falling, or moving around one's house.
One of the most salient features of SP is the REM sleep-related muscle atonia. The inability to move is
a striking and unusual experience for most individuals, and the mismatch between sensing the body and the
Positive felt presence experiences and doubling
Body experiences in sleep paralysis
6
loss of voluntary control over the body's movements may contribute to a range of somatosensory
experiences. As discussed above, some of the most intense SP episodes may involve a feeling of being
assaulted or touched by a supernatural entity. For instance, a 34-year-old man from USA describes the
following experience: "Felt my arms pinned across my chest in a strait jacket hold, felt hands on my chest
pinning me against a wall". Perception of not being able to fully breathe, often accompanied by feeling of
pressure on the chest, may be prevalent in as much as 57% of SP episodes (Sharpless et al, 2010).
Although most accounts of and research on SP experience have centered on paralysis accompanied
by terrifying mentation and by felt presence, not all SP experiences are characterized by imagery and many
are simply experiences of transient body paralysis during the transition between sleep and wakefulness,
without any other accompanying mental activity (American Academy of Sleep Medicine, 2001). Additionally,
SP episodes may be predominantly somatosensory in nature: Cheyne (1999) characterizes these experiences
as Vestibulo-Motor mentation.
Autoscopy, out-of-body experiences, vibrations, floating, falling and body doubling experiences
(Cheyne, 2002) are all possible within the SP framework due to its reliance on dream-supporting REM sleep
mechanisms. During a dream, especially a lucid dream (wherein one is aware of the fact that she is dreaming),
it is possible to have simultaneous experience of one's dream body and real body at the same time. Thompson
(2014) distinguishes between the dreaming self (I the dreamer) and the dream ego (I as dreamt) as two
coinciding modes of self-experience, which may sometimes be experienced in parallel. The dreaming self is
the sleeping self, it is the "I" of the waking life, now engaging in the practice of sleep and dreaming. The dream
ego, on the other hand, is the experiential self, immersed in the dream scenario. The I as dreamt is the
temporary "I" that takes on the first-person perspective as a subject (and sometimes an object) of the dream
world. Seen from this point of view, SP episodes may represent an intense experience of the dreaming ego,
lacking a dream body and temporarily "stuck" within her immobilized sleeping body of the dreaming self/I
the dreamer while experiencing dream-like mentation. This feeling of being stuck, coupled with awareness of
the overlap between states of vigilance, may then transform itself into a situation of perceptual doubling of
body imagery.
Contrary to most SP episodes with a felt presence component, some bodily experiences are described
in quite positive terms. For instance, a 20-year-old woman from England describes this characteristic of her
typical SP episode: "Generally, the experiences start with a low, pleasant vibration that moves through my
body in defined waves, from the feet up. I feel them most strongly in the throat and in my eardrums".
Out-of-body experiences are also relatively common in SP – as much as 39% of SP experiencers have
had one at some point (Cheyne, 2002). A 39-year-old woman from the United States writes: "I floated out of
my bed into the kitchen. But, as I floated over my bed, I saw like this beast figure crouched over on the front
of my bed. I floated over it down to the kitchen. That is where I saw this beautiful kaleidoscope-like leaves.
They were so vibrant … I then floated back to my room into my body". In this example there is a combination
of various SP characteristics: dream-like mentation superimposed onto the environment, a nocturnal visitor,
and an altered sense of the body.
SP experiences are also sometimes accompanied by false awakenings-dreams where one has a vivid
and realistic feeling of waking up in their own bed and engaging in usual activities only to realize that they are
still asleep (Buzzi, 2011). While false awakenings are typically characterized as dream experiences, their
phenomenology in terms of realism and possible state overlap is to a degree similar to SP. In Cheyne's report
(2002), 58% of people who experience SP also experienced false awakenings at least occasionally.
Additionally, false awakenings are often associated with feelings of dread, anxiety, and oppression (Green &
McCreery, 1994; Nielsen and Zadra, 2011), similarly to SP. The following two examples from our Internet-
based sample illustrate such cases: a 24-year old man from the United States reports: "… sometimes I think I
have moved... sometimes even gotten up and walked around only to find that I never got up at all." In a similar
vein, a 21-year old man from Jamaica describes his experience: "I will wake up into another dream inside my
bedroom and think I am awake and realize I am still sleeping minutes later and the same procedure repeats
several times".
communicate with those around me. (23-year-old woman, USA)
…Extreme anxiety and fear, mind is awake, but body is asleep. I feel as though I am trapped and cannot
Emotions
7
The most prevalent emotion associated with SP experiences is fear. Indeed, the most natural reaction
to waking up unable to move is panic, and the sensation of constricted breathing (consistent with REM sleep
physiology) may increase the state of distress. Sharpless and colleagues (2010) introduced the term fearful
isolated sleep paralysis to denote SP experiences characterized by an intense state of distress.
As much as 90% of reported SP episodes are described as fearful (Cheyne, Rueffer, & Newby-Clark,
1999). Similarly, in an Irish University students' sample, fear was found to be the most prevalent emotion,
with 82% of respondents stating that they have experienced fear at some point during a SP episode
(O'Hanlon, Murphy, & Di Blasi, 2011). Moreover, nightmare frequency was previously reported as a predictor
of SP occurrence (Liskova et al, 2016). This data suggests that SP, or at least the Fearful form of SP, can be
seen as an intensified form of a nightmare: a recent study by Robert and Zadra (2014) reported that about
65% of nightmares and 45% of bad dreams are characterized by fear.
One approach to classify the affective and personal impact of SP experiences is to assess not only
frequency or intensity of SP episodes, but also distress associated with SP experiences (Solomonova et al,
2008; Cheyne & Pennycook, 2013). To what extent is the individual affected by SP? To what extent do
negative emotions carry-over from an SP episode into waking life? Do SP experiences promote a negative
relationship with sleep? These questions have been successfully examined in previous research on
nightmares (Belicki, 1992; Blagrove, Farmer & Williams, 2004), showing that the individual impact of
negative and intense dream experiences depends more on a trait-like reactivity, sometimes referred to as
affect distress (Nielsen & Levine, 2007). This trait is thought to represent a general dysfunction of affect
regulation network, and it has been shown to be a better measure of how much nightmares influence waking
life emotional well-being than frequency or intensity of self-reported nightmare occurrence. Furthermore,
affect distress mediates reactivity, negative interpretation and degree of negative reaction to nightmares
(Belicki, 1992; Levin & Fireman, 2002). According to Nielsen and Levine (2007; Levine & Nielsen, 2009),
dreaming helps regulate emotional memory consolidation and emotional reactivity via fear extinction.
Nightmares, therefore, represent a case of problematic/dysfunctional processes of fear extinction. In
combination with other factors, affect distress is likely to play a role in formation, experience and
interpretation of SP.
Positive emotions associated with SP are much less studied, and it is not possible to accurately
estimate their prevalence. One possible reason for this is lack of appropriate screening (SP is often diagnosed
as an unpleasant phenomenon) and lack of medical/psychiatric concern: patients are not very likely to
describe such experiences to their health practitioner, since they are not bothered by them. In addition, the
current diagnostic criteria for a recurrent isolated SP as listed in the latest edition of the International
Classification of Sleep Disorders – 3d edition (American Academy of Sleep Medicine, 2014), include that the
episodes must cause "clinically significant distress including bedtime anxiety or fear of sleep". Such a
provision would effectively exclude all possible positive and non-distressing SP phenomena from
investigation and/or diagnosis. Nonetheless, in a web-based SP study Cheyne (2002) reports that in addition
to anger (30% of respondents) and sadness (23%), bliss (17%) and erotic sensations (17%) are also
sometimes present in SP.
Felt presence is the most prevalent, the most emotionally disturbing, and the most salient SP-related
experience. Therefore, it is unsurprising that most visual and auditory mentation during SP usually has
something to do with these unwelcome visitors. The entities, however, while felt in a very distinctive and
concrete way, are often described visually as rather general and vague shadowy beings. Visual experiences
are reported to occur in 54%– 56% and auditory experiences in 55%-60% (Solomonova et al, 2008; Cheyne,
2002) of SP sufferers.
SP may be accompanied by auditory experiences, ranging from abstract and mechanical sounds, such
as electric sounds and sounds of buzzing, to vivid auditory imagery, consistent with SP experience of an
Intruder or an Incubus. Sounds of footsteps and of voices are often reported (Cheyne, Rueffer & Newby-Clark,
1999; Cheyne, 2002; Solomonova et al, 2008).
Visual and auditory experiences
8
Cultural grounding of SP
Figure 2. Henry Fuseli. The Nightmare. 1781. Detroit Institute of Art. Public Domain image:
https://commons.wikimedia.org/w/index.php?curid=15453518
While sleep paralysis is a lesser-known sleep phenomenon in the West, it is quite prevalent and is
well-described in many other cultures. Due to the lack of general awareness of SP in the West, it is rarely
discussed in the context of family medicine or psychology. Cross-cultural work on SP revealed that it is rooted
in a variety of religious beliefs and cultural schemas, including interpretations of the experience and
techniques to engage with the nocturnal visitors. Some of the common qualities of SP across cultures (Adler,
2011) include: 1) sensation of being awake; 2) perception of the environment; 3) paralysis; 4) feeling of fear
and dread; 5) felt presence; 6) chest pressure/breathing difficulties; 7) supine position; 6) various unusual
body sensations. These apparently culturally-invariant qualities of SP-related occurrences of the experience
of a supernatural attack have been at the center of the phenomenological and cross-cultural cognitive
research on SP.
Figure 2 is a reproduction of an eighteenth-century work by Henri Fuseli entitled The Nightmare. It
represents a sleeping woman in a supine position being oppressed by a maleficent creature sitting on her
chest and with an ominous presence of the night-mare. It is likely that the early use of the English term
nightmare was to describe intense SP (Orly & Haines, 2014). Culture-specific presentations of SP-related felt
presence experiences typically involve a maleficent supernatural being, such as a witch or an evil spirit. Some
examples found across cultures include the kanashibari demon in Japan (Fukuda et al, 1987; Arikawa,
Templer & Brown, 1999); kokma in the West Indies (Ness, 1983); "old hag" in Newfoundland (Hufford, 1989);
pandafeche in Italy (Jalal, Romanelli & Hinton, 2015); uqumangirniq among the Inuit of Baffin Island (Law &
Kirmayer, 2005); and many others (for a comprehensive list of terms for SP experiences see Adler, 2011).
Figure 3 illustrates a possible SP representation (Orly & Haines, 2014): a Japanese demon Yamachichi
oppresses and inhales the breath of the sleeper.
The first systematic cultural exploration of SP was done by Hufford (1989): he described a
phenomenon specific to Newfoundland – the "old hag" witch attack. In his book Hufford discusses the tension
in situating SP experiences somewhere between the 'cultural source hypothesis', wherein cultural
interpretations and framing influence how an experience unfolds, and the 'experiential source hypothesis',
where some invariant lived experiences, such as SP, may influence the development of a spiritual
9
interpretation and formation of cultural beliefs (Hufford 1989, 2005). Similar to this notion, McNamara and
Bulkeley (2015) proposed an experiential hypothesis to describe how dreams and other dream-associated
experiences, including visions and transcendental experiences, can be seen as a cornerstone and a source of
religious belief (McNamara & Bulkeley, 2015). According to this view, a number of cultural, religious and
paranormal beliefs are shaped primarily by direct experience and then framed within a particular tradition,
which imbues them with existential and metaphysical meaning, a notion that is reminiscent of William James'
grounding of mystical experience in the phenomenology of lived experience (James, 1985).
The effect of framing such intense subjective experiences within a cultural tradition can have at least
two kinds of potentially opposing effects. On the one hand, many cultures provide not only supernatural
explanations of SP, but also remedies and protective rituals against it (some of which are described in the
later part of this chapter), thus rooting the SP in a framework which allows for shared narrative and for
practical interventions. On the other hand, intense and fearful SP, when interpreted as supernatural assault,
has a potential for traumatizing the sleeper, thus creating a vicious circle of anxiety, aversion to sleep,
facilitation of future SP episodes (Hinton et al, 2005; Sharpless et al, 2010), and increasing the level of distress
via "cultural fear priming" (Ohayon et al, 1999; Jalal, Romanelli, & Hinton, 2015). For instance, the Incubus
experience, when seen as part of the Christian tradition starting with the late Antique period, according to
Gordon (2015), gained additional stigmatizing power, with a connotation of an illicit supernatural sexual
experience. Not only were SP victims living through a waking nightmare of an encounter with a demonic
assailant, they were also seen as responsible for having summoned it due to their own sinful
predisposition/thoughts/impurities.
It is important to note that while SP can include a range of experiences, such as positive experiences,
neutral emotions, vestibulo-motor phenomena, out-of-body experiences, and others, most cultural
interpretations of SP deal specifically with overlapping aspects of Intruder and Incubus.
Figure 3. Takehara Shunsen. Yamachichi. Public domain image:
https://commons.wikimedia.org/w/index.php?curid=2074508
Human sleep is typically divided into two kinds: REM sleep and NREM sleep. Healthy adults alternate
between NREM and REM in cycles lasting about 90 minutes, for a total of 5-6 cycles over a night of sleep.
Neural basis, associated conditions, and precipitating factors
10
While it is possible to experience dreaming in all stages of sleep, REM sleep is typically characterized by the
most vivid, realistic, bizarre, and emotionally intense sleep mentation (Nielsen, 2000). Other vivid dream
experiences, such as nightmares (Nielsen & Levin, 2007) and lucid dreams (LaBerge, Levitan & Dement,
1986), are also typically associated with REM sleep.
Within the context of narcolepsy, sleep paralysis is a part of the diagnostic tetrad, alongside daytime
sleepiness, cataplexy, and hypnagogic hallucinations (Thorpy, 2016). There is not sufficient data to assess
whether there are significant differences in phenomenology between narcolepsy-associated SP and the
isolated form.
SP episodes are characterized by simultaneous presence of waking thought and of REM sleep
psychophysiology (Mahowald & Schenck, 1991, 2005; Terzaghi et al., 2012), and the sleeper can often open
her eyes and become relatively aware of her environment, while REM sleep-related spontaneous mentation -
vivid dreaming - superimposes onto otherwise awake consciousness. This imagery may occur at sleep onset
(hypnagogic) or upon awakening (hypnopompic). Other characteristics of REM sleep, such as airway
occlusion and rapid shallow respiration (Gould et al., 1988) may contribute to the feeling of being suffocated
or the perception of shortness of breath often reported by SP sufferers. Additionally, in one study obstructive
sleep apnea was found to be a possible precipitating factor for ISP (Hsieh et al, 2010).
Little research has been done on the sleep characteristics of SP sufferers. Some preliminary data
suggests that the SP sleep profile may be similar to that of frequent nightmare sufferers (Nielsen et al, 2010),
in that SP participants appear, paradoxically, to exhibit less REM sleep pressure, have more "skipped" REM
sleep periods, and show no increase in eye movement density (as opposed to healthy controls) throughout
the night (Solomonova, et al, 2012). SP participants also show higher delta power during sleep than non-SP
controls (Marquis et al, 2015), which suggests alteration of processes of wake-NREM-REM regulation.
Some of the vestibulo-motor characteristics, such as autoscopy, out-of-body experiences, and feelings
of physical transformation, may stem from disturbances in right parietal regions (Jalal & Ramachandran,
2014): the mismatch between intended motor movement and inability to move may contribute to unusual
physical sensations.
SP may be experimentally elicited in laboratory settings, but only using an arduous protocol of
repeated sleep interruption. For example, SP episodes were experimentally induced by letting participants
sleep uninterrupted for the first NREM period, thus eliminating most of the slow-wave sleep pressure (a
tendency of slow-wave NREM sleep to take precedence and occupy a large proportion of early night sleep),
and then repeatedly awakening participants after 5 minutes of REM sleep have elapsed, thus augmenting
REM sleep pressure and facilitating sleep-onset REM periods (SOREMPs). SOREMPs may be seen as a
facilitating factor in a REM-wake state dissociation thought to characterize SP experiences. It should be noted,
however, that even within such controlled settings and demanding protocols, rates of SP were relatively low:
6 episodes total in 16 participants who already had a tendency toward recurrent isolated SP (Takeuchi et al.,
1992), and 8 episodes from 184 sleep interruptions in 13 SP sufferers (Takeuchi et al, 2002). These results
suggest that incidence of SP at sleep onset may signify an individual's propensity to enter into REM sleep
directly upon falling asleep. This further supports the idea that SP may result from alterations in wake-REM-
NREM regulation patterns, resulting in state overlap.
Little is known about the epidemiology of SP, but growing evidence points to a combination of
genetic and experiential factors. The only study to date to examine genetic factors associated with SP has
reported moderate heritability and that this effect was associated with factors known to contribute to
disrupted sleep cycles (Denis et al., 2015). Sleep fragmentation and disruption in wake-NREM-REM
regulation are an important factor facilitating SP occurrence, but it is uncertain whether all types of SP can be
explained by a propensity for sleep fragmentation. Some ethnic groups seem to be more likely to experience
SP than others. The Hmong population in Wisconsin, for instance, had a significantly higher incidence of SP
than a non-Hmong cohort (Young et al, 2013), with as much as 31% of interviewed Hmong participant
reporting at least weekly occurrence of SP episodes. Individuals of African descent also seem to have elevated
rates of SP (Bell et al, 1984; Friedman & Paradis, 2002).
Links between affective disorders, especially depression and anxiety, and SP have also been reported.
A relationship have been found between SP and depression magnitude and anxiety (Szklo-Coxe, Young, Finn,
& Mignot, 2007), social phobia and panic disorder (Paradis & Friedman, 2005; Otto et al, 2006; Sharpless et
al, 2010) and social anxiety (Simard & Nielsen, 2005), especially with the sensation of being observed
Associated conditions
11
(Solomonova et al, 2008). Changes in REM sleep regulation are often found in mood disorders, especially in
depression (Arargun & Cartwright, 2003; Nofzinger et al, 1994)
The relationship between trauma, especially post-traumatic stress disorder (PTSD), and SP has been
noted by a number of researchers. McNally and Clancy found that there was a higher proportion of SP reports
in participants with a history of childhood sexual abuse (McNally & Clancy, 2005a), and Abrams and
colleagues (2008) reported that sexual abuse survivors report more distressing and more frequent SP
incidence. In addition, higher rates of SP were found in Hmong population in relation to traumatic Vietnam
War experiences (Young et al, 2013), as well as in Khmer (Hinton et al, 2005a) and Cambodian refugees
(Hinton et al, 2005b). Similarly, Sharpless and Grom (2013) report that some cases of SP onset in adolescents
begin after the loss of a family member. Considering that SP may be conceptualized as a nightmare spectrum
experience, this relationship may represent the same dysfunction in the affect regulation network (Levin and
Nielsen, 2007; Nielsen & Levin, 2007) as the one that has been proposed to be involved in nightmare
production. PTSD-related sleep disturbances have been extensively documented (Spoormaker &
Montgomery, 2008; Germain, Buysse & Nofzinger, 2008), including REM sleep dysregulation and increased
nightmares (Melman et al, 2002; Germain, 2013), which in itself may contribute to altered REM sleep
pressure, in turn facilitating occurrence of SP episodes.
Since SP is often associated with intense, detailed, and troubling visions, a link between SP and
psychiatric disorders has been hypothesized. Research, however, shows no consistent relationship between
psychiatric conditions and SP, with the exceptions of PTSD, panic disorder, and social anxiety. In one study a
number of links between SP and psychiatric conditions were found (Ohayon et al., 1999); these findings,were
challenged, however, by an internet-based study (Solomonova et al., 2008), with a larger sample size, in
which no strong links between psychopathology and SP were described. However, while isolated SP often
presents itself in the absence of psychopathology, higher rates of hypnagogic and hypnopompic experiences
(dream experiences occurring during the transition between sleep and wake: at sleep onset or upon
awakenings, respectively), some of which may be associated with SP, are often found in psychosis (Plante and
Winkelman, 2008).
In their recent book, Sharpless and Doghramji (2015) list a number of plausible precipitating factors
for SP occurrence in susceptible individuals. Sleep fragmentation and insufficient sleep are among the most
obvious factors. REM sleep deprivation has been shown to increase REM sleep pressure contributing to REM
rebound effect and intensified dreams at sleep onset (Nielsen et al, 2005). Poor sleep quality with frequent
awakenings and disruptions may also facilitate REM-wake overlap, creating fruitful conditions for the
occurrence of SOREMPs (Takeuchi et al, 1991, 2002; Spanos et al, 1995). Shift work, jet-lag, use of sleep
disrupting medication, stress, anxiety – all these factors affect sleep and may facilitate a SP episode. Alcohol
consumption was also reported to promote SP (Golzari & Ghabili, 2013; Munezawa et al, 2011), probably due
to its effect on altering sleep architecture (Roehrs & Roth, 2001). Sleeping in a supine position also appears to
enhance the risk of a SP episode (Sharpless et al, 2010).
While undoubtedly felt presences are a hallmark of SP, especially of the intense and frightening
episodes, presence experiences are not restricted to this parasomnia and are reported in a variety of
conditions, thus possibly representing a more general and basic social imagery process (Nielsen 2007,
Solomonova, Frantova & Nielsen, 2011). Arguably, the most salient and compelling felt presence occurs in the
context of mystical and spiritual experiences. Otto (1958) introduced the idea of the numinous as a
cornerstone of religious mystical experiences. Some of the recent work comes from anthropology: the ecstatic
presence of God is manifested in the community of Evangelical Christians in the USA (Luhrmann, 2012). Other
examples of felt presence have been documented in situations that are physically and emotionally straining
or novel. Some examples of these experiences include high altitude climbing (Brugger, Regard, Landis, & Oelz,
1999); feeling of the presence of a baby in postpartum mothers (Nielsen & Paquette, 2007); presence of
deceased relatives in the context of bereavement (Simon-Buller, Christopherson & Jones, 1989; Taylor, 2005;
Keen, Murray & Payne, 2013); in extreme environments, such as solitary sailing (see also chapter by Suedfeld
Neurocognitive considerations
Precipitating factors
A return to felt presence
12
Toward a cultural neurophenomenology of SP
in this volume), surviving in remote and hostile environments (Suedfeld & Mocellin, 1987), and others. While
in most cases felt presence is experienced spontaneously, in some cases it may be a product of sustained
mental practices (as in prayer and some forms of meditation). One contemporary non-religious phenomenon
is tulpamancy (Veissiere, 2016) – a long-term practice of conjuring up imaginary companions, that, over time,
may be experienced as almost as real as other people.
Additionally, being able to have a felt sense of others may be seen as a prerequisite for the
development of subjectivity. Recent work in phenomenology and enactivism suggests that development of
sense of self depends crucially on sensing others, as early as in utero (Gallagher, 2005; Ammaniti & Gallese,
2014), that the sense of one's own body depends on the sense of others (Maclaren, 2008) and that the self-
other dynamic is a necessary condition for the sense of self (Zahavi, 2014). Evidence from dream research too
suggests that dream processes are relational and intersubjective. The fact that dreams are most often about
other people has been conceptualized as simulations of social interaction (Revonsuo, 2016) and as
representations of individual attachment styles (McNamara et al, 2001). Additionally, dreams, similarly to
waking, can be seen as a dynamic interaction between the "self"-related and "non-self" elements of dream
content (everything extraneous to the dreamer). These non-self elements (non-human characters, dream
environment, even dream objects can be seen as a "dream other" due to their inherent relational property
(Solomonova et al 2015) and to the fact that dream environment in its entirety affectively motivates dreamer
to engage with it.
.
SP has often been characterized as dissociative (Terzaghi et al., 2012) state, since it effectively
combines characteristics of 'waking' consciousness (self-awareness, access to autobiographical memory,
ability to open eyes and perceive the environment) with REM-sleep phenomena, specifically muscle
atonia/paralysis and mentation/dreams. This notion of SP as dissociative has been at the heart of the
previous neurobiological work on the link between dreaming and REM sleep. The relative deactivation of the
dorsolateral prefrontal cortex characteristic of REM sleep (Hobson, Stickgold & Pace-Schott, 1988; Maquet,
2000) has been long hypothesized to be at the root of the loss of autobiographic memory and of the inability
to appreciate the contents of the dream as "bizarre" or implausible in relation to reality. This has led to the
hypothesis that in REM sleep dreaming one is effectively delusional and in a state of a transient psychosis
(Hobson, 2004). In SP, similarly, there is often incomplete autobiographical access. This association between
SP and REM sleep has also displaced the experience of SP from the psycho-spiritual domain of meaningful
encounters with menacing/unreal/supernatural others, into a more reductionist account of uncontrollable
and inescapable REM-initiated hallucinations.
In contrast, an account of SP in the context of an oneiric phenomenology and in a 4EA perspective
may allow for a more nuanced reading of these experiences. An emerging neurophenomenological framework
of sleep challenges strict distinctions between wake, NREM, and REM sleep. Indeed, while SP is one of the
examples of simultaneous presence of REM sleep and wake processes, it is not the only phenomenon that
attests to the fluidity and interpermeability of states of consciousness. Lucid dreaming is another example of
REM-wake co-occurrence (LaBerge, 1986); REM sleep behaviour disorder is characterized by preserved
motor output during REM dreaming (Peever, Luppi & Montplaisir, 2014); somnambulism episodes combine
NREM and wake physiology and phenomenology (Zadra et al, 2004); and a variety of dream-enacting
behaviours, such as laughing, simple movement, crying and looking for a baby in bed, are prevalent in normal
populations (Nielsen & Paquette, 2007; Nielsen, Svob & Kuiken, 2010).
A more continuous view of mentation in sleep includes viewing SP as a form of oneiric experience: as
a process of intensified mind-wandering (Fox et al, 2013), as a process of creativity (Hartmann & Kunzendorf,
2013), or as enactive imagination (Thompson, 2014), a process of sense-making in a rich, embodied and
intersubjective world (Solomonova & Sha, 2016). In his discussion of lucid dreaming, Thompson (2014)
proposes that in addition to seeing this state as a dissociative superimposition of two distinct states of
consciousness, it may be simultaneously approached as an integrative state, thus allowing for an integration
of two different yet related ways of self-experience.
While SP sufferers feel awake and in their own bed, the realism of the experience and the quality of
total immersion are completely overpowering to the dreamer, so that she is unable to appreciate the
dreamlike quality or the unreality of the SP episode. The high prevalence of tactile and physical sensations
13
waking. But their presence seems so real, I would compare the experience to having them
bizarre or terrifying event is happening all around me, and I am completely unable to respond
or defend myself. Sometimes I know it's not real, somewhere in my mind, but it looks real, and it
sounds real, and I'm terrified or revolted (or maybe just bemused), but I cannot wake myself up
probably contributes to this effect. There are, however, numerous accounts of long-time SP experiencers that
are characterized by a certain 'feel' for the experience as somewhere between real and unreal. SP-related
experiences may have a very compelling and realistic quality, but they are usually lived differently from
waking experiences, as a kind of a liminal state.
Consider the following example: while the participant is experiencing intense emotion and is quite
absorbed in the unfolding on the SP, he seems to have a kind of a dual awareness regarding the nature of his
SP:
"…Can't. Move. Not a muscle. Not an eyelash. It's often accompanied by hallucinations. So this
to stop it." (30-year-old man, USA)
Similarly, in another example the experiencer is also hesitant to ascribe any particular state to her
experience:
"… I might be answering wrong, because I see the beings in my dream-state immediately before
accompanying me in the room". (48-year-old woman, USA)
Grounding SP in its cultural context allows us to appreciate the variety of factors contributing to
qualities of the lived experience, and it may not be possible to dissect the relative contribution of the
multitude of neural, phenomenological and cultural narrative factors (Kirmayer, 2009). Importantly, in the
current medical context, reducing SP to a dysfunction of REM psychophysiology may also have an important
effect on reducing the potential for a deeper exploration of SP as a spiritual experience (Hufford, 2005).
The cultural neurophenomenology of SP is a powerful tool for investigating SP from the 4EA
cognition perspective. As neurophysiological, experiential accounts of SP show, the dreamer is in fact
embodied – the oneiric scenario is dependent on the dreamer's state of consciousness (REM intrusion) and on
the dreamer's physiological state (atonia, shallow rapid breathing). She is embedded in a physical (interprets
ambiguous stimuli around her) and in a cultural world (these ambiguous stimuli take on a familiar shape/are
infused with a deeper cultural and interpersonal signification). The sleeper is also extended into the world –
the whole environment, both dreamt and real, is part of her ongoing experience; and her experience is
enactive – there is a relational quality: she is not a passive observer of the oneiric drama unfolding before her
eyes, but rather she is deeply engaged (Solomonova & Sha, 2016).
In order to elucidate neurophenomenological qualities of SP in greater detail, future work may use
microdynamic phenomenology/elicitation interviews, aimed at uncovering the fine-grained temporal and
structural qualities of lived experience (Nielsen, this volume; Petitmengin, 2006; Petitmengin & Lachaux,
2013), in addition to neurophysiological data and deep awareness of the cultural, religious and spiritual
context of the experiencer.
While SP remains a relatively unknown phenomenon in much of Europe and America, a number of
practical culture-specific practices have been developed to protect the sleeper from the negative influence of
presumed supernatural forces. While some of these methods have deep roots in their respective metaphysical
contexts, and therefore need to be grounded in existing religious and mystical practices, a number of practical
and conceptually neutral recommendations have emerged, and seem beneficial for most SP sufferers,
regardless of background.
No established treatment for SP currently exists; its clinical management is instead often focused on
treating comorbid problems. According to a review by Sharpless & Doghramji (2015), psychoanalysis,
cognitive-behavioural therapy (CBT), hypnosis, and education in sleep hygiene have been investigated in
relation to SP, but no empirical consensus on efficacy of such interventions is currently available. Based on
the available evidence on SP and cognitive-behavioural approaches to treatment of sleep disorders, especially
insomnia, the authors propose a manual for CBT-ISP. This is a promising first step toward finding a
systematic method of dealing with SP. Sparse evidence for pharmacological interventions for SP also exists: in
one study it was suggested that REM sleep-suppressing antidepressants may provide temporary relief (Plante
and Winkelman, 2008), and treatment of narcolepsy may reduce SP frequency (Mamelak et al., 2004).
Sleep paralysis practices: prevention, disruption, treatment and exploration
14
Antidepressants and anxiolytics were also used in severe cases (Hsieh et al., 2010). Terrillon and Marques-
Bonham (2001) proposed that management of SP might benefit from administration of melatonin, which
would help normalize the circadian rhythm. The cost of side effects associated with these treatments,
however, may outweigh the benefit, and Shapless and Doghramji (2015) argue for a cautious approach,
tailored to each individual situation.
While methods for dealing with sleep paralysis have not been systematically explored by empirical
psychology or cognitive science, the contemporary context of Internet-facilitated support groups and
information sharing practices are changing the solitary and culture-bound nature of SP attacks. Furthermore,
a number of methods have been anecdotally reported and documented online and in print, that see SP
experiences as an opportunity rather than a nuisance, and promote exploration of one's own consciousness
via SP-supported lucid dreaming or even contemplative approaches to SP (Hurd, 2010). One popular support
group-mailing list is known as "Awareness during sleep paralysis" (ASP), and a reddit group on SP counts
over 4000 users, sharing information on the phenomenology of their experiences and methods of overcoming
them. Cultural and clinical practices associated with SP can be roughly separated into three kinds: 1)
preventative practices, focused on avoiding SP-enabling circumstances; 2) disruptive practices, designed to
stop SP in the middle of the experience; and 3) observational/explorative practices, aiming at observing SP
and possibly transforming it into a positive event, such as a lucid dream or an out-of-body experience.
Raising awareness of SP-associated phenomena itself may be one of the most important factors in
reducing fear and distress before, during and after SP occurrence (Otto et al., 2006; Sharpless et al, 2010).
Indeed, knowing that the experience is transient (will not last), benign (does not contain any real danger),
and common (is shared with many individuals across the world) are powerful tools for psychological
distancing and for facilitating an eventual observational, as opposed to fully immersive and fatalistic, attitude
toward SP. Knowing about SP phenomenology and neurophysiology and having access to cultural grounding
with available symbolic gestures helps prevent, disrupt and transform a negative experience into a tool for
self-exploration. Figure 4 summarizes the intricate links between precipitating factors and effects of SP
experiences in light of disruptive, and observational/transformational practices.
While undoubtedly helpful, simply knowing the basis of SP may not be enough to alleviate terror and
distress associated with the experiences, and disruption techniques are clearly warranted. A 25-year-old man
from the United States reports: "This happens sometimes every night, sometimes only once every few weeks.
Even though I 'know' what is happening, and that I am in no danger, it is always terrifying". The first study to
systematically assess prevention strategies for SP by Sharpless and Grom (2014) has suggested that while no
foolproof method for preventing SP is yet known, some strategies, such as avoiding sleeping on one's back
(supine position), maintaining optimal sleep hygiene (avoiding stimulants, noise, irregular sleep patterns and
anything that contributes to sleep fragmentation), and pre-sleep relaxation practices may help in preventing
SP.
A number of culture-specific preventative ritualistic measures to prevent SP exist. These include
placing a variety of defensive objects in the room or in the bed before going to sleep, such as a variety of
knives (Hufford, 1982, Law & Kirmayer, 2005); sprinkling salt (a common anti-witch remedy) (Roberts,
1998); putting a broom bottom-up (Paradis & Friedman, 2005) or a pile of sand at the bedroom door (Jalal,
Romanelli & Hinton, 2015); and many others. Putting a Bible in the room (Hufford, 1982) and saying a
protective prayer before bedtime are also thought of as effective deterrents. Other ritualistic actions,
designed to deter, divert and chase away unwelcome supernatural visitors were also documented in a variety
of contexts (Sharpless & Doghramji, 2015).
While preventative measures, whether culturally embedded or aimed at increased awareness and
promotion of sleep hygiene, may be effective in reducing the frequency of SP episodes, many methods for
dealing with an ongoing SP experience also exist. Considering that most SP experiences are characterized by
Methods for preventing sleep paralysis
Techniques for disrupting sleep paralysis
15
Observational/transformational practices
fear and other unpleasant sensations, it is not surprising that in one study the majority of participants
reported having attempted to disrupt the ongoing SP experience. Moving the extremities and self-monitoring
(raising awareness, promoting calm) may be helpful during the SP episode (Sharpless & Grom, 2014). Not all
attempts or all strategies are equally successful, but it seems that attempting micro movements, instead of
trying to get up or to scream, are most effective. Culture-bound rituals include saying a prayer (Hufford,
1982), making a sign of a cross with one's tongue (Davies, 2010), and asking someone to physically shake the
oppressed sleeper (Law and Kirmayer, 2005).
One may argue that ISP and lucid dreaming are polar opposites. However, they share the same
underlying psychophysiology and seem to involve similar mechanisms: both are dependent upon REM sleep
mechanisms; both are characterized by simultaneous presence of the dream state and by the feeling of being
awake, including activation of higher order metacognitive functions indicative of some degree of waking
thought processes (LaBerge, Levitan & Dement, 1986; Voss, Holzmann, Tuin, & Hobson, 2009; Dresler et al,
2012; Filevich et al, 2015); and in both cases muscle atonia is present. The crucial difference between the two
states is the quality and the focus of awareness and metacognition: in lucid dreaming one is aware of the
illusory nature of the dream scenario, whereas in SP the dreamer is often absorbed by the vision, not always
fully realizing that it is dreamlike, and, in case of fearful SP, is too absorbed in the panicky state of perceived
imminent danger.
The link between SP and lucid dreaming has not been systematically investigated in empirical
research, but two studies report a positive correlation between frequency of lucid dreaming and SP (Denis &
Poerio, 2016; Solomonova, Nielsen & Stenstrom, 2009), suggesting that the REM-wake intertwined state,
characterizing SP, may be a trait-like phenomenon predisposing individuals to SP on the one hand, and
facilitating lucidity in REM sleep dreams on the other.
Transforming SP into a positive experience, such as an OBE or a lucid dream, or utilizing SP
experiences as a means of contemplative insight into one's own mind, may become a practice in itself, since
not only techniques for disrupting and preventing SP exist in the contemporary digital culture, but also
techniques for inducing SP, with the hope that the experience will function as a portal to a desirable altered
state of consciousness (Hurd, 2010). The following two reports illustrate the transformative potential of SP:
"I have woken up from dreaming and found I can't move or open my eyes. I get the feeling of
lemonade bubbling in my body, especially my head. It is very frightening. But since I have been
having OBE3s I now relax and go with the flow of sleep paralysis and sometimes I actually
achieve an OBE" (40-year-old man, Australia)
ignore and now I've been trying to communicate with the presence". (40-year-old man, USA)
"At first I was very frightened until I found the ASP email group and found that I was not the
only one being "visited" by this being during sleep paralysis. … When it first started happening
it was more of an assault and I had to fight terribly to escape. But after years, I learned to
3 OBE = out-of-body experience
16
Further considerations and future directions
Figure 4. Predisposing, precipitating factors and experience and outcome of sleep paralysis episodes.
In terms of possible avenues for treatment, since SP can be conceptualized as a form of nightmare
occurring in a mixed state of consciousness, nightmare treatment techniques could be useful in approaching
SP. Currently, the most used and recommended technique for treating chronic nightmares is the Imagery
Rehearsal Therapy (Krakow et al, 1995; Krakow & Zadra, 2006), which consists of "rehearsing" and
transforming dysphoric oneiric imagery in a safe context. This method has been effective in treating PTSD-
related nightmares (Krakow et al, 2001; Germain et al, 2004; Cook et al, 2010; Casement & Swanson, 2012),
which seems particularly appropriate for intense and trauma-related SP experiences. Similarly, treatment of
nightmares by lucid dreaming is a promising avenue (Zadra & Pihl, 1997; Spoormaker & Van Den Bout, 2006;
LaBerge, 2009). Considering that neurophysiologically both states are characterized by an overlap between
REM sleep and wakefulness, and that a number of folk approaches treating SP as a portal to lucid dreams
already exist, mastering lucid dreaming could be an effective approach to transformation of an ongoing SP
episode. Such a strategy may also be highly effective in de-stigmatizing and desensitising the experiencer, and
especially in increasing her mastery and agency over her spontaneous oneiric experiences.
Contemplative practices, such as meditation or pranayama (yogic breathing) may also be useful in
dealing with recurring SP episodes. There is currently no empirical evidence for contemplative techniques
and SP management, with the exception of a case study by Jalal (2016), but anecdotal evidence from
practitioners as well as growing empirical literature linking contemplative practices with stress management,
emotion regulation, and increased self-awareness, provide grounds for future research.
Recent years have seen an important increase in empirical studies on the effects of meditation and
meditation-based mindfulness interventions. There are documented benefits of contemplative practice in
clinical populations including positive effects in mood disorders such as anxiety and depression (Hoffman et
al, 2010; Goyal et al, 2014), social anxiety (Goldin & Gross, 2010), and PTSD (Kearney et al, 2013). At least
four kinds of meditation are currently investigated in relation to mental health: focused attention, open
monitoring (Lutz et al, 2008), self-transcendence (Travis & Shear, 2010) and loving kindness meditation
(Hoffman, Grossman & Hinton, 2011). Different kinds of meditation practices may recruit different neural
networks (Fox et al, 2016), and particular psychological and neuroplastic changes, associated with meditation
practice, likely depend on the kind and duration of meditation experience (Lutz et al, 2015). These different
kinds of contemplative practice may be helpful in targeting different kinds of recurrent SP experiences,
promoting de-automatization (Kang, Gruber & Gray, 2013): deconstructing patterns of behaviour/reactivity.
17
Meditation may be effective in SP management as a way of cultivating a non-judgemental or 'non-sticky'
observational attitude to arising imagery, sensations and emotions, and in letting the experience unfold. In
addition, one important feature of most mindfulness-related practices is the focus on the experience of the
body (Kerr et al, 2013), and some evidence suggests that meditation practice may improve awareness of
one's own body states (Solomonova et al, 2016) and increase introspective accuracy for somatic experience
(Fox et al, 2012). Breathing practices, such as pranayama, may be particularly effective in transforming SP as
it is happening due to the fact that many SP episodes are characterised by a feeling of disordered/insufficient
breathing. A recent study (Seppälä et al, 2014) reported that breathing exercises were effective in decreasing
PTSD symptoms in war veterans. This implies that practicing techniques that improve awareness of body
sensations may lower the reactivity to SP episodes, thus lowering the distressing quality of the experience,
and increasing the potential for disrupting or transforming SP.
The author was supported by Social Sciences and Humanities Research Council (SSHRC) of Canada and by a J.-
A. DeSèves Sacre-Coeur Hospital Foundation doctoral scholarship. Thanks are due to Tore Nielsen, Philippe
Stenstrom and Michelle Carr for numerous conversations on sleep paralysis and its interpretation, and to the
members of the Dream and Nightmare Laboratory at the Center for Advanced Research in Sleep Medicine.
Additional thanks are due to Don Donderi and Elena Frantova, as well as to all participants who have
consented to share their experiences with us. Special thanks to Benjamen Samaha (artist) and to the
anonymous SP sufferer for generously offering the narrative and the drawing of a sleep paralysis episode.
Acknowledgements:
18
References:
Abrams, M. P., Mulligan, A. D., Carleton, R. N., & Asmundson, G. J. G. (2008). Prevalence and correlates of sleep
paralysis in adults reporting childhood sexual abuse. Journal of Anxiety Disorders, 22(8), 1535–1541.
doi:10.1016/j.janxdis.2008.03.007
Adler, S. R. (2011). Sleep paralysis: Night-mares, nocebos, and the mind-body connection. Rutgers University
Press.
American Academy of Sleep Medicine. (2014). International classification of sleep disorders-third edition
(ICSD-3). Darien, Illinois.
American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (DSM-5).
American Psychiatric Pub.
Ammaniti, M., & Gallese, V. (2014). The birth of intersubjectivity: Psychodynamics, neurobiology, and the self.
WW Norton & Company.
Antelmi, E., Ferri, R., Iranzo, A., Arnulf, I., Dauvilliers, Y., Bhatia, K. P., … Plazzi, G. (2016). From state
dissociation to status dissociatus. Sleep Medicine Reviews, 28, 1–13. doi:10.1016/j.smrv.2015.07.003
Agargun, M. Y., & Cartwright, R. (2003). REM sleep, dream variables and suicidality in depressed
patients. Psychiatry Research, 119(1), 33-39.
Arikawa, H., Templer, D. I., & Brown, R. (1999). The structure and correlates of kanashibari. The Journal of ….
Retrieved from http://www.tandfonline.com/doi/abs/10.1080/00223989909599749
Belicki, K. (1992). Nightmare frequency versus nightmare distress: relations to psychopathology and
cognitive style. Journal of abnormal psychology, 101(3), 592.
Belicki, K., Gulko, N., Ruzycki, K., & Aristotle, J. (2003). Sixteen years of dreams following spousal
bereavement. OMEGA-Journal of Death and Dying, 47(2), 93-106.
Bell, C. C., Shakoor, B., Thompson, B., Dew, D., Hughley, E., Mays, R., & Shorter-Gooden, K. (1984). Prevalence
of isolated sleep paralysis in black subjects. Journal of the National Medical Association, 76(5), 501.
Blagrove, M., Farmer, L., & Williams, E. (2004). The relationship of nightmare frequency and nightmare
distress to well-being. Journal of sleep research, 13(2), 129-136.
Bloom, J. D., & Gelardin, R. D. (1983). Uqamairineq and Uqumanigianiq: Eskimo Sleep Paralysis. The Culture-
Bound Syndromes, 117–122. doi:10.1007/978-94-009-5251-5_10
Brugger, P., Regard, M., & Landis, T. (1996). Unilaterally Felt" Presences": The Neuropsychiatry of One's
Invisible Doppelganger. Cognitive and Behavioral Neurology, 9(2), 114-122.
Brugger, P., Regard, M., & Landis, T. (1997). Illusory reduplication of one's own body: phenomenology and
classification of autoscopic phenomena.Cognitive Neuropsychiatry, 2(1), 19-38.
Brugger, P., Regard, M., Landis, T., & Oelz, O. (1999). Hallucinatory Experiences in Extreme-Altitude Climbers.
Cognitive and Behavioral Neurology, 12(1), 67. Retrieved from
http://journals.lww.com/cogbehavneurol/Fulltext/1999/01000/Hallucinatory_Experiences_in_Extre
me_Altitude.8.aspx
Buzzi, G. (2011). False awakenings in light of the dream protoconsciousness theory: a study in lucid dreamers.
International Journal of Dream Research, 4(2), 110-116.
Casement, M. D., & Swanson, L. M. (2012). A meta-analysis of imagery rehearsal for post-trauma nightmares:
effects on nightmare frequency, sleep quality, and posttraumatic stress. Clinical psychology
review, 32(6), 566-574.
Cheyne, J. A. (2001). The ominous numinous. Sensed presence and 'other' hallucinations. Journal of
Consciousness Studies, 8(5-6), 133-150.
Cheyne JA (2002) Waterloo Unusual Sleep Experiences Questionnaire-VIIIa: Technical report.
http://watarts.uwaterloo.ca/ *acheyne/spdoc/Techreport.pdf. Accessed 15 July 2016
Cheyne, J. A. (2005). Sleep paralysis episode frequency and number, types, and structure of associated
hallucinations. Journal of Sleep Research, 14(3), 319–324. doi:10.1111/j.1365-2869.2005.00477.x
Cheyne, J. A., & Girard, T. A. (2007). Paranoid delusions and threatening hallucinations: A prospective study of
sleep paralysis experiences, 16(4), 959–974. doi:10.1016/j.concog.2007.01.002
Cheyne, J. A., Rueffer, S. D., & Newby-Clark, I. R. (1999). Hypnagogic and Hypnopompic Hallucinations during
Sleep Paralysis: Neurological and Cultural Construction of the Night-Mare. Consciousness and Cognition,
8(3), 319–337. doi:10.1006/ccog.1999.0404
Cheyne, J. A., & Pennycook, G. (2013). Sleep Paralysis Postepisode Distress Modeling Potential Effects of
Episode Characteristics, General Psychological Distress, Beliefs, and Cognitive Style. Clinical
19
Psychological Science, 2167702612466656.
Clark, A., & Chalmers, D. (1998). The extended mind. analysis, 58(1), 7-19.
Colombetti, G. (2013). The feeling body: Affective science meets the enactive mind. MIT Press.
Davies, O. (2010). The Nightmare Experience, Sleep Paralysis, and Witchcraft Accusations. Folklore, 114(2),
181–203. doi:10.1080/0015587032000104211
Cook, J. M., Harb, G. C., Gehrman, P. R., Cary, M. S., Gamble, G. M., Forbes, D., & Ross, R. J. (2010). Imagery
rehearsal for posttraumatic nightmares: a randomized controlled trial. Journal of traumatic
stress, 23(5), 553-563.
Denis, D., French, C. C., Rowe, R., Zavos, H. M. S., Nolan, P. M., Parsons, M. J., & Gregory, A. M. (2015). A twin and
molecular genetics study of sleep paralysis and associated factors. Journal of Sleep Research, n/a.
doi:10.1111/jsr.12282
Denis, D., & Poerio, G. L. (2016). Terror and bliss? Commonalities and distinctions between sleep paralysis,
lucid dreaming, and their associations with waking life experiences. Journal of Sleep Research.
Dresler, M., Wehrle, R., Spoormaker, V. I., Koch, S. P., Holsboer, F., Steiger, A., ... & Czisch, M. (2012). Neural
correlates of dream lucidity obtained from contrasting lucid versus non-lucid REM sleep: a combined
EEG/fMRI case study. Sleep, 35(7), 1017-1020.
Farhall, J., Greenwood, K. M., & Jackson, H. J. (2007). Coping with hallucinated voices in schizophrenia: a
review of self-initiated strategies and therapeutic interventions. Clinical Psychology Review, 27(4),
476-493.
Filevich, Elisa, Martin Dresler, Timothy R. Brick, and Simone Kühn. "Metacognitive mechanisms underlying
lucid dreaming." The Journal of Neuroscience 35, no. 3 (2015): 1082-1088.
Fox, K. C., Zakarauskas, P., Dixon, M., Ellamil, M., Thompson, E., & Christoff, K. (2012). Meditation experience
predicts introspective accuracy. PloS one,7(9), e45370.
Fox, K. C., Nijeboer, S., Solomonova, E., Domhoff, G. W., & Christoff, K. (2013). Dreaming as mind wandering:
evidence from functional neuroimaging and first-person content reports. Frontiers in human
neuroscience, 7, 412.
Fox, K. C., Dixon, M. L., Nijeboer, S., Girn, M., Floman, J. L., Lifshitz, M., ... & Christoff, K. (2016). Functional
neuroanatomy of meditation: A review and meta-analysis of 78 functional neuroimaging
investigations. Neuroscience & Biobehavioral Reviews, 65, 208-228.
Friedman, S., & Paradis, C. (2002). Panic disorder in African-Americans: symptomatology and isolated sleep
paralysis. Culture, medicine and psychiatry, 26(2), 179-198.
Fukuda, K. (1993). One explanatory basis for the discrepancy of reported prevalences of sleep paralysis
among healthy respondents. Perceptual and Motor Skills. Retrieved from
http://www.amsciepub.com/doi/pdf/10.2466/pms.1993.77.3.803
Fukuda, K., Miyasita, A., Inugami, M., & Ishihara, K. (1987). High prevalence of isolated sleep paralysis:
Kanashibari phenomenon in Japan. SLEEP. (PAGES)
Fukuda, K., Ogilvie, R. D., & Takeuchi, T. (2000). Recognition of sleep paralysis among normal adults in Canada
and in Japan. Psychiatry and clinical neurosciences, 54(3), 292-293.
Gallagher, S. (2005). How the body shapes the mind (pp. 173-178). Oxford: Clarendon Press.
Gordon, S. (2015). Medical Condition, Demon or Undead Corpse? Sleep Paralysis and the Nightmare in
Medieval Europe. Social History of Medicine, 28(3), NaN–NaN. doi:10.1093/shm/hkv005
Garfield, P. (1996). Dreams and bereavement. In Trauma and Dreams, Barrett, D. editor. Pp.186-211
Germain, A., Krakow, B., Faucher, B., Zadra, A., Nielsen, T., Hollifield, M., ... & Koss, M. (2004). Increased
Mastery Elements Associated With Imagery Rehearsal Treatment for Nightmares in Sexual Assault
Survivors With PTSD. Dreaming, 14(4), 195.
Germain, A., Buysse, D. J., & Nofzinger, E. (2008). Sleep-specific mechanisms underlying posttraumatic stress
disorder: integrative review and neurobiological hypotheses. Sleep medicine reviews, 12(3), 185-195.
Germain, A. (2013). Sleep disturbances as the hallmark of PTSD: where are we now?. American Journal of
Psychiatry, 170(4), 372-382.
Goldin, P. R., & Gross, J. J. (2010). Effects of mindfulness-based stress reduction (MBSR) on emotion regulation
in social anxiety disorder. Emotion,10(1), 83.
Golzari, S. E., & Ghabili, K. (2013). Alcohol-mediated sleep paralysis: the earliest known description. Sleep
Medicine, 14(3), 298.
Gould, G. A., Gugger, M., Molloy, J., Tsara, V. M. S. C., Shapiro, C. M., & Douglas, N. J. (1988). Breathing pattern
and eye movement density during REM sleep in humans. Am Rev Respir Dis, 138(4), 874-877.
20
Green, C. E., & McCreery, C. (1994). Lucid dreaming: the paradox of consciousness during sleep. Psychology
Press.
Hartmann, E., & Kunzendorf, R. (2013). Thymophor in Dreams, Poetry, Art and Memory: Emotion Translated
into Imagery as a Basic Element of Human Creativity. Imagination, Cognition and Personality, 33(1),
165-191.
Hinton, D. E., Pich, V., Chhean, D., & Pollack, M. H. (2005a). 'The ghost pushes you down': sleep paralysis-type
panic attacks in a Khmer refugee population. Transcultural Psychiatry, 42(1), 46–77.
Hinton, D. E., Pich, V., Chhean, D., Pollack, M. H., & McNally, R. J. (2005b). Sleep paralysis among Cambodian
refugees: association with PTSD diagnosis and severity. Depression and Anxiety, 22(2), 47-51.
Hishikawa, Y., & Kaneko, Z. (1965). Electroencephalographic study on narcolepsy. Electroencephalography
and Clinical Neurophysiology, 18(3), 249-259.
Hobson, J. A., Stickgold, R., & Pace-Schott, E. F. (1998). The neuropsychology of REM sleep
dreaming. Neuroreport, 9(3), R1-R14.
Hobson J. A., Pace-Schott E. F. & Stickgold R. (2000) Dreaming and the brain: Toward a cognitive neuroscience
of conscious states. Behavioral and Brain Science 23(6): 793–842; discussion 904–1121
Hobson, A. (2004). A model for madness?. Nature, 430(6995), 21-21.
Hofmann, S. G., Sawyer, A. T., Witt, A. A., & Oh, D. (2010). The effect of mindfulness-based therapy on anxiety
and depression: A meta-analytic review. Journal of consulting and clinical psychology, 78(2), 169.
Hofmann, S. G., Grossman, P., & Hinton, D. E. (2011). Loving-kindness and compassion meditation: Potential
for psychological interventions. Clinical psychology review, 31(7), 1126-1132.
Howell, M. J. (2012). Parasomnias: an updated review. Neurotherapeutics : The Journal of the American Society
for Experimental NeuroTherapeutics, 9(4), 753–775. doi:10.1007/s13311-012-0143-8
Hsieh, S.-W., Lai, C.-L., Liu, C.-K., Lan, S.-H., & Hsu, C.-Y. (2010). Isolated sleep paralysis linked to impaired
nocturnal sleep quality and health-related quality of life in Chinese-Taiwanese patients with
obstructive sleep apnea. Quality of Life Research : An International Journal of Quality of Life Aspects of
Treatment, Care and Rehabilitation, 19(9), 1265–1272. doi:10.1007/s11136-010-9695-4
Hufford, D. (1989). The terror that comes in the night: An experience-centered study of supernatural assault
traditions (Vol. 7). University of Pennsylvania Press.
Hufford, D. J. (2005). Sleep paralysis as spiritual experience. Transcultural Psychiatry, 42(1), 11–45.
Hurd, R. (2010). Sleep Paralysis: A Guide to Hypnagogic Visions and Visitors of the Night. Hyena Press.
Goyal, M., Singh, S., Sibinga, E. M., Gould, N. F., Rowland-Seymour, A., Sharma, R., ... & Ranasinghe, P. D. (2014).
Meditation programs for psychological stress and well-being: a systematic review and meta-analysis.
JAMA internal medicine, 174(3), 357-368.
Jalal, B., & Ramachandran, V. S. (2014). Sleep paralysis and 'the bedroom intruder': the role of the right
superior parietal, phantom pain and body image projection. Medical Hypotheses, 83(6), 755–757.
doi:10.1016/j.mehy.2014.10.002
Jalal, B., Simons-Rudolph, J., Jalal, B., & Hinton, D. E. (2014). Explanations of sleep paralysis among Egyptian
college students and the general population in Egypt and Denmark. Transcultural Psychiatry, 51(2),
158–175. doi:10.1177/1363461513503378
Jalal, B., Romanelli, A., & Hinton, D. E. (2015). Cultural Explanations of Sleep Paralysis in Italy: The Pandafeche
Attack and Associated Supernatural Beliefs. Culture, Medicine, and Psychiatry, 1–14.
doi:10.1007/s11013-015-9442-y
Jalal, B. (2016). How to make the ghosts in my bedroom disappear? Focused-attention meditation combined
with muscle relaxation (MR therapy)-a direct treatment intervention for sleep paralysis. Frontiers in
psychology,7.
James, W. (1985). The varieties of religious experience (Vol. 15). Harvard University Press.
Jiménez-Genchi, A., Ávila-Rodríguez, V. M., Sánchez-Rojas, F., Vargas Terrez, B. E., & Nenclares-Portocarrero,
A. (2009). Sleep paralysis in adolescents: the 'a dead body climbed on top of me' phenomenon in
Mexico.Psychiatry and clinical neurosciences, 63(4), 546-549.
Kang, Y., Gruber, J., & Gray, J. R. (2013). Mindfulness and de-automatization. Emotion review, 5(2), 192-201.
Kearney, D. J., Malte, C. A., McManus, C., Martinez, M. E., Felleman, B., & Simpson, T. L. (2013). Loving-kindness
meditation for posttraumatic stress disorder: A pilot study. Journal of Traumatic Stress, 26(4), 426-434.
Keen, C., Murray, C. D., & Payne, S. (2013). A qualitative exploration of sensing the presence of the deceased
following bereavement. Mortality,18(4), 339-357.
Kerr, C. E., Sacchet, M. D., Lazar, S. W., Moore, C. I., & Jones, S. R. (2013). Mindfulness starts with the body:
21
somatosensory attention and top-down modulation of cortical alpha rhythms in mindfulness
meditation. Frontiers in human neuroscience, 7, 12.
Kirmayer, L. J. (2009). Nightmares, neurophenomenology and the cultural logic of trauma. Culture, medicine
and psychiatry, 33(2), 323-331.
Krakow, B., Kellner, R., Pathak, D., & Lambert, L. (1995). Imagery rehearsal treatment for chronic
nightmares. Behaviour Research and Therapy, 33(7), 837-843.
Krakow, B., & Zadra, A. (2006). Clinical management of chronic nightmares: imagery rehearsal
therapy. Behavioral sleep medicine, 4(1), 45-70.
LaBerge, S., Levitan, L., & Dement, W. C. (1986). Lucid dreaming: Physiological correlates of consciousness
during REM sleep. Journal of Mind and Behavior.
LaBerge, S. (2009). Lucid dreaming: A concise guide to awakening in your dreams and in your life.
ReadHowYouWant. com.
Law, S., & Kirmayer, L. J. (2005). Inuit interpretations of sleep paralysis.Transcultural psychiatry, 42(1), 93-
112.
Levin, R., & Fireman, G. (2002). Nightmare prevalence, nightmare distress, and self-reported psychological
disturbance. SLEEP-NEW YORK-, 25(2), 205-212.
Levin, R., & Nielsen, T. A. (2007). Disturbed dreaming, posttraumatic stress disorder, and affect distress: a
review and neurocognitive model. Psychological bulletin, 133(3), 482.
Liddon, S. C. (1967). Sleep paralysis and hypnagogic hallucinations: Their relationship to the
nightmare. Archives of General Psychiatry, 17(1), 88.
Liskova, M., Janeckova, D., Kluzova Kracmarova, L., Mlada, K., & Buskova, J. (2016). The occurrence and
predictive factors of sleep paralysis in university students. Neuropsychiatr Dis Treat, 12, 2957-2962.
doi:10.2147/NDT.S115629
Luhrmann, T. M. (2012). When God talks back: Understanding the American evangelical relationship with God.
Vintage.
Lutz, A., Slagter, H. A., Dunne, J. D., & Davidson, R. J. (2008). Attention regulation and monitoring in
meditation. Trends in cognitive sciences, 12(4), 163-169.
Lutz, A., Jha, A. P., Dunne, J. D., & Saron, C. D. (2015). Investigating the phenomenological matrix of
mindfulness-related practices from a neurocognitive perspective. American Psychologist, 70(7), 632.
Mamelak, M., Black, J., Montplaisir, J., & Ristanovic, R. (2004). A pilot study on the effects of sodium oxybate on
sleep architecture and daytime alertness in narcolepsy. Sleep, 27, 1327-1336.
Mahowald, M. W., & Schenck, C. H. (1991). Status dissociatus: A perspective on states of being. SLEEP.
Retrieved from http://psycnet.apa.org/psycinfo/1991-30903-001
Mahowald, M. W., & Schenck, C. H. (2005). Insights from studying human sleep disorders. Nature, 437(7063),
1279-1285.
Maclaren, K. (2008). Embodied perceptions of others as a condition of selfhood?. Journal of Consciousness
Studies, 15(8), 63-93.
Maquet, P. (2000). Functional neuroimaging of normal human sleep by positron emission
tomography. Journal of sleep research, 9(3), 207-232.
McNally, R. (2011). Explaining 'Memories' of Space Alien Abduction and Past Lives: An Experimental
Psychopathology Approach. Journal of Experimental Psychopathology. doi:10.5127/jep.017811
McNally, R. J., & Clancy, S. A. (2005a). Sleep paralysis in adults reporting repressed, recovered, or continuous
memories of childhood sexual abuse. Journal of Anxiety Disorders, 19(5), 595–602.
doi:10.1016/j.janxdis.2004.05.003
McNally, R. J., & Clancy, S. A. (2005b). Sleep paralysis, sexual abuse, and space alien abduction. Transcultural
Psychiatry, 42(1), 113–122.
McNamara, P., Andresen, J., Clark, J., Zborowski, M., & Duffy, C. A. (2001). Impact of attachment styles on
dream recall and dream content: a test of the attachment hypothesis of REM sleep. Journal of sleep
research, 10(2), 117-127.
McNamara, P., & Bulkeley, K. (2015). Dreams as a source of supernatural agent concepts. Frontiers in
Psychology, 6, 283. doi:10.3389/fpsyg.2015.00283
Mellman, T. A., Bustamante, V., Fins, A. I., Pigeon, W. R., & Nolan, B. (2002). REM sleep and the early
development of posttraumatic stress disorder.American Journal of Psychiatry, 159(10), 1696-1701.
Menary, R. (2010). The extended mind. Cambridge, MA: The MIT Press.
Morris D. (2010) Empirical and phenomenological studies of embodied cognition. In: Gallagher S. &
22
Schmicking D. (eds.) Handbook of phenomenology and cognitive science. Springer, New York: 235–252.
Munezawa, T., Kaneita, Y., Osaki, Y., Kanda, H., Ohtsu, T., Suzuki, H., . . . Ohida, T. (2011). Nightmare and sleep
paralysis among Japanese adolescents: a nationwide representative survey. Sleep Med, 12(1), 56-64.
doi: 10.1016/j.sleep.2010.04.015
Ness, R. C. (1983). The Old Hag Phenomenon as Sleep Paralysis: A Biocultural Interpretation. The Culture-
Bound Syndromes, 123–145. doi:10.1007/978-94-009-5251-5_11
Nielsen, T. A. (2000). A review of mentation in REM and NREM sleep: "covert" REM sleep as a possible
reconciliation of two opposing models. Behavioral and Brain Sciences, 23(06), 851-866.
Nielsen, T., Stenstrom, P., Takeuchi, T., Saucier, S., Lara-Carrasco, J., Solomonova, E., & Martel, E. (2005).
Partial REM-sleep deprivation increases the dream-like quality of mentation from REM sleep and sleep
onset. Sleep 28(9), 1083.
Nielsen, T. (2007). Felt presence: paranoid delusion or hallucinatory social imagery? Consciousness and
Cognition, 16(4), 975. doi:10.1016/j.concog.2007.02.002
Nielsen, T., & Levin, R. (2007). Nightmares: a new neurocognitive model. Sleep Medicine Reviews, 11(4), 295-
310.
Nielsen, T., & Paquette, T. (2007). Dream-associated behaviors affecting pregnant and postpartum
women. Sleep, 30(9), 1162.
Nielsen, T., Svob, C., & Kuiken, D. (2009). Dream-enacting behaviors in a normal population. Sleep, 32(12),
1629-36.
Nielsen, T. A., Paquette, T., Solomonova, E., Lara-Carrasco, J., Popova, A., & Levrier, K. (2010). REM sleep
characteristics of nightmare sufferers before and after REM sleep deprivation. Sleep medicine, 11(2),
172-179.
Noë, A. (2004). Action in perception. Cambridge, MA: The MIT press.
Nofzinger, E. A., Schwartz, R. M., Reynolds, C. F., Thase, M. E., Jennings, J. R., Frank, E., ... & Kupfer, D. J. (1994).
Affect intensity and phasic REM sleep in depressed men before and after treatment with cognitive-
behavioral therapy. Journal of Consulting and Clinical Psychology, 62(1), 83.
O'Hanlon, J., Murphy, M., & Di Blasi, Z. (2011). Experiences of sleep paralysis in a sample of Irish university
students. Irish Journal of Medical Science, 180(4), 917–919. doi:10.1007/s11845-011-0732-2
Ohayon, M. M., Zulley, J., Guilleminault, C., & Smirne, S. (1999). Prevalence and pathologic associations of sleep
paralysis in the general population. Neurology, 52(6), 1194–1200. doi:10.1212/WNL.52.6.1194
Olry, R., & Haines, D. E. (2014). Kanashibari : a ghost's business. Journal of the History of the Neurosciences,
23(2), 192–197. doi:10.1080/0964704X.2013.862132
Otto, R. (1958). The idea of the holy (Vol. 14). Oxford University Press.
Otto, M. W., Simon, N. M., Powers, M., Hinton, D., Zalta, A. K., & Pollack, M. H. (2006). Rates of isolated sleep
paralysis in outpatients with anxiety disorders. Journal of Anxiety Disorders, 20(5), 687–693.
doi:10.1016/j.janxdis.2005.07.002
Paradis, C. M., & Friedman, S. (2005). Sleep paralysis in African Americans with panic disorder. Transcultural
psychiatry, 42(1), 123-134.
Peever, J., Luppi, P. H., & Montplaisir, J. (2014). Breakdown in REM sleep circuitry underlies REM sleep
behavior disorder. Trends in neurosciences,37(5), 279-288.
Petitmengin, C. (2006). Describing one's subjective experience in the second person: An interview method for
the science of consciousness.Phenomenology and the Cognitive Sciences, 5(3-4), 229-269.
Petitmengin, C., & Lachaux, J. P. (2013). Microcognitive science: bridging experiential and neuronal
microdynamics. Frontiers in human neuroscience,7, 617.
Pessoa, L. (2013). The cognitive-emotional brain: From interactions to integration. MIT press.
Plante, D. T., & Winkelman, J. W. (2008). Parasomnias: Psychiatric Considerations. Sleep Medicine Clinics, 3(2),
217–229. doi:10.1016/j.jsmc.2008.01.005
Portocarrero A (2009): Sleep paralysis in adolescents: The "a dead body climbed on top of me" phenomenon
in Mexico. Psychiatry and Clinical Neurosciences 63: 546-549.
Protevi, J. (2012). One More 'Next Step': Deleuze and Brain, Body and Affect in Contemporary Cognitive
Science. Revisiting Normativity with Deleuze, 25.
Rechtschaffen A. (1978) The single-mindedness and isolation of dreams. Sleep 1: 97–109
Revonsuo, A. (2006). Inner presence: Consciousness as a biological phenomenon. Mit Press.
Revonsuo, A., Tuominen, J., Valli, K., Dresler, M., Metzinger, T. & Windt, J.M. "The Avatars in the Machine." Open
MIND. Frankfurt a. M., GER: MIND Group (2015).
23
Roehrs, T., & Roth, T. (2001). Sleep, sleepiness, sleep disorders and alcohol use and abuse. Sleep medicine
reviews, 5(4), 287-297.
Robert, G., & Zadra, A. (2014). Thematic and content analysis of idiopathic nightmares and bad
dreams. Sleep, 37(2), 409-417
Roberts, K. (1998). Contemporary cauchemar: Experience, belief, prevention. Louisiana Folklore
Miscellany, 13, 15-26.
Seppälä, E. M., Nitschke, J. B., Tudorascu, D. L., Hayes, A., Goldstein, M. R., Nguyen, D. T., ... & Davidson, R. J.
(2014). Breathing-based meditation decreases posttraumatic stress disorder symptoms in US Military
veterans: A randomized controlled longitudinal study. Journal of traumatic stress, 27(4), 397-405.
Sharpless, B. A., McCarthy, K. S., Chambless, D. L., Milrod, B. L., Khalsa, S.-R., & Barber, J. P. (2010)., 66(12),
1292–1306. doi:10.1002/jclp.20724
Sharpless, B. A., & Barber, J. P. (2011). Lifetime prevalence rates of sleep paralysis: a systematic review. Sleep
Medicine Reviews, 15(5), 311–315. doi:10.1016/j.smrv.2011.01.007
Sharpless, B. A., & Grom, J. L. (2014). Isolated Sleep Paralysis: Fear, Prevention, and Disruption. Behavioral
Sleep Medicine, 1–6. doi:10.1080/15402002.2014.963583
Sharpless, B., & Doghramji, K. (2015). Sleep paralysis: Historical, psychological, and medical perspectives.
Oxford University Press.
Simard, V., & Nielsen, T. A. (2005). Sleep paralysis-associated sensed presence as a possible manifestation of
social anxiety. Dreaming, 15(4), 245.
Simon-Buller, S., Christopherson, V. A., & Jones, R. A. (1989). Correlates of sensing the presence of a deceased
spouse. OMEGA-Journal of Death and Dying, 19(1), 21-30.
Solomonova, E., Nielsen, T., Stenstrom, P., Simard, V., Frantova, E., & Donderi, D. (2008). Sensed presence as a
correlate of sleep paralysis distress, social anxiety and waking state social imagery. Consciousness and
Cognition, 17(1), 49–63. doi:10.1016/j.concog.2007.04.007
Solomonova, E., Nielsen, T., & Stenstrom, P. (2009). Lucid dreaming is associated with sleep paralysis but not
nightmares. Third Annual Conference of the Canadian Sleep Society. April 26-28, Toronto, Canada
Solomonova, E., Frantova, E., & Nielsen, T. (2010). Felt presence: the uncanny encounters with the numinous
Other. AI & SOCIETY, 26(2), 171–178. doi:10.1007/s00146-010-0299-x
Solomonova, E., Nielsen, T., Takeuchi, T. Bezinger, E. & Carr, M. (2012). The relationship between REM sleep
eye movement density and dream content: differences between idiopathic sleep paralysis sufferers
and healthy controls. International Association for the Study of Dreams annual meeting. June 24, 2012.
Berkeley, CA.
Solomonova, E., Stenstrom, P., Paquette, T., & Nielsen, T. (2015). Different temporal patterns of memory
incorporations into dreams for laboratory and virtual reality experiences: relation to dreamed locus of
control. International journal of dream research, 8(1), 10-26.
Solomonova, E. & Sha, X.W. (2016). Exploring the depth of dream experience: The enactive framework and
methods for neurophenomenological research. Constructivist Foundations. Special Issue: Exploring the
Diversity within Enactivism and Neurophenomenology. Tom Froese, Alexander Riegler and Sebastian
Voros, editors. 11(2), 407-416.
Solomonova, E., Dubé, S., Blanchette-Carrière, C., Samson, A., Picard-Deland, C. & Nielsen, T. (2016). Effect of
Vipassana meditation on sleep-dependent procedural memory consolidation and postural balance
following a daytime nap: Making implicit memories explicit. Mind and Life Summer Research Institute,
June 2016, Garrison, NY, USA.
Spanos, N. P., McNulty, S. A., DuBreuil, S. C., Pires, M., & Burgess, M. F. (1995). The frequency and correlates of
sleep paralysis in a university sample. Journal of Research in Personality, 29(3), 285-305.
Spoormaker, V. I., & Van Den Bout, J. (2006). Lucid dreaming treatment for nightmares: a pilot
study. Psychotherapy and psychosomatics, 75(6), 389-394.
Spoormaker, V. I., & Montgomery, P. (2008). Disturbed sleep in post-traumatic stress disorder: secondary
symptom or core feature?. Sleep medicine reviews, 12(3), 169-184.
Stewart, J. R., Gapenne, O., & Di Paolo, E. A. (2010). Enaction: Toward a new paradigm for cognitive science.
MIT Press.
Suedfeld, P., & Mocellin, J. S. (1987). The" Sensed Presence" in Unusual Environments Peter
Suedfeld. Environment and Behavior, 19(1), 33-52.
Szelenberger, W., Niemcewicz, S., & Dabrowska, A. J. (2005). Sleepwalking and night terrors:
psychopathological and psychophysiological correlates. International Review of Psychiatry, 17(4), 263-
24
270.
Szklo-Coxe, M., Young, T., Finn, L., & Mignot, E. (2007). Depression: relationships to sleep paralysis and other
sleep disturbances in a community sample. Journal of sleep research, 16(3), 297-312.
Takeuchi, T., Miyasita, A., Sasaki, Y., Inugami, M.. (1992). Isolated sleep paralysis elicited by sleep interruption.
SLEEP. (PAGES)
Takeuchi, T., Fukuda, K., Sasaki, Y., Inugami, M., & Murphy, T. I. (2002). Factors related to the occurrence of
isolated sleep paralysis elicited during a multi-phasic sleep-wake schedule. Sleep, 25(1), 89-96.
Taylor, S. F. (2005). Between the idea and the reality: A study of the counselling experiences of bereaved
people who sense the presence of the deceased. Counselling and Psychotherapy Research, 5(1), 53-61.
Terrillon, J. C., & Marques-Bonham, S. (2001). Does recurrent isolated sleep paralysis involve more than
cognitive neurosciences. Journal of Scientific Exploration, 15(1), 97-123.
Terzaghi, M., Ratti, P. L., Manni, F., & Manni, R. (2012). Sleep paralysis in narcolepsy: more than just a motor
dissociative phenomenon? Neurological Sciences : Official Journal of the Italian Neurological Society and
of the Italian Society of Clinical Neurophysiology, 33(1), 169–172. doi:10.1007/s10072-011-0644-y
Thompson, E. (2005). Sensorimotor subjectivity and the enactive approach to experience. Phenomenology and
the cognitive sciences, 4(4), 407-427.
Thompson, E. (2014). Waking, Dreaming, Being: Self and Consciousness in Neuroscience, Meditation, and
Philosophy. New York: Columbia University Press.
Thompson E. (2015a) Steps toward a neurophenomenology of conscious sleep. A Reply to Jennifer M. Windt.
In: Metzinger T. & Windt J. M. (eds.) Open MIND. MIND Group, Frankfurt am Main: 37(R). http://open-
mind.net/papers/steps-toward-a-neurophenomenology-of-consciousness-in-sleep-a-reply-to-jennifer-
m-windt
Thompson E. (2015b) Dreamless sleep, the embodied mind, and consciousness: The Relevance of a classical
Indian debate to cognitive science. In: Metzinger T. & Windt J. M. (eds.) Open MIND. MIND Group,
Frankfurt am Main: 37(T). http://open-mind.net/papers/dreamless-sleep-the-embodied-mind-and-
consciousness-the-relevance-of-a-classical-indian-debate-to-cognitive-science
Thorpy, M. J. (2016). Diagnostic Criteria and Delay in Diagnosis of Narcolepsy. In Narcolepsy (pp. 45-49).
Springer International Publishing.
Travis, F., & Shear, J. (2010). Focused attention, open monitoring and automatic self-transcending: categories
to organize meditations from Vedic, Buddhist and Chinese traditions. Consciousness and
cognition, 19(4), 1110-1118.
Varela F. J., Thompson E. & Rosch E. (1991) The embodied mind: Cognitive science and human experience.
MIT Press, Cambridge MA.
Veissière, S. (2016). Varieties of Tulpa Experiences: The Hypnotic Nature of Human Sociality, Personhood,
and Interphenomenality. In Raz and Lifshitz, eds. Hypnosis and Meditation Towards an Integrative
Science of Conscious Planes. Oxford University Press.
Wheeler, M. (2005). Reconstructing the cognitive world: The next step. MIT press.
Wing, Y.-K., Chiu, H., Leung, T., & Ng, J. (1999). Sleep paralysis in the elderly. Journal of Sleep Research, 8(2),
151–155. doi:10.1046/j.1365-2869.1999.00143.x
Wing, Y.-K., Lee, S. T., & Chen, C.-N. (1994). Sleep paralysis in Chinese: Ghost oppression phenomenon in Hong
Kong. Sleep.
Zadra, A. L., & Pihl, R. O. (1997). Lucid dreaming as a treatment for recurrent nightmares. Psychotherapy and
Psychosomatics, 66(1), 50-55.
Zadra, A., Pilon, M., Joncas, S., Rompré, S., & Montplaisir, J. (2004). Analysis of postarousal EEG activity during
somnambulistic episodes.Journal of sleep research, 13(3), 279-284.
Zahavi, D. (2014). Self and other: Exploring subjectivity, empathy, and shame. OUP Oxford.
25
|
1504.06290 | 1 | 1504 | 2015-04-23T18:36:01 | Entrainment in up and down states of neural populations: non-smooth and stochastic models | [
"q-bio.NC"
] | We study the impact of noise on a neural population rate model of up and down states. Up and down states are typically observed in neuronal networks as a slow oscillation, where the population switches between high and low firing rates (Sanchez-Vivez and McCormick, 2000). A neural population model with spike rate adaptation is used to model such slow oscillations, and the timescale of adaptation determines the oscillation period. Furthermore, the period depends non-monotonically on the background tonic input driving the population, having long periods for very weak and very strong stimuli. Using both linearization and fast-slow timescale separation methods, we can compute the phase sensitivity function of the slow oscillation. We find that the phase response is most strongly impacted by perturbations to the adaptation variable. Phase sensitivity functions can then be utilized to quantify the impact of noise on oscillating populations. Noise alters the period of oscillations by speeding up the rate of transition between the up and down states. When common noise is presented to two distinct populations, their transitions will eventually become entrained to one another through stochastic synchrony. | q-bio.NC | q-bio | Journal of Mathematical Biology manuscript No.
(will be inserted by the editor)
Entrainment in up and down states of neural
populations: non-smooth and stochastic models
Zachary McCleney · Zachary P. Kilpatrick
5
1
0
2
r
p
A
3
2
]
.
C
N
o
i
b
-
q
[
1
v
0
9
2
6
0
.
4
0
5
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract We study the impact of noise on a neural population rate model of up
and down states. Up and down states are typically observed in neuronal networks
as a slow oscillation, where the population switches between high and low fir-
ing rates (Sanchez-Vives and McCormick, 2000). A neural population model with
spike rate adaptation is used to model such slow oscillations, and the timescale
of adaptation determines the oscillation period. Furthermore, the period depends
non-monotonically on the background tonic input driving the population, hav-
ing long periods for very weak and very strong stimuli. Using both linearization
and fast-slow timescale separation methods, we can compute the phase sensitivity
function of the slow oscillation. We find that the phase response is most strongly
impacted by perturbations to the adaptation variable. Phase sensitivity functions
can then be utilized to quantify the impact of noise on oscillating populations.
Noise alters the period of oscillations by speeding up the rate of transition be-
tween the up and down states. When common noise is presented to two distinct
populations, their transitions will eventually become entrained to one another
through stochastic synchrony.
Keywords neural adaptation · phase sensitivity function · stochastic synchrony
1 Introduction
Cortical networks can generating a wide variety oscillatory rhythms with frequen-
cies spanning five orders of magnitude (Buzs´aki and Draguhn, 2004). Slow os-
cillatory activity (0.1-1Hz) has been observed in vivo during decreased periods
of alertness, such as slow wave sleep and anesthesia (Steriade et al., 1993). Fur-
thermore, such activity can be produced in vitro when bathing cortical slices in
This work was supported by NSF DMS-1311755.
Z. McCleney · Z.P. Kilpatrick
Department of Mathematics, University of Houston, Houston TX 77204
E-mail: [email protected]
Z. McCleney
E-mail: [email protected]
2
McCleney, Kilpatrick
a medium with typical extracellular ion concentrations (Sanchez-Vives and Mc-
Cormick, 2000). A key feature of these slow oscillations is that they tend to be an
alternating sequence of two bistable states, referred to as the up and down states.
Up states in networks are characterized by high levels of firing activity, due to
depolarization in single cells. Down states in networks typically appear quiescent,
due to hyperpolarization in single cells. There is strong evidence that up states
are neural circuit attractors, that emerge due to synaptic feedback (Cossart et al.,
2003). This suggests up states may be spontaneous remnants of stimulus-induced
persistent states utilized for working memory (Wang, 2001) and other network
computations (Major and Tank, 2004).
Several different cellular and synaptic mechanisms have been suggested to un-
derlie the transitions between up and down states. One possibility is that the
network is recurrently coupled with excitation, stabilizing both a quiescent and
active state (Amit and Brunel, 1997; Renart et al., 2007). Fluctuations due to prob-
abilistic synapses, channel noise, and randomness in network connectivity can then
lead to spontaneous transitions between the quiescent and active state (Bressloff,
2010; Litwin-Kumar and Doiron, 2012). Alternatively, switches between low and
high activity states may arise by some underlying systematic slow process. For in-
stance, it has been shown that competition between recurrent excitation and the
negative feedback produced by activity-dependent synaptic depression can lead to
slow oscillations in firing rate whose timescale is set by the depression timescale
(Bart et al., 2005; Kilpatrick and Bressloff, 2010). Excitatory-inhibitory networks
with facilitation can produce slow oscillations, due to the slow facilitation of feed-
back inhibition that terminates the up state, the down state is then rekindled
due to positive feedback from recurrent excitation (Melamed et al., 2008). These
neural mechanisms utilize dynamic changes in the strength of neural architecture.
However, Compte et al. (2003) proposed that single cell mechanisms can also shift
network states between up and down states. The up state is maintained by strong
recurrent excitation balanced by inhibition, and transitions to the down state occur
due to a slow adaptation current. Once in the down state, the adaptation current is
inactivated, and excitation reinitiates the up state. A similar mechanism has been
utilized in models of perceptual rivalry, where dominance switches between two
mutually inhibiting populations are due to the build up of a rate-based adaptation
current (Laing and Chow, 2002; Moreno-Bote et al., 2007).
In this paper, we utilize a rate-based model of an excitatory network with spike
rate adaptation to explore the impact that noise perturbations have upon the rel-
ative phase and duration of slow oscillations. We find that, as in the spiking model
studied by Compte et al. (2003), the interplay between recurrent excitation and
adaptation produces a slow oscillation in the firing rate of the network. In fact, for
slow timescale adaptation currents, the oscillations evolve as fast switches between
a low and high activity state, stable fixed points of the adaptation-free system.
Since the timescale and slow dynamics of the oscillation are set by the adaptation
current, we mainly focus on the impact of perturbation to the adaptation variable
in our model. As we will show, perturbations of the activity variable have much
lower impact on the oscillation phase. Introducing noise into the adaptation vari-
able of the population model leads to a speeding up of the slow oscillation, due to
early switching between the low and high activity state.
Another remarkable feature of slow oscillations, observed during slow-wave
sleep and anesthesia, is that the up and down states tend to be synchronized
Population models of up and down states
3
across different regions of cortex and thalamus (Steriade et al., 1993; Massimini
et al., 2004). Specifically, both the up and down states start near synchronously
in cells located up to 12mm apart (Volgushev et al., 2006). Such remarkable co-
herence between distant network activity cannot be accomplished by single cell
mechanisms, but require either long range network connectivity or some external
signal forcing entrainment (Traub et al., 1996; Smeal et al., 2010). Activity tends
to originate from several different foci in the network, quickly spreading across
the rest of the network on a timescale orders of magnitude faster than the oscil-
lation itself (Compte et al., 2003; Massimini et al., 2004). The fact that the onset
of quiescence is fast and well synchronized means there must be either a rapid
relay signal between all foci or there is some global signal cueing the down state.
Rather than suggest a disynaptic relay, using long range excitation acting on local
inhibition, we suggest that background noise can serve as a synchronizing global
signal (Ermentrout et al., 2008). Noisy but correlated inputs have been shown to
be capable of synchronizing uncoupled populations of phase oscillators (Teramae
and Tanaka, 2004) as well as experimentally recorded cells in vitro (Gal´an et al.,
2006). Here we will show correlated noise is a viable mechanisms for coordinating
slow oscillations in distinct uncoupled neural populations.
The paper is organized as follows. We introduce the neural population model
in section 2, indicating the way external noise is incorporated into the model.
In section 3, we demonstrate the periodic solutions that emerge in the noise-free
model, demonstrating it is possible to derive analytical expressions for the oscil-
lation period in the case of steep firing rate functions. Then, in section 4 we show
how to derive phase sensitivity functions that describe how external perturbations
to the periodic solution impact the asymptotic phase of the oscillation. As demon-
strated, the impact of perturbations to the adaptation variable is much stronger
than activity variable perturbations, especially for longer adaptation timescales.
Thus, our studies of the impact of noise mainly focus on the effects of fluctuations
in the adaptation variable. We find, in section 5, that adding noise to the adap-
tation variable leads to up and down state durations that are shorter and more
balanced, so that the up and down state last for similar lengths of time. In sec-
tion 6, we demonstrate that slow oscillations in distinct populations can become
entrained to one another when both populations are forced by the same common
noise signal. This phenomenon is robust to the introduction of independent noise
in each population, as we show in section 7.
2 Adaptive neural populations: deterministic and stochastic models
We begin by describing the models we will use to explore the impact of external
perturbations on slow oscillations. Motivated by Compte et al. (2003), we will focus
on a neural population model with spike rate adaptation, akin to mutual inhibitory
models used to study perceptual rivalry (Laing and Chow, 2002; Moreno-Bote
et al., 2007).
Single population model. In a single population, neural activity u(t) receives
negative feedback due to a subtractive spike rate adaptation term (Benda and
4
Herz, 2003)
McCleney, Kilpatrick
u(t) = −u(t) + f (αu(t) − a(t) + I),
τ a(t) = −a(t) + φu(t).
(1a)
(1b)
Here, u represents the mean firing rate of the neural population with excitatory
connection strength α. The negative feedback variable a is spike frequency adap-
tation with strength φ and time constant τ . For some of our analysis we will utilize
the assumption τ (cid:29) 1, based on the fact that many form for spike rate adaptation
tend to be much slower than neural membrane time constants (Benda and Herz,
2003). The constant tonic drive I initiates the high firing rate (up) state, and slow
adaptation eventually attenuates activity to a low firing rate (down) state. Weak
but positive drive I > 0 is meant to model the presence of low spiking threshold
cells that spontaneously fire, utilized as a mechanism for initiating the up state
in Compte et al. (2003). The firing rate function f is monotone and saturating
function such as the sigmoid
f (x) =
1
1 + e−γx .
(2)
Commonly, in studies of neural field models, the high gain limit of (2) is taken to
yield the Heaviside firing rate function (Amari, 1977; Laing and Chow, 2002)
(cid:26) 1 : x ≥ 0,
0 : x < 0,
H(x) =
(3)
which often allows for a more straightforward analytical study of model dynamics.
We exploit this fact extensively in our study. Nonetheless, we have also carried
out many numerical simulations of the model for a smooth firing rate function
(2), and they correspond to the results we presentfor sufficiently high gain. Note,
this form of adaptation is often referred to as subtractive negative feedback, as
current is subtracted from the population input. Alternative models of slow neural
population oscillations have employed short term synaptic depression (Tabak et al.,
2000; Bart et al., 2005; Kilpatrick and Bressloff, 2010), a form of divisive negative
feedback.
A primary concern of this work is the response of (1) to external perturbations,
acting on the activity u and adaptation a variables. To do so, we will use both
an exact method and a linearization to identify the phase response curve of the
limit cycle solutions to (1). Understanding the susceptibility of limit cycles (1) to
inputs will help us understand ways in which noise will influence the frequency
and regularity of oscillations.
Stochastic single population model. Following our analysis of the noise-
free system, we will consider how fluctuations influence oscillatory solutions to (1).
To do so, we will employ the following Langevin equation for (1) forced by white
noise
du(t) = [−u(t) + f (αu(t) − a(t) + I)] dt + dξu(t)
da(t) = [−a(t) + φu(t)] dt/τ + dξa(t),
(4a)
(4b)
where we have introduced the independent Gaussian white noise processes ξu(t)
and ξa(t) with zero mean (cid:104)ξu(t)(cid:105) = (cid:104)ξa(t)(cid:105) = 0 and variances (cid:104)ξu(t)2(cid:105) = σ2
ut
Population models of up and down states
5
and (cid:104)ξa(t)2(cid:105) = σ2
at. Extending our results concerning the phase response curve,
we will explore how noise forcing impacts the statistics of the resulting stochastic
oscillations in (4). In particular, since we find noise tends to impact the phase of
the oscillation more strongly when applied to the adaptation variable, we will tend
to focus on the case ξu ≡ 0.
Stochastic dual population model. Finally, we will focus on how correla-
tions in noise-forcing impact the coherence of two distinct uncoupled populations
du1 = [−u1(t) + f (αu1(t) − a1(t) + I)] dt + dξu
da1 = [−a1(t) + u1(t)] dt/τ + dξa
du2 = [−u2(t) + f (αu2(t) − a2(t) + I)] dt + dξu
da2 = [−a2(t) + u2(t)] dt/τ + dξa.
(5a)
(5b)
(5c)
(5d)
Thus, the system (5) describes the dynamics of two distinct neural populations
u1 and u2, with inputs I. Our main interest lies in the impact the noise terms
have upon the phase relationship between the two systems' states. In this version
of the model, noise to the activity variables ξu is totally correlated, as is noise
to the adaptation variables ξa. Thus, all means are zero and (cid:104)ξ2
u =
D11t. Furthermore, (cid:104)ξ2
at = D22t. For this study, we assume there are no
correlations between activity and adaptation noise, so (cid:104)ξu(t)ξa(t)(cid:105) = 0. A more
general version of the model (5) would consider the possibility of independent noise
in each population
u(t)(cid:105) = σ2
a(t)(cid:105) = σ2
1 − χ2
udξu1
1 − χ2
du1 = [−u1(t) + f (αu1(t) − a1(t) + I)] dt + χudξuc +
da1 = [−a1(t) + u1(t)] dt/τ + χadξac +
adξa1
du2 = [−u2(t) + f (αu2(t) − a2(t) + I)] dt + χudξuc +
da2 = [−a2(t) + u2(t)] dt/τ + χadξac +
adξa2.
Noise terms all have zero mean and variances defined (cid:104)ξ2
ujt = Dujt and
(cid:104)ξ2
aj(t)(cid:105) = σ2
ajt = Dajt (j = 1, 2, c). To ease calculations, we take Du1 = Du2 ≡
Dul = σ2
a. The degree of noise correlation between
populations is controlled by the parameters χu and χa, so in the limit χu,a → 1,
the model (6) becomes (5).
u and Da1 = Da2 ≡ Dal = σ2
uj(t)(cid:105) = σ2
1 − χ2
(6a)
(6b)
1 − χ2
udξu2
(6c)
(6d)
(cid:112)
(cid:112)
(cid:112)
(cid:112)
3 Periodic solutions of a single population
We begin by studying periodic solutions of the single population system (1), as
demonstrated in Fig. 1A. First, we note that for firing rate functions f with finite
gain, we can identify the emergence of oscillations by analyzing the stability of the
equilibria of (1). That is, we assume ( u, a) = (0, 0), so the system becomes
¯u = f (α¯u − ¯a + I),
¯a = φ¯u,
which can be reduced to the single equation
¯u = f ((α − φ)¯u + I) = g(¯u).
(7)
6
McCleney, Kilpatrick
Fig. 1 Single adapting neural population (1) generates slow oscillations. (A) Numerical simu-
lation of (1) for adaptation timescale τ = 100 (1s) and input I = 0.2. (B) Partitioning of (τ, I)
parameter space shows the range of inputs I leading to oscillations expands as the adaptation
timescale τ is increased, according to (9). (C,D) Bifurcation diagram showing supercritical
(I−H ) and subcritical (I+H ) Hopf bifurcations that arise as the input is increased for (C)
τ = 10 and (D) τ = 100. Firing rate function is sigmoidal (2). Other parameters are φ = 1,
α = 0.5, and γ = 15
Roots of (7), defining fixed points of (1) are plotted as a function of the input I
in Fig. 1C,D. Utilizing the sigmoidal firing rate function f given by (2), we can
show that there will be a single fixed point as long as φ > α. In this case, we can
compute
dg(¯u)
d¯u
= −(φ − α)f
(cid:48)
((α − φ)¯u + I) = − (φ − α)e−γ((α−φ)¯u+I)
(cid:0)1 + e−γ((α−φ)¯u+I)(cid:1)2 < 0.
Since ¯u is monotone increasing, then ¯u − g(¯u) is monotone increasing. Further,
noting lim¯u→±∞ [¯u − g(¯u)] = ±∞, it is clear ¯u− g(¯u) crosses zero once, so (7) has
a single root when φ > α. Stability of this equilibrium is given by the eigenvalues
of the associated Jacobian
(cid:18)−1 + αf(cid:48)((α − φ)¯u + I) −f(cid:48)((α − φ)¯u + I)
(cid:19)
−1/τ
.
J(¯u, ¯a) =
φ/τ
Population models of up and down states
7
We note that the sigmoid (2) satisfies the Ricatti equation f(cid:48) = γf (1 − f ), so we
can use (7) to write
(cid:18)−1 + αγ ¯u(1 − ¯u) −γ ¯u(1 − ¯u)
(cid:19)
φ/τ
−1/τ
.
J(¯u, ¯a) =
Oscillations arise when stable spiral equilibria destabilize through a Hopf bifurca-
tion. Hopf bifurcations will occur when complex eigenvalues associated with fixed
points (¯u, ¯a) cross from the left to the right half plane. We can require this with the
pair of expression: tr(J) = 0 and tr(J)2 < 4det(J). Thus, a necessary condition
for the Hopf bifurcation point is that the equilibrium value ¯u satisfy
αγ ¯u(1 − ¯u) = 1 + 1/τ.
Solving this for ¯u yields
¯u±H =
1
2
(cid:104)
1 ±(cid:112)1 − 4χ
(cid:105)
,
χ =
1 + 1/τ
αγ
.
(8)
Thus, Hopf bifurcations will only occur when the timescale of adaptation is suffi-
ciently large τ > [αγ/4 − 1]
−1. Plugging the formula (8) back into the fixed point
equation (7) and solving for the input I, we can parameterize Hopf bifurcation
curves based upon the equation
(cid:20)
(cid:21)
I±H =
1
γ
ln
¯u±H
1 − ¯u±H
− (α − φ)¯u±H ,
(9)
(10)
along with the additional condition tr(J)2 < 4det(J) which becomes
4
τ 2 <
4φ
ατ 2 +
4φ
ατ
,
which will always hold as long as φ > α. We partition the parameter space (τ, I)
using our formula for the Hopf curve (9) in Fig. 1B. As demonstrated, there tend
to be either two or zero Hopf bifurcation points for a given timescale τ , and the
coalescence of the two Hopf points is given by the point where τ = [αγ/4 − 1]
−1.
In the limit of slow adaptation τ (cid:29) 1, we can separate the timescales of the
activity u and adaptation a variables, finding u will equilibrate according to the
equation
u(t) = f (αu(t) − a(t) + I),
and subsequently a will slowly evolve according to the equation
a(t) = [φu(t) − a] /τ.
(11)
(12)
We always have an implicit formula for u(t) in terms of a(t), so the dynamics
will tend to slowly evolve along the direction of the a variable. This demonstrates
why periodic solutions to (1) are comprised of a slow rise and decay phase of
a, punctuated by fast excursions in the activity variable u. In general, it is not
straightforward to analytically treat the pair of equations (11) and (12), but we
will show how computing solutions of the singular system becomes straightforward
when we take the high gain limit γ → ∞.
8
McCleney, Kilpatrick
Fig. 2 Analytical approximations to periodic solutions of (1) with a Heaviside firing rate func-
tion (3). (A) Numerical simulation (solid lines) of the periodic solution is well approximated
by the analytical approximation (dashed lines) given by (15) when I = 0.2 and τ = 100. (B)
The period of the oscillation T computed from numerical simulations (dots) is accurately ap-
proximated by the analytical formula (solid lines) given by (14). Other parameters are α = 0.5
and φ = 1.
Having established the existence of periodic solutions to (1) in the case of
sigmoid firing rates (2), we now explore the system in the high gain limit γ → ∞
whereby the firing rate function becomes a Heaviside (3). In this case, fixed points
(¯u, ¯a) satisfy the equations
¯u = H((α − φ)¯u + I) =
(cid:26) 1 : ¯u < I/(φ − α),
0 : ¯u > I/(φ − α),
and ¯a = φ¯u. Thus, assuming φ > α, then ¯u = 0 when I < 0 and ¯u = 1 when I >
(φ − α). In both cases, the fixed points are linearly stable. When 0 < I < (φ − α),
there are no fixed points and we expect to find oscillatory solutions. Assuming
τ (cid:29) 1, we can exploit a separation of timescales to identify the shape and period
of these limit cycles. To begin, we note that on fast timescales
u(t) = −u(t) + H(I + αu(t) − a0),
where a0 is a quasi steady state. On slow timescales on the order of τ , then u(t)
quickly equilibrates and
u(t) = H(I + αu(t) − a(t)),
τ a(t) = −a(t) + φu(t).
(13a)
(13b)
Periodic solutions to (1) must obey the condition (u(t), a(t)) = (u(t + nT ), a(t +
nT )) for t ∈ [0, T ] and n ∈ Z, so we focus on the domain t ∈ [0, T ]. Examining
(13), we can see oscillations in (1) involve switches between u(t) ≈ 1 and u(t) ≈ 0.
We translate time so that u(t) ≈ 1 on t ∈ [0, T1) and u(t) ≈ 0 on t ∈ [T1, T ).
Subsequently, this means for t ∈ [0, T1] the system (13) becomes u ≡ 1 and
τ a = −a + φ so a(t) = φ − (φ − I)e−t/τ . We know a(0) = I because u(0−) ≡ 0
in (13), and the argument of H(x) must have crossed zero at t = 0. In a similar
way, we find on t ∈ [T1, T ) that u ≡ 0 and a(t) = (I + α)e−(t−T1)/τ . Using the
Population models of up and down states
9
conditions a(T1) = I +α and a(T ) = I, we find that the rise time of the adaptation
variable (or the duration of the up state) is
T1 = τ ln
(cid:21)
,
φ − α − I
(cid:20) φ − I
(cid:21)
(cid:20) I + α
(cid:20) (I + α)(φ − I)
I
T2 = τ ln
,
and the decay time (or the duration of the down state) is
and the total period of the oscillation is
(cid:21)
.
(14)
Thus, approximate periodic solutions to (1) in the case of a Heaviside firing rate
(3) take the form
T = τ ln
I(φ − α − I)
(cid:26) 1 : t ∈ [0, T1),
(cid:26) φ − (φ − I)e−t/τ
0 : t ∈ [T1, T ],
u(t) =
a(t) =
: t ∈ [0, T1),
(I + α)e−(t−T1)/τ : t ∈ [T1, T ].
(15a)
(15b)
We demonstrate the accuracy of the approximation (15) in Fig. 2A. Furthermore,
we show that relationship between the period T and model parameters is well
captured by the formula (14). Notice there is a non-monotonic relationship between
the period T and the input I. We can understand this further by noting that the
rise time T1 of the adaptation variable a increases monotonically with input
dT1
dI
=
τ α
(φ − I)(φ − α − I)
> 0,
when 0 < I < (φ − α). Furthermore, the decay time T2 of the adaptation variable
a decreases monotonically with input
dT2
dI
= − τ α
I(I + α)
< 0,
when 0 < I < (φ − α). Thus, as I → 0+, the slow oscillation's period T is
dominated by very long decay times T2 (cid:29) 1 and as I → (φ− α)−, it is dominated
by very long rise times T1 (cid:29) 1. We can identify the minimal period as a function
of the input I by finding the critical point of T (I). To do so, we differentiate and
simplify
= −
dT
dI
τ αφ(2I − (φ − α))
I(I + α)(φ − I)(φ − α − I)
,
so the critical point of T (I) is Icrit = (φ− α)/2, which corresponds to the minimal
value of the period Tmin(I) = 2τ ln [(φ + α)/(φ − α)] as pictured in Fig. 2B.
10
McCleney, Kilpatrick
Fig. 3 (A, B, C) Periodic solution (u, a) and (D, E, F) phase sensitivity function (Zu, Za)
of (1) plotted as a function of phase θ = t/T for a sigmoidal firing rate function (2). (A,D)
For shorter adaptation timescale τ = 10 and input I = 0.2, the activity variable u has a more
rounded trajectory, so perturbations to activity influence the oscillation phase more (note size
of lobes in on Zu in (D). (B,E) As the adaptation timescale is increased to τ = 100, with
I = 0.2, the influence of perturbations to the activity variable decrease (compare lobes of Zu
to those in (D). Perturbations of the adaptation variable influence the phase more strongly
as shown by the change in the relative amplitude of Za. (C,F) Increasing the input I = 0.4,
with τ = 10, increases the relative duration of the rise time of a. As a result, there is a wider
region where perturbations to a advance the phase. Other parameters are α = 0.5, φ = 1, and
γ = 15.
4 Phase response curves
We can further understand the dynamics of the slow oscillations in (1) by comput-
ing phase response curves for both the case of a sigmoidal firing rate (2) and the
Heaviside firing rate (3). As we will show, perturbations of the activity variable u
have decreasing impact as the timescale of adaptation τ and the gain γ of the firing
rate are increased. Perturbations of the adaptation variable a tend to dominate
the resulting dynamics, as it is the evolution of this slow variable that primarily
determines the phase of the oscillation.
(cid:19)
L
To begin, we derive a general formula that linearly approximates the influence
of small perturbations on limit cycle solutions (u0(t), a0(t)) to (1). Essentially, we
utilize the fact that solutions Z(t) to the adjoint equation associated with lineariza-
tion about the limit cycle solution (u0(t), a0(t)) provide a complete description of
how infinitesimal perturbations of the limit cycle impact its phase (Ermentrout,
1996; Brown et al., 2004). To start, we note that
(cid:18) u1 + u1 − αf(cid:48)(αu0 − a0 + I)u1 + f(cid:48)(αu0 − a0 + I)a1,
(cid:18) u1
product on T -periodic functions in R2 as (cid:104)F (t), G(t)(cid:105) =(cid:82) T
is the linearization of (1) about the limit cycle (u0(t), a0(t)). Defining the inner
0 F (t) · G(t)dt, we can
find the adjoint operator L∗ by noting it satisfies (cid:104)F,LG(cid:105) = (cid:104)L∗F, G(cid:105) for all L2
a1 − φu1/τ + a1/τ
(cid:18) 0
(cid:19)
,
0
=
a1
(cid:19)
=
Population models of up and down states
integrable vector functions F, G. We can then compute
L∗(cid:18) v
b
(cid:19)
=
(cid:18)− v + v − αf(cid:48)(αu0 − a0 + I)v − φb/τ
−b + f(cid:48)(αu0 − a0 + I)v + b/τ
(cid:19)
.
11
(16)
It can be shown that the null space of L∗ describes the response of the phase of
the limit cycle (u0(t), a0(t)) to infinitesimal perturbations (Brown et al., 2004).
Note that if (u0(t), a0(t)) is a stable limit cycle then the nullspace of L is spanned
0(t), a(cid:48)
by scalar multiples of (u(cid:48)
0(t)). Furthermore, appropriate normalization re-
quires that Z(t) · (u(cid:48)
0(t)) = 1 along with L∗Z = 0 (Ermentrout, 1996). To
0(t), a(cid:48)
numerically compute Z(t) = (Zu(t), Za(t)), we thus integrate the system
Zu(t) = Zu(t) − αf
Za(t) = f
(cid:48)
(cid:48)
(αu0(t) − a0(t) + I)Zu(t) − φZa(t)/τ,
(17a)
(αu0(t) − a0(t) + I)Zu(t) − Za(t)/τ,
(17b)
backward in time, taking the long time limit to find (Zu(t), Za(t)) on t ∈ [0, T ],
and normalizing (cid:104)(Zu(t), Za(t)), (u(cid:48)
0(t))(cid:105) = 1 by rescaling appropriately. We
demonstrate this result in Fig. 3, showing the relationship between the shape and
relative amplitude of the phase sensitivity functions (Zu, Za) and the parameters.
Notably, perturbing the activity variable u become less influential as the timescale
of adaptation τ is increased (Zu). Furthermore, there is a sharper transition be-
tween phase advance and phase delay region of the adaptation phase response (Za)
for larger timescales τ .
0(t), a(cid:48)
In addition to a general formula for the phase sensitivity functions (Zu(t), Za(t)),
we can derive an amplitude dependent formula for the response of limit cycle so-
lutions (u0(t), a0(t)) of (1) with a Heaviside firing rate (3), assuming τ (cid:29) 1. In
this case, we utilize the formula for the period (14) and limit cycle (15), derived
using a separation of timescales assumption. Then, we can compute the change
to the variables (u, a) as a result of a perturbation (δu, δa), which we denote
(cid:55)−→ (u0(t), a0(t)). We are primarily interested in how the relative
(u0(t), a0(t))
time in the limit cycle is altered by a perturbation δu - how much closer or further
the limit cycle is to the end of the period T after being perturbed. We can readily
determine this by first inverting the formula we have for (u0(t), a0(t)), given by
(15), to see how this value determines the time t0 along the limit cycle
(δu,δa)
t0(u0, a0) =
τ ln [(φ − I)/(φ − a0)]
: u0 = 1,
τ ln [(φ − I)(I + α)/a0/(φ − α − I)] : u0 = 0.
(18)
(cid:26)
Using this formula, we can now map the value (u0, a0) to an associated updated
relative time t0 along the oscillation.
Here, we decompose the impact of perturbations to the u and a variables. We
begin by studying the impact of perturbations δu to the activity variable u. We
can directly compute
u0(t) = H(I + α [u0(t) + δu] − a0(t)).
Thus, the singular system (13) will be unaffected by such perturbations if sgn(I +
α[u + δu]− a) = sgn(I + αu− a). This is related to the flatness of the susceptibility
function Zu over much of the time domain in Fig. 3D,E,F. However, if sgn(I +
12
McCleney, Kilpatrick
Fig. 4 Phase response curves of the fast-slow timescale separated system τ (cid:29) 1. (A,B)
Amplitude δu- and δa-dependent phase response curves Gu(θ, δu) and Ga(θ, δu) characterizing
phase advances/delays resulting from perturbation of neural activity u and adaptation a. We
compare analytical formulae (solid lines) to numerically computed PRCs (dashed lines). (C)
Phase response curve associated with perturbations of the adaptation variable a in the small
amplitude 0 < δa (cid:28) 1 limit. We compare the large amplitude formula (solid line) determined
by (24) to the linear approximation (dotted line) given by (25) to numerical computations
(dashed line).
α[u + δu]− a) (cid:54)= sgn(I + αu− a), then u0(t) = 1− u0(t), as detailed in the following
piecewise smooth map:
u0(t) = 0 (cid:55)→ u0(t) = 1 : δu > −(I − a0(t))/α > 0,
u0(t) = 0 (cid:55)→ u0(t) = 0 : δu < −(I − a0(t))/α > 0,
u0(t) = 1 (cid:55)→ u0(t) = 0 : −δu < −(I + α − a0(t))/α < 0,
u0(t) = 1 (cid:55)→ u0(t) = 1 : −δu > −(I + α − a0(t))/α < 0,
where (u0(t), a0(t)) are defined by (15). The formula (18) can then be utilized to
compute the updated relative time t0 := t0(u0, a0), finding
τ ln [(φ − I)(I + α)/a0/(φ − α − I)]
T + τ ln [(φ − I)/(φ − a0)]
t0(u0, a0)
t0 =
: δu > (I + α − a0)/α > 0, u0 = 1
: −δu > (a0 − I)/α > 0, u0 = 0
: otherwise,
(19)
where a0 = φ − (φ − I)e−t0/τ if u0 = 1 and a0 = (I + α)e−(t0−T1)/τ if u0 = 0. We
can refer to the function t0/T , where t0 is defined by (19), as the phase transition
curve for u perturbations. Thus, the function Gu(θ, δu) = (t0 − t0)/T will be
the phase response curve, where θ = t0/T , and phase advances occur for positive
values and phase delays occur for negative values. We plot the function Gu(θ, δu)
in Fig. 4A for different values of δu, demonstrating the nontrivial dependence
on the perturbation amplitude is not simply a rescaling but an expansion of the
non-zero phase shift region. Due to the singular nature of the fast-slow limit cycle
(15), the size of the phase perturbation has a piecewise constant dependence on the
amplitude of the u perturbation. Note, this formulation allows us to quantify phase
shifts that would not be captured by a perturbative theory for phase sensitivity
functions, as computed for the general system in (17).
For perturbations δa of the adaptation variable a, there is a more graded de-
pendence of the phase advance/delay amplitude on the perturbation amplitude δa.
We expect this, as it was a property we observed in Za as we varied parameters in
Population models of up and down states
13
Fig. 3. We can partition the limit cycle (u0(t), a0(t)) into four different regions: two
advance/delay regions of exponential saturation and two early threshold crossings.
First, note if u0(t) = 1 and a0(t) + δa < I + α, then
a0(t) = φ − (φ − I)e
(20)
so t0 = T1−tw with tw = τ ln [(φ − a0 − δa)/(φ − I − α)], but if a0(t)+δa > I +α,
then
−t/τ + δa,
u0(t) = 1,
u0(t) = 0,
a0(t) = φ − (φ − I)e
−t/τ + δa.
(21)
Determining the relative time of the perturbed variables (u0(t), a0(t)) in (20) is
straightforward using the mapping (18). However, to determine the relative time
described by (21), we compute the time, after the perturbation, until a0(t) = I +α,
which will be tw = τ ln [(a0 + δa)/(I + α)], so t0 = T1− tw. Second, note if u0(t) =
0 and a0(t) + δa > I, then
u0(t) = 0,
(22)
so t0 = T − tw with tw = τ ln [(a0 + δ0)/I], but if a0(t) + δa < I, so that it is
necessary that δa < 0, then
a0(t) = (I + α)e
−(t−T1)/τ + δa,
u0(t) = 1,
(23)
In the case of (23), we note that tw = τ ln [(φ − a0 − δa)/(φ − I)], so t0 = T − tw.
Combining our results, we find we can map the relative time to the perturbed
relative time as
a0(t) = (I + α)e
−(t−T1)/τ + δa.
T1 − τ ln [(φ − a0 − δa)/(φ − I − α)]
T1 − τ ln [(a0 + δa)/(I + α)]
T − τ ln [(a0 + δa)/I]
T − τ ln [(φ − a0 − δa)/(φ − I)]
: δa < I + α − a0, u0 = 1,
: δa > I + α − a0, u0 = 1,
: δa > I − a0, u0 = 0,
: δa < I − a0, u0 = 0,
(24)
t0 =
where a0 = φ − (φ − I)e−t0/τ if u0 = 1 and a0 = (I + α)e−(t0−T1)/τ if u0 = 0.
Again, we have a phase transition curve given by the function t0/T and phase
response curve given by Ga(θ, δa) = (t0 − t0)/T , where θ = t0/T . As opposed to
the case of u perturbations, the phase perturbation here depends smoothly on the
amplitude of the a perturbation δa.
Furthermore, we can obtain a perturbative description of the phase response
curve for the singular system (13) in two ways: (a) Taylor expand the amplitude-
dependent phase response curve expressions defined by (19) and (24) and truncate
to linear order or (b) solving the adjoint equation (17) in the case of a Heaviside
firing rate (3) and long adaptation timescale τ (cid:29) 1. We begin with the first
derivation, which simply requires differentiating (19) to demonstrate that the in-
finitesimal phase response curve (iPRC) associated with perturbations of the u
variable is zero almost everywhere. However, differentiating (24) reveals that the
iPRC associated with perturbations of the adaptation variable a is given by the
piecewise smooth function
Za(t) =
τ
T (φ − I)s
−
τ
T (I + α)
: t ∈ (0, T1),
et/τ
e(t−T1)/τ : t ∈ (T1, T ).
(25)
14
McCleney, Kilpatrick
Furthermore, note we could derive the same result by solving the adjoint equations
(17) in the case of Heaviside firing rate (3), so that
Zu(t) = −Zu(t) + αδ(αu0(t) − a0(t) + I)Zu(t) + φZa(t)/τ,
Za(t) = −δ(αu0(t) − a0(t) + I)Zu(t) + Za(t)/τ.
(26b)
Note, we have reversed time t (cid:55)→ −t, so we can simply solve the system forward.
Furthermore, we can use the identity
(26a)
δ(αu0(t) − a0(t) + I) =
δ(t)
u(cid:48)(0) − a(cid:48)(0)
+
δ(t − T1)
u(cid:48)(T1) − a(cid:48)(T1)
.
(27)
Utilizing the separation of timescales, τ (cid:29) 1, we find that almost everywhere
(except where t = 0, T1, T ), we have that (26) becomes the system
Zu(t) = −Zu(t),
τ Za(t) = Za(t).
(28)
As before Zu(t) will be zero almost everywhere, whereas Za(t) = A(t)et/τ , where
A(t) is a piecewise constant function taking two different values on t ∈ (0, T1) and
t ∈ (T1, T ), determined by considering the δ distribution terms. This indicates
how one would derive the formula (25) using the adjoint equations (26).
Note, in previous work (Jayasuriya and Kilpatrick, 2012), we explored the en-
trainment of slowly adapting populations to external forcing, comprised of smooth
and non-smooth inputs to the system (1). In the next section, we explore the im-
pact of external noise forcing on the slow oscillations of (1), subsequently demon-
strating that noise can be utilized to entrain the up and down states of two distinct
networks.
5 Impact of noise on the timing of up/down states
We now study the effects of noise on the duration of up and down states of the
single population model (1). Switches between high and low firing rate states
can occur at irregular intervals (Sanchez-Vives and McCormick, 2000), suggesting
internal or external sources of noise determine state changes. This section focuses
on how noise can reshape the mean duration of up and down residence times. Due
to our findings in the previous sections, we focus on noise applied to the adaptation
variable in this section. As we have shown, very weak perturbations to the neural
activity variable have a negligible effect on the phase of oscillations. Analytic
calculations are presented for the piecewise smooth system with Heaviside firing
rate (3), as accurate approximations of the mean up and down state durations can
be computed.
Our approach is to derive expressions for the mean first passage times of both
the up and down state ( ¯T1 and ¯T2) of the stochastic population model (4). Focusing
on adaptation noise allows us to utilize the separation of fast-slow timescales, and
recast the pair of equations as a stochastic-hybrid system
u(t) = H((αu(t) + I − a(t)),
da(t) = [−a(t) + φu(t)] dt/τ + dξa(t),
Population models of up and down states
15
Fig. 5 Noise alters the duration of up and down states. (A) Numerical simulation of the
stochastically driven population model (4) demonstrates up and down state durations (e.g.,
T1 and T2) are variable when driven by adaption noise ξa with (cid:104)ξ2
at, σa = 0.01. Switches
are determined by the threshold crossings of the adaptation variable a(t) = I and a(t) = I + α.
(B) Up/down states become more variable when the noise amplitude σa = 0.02. (C) Mean
durations of the up and down state, (cid:104)T1(cid:105) and (cid:104)T2(cid:105), decrease as a function noise amplitude
σa. (D) Impact of noise σa on the balance of up to down state durations ¯T1/ ¯T2 as input I
is varied. Firing rate is given by the Heaviside function (3). Other parameters are α = 0.5,
φ = 1, and τ = 50.
a(cid:105) = σ2
where ξa is white noise with mean (cid:104)ξa(cid:105) = 0 and variance (cid:104)ξ2
at. To begin,
assume the system has just switched to the up state, so the initial conditions are
u(0) = 1 and a(0) = I. Determining the amount of time until a switch to the down
state requires we calculate the time T1 until the threshold crossing a(T1) = I + α
where a(t) is determined by the stochastic differential equation (SDE)
a(cid:105) = σ2
da(t) = [−a(t) + φ] dt/τ + dξa,
which is the well-known threshold crossing problem for an Ornstein-Uhlenbeck
process Gardiner (2004). The mean ¯T1 of the passage time distribution is thus
given by defining the potential V (a) = a2
τ and computing the integral
(cid:90) I+α
(cid:90) x
2τ − φa
e[V (x)−V (y)]/σ2
a dydx.
¯T1 =
1
σ2
a
I
−∞
Next, note that the duration of the down state T2 will be the amount of time until
the threshold crossing a(T2) = I given u(0) = 0 and a(0) = I + α, where a(t)
obeys the SDE
da(t) = [−a(t)] dt/τ + dξa(t).
16
McCleney, Kilpatrick
Again, defining the potential V (a) = a2
2τ , we can compute
e[V (x)−V (y)]/σ2
a dydx.
(cid:90) −I
(cid:90) x
−I−α
−∞
¯T2 =
1
σ2
a
We compare the theory we have developed utilizing passage time problems to
residence times computed numerically in Fig. 5C. Notice that increasing the noise
amplitude tends to shorten both up and down state durations on average, due to
early threshold crossings of the variable a(t).
Furthermore, we can examine how noise reshapes the relative balance of up
versus down state durations. Specifically, we will explore how the relative fraction
of time the up state persists ¯T1/( ¯T1 + ¯T2) changes with noise intensity σa and
input I. First, notice that, in the absence of noise the ratio
ln [(φ − I)/(φ − α − I)]
T1
T1 + T2
=
ln [(I + α)(φ − I)/I(φ − α − I)]
.
(29)
The up and down state have equal duration when T1/(T1 + T2) = 1/2, or when
the input I = (φ − α)/2, as shown in Fig. 5D. Interestingly, this is the precise
input value at which the period obtains a minimum, as we demonstrated in section
3. Along with our plot of (29) in the noise-free case (σa = 0), we also study the
impact of noise on this measure of up-down state balance. Noise leads to up and
down state durations becoming more similar, so the ratio (29) of the means ¯T1 and
¯T2 flattens as a function of the input I. This is due to the fact that long durations,
wherein the variable a(t) occupies the tail of exponentially saturating functions
A0 + A1e−t/τ , are shortened by early threshold crossings due to the external noise
forcing.
6 Synchronizing two uncoupled populations
Now we demonstrate that common noise can synchronize the up and down states
of two distinct and uncoupled populations. We begin with the case of identical
noise and then, in section 7, relax these assumptions to show that some level of
coherence is still possible when each population has an intrinsic and independent
source of noise. This is motivated by the observation that the neural Langevin
equation derived in the large system-size limit of a neural master equation tends
to possess intrinsic noise in each population, in addition to an extrinsic common
noise term (Bressloff and Lai, 2011). As we will show, intrinsic noise tends to
disrupt the phase synchronization due to extrinsic noise.
To begin, we recast the stochastic system (5), describing a pair of adapting
noise-driven neural populations, as a pair of phase equations:
dθ1(t) = ωdt + Z(θ1(t)) · dξ(t),
dθ2(t) = ωdt + Z(θ2(t)) · dξ(t),
(30a)
(30b)
where θ1 and θ2 are the phase of the first and second neural populations. As
we demonstrate in Fig. 6A, this introduction of common noise tends to drive
the oscillation phases θ1(t) and θ2(t) toward one another. Note that since the
governing equations of both populations are the same, then the phase sensitivity
Population models of up and down states
17
Fig. 6 Synchronizing slow oscillations in two uncoupled populations described by (5) with
sigmoidal firing rate (2). (A) Single realization of the system (5) driven by common noise ξa
to the adaptation variable ((cid:104)ξ2
a(cid:105) = ε2t, ε = 0.01) with input I = 0.2 and adaptation timescale
τ = 50. Notice that the phase difference ψ(t) = ∆1(t) − ∆2(t) roughly decreases over time.
(B) Plot of the log of the phase difference y(t) = ln ψ(t) for several realizations (thin lines)
compared with the theory (thick line) of the mean y(0) + λt computed using the Lyapunov
exponent (35). (C) Lyapunov exponent λ decreases as a function of the adaptation timescale
τ , for I = 0.2. We compare numerical simulations (dots) to theory (solid). (D) Lyapunov
exponent λ varies non-monotonically with the strength of the input I. Other parameters are
α = 0.5, γ = 15, and φ = 1.
function Z(θ) will be the same for both. Furthermore, the synchronized solution
θ1(t) = θ2(t) is absorbing -- once the phases synchronize, they remain so. We
can analytically calculate the Lyapunov exponent λ of the synchronized state to
determine its stability. In particular, we will be interested in how this stability
depends on the parameters that shape the dynamics of adaptation.
We next convert the pair of Stratonovich differential equations into a equivalent
pair of Ito differential equations:
(cid:105)
(cid:105)
dt + Z(θ1(t)) · dξ(t),
dt + Z(θ2(t)) · dξ(t),
(31a)
(31b)
(cid:104)
(cid:104)
dθ1(t) =
dθ2(t) =
(cid:48)
(cid:48)
ω + Z
ω + Z
(θ1(t))T DZ(θ1(t))
(θ2(t))T DZ(θ2(t))
introducing a drift term due to our change in definition of the noise term. Now,
to determine stability of the synchronized state θ1(t) = θ2(t), we assume we are
infinitesimally close to it. We define ψ(t) = θ1(t) − θ2(t) and require ψ(t) (cid:28) 1.
Linearizing the system of Ito differential equations (31) with respect to ψ(t) then
18
yields
(cid:20)(cid:16)
(cid:48)
Z
dψ(t) = ψ(t)
(cid:17)(cid:48)
dt + Z
(cid:48)
(θ(t)) · dξ
(cid:21)
,
(θ(t))T DZ(θ(t))
(32)
McCleney, Kilpatrick
where θ(t) obeys either one of the equations in (31). Applying the change of
variables y(t) = ln ψ(t), we can rewrite (32) as
(cid:48)T DZ
dy(t) =(cid:0)Z
dt −(cid:16)
DZ(cid:1)(cid:48)
(cid:48) · dξ(t).
(cid:48)(cid:17)
dt + Z
(33)
Z
(cid:48)
(cid:90) t
0
(cid:90) 1
Notice, on average, the log of the phase difference y(t) tends to decrease over time
(Fig. 6B), indicating the phases θ1 and θ2 move toward one another. Subsequently,
we can integrate equation (33) to determine the mean drift of y(t)
(cid:104)(cid:0)Z
(cid:48)
(θ(s))DZ(θ(s))(cid:1)(cid:48) −(cid:16)
(cid:17)(cid:105)
(cid:48)T (θ(s))DZ
(cid:48)
(θ(s))
Z
ds,
λ := lim
t→∞
so the phase difference ψ(t) will tend to decay grow if the Lyapunov exponent
λ < 0 (λ > 0), and the synchronous state will be stable (unstable). Now, utilizing
ergodicity of (33), we can compute λ using the ensemble average across realizations
of Z(cid:48)(θ(t)) · ξ(t), so
(cid:104)(cid:0)Z
(cid:48)
(θ)DZ(θ)(cid:1)(cid:48) −(cid:16)
(cid:48)T (θ)DZ
(cid:48)
(cid:17)(cid:105)
0
λ =
Ps(θ)
(34)
where Ps(θ) is the steady state distribution of θ. Since noise is weak (Djk (cid:28) 1,
j, k = 1, 2), we can approximate the distribution as constant Ps(θ) = 1. Upon
applying this to the integrand of (34) and noting the periodicity of Z(θ), we find
we can approximate the Lyapunov exponent
dθ,
(θ)
Z
λ = −
(cid:48)T (θ)DZ
(cid:48)
Z
(θ)dθ.
(35)
Assuming noise to the activity variable u and adaptation variable a is not cor-
related, D will be diagonal. In this case, we can further decompose the phase
sensitivity function into its Fourier expansion
(cid:90) 1
0
∞(cid:88)
k=0
∞(cid:88)
2π2k2(cid:104)(cid:16)
k=0
Z(θ) =
ak sin(2πkθ) + bk cos(2πkθ),
where ak = (ak1, ak2)T and bk = (bk1, bk2)T are vectors in R2 so that
(cid:48)
Z
(θ) =
2πk [ak cos(2πkθ) − bk sin(2πkθ)] ,
and we can expand the terms in (35) to yield
∞(cid:88)
λ = −
(cid:17)
(cid:16)
a2
k2 + b2
k2
(cid:17)
(cid:105)
.
D22
a2
k1 + b2
k1
D11 +
k=0
Thus, as long as Z(θ) is continuous and non-constant, the Lyapunov exponent λ
will be negative, so the synchronous state θ1 = θ2 will be stable. Note, continuity is
Population models of up and down states
19
Fig. 7 Stationary density M0(ψ) of the phase difference ψ = θ1 − θ2 for two slowly oscillating
neural population driven by both common and independent noise (6). As the degree of noise
correlation is decreased from (A) χa = 0.95 to (B) χa = 0.90, the density spreads, but there
is still a peak at ψ = 0, the phase-locked state. We focus on noise in the adaptation variable,
so σu = 0 and σa = 0.01. Other parameters are α = 0.5, γ = 15, φ = 1, and τ = 20.
not satisfied in the case of our singular approximation to Z(θ). We demonstrate the
accuracy of our theory (35) in Fig. 6C,D, showing that λ decreases as a function
of τ and is non-monotonic in I. Thus, slow oscillations with longer periods are
synchronized more quickly, relative to the number of oscillation cycles. Since the
Lyapunov exponent has highest amplitude λ for both low and high values of the
tonic input I, we also suspect this is related to the period of the oscillation T .
7 Impact of intrinsic noise on stochastic synchronization
We now extend our results from the previous section by studying the impact
of independent noise in each population. Independent noise is incorporated into
the modified model (6). Noting, again there is a periodic solution to the noise-free
version of this system, phase-reduction methods can be used to obtain approximate
Langevin equations for the phase variables (Nakao et al., 2007)
dθ1 = ωdt + Z(θ1(t)) · [dξc(t) + dξ1(t)] ,
dθ2 = ωdt + Z(θ2(t)) · [dξc(t) + dξ2(t)] ,
where the noise vectors ξc = (χuξuc, χaξac)T and ξj = ((cid:112)1 − χ2
aξaj)T
(j = 1, 2). We can reformulate the pair of Stratonovich differential equations as
Ito stochastic differential equations given by the system
uξuj,(cid:112)1 − χ2
(36b)
(36a)
dθ1 = A1(θ)dt + dζ1(θ, t),
dθ2 = A2(θ)dt + dζ2(θ, t),
where the statistics of the noise terms dζj(θ, t) = Z(θj(t)) ·(cid:2)dξc(t) + dξj(t)(cid:3) (j =
(cid:17)
1, 2) are specified by (cid:104)dζj(θ, t)(cid:105) = 0 and (cid:104)dζj(θ, t)dζk(θ, t)(cid:105) = Cjk(θ)dt where
(cid:17)(cid:16)
(37b)
(37a)
(cid:16)
ac Za(θj)
χuD1/2
uc Zu(θk) + χaD1/2
ac Za(θk)
Cjk(θ) =
χuD1/2
(cid:16)(cid:112)
+
uc Zu(θj) + χaD1/2
1 − χ2
ul Zu(θj) +
uD1/2
(cid:112)
(cid:17)2
1 − χ2
aD1/2
al Za(θj)
δj,k,
(38)
20
McCleney, Kilpatrick
separating the impact of correlated and local sources of noise. The drift terms
can thus be calculated Aj(θ) = ω + 1
Cjj(θ). The Fokker-Planck equation
4
describing the evolution of the probability density function P (θ, t) of the phases
is given
∂
∂θj
= − 2(cid:88)
j=1
∂P
∂t
2(cid:88)
2(cid:88)
j=1
k=1
∂
∂θj
[Aj(θ)P ] +
1
2
∂2
∂θj∂θk
[Cjk(θ)P ] .
(39)
Now, we apply a change of variables to the Fokker-Planck equation (39) defined
θj = ωt + ϑj. Assuming noise is weak, the function Q(ϑ, t) varies slowly compared
with the period of the phase oscillators θj. Thus, we average the drifts Aj(θ)
and correlation function Cjk(θ) over a single period. The resulting Fokker-Planck
equation is then
2(cid:88)
2(cid:88)
j=1
k=1
∂Q(ϑ, t)
∂t
=
1
2
∂2
∂ϑj∂ϑk
[Bjk(ϑ)Q]
where the averaged correlation function is given by the formula
Bjk(ϑ) = g(θ1 − θ2) + h(0)δj,k,
(cid:90) 1
(cid:104)
(cid:105)
(cid:48)
dθ
where the correlation functions are defined
(cid:20) ∂2Q
∂ϑ2
1
(cid:21)
g(θ) =
χ2
uDucZu(θ
(cid:48)
(cid:48)
)Zu(θ
+ θ) + χ2
aDacZa(θ
(cid:48)
(cid:48)
)Za(θ
+ θ)
and
h(θ) =
(cid:90) 1
(cid:104)
0
0
(1 − χ2
u)DulZu(θ
(cid:48)
(cid:48)
)Zu(θ
+ θ) + (1 − χ2
a)DalZa(θ
(cid:48)
(cid:48)
)Za(θ
+ θ)
(cid:105)
(cid:48)
.
dθ
We study the relationship between the phases ϑ1 and ϑ2 by substituting the for-
mula for the averaged correlation matrix
∂Q(ϑ, t)
∂t
=
1
2
[g(0) + h(0)]
+
∂2Q
∂ϑ2
2
+
∂2
∂ϑ1∂ϑ2
[g(ϑ1 − ϑ2)Q] .
(40)
We can write (40) as a separable equation by employing a change of variables that
tracks the average ρ = (ϑ1 + ϑ2)/2 and phase difference ψ = ϑ1− ϑ2 of the original
position variables, so
∂U (ρ, t)
∂t
∂M (ψ, t)
∂t
=
=
1
4
∂2
[g(0) + g(ψ) + h(0)]
∂2U (ρ, t)
∂ρ2
,
∂ψ2 [g(0) − g(ψ) + h(0)] M (ψ, t).
(41a)
(41b)
Thus, we can solve for the stationary solution of the system (41a) by serving Ut =
Mt ≡ 0 and requiring periodic boundary conditions. We find that the stationary
distribution of the position average is U0(ρ) = 1. In addition, we can integrate the
stationary equation for M (ψ, t) to find
M0(ψ) =
u [(2 − χ2
σ2
u)gu(0) − χ2
ugu(ψ)] + σ2
a [(2 − χ2
a)ga(0) − χ2
aga(ψ)]
m0
,
(42)
Population models of up and down states
21
where m0 = 1/(cid:82) 1
simplified the expression using Du1 = Du2 ≡ Dul = σ2
σ2
a and defined
0 [g(0) − g(x) + h(0)]
(cid:90) 1
−1 dx is a normalization factor and we have
u and Da1 = Da2 ≡ Dal =
gj(ψ) =
Zj(θ)Zj(θ + φ)dθ.
0
When noise to each layer is independent (χu, χa → 0), then M0(ψ) = 1 is constant
in space. When noise is totally correlated (χu, χa → 1), then M0(ψ) = δ(φ).
The stationary distribution M0(ψ) will broaden as the correlations χu and χa
are decreased from unity, with a peak remaining at φ = 0. We demonstrate the
accuracy of the formula (42) for the stationary density of the phase difference ψ
in Fig. 7, showing that it widens as the level noise correlation is decreased. Again,
we focus on the impact of adaptation noise. Thus, even when independent noise
is introduced, there is some semblance of synchronization in the phases of two
noise-driven neural populations (6).
8 Discussion
We have studied the impact of deterministic and stochastic perturbations to a
neural population model of slow oscillations. The model was comprised of a sin-
gle recurrently coupled excitatory population with negative feedback from a slow
adaptive current (Laing and Chow, 2002; Jayasuriya and Kilpatrick, 2012). By
examining the phase sensitivity function (Zu, Za), we found that perturbations of
the adaptation variable lead to much larger changes in oscillation phase than per-
turbations of neural activity. Furthermore, this effect becomes more pronounced as
the timescale τ of adaptation is increased. Introducing noise in the model decreases
the oscillation period and helps to balance the mean duration of the oscillation's
up and down states. When two uncoupled populations receive common noise, their
oscillation phases θ1 and θ2 eventually become synchronized, which can be shown
by deriving a formula for the Lyapunov exponent of the absorbing state θ1 ≡ θ2
(Teramae and Tanaka, 2004). When independent noise is introduced to each popu-
lation, in addition to common noise, the long-term state of the system is described
by a probability density for ψ = θ1 − θ2, which peaks at ψ ≡ 0.
Our study was motivated by the observation that recurrent cortical networks
can spontaneously generate stochastic oscillations between up and down states.
Guided by previous work in spiking models (Compte et al., 2003), we explored a
rate model of a recurrent excitatory network with slow spike frequency adaptation.
One of the open questions about up and down state transitions concerns the degree
to which they are generated by noise or by more deterministic mechanisms, such
as slow currents or short term plasticity (Cossart et al., 2003). Here, we have
provided some characteristic features that emerge as the level of noise responsible
for transitions is increased. Similar questions have been probed in the context
of models of perceptual rivalry (Moreno-Bote et al., 2007). In addition, we have
provided a plausible mechanism whereby the onset of up and down states could
be synchronized in distinct networks (Volgushev et al., 2006).
There are several potential extensions to this work. For instance, we could ex-
amine the impact of long-range connections between networks to see how these
22
McCleney, Kilpatrick
interact with common and independent noise to shape the phase coherence of
oscillations. Similar studies have been performed in spiking models by Ly and Er-
mentrout (2009). Interestingly, shared noise can actually stabilize the anti-phase
locked state in this case, even though it is unstable in the absence of noise. Fur-
thermore, it is known that coupling spanning long distances can be subject to
axonal delays. In spite of this, networks of distantly coupled clusters of cells can
still sustain zero-lag synchronized states (Vicente et al., 2008). Thus, we could also
explore the impact of delayed coupling, determining how features of phase sensi-
tivity function interact with delay to promote in-phase or anti-phase synchronized
states.
References
Amari S (1977). Dynamics of pattern formation in lateral-inhibition type neural
fields. Biol Cybern 27:77 -- 87.
Amit DJ, Brunel N (1997). Model of global spontaneous activity and local struc-
tured activity during delay periods in the cerebral cortex. Cerebral cortex 7:237 --
252.
Bart E, Bao S, Holcman D (2005). Modeling the spontaneous activity of the
auditory cortex. Journal of computational neuroscience 19:357 -- 378.
Benda J, Herz AVM (2003). A universal model for spike-frequency adaptation.
Neural Comput 15:2523 -- 2564.
Bressloff PC (2010). Metastable states and quasicycles in a stochastic wilson-cowan
model of neuronal population dynamics. Physical Review E 82:051903.
Bressloff PC, Lai YM (2011). Stochastic synchronization of neuronal populations
with intrinsic and extrinsic noise. The Journal of Mathematical Neuroscience
(JMN) 1:1 -- 28.
Brown E, Moehlis J, Holmes P (2004). On the phase reduction and response
dynamics of neural oscillator populations. Neural computation 16:673 -- 715.
Buzs´aki G, Draguhn A (2004). Neuronal oscillations in cortical networks. science
304:1926 -- 1929.
Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003). Cellular and
network mechanisms of slow oscillatory activity (¡ 1 hz) and wave propagations
in a cortical network model. Journal of neurophysiology 89:2707 -- 2725.
Cossart R, Aronov D, Yuste R (2003). Attractor dynamics of network up states
in the neocortex. Nature 423:283 -- 8.
Ermentrout B (1996). Type i membranes, phase resetting curves, and synchrony.
Neural computation 8:979 -- 1001.
Ermentrout GB, Gal´an RF, Urban NN (2008). Reliability, synchrony and noise.
Trends in neurosciences 31:428 -- 434.
Gal´an RF, Fourcaud-Trocm´e N, Ermentrout GB, Urban NN (2006). Correlation-
induced synchronization of oscillations in olfactory bulb neurons. The Journal
of neuroscience 26:3646 -- 3655.
Gardiner CW (2004). Handbook of stochastic methods for physics, chemistry, and
the natural sciences. Berlin: Springer-Verlag.
Jayasuriya S, Kilpatrick ZP (2012). Effects of time-dependent stimuli in a com-
petitive neural network model of perceptual rivalry. Bulletin of mathematical
biology 74:1396 -- 1426.
Population models of up and down states
23
Kilpatrick ZP, Bressloff PC (2010). Effects of synaptic depression and adaptation
on spatiotemporal dynamics of an excitatory neuronal network. Physica D:
Nonlinear Phenomena 239:547 -- 560.
Laing CR, Chow CC (2002). A spiking neuron model for binocular rivalry. J
Comput Neurosci 12:39 -- 53.
Litwin-Kumar A, Doiron B (2012). Slow dynamics and high variability in balanced
cortical networks with clustered connections. Nature neuroscience 15:1498 -- 1505.
Ly C, Ermentrout GB (2009). Synchronization dynamics of two coupled neural os-
cillators receiving shared and unshared noisy stimuli. Journal of computational
neuroscience 26:425 -- 443.
Major G, Tank D (2004). Persistent neural activity: prevalence and mechanisms.
Current opinion in neurobiology 14:675 -- 684.
Massimini M, Huber R, Ferrarelli F, Hill S, Tononi G (2004). The sleep slow
oscillation as a traveling wave. The Journal of Neuroscience 24:6862 -- 6870.
Melamed O, Barak O, Silberberg G, Markram H, Tsodyks M (2008). Slow oscil-
lations in neural networks with facilitating synapses. Journal of computational
neuroscience 25:308 -- 316.
Moreno-Bote R, Rinzel J, Rubin N (2007). Noise-induced alternations in an at-
tractor network model of perceptual bistability. J Neurophysiol 98:1125 -- 39.
Nakao H, Arai K, Kawamura Y (2007). Noise-induced synchronization and clus-
tering in ensembles of uncoupled limit-cycle oscillators. Physical review letters
98:184101.
Renart A, Moreno-Bote R, Wang XJ, Parga N (2007). Mean-driven and
fluctuation-driven persistent activity in recurrent networks. Neural computa-
tion 19:1 -- 46.
Sanchez-Vives MV, McCormick DA (2000). Cellular and network mechanisms of
rhythmic recurrent activity in neocortex. Nat Neurosci 3:1027 -- 34.
Smeal RM, Ermentrout GB, White JA (2010). Phase-response curves and syn-
chronized neural networks. Philosophical Transactions of the Royal Society B:
Biological Sciences 365:2407 -- 2422.
Steriade M, Nunez A, Amzica F (1993). A novel slow (¡ 1 hz) oscillation of
neocortical neurons in vivo: depolarizing and hyperpolarizing components. The
Journal of Neuroscience 13:3252 -- 3265.
Tabak J, Senn W, O'Donovan MJ, Rinzel J (2000). Modeling of spontaneous
activity in developing spinal cord using activity-dependent depression in an ex-
citatory network. The Journal of Neuroscience 20:3041 -- 3056.
Teramae Jn, Tanaka D (2004). Robustness of the noise-induced phase synchro-
nization in a general class of limit cycle oscillators. Physical review letters
93:204103.
Traub RD, Whittington MA, Stanford IM, Jefferys JG (1996). A mechanism for
generation of long-range synchronous fast oscillations in the cortex .
Vicente R, Gollo LL, Mirasso CR, Fischer I, Pipa G (2008). Dynamical relay-
ing can yield zero time lag neuronal synchrony despite long conduction delays.
Proceedings of the National Academy of Sciences 105:17157 -- 17162.
Volgushev M, Chauvette S, Mukovski M, Timofeev I (2006). Precise long-range
synchronization of activity and silence in neocortical neurons during slow-wave
sleep. The Journal of Neuroscience 26:5665 -- 5672.
Wang XJ (2001). Synaptic reverberation underlying mnemonic persistent activity.
Trends Neurosci 24:455 -- 63.
|
1809.05196 | 1 | 1809 | 2018-09-13T22:27:24 | Statistical Data Assimilation: Formulation and Examples from Neurobiology | [
"q-bio.NC",
"physics.bio-ph"
] | For the Research Topic Data Assimilation and Control: Theory and Applications in Life Sciences we first review the formulation of statistical data assimilation (SDA) and discuss algorithms for exploring variational approximations to the conditional expected values of biophysical aspects of functional neural circuits. Then we report on the application of SDA to (1) the exploration of properties of individual neurons in the HVC nucleus of the avian song system, and (2) characterizing individual neurons formulated as very large scale integration (VLSI) analog circuits with a goal of building functional, biophysically realistic, VLSI representations of functional nervous systems. Networks of neurons pose a substantially greater challenge, and we comment on formulating experiments to probe the properties, especially the functional connectivity, in song command circuits within HVC. | q-bio.NC | q-bio |
Statistical Data Assimilation: Formulation and
Examples from Neurobiology
Anna Miller 1,∗, Dawei Li 1 Jason Platt 1 Arij Daou 4 Daniel Margoliash 2 and
Henry D. I. Abarbanel 1,3
1Department of Physics, University of California San Diego, La Jolla , CA, USA
2 Department of Anatomy and Organismal Biology, University of Chicago, Chicago ,
IL, USA
3 Marine Physical Laboratory (Scripps Institution of Oceanography) and Department
of Physics, University of California, La Jolla, CA, USA
4 Department of Biomedical Engineering, American University of Beirut, Beirut,
Lebanon
Correspondence*:
Anna Miller
[email protected]
ABSTRACT
For the Research Topic Data Assimilation and Control: Theory and Applications in Life Sciences
we first review the formulation of statistical data assimilation (SDA) and discuss algorithms for
exploring variational approximations to the conditional expected values of biophysical aspects
of functional neural circuits. Then we report on the application of SDA to (1) the exploration
of properties of individual neurons in the HVC nucleus of the avian song system, and (2)
characterizing individual neurons formulated as very large scale integration (VLSI) analog circuits
with a goal of building functional, biophysically realistic, VLSI representations of functional
nervous systems. Networks of neurons pose a substantially greater challenge, and we comment
on formulating experiments to probe the properties, especially the functional connectivity, in song
command circuits within HVC.
Keywords: Data Assimilation, Neuronal Dynamics, HVC, Ion Channel Properties, Variational Annealing, Neuromorphic, VLSI
1 INTRODUCTION
A broad class of 'inverse' problems presents itself in many scientific and engineering inquiries. The overall
question addressed by these is how to transfer information from laboratory and field observations to
candidate models of the processes underlying those observations.
The existence of large,
information rich, well curated data sets from increasingly sophisticated
observations on complicated nonlinear systems has set new challenges to the information transfer task.
Assisting with this challenge are new substantial computational capabilities.
Together they have provided an arena in which principled formulation of this information transfer along
with algorithms to effect the transfer have come to play an essential role. This paper reports on some efforts
to meet this class of challenge within neuroscience. Many of the ideas are applicable much more broadly
than our focus, and we hope the reader will find this helpful in their own inquiries.
1
Miller et al.
In this special issue entitled Data Assimilation and Control: Theory and Applications in Life Sciences,
of the journal Frontiers in Applied Mathematics and Statistics -- Dynamical Systems, we participate in the
broader quantitative setting for this information transfer. The procedures are called "data assimilation"
following its use in the effort to develop realistic numerical weather prediction models [1, 2] over many
decades. We prefer the term 'statistical data assimilation' (SDA) to emphasize that key ingredients in the
procedures involved in the transfer rest on noisy data and on recognizing errors in the models to which
information in the noisy data is to be transferred.
This article begins with a formulation of SDA with some additional clarity beyond the discussion
in [3]. We also discuss some algorithms helpful for implementing the information transfer, testing model
compatibility with the available data, and quantitatively identifying how much information in the data can
be represented in the model selected by the SDA user. Using SDA will also remind us that data assimilation
efforts are well cast as problems in statistical physics [4].
After the discussion of SDA, we turn to some working ideas on how to perform the high dimensional
integrals involved in SDA. In particular we address the 'standard model' of SDA where data is contaminated
by Gaussian noise and model errors are represented by Gaussian noise, though the integrals to be performed
are, of course, not Gaussian. The topics include the approximation of Laplace [5, 6] and Monte Carlo
methods.
With these tools in hand, we turn to neurobiological questions that arise in the analysis of individual
neurons and, in planning, for network components of the avian song production pathway. These questions
are nicely formulated in the general framework, and we dwell on specifics of SDA in a realistic biological
context. The penultimate topic we address is the use of SDA to calibrate VLSI analog chips designed and
built as components of a developing instantiation of the full songbird song command network, called HVC.
Lastly, we discuss the potential utlization of SDA for exploring biological networks.
At the outset of this article we may expect that our readers from Physics and Applied Mathematics
along with our readers from Neurobiology may find the conjunction of the two "strange bedfellows" to be
incongruous. For the opportunity to illuminate the natural melding of the facets of both kinds of questions,
we thank the editors of this special issue.
2 MATERIALS AND METHODS
2.1 General Overview of Data Assimilation
We will provide a structure within which we will frame our discussion of transfer of information from
data to a model of the underlying processes producing the data.
We start with a observation window in time [t0, tF ] within which we make a set of measurements at
times t = {τ1, τ2, ..., τk, ..., τF}; t0 ≤ τk ≤ tF . At each of these measurement times, we observe L
quantities y(τk) = {y1(τk), y2(τk), ..., yL(τk)}. The number L of observations at each measurement time
τk is typically less, often much less, than the number of degrees of freedom D in the observed system;
D (cid:29) L.
These are views into the dynamical processes of a system we wish to characterize. The quantitative
characterization is through a model we choose. It describes the interactions among the states of the observed
system. If we are observing the time course of a neuron, for example, we might measure the membrane
voltage y1(τk) = Vm(τk) and the intracellular Ca2+ concentration y2(τk) = [Ca2+](τk). From these data
This is a provisional file, not the final typeset article
2
Miller et al.
we want to estimate the unmeasured states of the model as a function of time as well as estimate biophysical
parameters in the model.
The processes characterizing the state of the system (neuron) we call xa(t); a = 1, 2, ..., D ≥ L, and they
are selected by the user to describe the dynamical behavior of the observations through a set of equations
in continuous time
= Fa(x(t), q),
dxa(t)
dt
or in discrete time tn = t0 + n∆t; n = 0, 1, ..., N ; tN = tF via
xa(tn+1) = xa(n + 1) = fa(x(tn), q) = fa(x(n), q),
(1)
(2)
where q is a set of fixed parameters associated with the model. f (x(n), q) is related to F(x(t), q) through
the choice the user makes for solving the continuous time flow for x(t) through a numerical solution
method of choice [7].
Considering neuronal activity, Eq. (1) could be coupled Hodgkin-Huxley (HH) equations [8, 9] for
voltage, ion channel gating variables, constituent concentrations, and other ingredients. If the neuron is
isolated in vitro, such as by using drugs to block synaptic transmission, then there would be no synaptic
input to the cell to describe. While if it is coupled to a network of neurons, their functional connectivity
would be described in F(x(t), q) or f (x(n), q). Typical parameters might be maximal conductances of the
ion channels, reversal potentials, and other time-independent numbers describing the kinetics of the gating
variables. In many experiments L is only 1, namely, the voltage across the cell membrane, while D may be
on the order of 100; Hence D (cid:29) L.
As we proceed from the initiation of the observation window at t0 we must move our model equation
variables x(0), Eq. (2), from t0 to τ1 where a measurement is made. Then using the model dynamics we
move along to τ2 and so forth until we reach the last measurement time τF and finally move the model
from x(τF ) to x(tF ). In each stepping of the model equations Eq. (2) we may make many steps of ∆t in
time to achieve accuracy in the representation of the model dynamics. The full set of times tn at which we
evaluate the model x(tn) we collect into the path of the state of the model through D-dimensional space:
X = {x(0), x(1), ..., x(n), ..., x(N ) = x(F )}. The dimension of the path is (N + 1)D + Nq, where Nq is
the number of parameters q in our model.
It is worth a pause here to note that we have now collected two of the needed three ingredients to effect
our transfer of the information in the collection of all measurements Y = {y(τ1), y(τ2), ..., y(τF )} to the
model f (x(n), q) along the path X through the observation window [t0, tF ]: (1) data Y and (2) a model of
the processes in Y, devised by our experience and knowledge of those processes. The notation and a visual
presentation of this is found in Fig. (1).
The third ingredient, comprised of methods to generate the transfer from Y to properties of the model,
will command our attention throughout most of this paper. If the transfer methods are successful and,
according to some metric of success, we arrange matters so that at the measurement times τk, the L model
variables x(t) associated with y(τk) are such that xl(τk) ≈ yl(τk), we are not finished. We have then only
demonstrated that the model is consistent with the known data Y. We must use the model, completed by
the estimates of q and the state of the model at tF , x(tF ), to predict forward for t > tF , and we should
succeed in comparison with measurements for y(τr) for τr > tF . As the measure of success of predictions,
we may use the same metric as utilized in the observation window.
Frontiers
3
Miller et al.
Figure 1. A visual representation of the window t0 ≤ t ≤ tF in time during which L-dimensional
observations y(τk) are performed at observation times t = τk; k = 1, ..., F . This also shows times at
which the D-dimensional model developed by the user x(n + 1) = f (x(n), q) is used to move forward
from time n to time n + 1: tn = t0 + n∆t; n = 0, 1, ..., N. D ≥ L. The path of the model X =
{x(0), x(1), ..., x(n), ..., x(N ) = x(F )} and the collection Y of L-dimensional observations at each
observation time τk, Y = {y(τ1), y(τ2), ..., y(τk), ..., y(τF} (y = {y1, y2, ..., yL}) is also indicated.
As a small aside, the same overall setup applies to supervised machine learning networks [10] where the
observation window is called the training set; the prediction window is called the test set, and prediction is
called generalization.
2.1.1 The Data are Noisy; the Model has Errors
Inevitably, the data we collect is noisy, and equally the model we select to describe the production of
those data has errors. This means we must, at the outset, address a conditional probability distribution
P (XY) as our goal in the data assimilation transfer from Y to the model. In [3] we describe how to use the
Markov nature of the model x(n) → x(n + 1) = f (x(n), q) and the definition of conditional probabilities
to derive the recursion relation:
P (X(n + 1)Y(n + 1)) =
P (y(n + 1), x(n + 1), X(n)Y(n))
P (y(n + 1)Y(n)) P (x(n + 1), X(n + 1)Y(n))
P (x(n + 1)x(n))P (X(n)Y(n))
= exp[CM I(y(n + 1), x(n + 1), X(n)Y(n))] •
P (x(n + 1)x(n))P (X(n)Y(n)),
This is a provisional file, not the final typeset article
•
(3)
4
Miller et al.
(cid:104) (P (a,bc)
(cid:105)
where we have identified CM I(a, bc) = log
. This is Shannon's conditional mutual
information [11] telling us how many bits (for log2) we know about a when observing b conditioned
on c. For us a = {y(n + 1)}, b = {x(n + 1), X(n + 1)}, c = {Y(n)}.
P (ac) P (ac)
Using this recursion relation to move backwards through the observation window from tF = t0 + N ∆t
through the measurements at times τk to the start of the window at t0, we may write, up to factors
independent of X
P (XY) =
P (y(τk)X(τk))
P (x(n + 1)x(n))
P (x(0)).
(4)
k=1
n=0
If we now write P (XY) ∝ exp[−A(X)] where A(X), the negative of the log likelihood, we call the action,
then conditional expected values for functions along the path X are defined by
(cid:26) F(cid:89)
N−1(cid:89)
E[G(X)Y] = (cid:104)G(X)(cid:105) =
,
(5)
(cid:82) dX G(X)e−A(X)
(cid:82) dX e−A(X)
dX =(cid:81)N
convenient expression
(cid:26) F(cid:88)
N−1(cid:88)
(cid:27)
(cid:27)
n=0 dDx(n), and all factors in the action independent of X cancel out here. The action takes the
A(X) = −
log[P (y(τk)X(τk)] +
log[P (x(n + 1)x(n))]
− log[P (x(0))],
(6)
k=1
n=0
which is the sum of the terms which modify the conditional probability distribution when an observation
is made at t = τk and the sum of the stochastic version of x(n) → x(n + 1) − f (x(n), q) and finally the
distribution when the observation window opens at t0.
What quantities G(X) are of interest? One natural one is the path G(X) = Xµ; µ = {a, n} itself; another
is the covariance around that mean (cid:104)Xµ(cid:105) = ¯Xµ = (cid:104)Xµ(cid:105) : (cid:104)(Xµ − ¯Xµ)(Xν − ¯Xν)(cid:105). Other moments are of
interest, of course. If one has an anticipated form for the distribution at large X, then G(X) may be chosen
as a parametrized version of that form and those parameters determined near the maximum of P (XY).
The action simplifies to what we call the 'standard model' of data assimilation when (1) observations y
are related to their model counterparts via Gaussian noise with zero mean and diagonal precision matrix
Rm, and (2) model errors are associated with Gaussian errors of mean zero and diagonal precision matrix
Rf :
A(X) =
Rm(k)
2
(xl(τk) − yl(τk))2 +
Rf (a)
2
(xa(n + 1) − fa(x(n), q))2.
(7)
If we have knowledge of the distribution P (x(0)) at t0 we may add it to this action. If we have no knowledge
of P (x(0)), we may take its distribution to be uniform over the dynamic range of the model variables, then
it, as here, is absent, canceling numerator and denominator in Eq. (5).
Our challenge is to perform integrals such as Eq. (5). One should anticipate that the dominant contribution
to the expected value comes from the maxima of P (XY) or, equivalently the minima of A(X). At such
Frontiers
5
N−1(cid:88)
D(cid:88)
n=0
a=1
F(cid:88)
L(cid:88)
k=1
l=1
Miller et al.
minima, the two contributions to the action, the measurement error and the model error, balance each other
to accomplish the explicit transfer of information from the former to the latter.
We note, as before, that when f (x(n), q) is nonlinear in X, as it always is in interesting examples, the
expected value integral Eq. (5) is not Gaussian. So, some thinking is in order to approximate this high
dimensional integral. We turn to that now. After consideration of methods to do the integral, we will return
to a variety of examples taken from neuroscience.
The two generally useful methods available for evaluating this kind of high dimensional integral are
Laplace's method [5, 6] and Monte Carlo techniques [7, 12, 13]. We address them in order. We also add
our own new and useful versions of the methods.
2.1.2 Laplace's Method
To locate the minima of the action A(X) = − log[P (XY)] we must seek paths X(j); j = 0, 1, ... such
that ∂A(X)/∂XX(j) = 0, and then check that the second derivative at X(j), the Hessian, is a positive
definite matrix in path coordinates. The vanishing of the derivative is a necessary condition.
Laplace's method [5] expands the action around the X(j) seeking the path X(0) with the smallest value
of A(X). The contribution of X(0) to the integral Eq. (5) is approximately exp[A(X(1)) − A(X(0))] bigger
than that of the path with the next smallest action.
This sounds more or less straightforward; however, finding the global minimum of a nonlinear function
such as A(X) is an NP-complete problem [14]. In a practical sense one cannot expect to succeed with such
problems. However there is an attractive feature of the form of A(X) that permits us to accomplish more.
We now discuss two algorithmic approaches to implementing Laplace's method.
2.1.3 Precision Annealing for Laplace's Method
Looking at Eq. (7) we see that if the precision of the model is zero, Rf = 0, the action is quadratic in
the L measured variables xl(n) and independent of the remaining states. The global minimum of such
an action comes with xl(τk) = yl(τk) and any choice for the remaining states and parameters. Choose
the path with these values of x(τk) and values from a uniform distribution of the other state variables and
parameters covering the expected dynamic range of those, and call it path Xinit. In practice, we recognize
that the global minimum of A(X) is degenerate at Rf = 0, so we select many initial paths. We choose NI
of them, and initialize whatever numerical optimization program we have selected, to run on each of them.
We continue to call the collection of NI paths Xinit.
• Now we increase Rf from Rf = 0 to a small value Rf 0. Use each of the NI paths in Xinit as initial
conditions for our numerical optimization program chosen to find the minima of A(X), and we arrive
at NI paths X0. Evaluate A(X0) on all NI paths X0.
• We increase Rf = Rf 0 → Rf 0α; α > 1, and now use the NI paths X0 as the initial conditions for
our numerical optimization program chosen to find the minima of A(X), we arrive at NI paths X1.
Evaluate A(X1) on all NI paths X1.
• We increase Rf = Rf 0α → Rf 0α2. Now use the NI paths X1 as the initial conditions for our numerical
optimization program chosen to find the minima of A(X), we arrive at NI paths X2. Evaluate A(X2)
on all NI paths X2.
• Continue in this manner increasing Rf to Rf = Rf 0αβ; β = 0, 1, ..., then using the selected numerical
optimization program to arrive at NI paths Xβ. Evaluate A(Xβ) on all NI paths Xβ.
This is a provisional file, not the final typeset article
6
Miller et al.
• As a function of logα
(cid:105)
(cid:104) Rf
Rf 0
display all NI values of A(Xβ) versus β for all β = 0, 1, 2, ...βmax.
We call this method precision annealing (PA) [15, 16, 17, 18]. It slowly turns up the precision of the
model collecting paths at each Rf that emerged from the degenerate global minimum at Rf = 0. In practice
it is able to track NI possible minima of A(X) at each Rf . When not enough information is presented to
the model, that is L is too small, there are many local minima at all Rf . This is a manifestation of the
NP-completeness of the minimization of A(X) problem. None of the minima may dominate the expected
value integral of interest.
As L increases, and enough information is transmitted to the model, for large Rf one minimum appears
to stand out as the global minimum, and the paths associated with that smallest minimum yields good
√
predictions. We note that there are always paths, not just a single path, as we have a distribution of paths,
NI of which are sampled in the PA procedure, within a variation of size 1/
Rm. A clear example of this
is seen in [19] in a small, illustrative model.
2.1.4 "Nudging" within Laplace's Method
In meteorology one approach to data assimilation is to add a term to the deterministic dynamics which
move the state of a model towards the observations [20]
xa(n + 1) = fa(x(n), q) + u(n)(yl(n) − xl(n))δal,
(8)
where u(n) > 0 and vanishes except where a measurement is available. This is referred to as 'nudging'. It
appears in an ad hoc, but quite useful, manner.
Within the structure we have developed, one may see that the 'nudging term' arises through the balance
between the measurement error term and the model error term in the action. This is easy to see when we
look at the continuous time version of the data assimilation standard model
A(x(t), x(t)) =
dt
Rm(t, l)
(xl(t) − yl(t))2
(cid:90) tF
(cid:26) L(cid:88)
t0
l=1
2
(cid:27)
Rf (a)
2
( xa(t) − Fa(x(t), q))2
.
(9)
D(cid:88)
a=1
+
(cid:20)
The extremum of this action is given by the Euler-Lagrange equations for the variational problem [21]
(cid:21)(cid:20)
(cid:21)
δab
d
dt
+
=
Rm(a, t)
Rf (a)
xb(t) − Fb(x(t))
∂Fb(x(t)
∂xa(t)
δal(xl(t) − yl(t)),
(10)
in which the right hand side is the 'nudging' term appearing in a natural manner. Approximating the
operator δab
∂xa(t) we can rewrite this Euler-Lagrange equation in 'nudging' form
d
dt + ∂Fb(x(t)
Frontiers
dxa(t)
dt
= Fa(x(t)) + u(t)δal(xl(t) − yl(t)).
(11)
7
Miller et al.
We will use both the full variation of the action, in discrete time, as well as its nudging form in our
examples below.
2.1.5 Monte Carlo Methods
Monte Carlo methods [22, 12, 18, 7] are well covered in the literature. We have not used them in
the examples in this paper. However, the development of a precision annealing version of Monte Carlo
techniques promises to address the difficulties with large matrices for the Jacobian and Hessians required
in variational principles [23]. When one comes to network problems, about which we comment later, this
method may be essential.
3 RESULTS
3.1 Using SDA to Analyze the Avian Song System
We take our examples of the use of SDA in neurobiology from experiments on the avian song system.
These have been performed in the University of Chicago laboratory of Daniel Margoliash, and we do
not plan to describe in any detail the experiments nor the avian song production pathways in the avian
brain. We give the essentials of the experiments and direct the reader to our references to develop the full
biologically oriented idea why this system is enormously interesting.
Essentially, however, the manner in which songbirds learn and produce their functional vocalization --
song -- is an elegant non-human example of a behavior that is cultural: the song is determined both by a
genetic substrate and, interestingly, by refinement on top of that substrate by juveniles learning the song
from their (male) elders. The analogs to learning speech in humans [24] are striking.
Our avian singer is a zebra finch. They, as do most other songbirds, learn vocal expression through
auditory feedback [25, 26, 27, 28, 24]. This makes the study of the song system a good model for learning
complex behavior [29, 27, 30]. Parts of the song system are analogous to the mammalian basal ganglia and
regions of the frontal cortex [27, 31, 32]. Zebra finch in particular have the attractive property of singing
only a single learned song, and with high precision, throughout their adult life.
Beyond the auditory pathways themselves, two neural pathways are principally responsible for song
acquisition and production in zebra finch. The first is the Anterior Forebrain Pathway (AFP) which
modulates learning. The second is a posterior pathway responsible for directing song production: the Song
Motor Pathway (SMP) [26, 28, 33]. The HVC nucleus in the avian brain uniquely contributes to both of
these [28].
There are two principal classes of projection neurons which extend from HVC: neurons which project to
the robust nucleus of the arcopallium (HVCRA), and neurons which project to Area X (HVCX). HVCRA
neurons extend to the SMP pathway and HVCX neurons extend to the AFP [28, 34]. These two classes
of projection neurons combined with classes of HVC interneurons, make up the three broad classes of
neurons within HVC. Fig. (2) [35] displays these structures in the avian brain.
In vitro observations of each HVC cell type have been obtained through patch-clamp techniques making
intracellular voltage measurements in a reduced, brain slice preparation [25]. In this configuration, the
electrode can simultaneously inject current into the neuron while measuring the whole cell voltage
response [36]. From these data, one can establish the physical parameters of the system [25]. Traditionally
this is done using neurochemicals to block selected ion channels and measuring the response properties of
others [37]. Single current behavior is recorded and parameters are found through mathematical fits of the
This is a provisional file, not the final typeset article
8
Miller et al.
data. This procedure has its drawbacks, of course. There are various technical problems with the choice of
channel blockers. Many of even the modern channel blockers are not subtype specific [38] and may only
partially block channels [39]. A deeper conceptual problem is that is difficult to know what channels one
may have missed altogether. Perhaps there are channels which express themselves only outside the bounds
of the experimental conditions.
Our solution to such problems is the utilization of statistical data assimilation (SDA). This a method
developed by meteorologists and others as computational models of increasingly large dynamical systems
have been desired. Data assimilation has been described in our earlier sections.
In this paper, we focus on the song learning pathway, reporting on experiments involving the HVCX
neuron. The methods are generally applicable to the other neurons in HVC, and actually, to neurons seen
as dynamical systems in general.
We start with a discussion about the neuron model. First we demonstrate the utility of our precision
annealing methods through the use of twin experiments. These are numerical experiments in which 'data' is
generated through a known model (of HVCX), then analyzed via precision annealing. In a twin experiment,
we know everything, so we can verify the SDA method by looking at predictions after a observation
window in which the model is trained, and we may also compare the estimations of unobserved state
variables and parameters to the ingredients and output of the model.
Twin experiments are meant to mirror the circumstances of the real experiment. We start by taking the
model that we think describes our physical system. Initial points for the state variables and parameters
are chosen, which are used along with the model to numerically integrate forward in time. This leaves us
with complete information about the system. Noise is added to a subset of the state variables to emulate
the data to be collected in a lab experiment. We then perform PA on these simulated data, as if they were
real data. The results of these numerical experiments can be used to inform laboratory experiments, and
indeed help design them, by identifying the necessary measurements and stimulus needed to accurately
electrophysiologically characterize a neuron.
The second set of SDA analyses we report on using 'nudging', as described above, to estimate some key
biophysical properties of HVCX neurons from laboratory data. This SDA procedure is applied to HVCX
neurons in two different birds. The results show that though each bird is capable of normal vocalization,
their complement of ion channel strengths is apparently different. We report on a suggestive example of
this, leaving a full discussion to [40].
In order to obtain good estimation results, we must choose a forcing or stimulus with the model in mind:
the dynamical range of the neuron must be thoroughly explored. This suggests a few key properties of the
stimulus:
• The current waveform of Iinjected(t) must have sufficient amplitudes (±) and must be applied
sufficiently long in time that it explores the full range of the neuron variation.
• The frequency content of the stimulus current must be a low enough that it does not exceed the low-pass
cutoff frequency associated with the RC time constant of the neuron. This cutoff is typically in the
neighborhood of 50-100Hz.
• The current must explore all time scales expressed in the neuron's behavior.
Frontiers
9
Miller et al.
3.2 Analysis of HVCX Data
The model for an HVCX neuron is substantially taken from [25] and described in our Appendix. We now
use this model in a 'twin experiment' in which PA is utilized, and then using 'nudging' we present the
analysis of experimental data on two Zebra Finch.
3.2.1 Twin Experiment on HVCX Neuron Model
A twin experiment is a synthetic numerical experiment meant to mirror the conditions of a laboratory
experiment. We use our mathematical model with some informed parameter choices in order to generate
numerical data. Noise is added to observable variables in the model, here V (t). These data are then put
through our SDA procedure to estimate parameters and unobserved states of the model. The neuron model
is now calibrated or completed.
Using the parameters and the full state of the model at the end tF of an observation window [t0, tF ], we
take a current waveform Iinjected(t ≥ tF ) to drive the model neuron and predict the time course of all
dynamical variables in the prediction window [tF , ...]. This validation of the model is the critical test of our
SDA procedure, here PA. In a laboratory experiment we have no specific knowledge of the parameters
in the model and, by definition, cannot observe the unobserved state variables; here we can do that. So,
'fitting' the observed data within the observation window [t0, tF ] is not enough, we must reproduce all
states for t ≥ tF to test our SDA methods.
We use the model laid out in the Appendix. We assume that the neuron has a resting potential of −70
mV and set the initial values for the voltage and each gating variable accordingly. We assume that the
internal calcium concentration of the cell is Cin = 0.1 µM. We use an integration time step of 0.02 ms and
Figure 2. A drawing of the Song Production Pathway and the Anterior Forebrain Pathway of avian
songbirds. Parts of the auditory pathways are shown in grey. Pathways from the brainstem that ultimately
return to HVC are not shown.
The HVC network image is taken from [35]. Reprinted by permission from Copyright Clearance Center:
Springer Nature, Nature Reviews Neuroscience, Auditory Feedback in Learning and Maintenance of Vocal
Behavior, M. S. Brainard and A. J. Doupe, 2000.
This is a provisional file, not the final typeset article
10
Miller et al.
integrate forward in time using an adaptive Runge-Kutta method [7]. Noise is added to the voltage time
course by sampling from a Gaussian distribution N (0, 2) in units of mV.
Figure 3. Stimulating current Iinjected(t) presented to the HVCX model.
The waveform of the injected current was chosen to have three key attributes: (1) It is strong enough
to cause spiking in the neuron, (2) it dwells a long time in a hyperpolarizing region, and (3) its overall
frequency content is low enough to not be filtered out by the neuron's RC time constant. On this last point,
a neuron behaves like an RC circuit, it has a cut off frequency limited by the time constant of the system.
Any input current which has a frequency higher than that of the cut off frequency won't be 'seen' by the
neuron. The time constant is taken to be the time it takes to spike and return back to 37% above its resting
voltage. We chose a current where the majority of the power spectral density exists below 50 Hz.
The first two seconds of our chosen current waveform is a varying hyperpolarizing current. In order to
characterize Ih(t) and ICaT (t), it is necessary to thoroughly explore the region where the current is active.
Ih(t) is only activated when the neuron is hyperpolarized. The activation of Ih(t) deinactivates ICaT (t),
thereby allowing its dynamics to be explored. In order to characterize IN a(t) and IK(t), it is necessary to
cause spiking in the neuron. The depolarizing current must be strong enough to hit the threshold potential
for spike activation.
The parameters used to generate the data used in the twin experiment are in Table (1), and the injection
current data and the membrane voltage response may be seen in figures Fig. (3) and Fig. (4).
The numbers chosen for the data assimilation procedure in this paper are α = 1.4 and β ranging from
1 to 100. Rf,0,V = 10−4 for voltage and Rf,0,j = 1 for all gating variables. These numbers are chosen
Frontiers
11
Miller et al.
Figure 4. Response of the HVCX model membrane voltage to the selected Iinjected(t). The displayed
time course V (t) has no added noise.
because the voltage range is 100 times large than the gating variable range. Choosing a single Rf,0 value
would result in the gating variable equations being less enforced than the voltage equation by a factor of
104. The α and β numbers are chosen because we seek to make Rf
sufficiently large. The α and β values
Rf 0
chosen allow Rf
Rf 0
to reach 1015.
During estimation we instructed our methods to estimate the inverse capacitance and estimate the ratio
g(cid:48) = g
instead of g and Cm independently. That separation can present a challenge to numerical
procedures. We also estimated the reversal potentials as a check on the SDA method.
Cm
Within our computational capability we can reasonably perform estimates on 50,000 data points. This
captures a second of data when ∆t = 0.02 ms. However, there are time constants in the model neuron
which are on order 1 second. In order for us to estimate the behavior of these parameters accurately, we
need to see multiple instances of the full response. We need a window on the order of 2-3 seconds. We
can obtain this by downsampling the data. We know from previous results that downsampling can lead to
better estimations [41]. We take every ith data point, depending on the level of down sampling we want to
do. In this data assimilation run, we downsampled by a factor of 4 to incorporate 4 seconds of data in the
estimation window.
We look at a plot of the action as a function of β; that is, log[Rf /Rf 0]. We expect to see a leveling of of
the action [17] as a function of Rf . If the action becomes independent of Rf , we can then explore how
well our parameter estimations perform when integrating them forward as predictions of the calibrated
This is a provisional file, not the final typeset article
12
Miller et al.
Figure 5. Action Levels of the standard model action for the HVCX neuron model discussed in the text.
We see that the action rises to a 'plateau' where it becomes quite independent of Rf . The calculation of the
action uses PA with α = 1.4 and Rf 0 = Rm. NI = 100 initial choices for the path Xinit were used in this
calculation. For small Rf one can see the slight differences in action level associated with local minima of
A(X).
model. Looking at the action plot in Fig. (5), we can see there is a region in which the action appears to
level off, around β = 40. It is in this region where we look for our parameter estimates.
We examine all solutions around this region of β and utilize their parameter estimates in our predictions.
We compare our numerical prediction to the "real" data from our synthetic experiment. We evaluate good
predictions by finding the correlation coefficient between these two curves. This metric is chosen instead
of a simple root mean square error because slight variations in spike timings yield a high amount of error
even if the general spiking pattern is correct. The prediction plot and parameters for the best prediction can
be seen in Fig. (6) and Table (2). The voltage trace in red is the estimated voltage after data assimilation is
completed. It is overlayed on the synthetic input data in black. The blue time course is a prediction, starting
at the last time point of the red estimated V (t) trace and using the parameter estimates for t ≤ 4000 ms.
The red curve matches the computed voltage trace quite well. There is no deviation in the frequency
of spikes, spike amplitudes, or the hyperpolarized region of the cell. Looking at the prediction window,
we can see that there is some deviation in the hyperpolarized voltage trace around 9000 ms. This is an
indication that our parameter estimates for currents activated in this region are not correct. Comparing
parameters, we can see that Eh is estimated as lower than its actual value. Despite that, we still are able to
reproduce neuron behavior fairly well.
Frontiers
13
Miller et al.
Figure 6. Results of the 'twin experiment' using the model HVCX neuron described in the Appendix.
Noise was added to data developed by solving the dynamical equations. The noisy V (t) was presented
to the precision annealing SDA calculation along with the Iinjected(t) in the observation window t0 =
0ms, tF = 4000 ms. The noisy model voltage data is shown in black, and the estimated voltage is shown in
red. For t ≥ 4000ms we show the predicted membrane voltage, in blue, generated by solving the HVCX
model equations using the parameters estimated during SDA within the observation window.
3.2.2 Analysis of Biophysical Parameters from HVCX Neurons in two Zebra Finch
Our next use of SDA employs the 'nudging' method described in Eq. (8). In this section we used some of
the data [40] taken in experiments on multiple HVCX neurons from different zebra finches. The questions
we asked was whether we could, using SDA, identify differences in biophysical characteristics of the birds.
This question is motivated by prior biological observations [40].
Using the same HVCX model as before, we estimated the maximal conductances {gN a, gK, gCaT , gSK, gh}
holding fixed other kinetic parameters and the Nernst/Reversal potentials. The baseline characteristics
of an ion channel are set by the properties of the cell membrane and the complex proteins penetrating
the membrane forming the physical channel. Differences among birds would then come from the density
or numbers of various channels as characterized by the maximal conductances. If such differences were
identified, it would promote further investigation of the biologically exciting proposition that these
differences arise in relation to some aspect of the song learning experience of the birds [40].
In Fig. (7) and Fig. (8) we display the stimulating current and membrane voltage response from one of 9
neurons in our large sample. The analysis using SDA was of four neurons from one bird and seven neurons
from another. The results for {gCaT , gN a, gSK} is displayed in Fig. (9). The maximal conductances from
This is a provisional file, not the final typeset article
14
Miller et al.
one bird are shown in blue and from the other bird, in red. There is a striking difference between the
distributions of maximal conductances.
We do not propose here to delve into the biological implications of these results. Nevertheless, we note
that the neurons from each bird occupy a small but distinct region of the parameter space (Fig. 9). This
result and its implications for birdsong learning, and more broadly for neuroscience, are described in [40].
Here, however, we display this result as an example of the power of SDA to address a biologically important
question in a systematic, principled manner beyond what is normally achieved in analyses of such data.
To fully embrace the utility of SDA for these experiments, however, further work is needed. A limitation
of the present result is that the SDA estimates for gSK for a subset of the neurons/observations for Bird
One reach the bounds of the observation window (Fig. 9). Addressing such issues would be prelude to
the exciting possibility of estimating more parameters than just the principle ion currents in the Hodgkin-
Huxley equations. This could use SDA numerical techniques to calculate, over hours or days, estimates
of parameters that could require months or years of work to measure with traditional biological and
biophysical approaches, in some cases requiring specialized equipment beyond that available for most in
vitro recording set ups. In contrast, applying SDA to such data sets requires only a computer.
Figure 7. One of the library of stimuli used in exciting voltage response activity in an HVCX neuron. The
cell was prepared in vitro, and a single patch clamp electrode injected Iinjected(t) (this waveform) and
recorded the membrane potential.
Frontiers
15
05001000150020002500300035004000-150-100-50050100150200Iapplied(t) (pA)Time (Units of 0.02 ms) Stimulating Current #7 Miller et al.
Figure 8. The voltage response from Iapplied, Fig. (7). One of the library of stimuli used in exciting
voltage response activity in an HVCX neuron. The cell was prepared in vitro, and a single patch clamp
electrode injected Iinjected(t) (this waveform) and recorded the membrane potential.
3.3 Analysis of Neuromorphic VLSI Instantiations of Neurons
An ambitious effort in neuroscience is the creation of low power consumption analog neural-emulating
VLSI circuitry. The goals for this effort range from the challenge itself to the development of fast,
reconfigurable circuitry on which to incorporate information revealed in biological experiments for use
in
• creating model neural circuits of 'healthy' performance to be compared to subsequent observations
on the same circuitry informed by 'unhealthy' performances. If the comparison can be made rapidly
and accurately, the actual instantiations in the VLSI circuitry could be used to diagnose the changes in
neuron properties and circuit connectivity perhaps leading to directions for cures, and
• in creating VLSI realizations of neural circuits with desired functions -- say, learning syntax in interesting
sequences -- might allow those functions to be performed at many times increased speed than seen in
the biological manifestation. If those functionalities are of engineering utility, the speed up could be
critical in applications.
One of the curious roadblocks in achieving critical steps of these goals is that after the circuitry is
designed and manufactured into VLSI chips, what comes back from a fabrication plant is not precisely
what we designed. This is due to the realities of the manufacturing processes and not inadequacies of the
designers.
This is a provisional file, not the final typeset article
16
0500100015002000250030003500-100-80-60-40-2002040HVCX Voltage Response Stimulus 7V(t) (mV)Time (Units of 0.02 ms) Miller et al.
Figure 9. A three dimensional plot of three of the maximal conductances estimated from HVCX cells
using the stimulating current shown in Fig. (7). Membrane voltage responses from five neurons from one
bird were recorded many times, and membrane voltage responses from four neurons from a second bird
were recorded many times. One set of maximal conductances {gN a, gCaT , gSK} are shown. The estimates
from Bird 1 are in red-like colors, and the estimates from Bird 2 are in blue-like colors. This is just one out
of a large number of examples discussed in detail in [40].
To overcome this barrier in using the VLSI devices in networks, we need an algorithmic tool to determine
just what did return from the factory, so we know how the nodes of a silicon network are constituted. As
each chip is an electronic device built on a model design, and the flaws in manufacuring are imperfections
in the realization of design parameters, we can use data from the actual chip and SDA to estimate the actual
parameters on the chip.
SDA has an advantageous position here. If we present to the chip input signals with much the same
design as we prepared for the neruobiological experimets discussed in the previous section, we can measure
everything about each output from the chip and use SDA to estimate the actual parameters produced in
manufacturing. Of course, we do not know those paramters a priori so after estimating the parameters, thus
'calibrating' the chip, we must use those estimated parameters to predict the response of the chip to a new
stimulating currents. That will validate (or not) the completion of the model of the actual circuitry on the
chip and permit confidence in using it in building interesting networks.
We have done this on chips produced in the laboratory of Gert Cauwenberghs at UCSD using PA [42, 41]
and using 'nudging' as we now report.
The chip we worked with was designed to produce the simplest spiking neuron, namely one having just
Na, K, and leak channels [8, 9] as in the original HH experiments. This neuron has four state variables
Frontiers
17
0.00.20.40.60.81.01.21.41.61.82.03004005006007002468101214gNagSKgCaT Bird One Bird Two Miller et al.
{V (t), m(t), h(t), n(t)}:
dt
= gN am3(t)h(t)(EN a − V (t)) + gKn4(t)(EK − V (t))
dV (t)
C
+gL(EL − V (t)) + Iinjected(t)
in which the gating variables w(t) = {m(t), h(t), n(t)} satisfy
dw(t)
(w∞(V (t)) − w(t))
=
dt
τw(V (t))
,
(12)
and the functions w∞(V ) are discussed in depth in [8, 9].
In our experiments on a 'NaKL' chip we used the stimulating current displayed in Fig. (10),
Figure 10. This waveform for Iinjected(t) was used to drive the VLSI NaKL neuron after receipt from the
fabrication facility.
and measured all the neural responses {V (t), m(t), h(t), n(t)}. These observations were presented to the
designed model within SDA to estimate the parameters in the model.
We then tested/validated the estimations by using the calibrated model to predict how the VLSI chip
would respond to a different injected current. In Fig. (11) we show the observed Vdata(t) in black, the
estimation of the voltage through SDA in red, and the prediction of V (t) in blue for times after the end of
the observation window.
This is a provisional file, not the final typeset article
18
Miller et al.
While one can be pleased with the outcome of these procedures, for our purposes we see that the use of
our SDA algorithms gives the user substantial confidence in the functioning characteristics of the VLSI
chips one wishes to use at the nodes of a large, perhaps even very large, realization of a desired neural
circuit in VLSI. We are not unaware of the software developments to allow efficient calibration of very
large numbers of manufactured silicon neurons. A possible worry about also determining the connectivity,
both the links and their strength and time constants, may be alleviated by realizing these links through a
high bandwidth bi-directional connection of the massive array of chips and the designation of connectivity
characteristics on an off-chip computer.
Figure 11. The NaKL VLSI neuron was driven by the waveform for Iinjected(t) seen in Fig. (10). The four
state variables {V (t), m(t), h(t), n(t)} for the NaKL model were recorded and used in an SDA 'nudging'
protocol to estimate the parameters of the model actually realized at the manufacturing facility. Here we
display the membrane voltages: {Vdata(t), Vest(t), Vpred(t)} -- the observed membrane voltage response
when Iinjected(t) was used, the estimated voltage response using SDA, and finally the predicted voltage
response Vpred(t) from the calibrated model actually on the VLSI chip. In a laboratory experiment, only
this attribute of a neuron would be observable.
Part of the same analysis is the ability to observe, estimate and predict the experimentally unobservable
gating variables. This serves, in this context, as a check on the SDA calculations. The Na inactivation
variable h(t) is shown in Fig. (12) as its measured time course hdata(t) in black, its estimated time course
hest(t) in red, and its predicted time course hpred(t) in blue.
Frontiers
19
Miller et al.
Figure 12. The NaKL VLSI neuron was driven by the waveform for Iinjected(t) seen in Fig. (10. The four
state variables {V (t), m(t), h(t), n(t)} for the NaKL model were recorded and used in an SDA 'nudging'
protocol to estimate the parameters of the model actually realized at the manufacturing facility. Here we
display the Na inactivation variable h(t): {hdata(t), hest(t), hpred(t)} -- the observed h(t) time course when
Iinjected(t) was used, the estimated h(t) time course using SDA, and finally the predicted h(t) time course
from the calibrated model actually on the VLSI chip. In a laboratory experiment, this attribute of a neuron
would be unobservable. Note we have rescaled the gating variable from its natural range 0 ≤ h(t) ≤ 1) to
the range within the VLSI chip. The message of this Figure is in the very good accuracy and prediction of
an experimentally unobservable time course.
4 DISCUSSION
Our review of the general formulation of statistical data assimilation (SDA) started our remarks. Many
details can be found in [3] and subsequent papers by the authors. Recognizing that the core problem is
to perform, approximately of course, the integral in Eq. (5) is the essential take away message. This task
requires well 'curated' data and a model of the processes producing the data. In the context of experiments
in life sciences or, say, aquisition of data from earth system sensors, curation includes an assessment of
errors and the properties of the instruments as well.
One we have the data and a model, we still need a set of procedures to transfer the information from
the data to the model, then test/validate the model on data not used to train the model. The techniques we
covered are general. Their application to examples from neuroscience comprise the second part of this
paper.
In the second part we first address properties of the avian songbird song production pathway and a neural
control pathway modulating the learning and production of functional vocalization -- song. We focus our
This is a provisional file, not the final typeset article
20
Miller et al.
attention on one class of neurons, HVCX, but have also demonstrated the utility of SDA to describe the
response properties of other classes of neurons in HVC, such as HVCRA [43] and HVCI [44]. Indeed, the
SDA approach is generally applicable wherever there is insight to relate biophysical properties of neurons
to their dynamics through Hodgkin Huxely equations.
Our SDA methods considered variational algorithms that seek the highest conditional probability
distributions of the model states and parameters conditioned on the collection of observations over a
measurement window in time. Other approaches, especially Monte Carlo algorithms were not discussed
here, but are equally attractive.
We discussed methods of testing models of HVCX neurons using "twin experiments" in which a model
of the individual neuron produces synthetic data to which we add noise with a selected signal-to-noise-ratio.
Some state variable time courses from the library of these model produced data, for us the voltage across
the cell membrane, is then part of the action Eq. (7), specifically in the measurement error term. Errors in
the model are represented in the model error term of the action.
Using a precision annealing protocol to identify and track the global minimum of the action, the successful
twin experiment gives us confidence in this SDA method from information trans from data to the model.
We then introduced a 'nudging' method as an approximation to the Euler-Lagrange equations derived
from the numerical optimization of the action Eq. (7) -- this is Laplace's method in our SDA context. The
nudging method, introduced in meteorology some time ago, was used to distinguish between two different
members of the Zebra Finch collection. We showed, in a quite preliminary manner, that the two, unrelated
birds of the same species, express different HVC network properties as seen in a critical set of maximal
conductances for the ion channels in their dynamics.
Finally we turned to a consideration of the challenge of implementing in VLSI technology the neurons in
HVC towards the goal of building a silicon-HVC network. The challenge at the design and fabrication
stages of this effort where illuminated by our use of SDA to determine what was actually returned from the
manufacturing process for our analog neurons.
4.1 Moving Forward to Network Analysis
Finally, we have a few comments associated with the next stage of analysis of HVC. In this, and previous
papers, we analyzed individual neurons in HVC. These analyses were assisted by our using SDA, through
twin experiments, to design laboratory experiments though the selection of effective stimulilaing injected
currents.
Having characterized the electrophysiology of an individual neuron within the framework of Hodgkin-
Huxley (HH) models, we may now proceed beyond the study of individual neurons [45] in vitro. Once we
have characterized an HVC neuron through a biophysical HH model, we may then use it in vivo as a sensor
of the activity of the HVC network where it is connected to HVCRA, HVCI, and other HVCX neurons.
The schema for this kind of experiment is displayed in Fig. (13). These experiments require the capability
to perform measurements on HVC neurons in the living bird. That capability is available, and experiments
as suggested in our graphic are feasible, if challenging.
The schematic indicates that the stimulating input to the experiments is auditory signals, chosen by the
user, presented to the bird's ear and reaching HVC through the auditory pathway. The stimuli from this
signal is then distributed in a manner to be deduced from experiment and then produces activity in the
HVC network that we must model. The goal is, at least initially, to establish, again within the models we
Frontiers
21
Miller et al.
develop, the connectivity of HVC neuron classes as it manifests itself in the function of the network. We
have some information about this [46, 47], and these results will guide the development of the HVC model
used in these whole-network experiments. An important point to address is what changes to the in vitro
model might be necessary to render it a model for in vivo activity.
Figure 13. A cartoon-like idea of an experiment to probe the HVC network. In this graphic three neuron
populations of {HV CX , HV CI , HV CRA} neurons are stimulated by auditory signals P (t) presented to
the bird in vivo. This drives the auditory pathway from the ear to HVC, and the network activity is recorded
from a calibrated, living HVCX neuron, used here as a sensor for network activity. While the experiment is
now possible, the construction of a model HVC network will proceed in steps of complexity using simple
then more biophysiclly realistic neuron models and connections among the nodes (neurons) of the network.
From libraries of time courses of P (t), chosen by the user, and responses of V (t) in the 'sensor' HVCX
neuron, we will use SDA to estimate properties of the network.
This is a provisional file, not the final typeset article
22
Miller et al.
APPENDIX
The equations describing the HVCX neuron dynamics are taken from the work of Daou [25]. That paper
also has an extensive account of his experiments on the other two major classes of neurons in HVC. The
equations are of Hodgkin-Huxley (HH) form for a neuron without spatial extent; this is called a one-
compartment model. It is meant to apply to neurons in isolation of the network, here HVC, in which they sit
in vivo. The dynamical variables include the observable quantities: voltage across the cell membrane, V (t)
and the intracellular concentration of [Ca2+]in(t) = C(t). V (t) is directly connected to action potentials
or voltage spikes that communicate among cells in a network; the time scale of these spikes is a few ms.
C(t) provides a slow background modulation that raises the cells potential (depolarizes the cell) or lowers
it (hyperpolarizes the cell) on time scales as long as 10's of ms.
The voltage equation, conservation of charge, relates the capacitance of the cell membrane Cm as it
separates concentrations of ions within and without the cells to the various currents which contain the
nonlinear voltage dependence of the permeability of ions to passing into and out of the cell. The model
represents these ion currents: {N a, K, Ca} in several different ways.
The general form of an HH current is
(t)(Erev−ion − V (t)),
ion
ion
(t)hinteger2
Iion(t) = gionminteger1
(13)
where the reversal potential is the equilibrium Nernst potential [8, 9]. The gating variables {m(t), h(t)} lie
between zero and one and represent the amount the ion channel is open relative to the maximum opening it
may have. The maximal conductance gion represent the number or density of ion channels in the neuron
model. This form of ion current applies when the concentrations of the ion are not significantly different
outside and inside the cell. This is not so for Ca2+ ions, so there we use the Goldman-Hodgkin-Katz
(GHK) form of the current [48].
(cid:32)
(cid:33)
Ia(t) = −Paz2
aF 2
[ion]in(t) − [ion]oute−zaF V (t)/RT
1 − e−zaF V (t)/RT
,
(14)
for iona. za is the charge on the ion, F the Faraday constant, R the gas constant, T the temperature, and Pa
the permeability of the cell membrane to ion a.
We use an approximation to the GHK equations for the two types of Calcium currents selected in [25]
(cid:18)
(cid:19)
[Ca]out
e2F V (t)/RT − 1
(cid:18)
ICaL(t) = gCaLV s2∞(V (t))
ICaT (t) = gCaT V (t)[aT ]3∞(V )[bT ]3∞(rA
T )
(cid:17) −
bT∞(rT ) =
(cid:16) rT −θb
1
[Ca]out
e2F V (t)/RT − 1
(cid:16)−θb
1
(cid:17)
1 + e
σb
1 + e
σb
(cid:19)
(15)
aT and bT are instantaneous activating and inactivating gating variables, respectively. rT is a slow gating
variable which takes the same functional form as aT and other gating variables m(t) and h(t). These gating
Frontiers
23
Miller et al.
variables w(t) satisfy a first order kinetic equation
in which
w∞(V ) =
1
2
1 − tanh
dw(t)
dt
=
w∞(V (t)) − w(t))
τw(V (t))
(cid:20)
,
(cid:18) V − θw
(cid:19)(cid:21)
(16)
(17)
,
2σw
for all gating variables except h∞(V ) appearing in IN a(t) [25]. θw is the half-activation voltage and σw
controls the slope of the activation function. For fast gating variables, such as m of IN a, and s of ICaL we
replace the time dependence by W∞(V ).
τw(V ) is the time constant of each gating variable. Time constants for the n and hp gating variables (these
names refer to [25]) are given below, where ¯τw is an average time constant. Our model differs from [25] by
one time constant. Instead, τrs(V ) takes the form presented here:
τw(V ) =
cosh
for n or hp
(cid:16) V −θw
¯τw
2σw
(cid:17)
(cid:18)
(cid:19)(cid:19)
(cid:18) V (t) + 80
−21
1 − tanh2
τrs(V ) = 0.1 + 193.0
τrf =
prf
−7.4(V +70)
V +70−0.8 −1
e
+ 65e
V +56−23
τrT (V ) = τr0 +
1 + e
(cid:17)
τr1
(cid:16) V −θrT
σrT
For our choice of ion currents we follow the results of experimental data [49, 50, 51] and generally
reproduce the model listed in [25]. HVCX spiking properties include fast rectifying current, sag in response
to hyperpolarizing current, and spike frequency adaption in response to depolarizing current.
C
dV (t)
dt
= IN a(t) + IK(t) + IL(t) + ICaT (t) + ICaL(t)
+IA(t) + ISK(t) + Ih(t) + IN ap(t) + Iinjected(t)
(18)
IN a(t) and IK(t) are the standard HH currents. They produce fast spiking in response to injected currents.
IL(t) is a leak current meant to capture all linear currents of the neuron. ICaT (t) is a low threshold T-type
calcium current that causes rebound depolarization in cooperation with Ih(t). ICaL(t) is a high threshold
L-type calcium current. ICaL(t) works in conjunction with ISK(t), a calcium concentration dependent
potassium current, to create frequency adaptation in neuron spiking. IA(t) is an A-type potassium current.
IN ap(t) is a persistent sodium current. From the model presented in [25], we eliminate IKN a(t), a sodium
dependent potassium current, and rewrite all sigmoidal functions as hyperbolic tangents.
The mass conservation equation for Ca2+ is written as
dC(t)
dt
= (ICaT (t) + ICaL(t)) + kCa(bCa − C(t)),
again following [25].
This is a provisional file, not the final typeset article
(19)
24
Miller et al.
TABLES
Parameter
gN a
EN a
gK
EK
gCaT
gSK
gH
EH
Cm
θa
σa
θrs
σrs
θrT
σrT
τr0
kCa
Parameter
Value
450 nS
gL
50 mV
EL
50 nS
gN ap
-90 mV gCaL
2.65 nS
θm
6 nS
σm
4 nS
θn
-30 mV σn
100 pF
¯τn
-20 mV θrf
-10 mV σrf
-105 mV θaR
25 mV
σaT
-67 mV θrrT
2 mV
σrrT
200 ms
τr1
0.3 ms−1
bCa
Value
Parameter
2 nS
kf
-70 mV θmp
1 nS
σmp
19 nS
θs
-35 mV σs
-5 mV
θhp
-30 mV σhp
-5 mV
¯τhp
10 ms
θe
-105 mV σe
5 mV
τe
-65 mV θb
-7.8 mV σb
68 mV f
2.2 mV
87.5 ms
prf
0.1 µM ks
Value
0.3
-40 mV
-6 mV
-20 mV
-0.05 mV
-48 mV
6 mV
1000 ms
-60 mV
5 mV
20 ms
0.4 mV
-0.1 mV
0.1
0.0015 µM
pA·ms
100
0.5 µM
Table 1. Parameter values used to numerically generate the HVCX data. The source of these values comes
from [25]. Data was generated using an adaptive Runge-Kutta method, and can be seen in Fig. (3) and Fig.
(4).
Parameter
g(cid:48)
N a
EN a
g(cid:48)
K
EK
g(cid:48)
g(cid:48)
CaT
g(cid:48)
h
Eh
Cinv
SK
Bounds
0.1, 10
1, 100
0.01, 5
-140, -10
0.001, 1
0.001, 1
0.001, 1
-100, -1
0.001, 0.5
4.5
50
0.5
-90
Best Estimate Actual Value Units
nS/pF
mV
nS/pF
mV
nS/pF
nS/pF
nS/pF
mV
pF−1
4.98
43.2
0.907
-127.4
0.0326
0.0373
0.0432
-44.1
0.011
0.0265
0.06
0.04
-30
0.01
Table 2. Parameter Estimates from the Best Predictions. The best prediction is chosen by finding the
highest correlation coefficient between the predicted voltage and "real" voltage. This comparison can
be made on experimental data. It represents an attractive alternative to the familiar least squares metric
commonly used. That metric is very sensitive to spike times in data with action potentials: small errors in
spike times may result in large errors in a least squares metric.
Frontiers
25
Miller et al.
REFERENCES
[1] Lorenz EN. Predictability: A problem partly solved. Palmer T, Hagedorn R, editors, Predictability of
weather and climate (Cambridge) (2006).
[2] Evensen G. Data Assimilation: The Ensemble Kalman Filter (Springer) (2009).
[3] Abarbanel HDI. Predicting the Future: Completing Models of Observed Complex Systems (Springer)
(2013).
[4] Tong D. Statistical physics (2011). Available at www.damtp.cam.ac.uk/user/tong/statphys/sp.pdf.
[5] Laplace PS. Memoir on the probability of causes of events. Math´ematique et de Physique,Tome
Sixi´eme (1774) 621 -- 656.
[6] Laplace P. Memoir of the probability of causes of events. Statistical Science 1 (1986) 365 -- 378.
Translation to English by S. M. Stigler.
[7] Press WH, Teukolsky SA, Vetterling WT, Flannery BP. Numerical Recipes: The Art of Scientific
Computing, Third Edition (Cambridge University Press) (2007).
[8] Johnston D, Wu SMS. Foundations of Cellular Neurophysiology (Bradford Books, MIT Press) (1995).
[9] Sterratt D, Graham B, Gillies A, Willshaw D. Principles of Computational Modelling in Neuroscience
(Cambridge University Press) (2011), 390 .
[10] Abarbanel HDI, Rozdeba PJ, Shirman S. Machine learning: Deepest learning as statistical data
assimilation problems. Neural Computation 30 (2018) 2025 -- 2055.
[11] Fano RM. Transmission of Information; A Statistical Theory of Communication (MIT Press) (1961).
[12] Kostuk M, Toth BA, Meliza CD, Margoliash D, Abarbanel HDI. Dynamical estimation of neuron and
network properties ii: path integral monte carlo methods. Biological Cybernetics 106 (2012) 155 -- 167.
[13] Neal R. Mcmc using hamiltonian dynamics. Gelman A, Jones G, Meng XL, editors, Handbook of
Markov Chain Monte Carlo (Chapman and Hall; CRC Press), chap. 5 (2011).
[14] Murty KG, Kabadi SN. Some np-complete problems in quadratic and nonlinear programming.
Mathematical Programming 39 (1987) 117 -- 129.
[15] Ye J, Kadakia N, Rozdeba PJ, Abarbanel HDI, Quinn JC.
Improved variational methods in
statistical data assimilation. Nonlinear Processes in Geophysics 22 (2015) 205 -- 213. doi:10.5194/
npg-22-205-2015.
[16] Ye J, Rey D, Kadakia N, Eldridge M, Morone UI, Rozdeba P, et al. Systematic variational method
for statistical nonlinear state and parameter estimation. Phys. Rev. E 92 (2015) 052901. doi:10.1103/
PhysevE.92.052901.
[17] Ye J. Systematic Annealing Approach for Statistical Data Assimilation. Ph.D. thesis, University of
California San Diego (2016).
[18] Quinn JC. A path integral approach to data assimilation in stochastic nonlinear systems. Ph.D. thesis,
University of California San Diego (2010).
[19] Shirman S. Strategic Monte Carlo and Variational Methods in Statistical Data Assimilation for
Nonlinear Dynamical Systems. Ph.D. thesis, University of California San Diego (2018).
[20] Anthes R. Data assimilation and initialization of hurricane prediction models. J. Atmos. Sci. 31
(1974).
[21] Gelfand IM, Fomin SV. Calculus of Variations (Dover Publications, Inc.) (1963).
[22] Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E. Equation of State Calculations
by Fast Computing Machines. J. Chem. Phys. 21 (1953) 1087 -- 1092. doi:10.1063/1.1699114.
[23] Wong AS, Hao K, Abarbanel HDI. Precision annealing for monte carlo methods: Application to the
shallow water equations (2018). In Preparation, 2018.
This is a provisional file, not the final typeset article
26
Miller et al.
[24] Doupe AJ, Kuhl PK. Birdsong and human speech: Common themes and mechanisms. Annual Review
of Neuroscience 22 (1999) 567 -- 631. doi:10.1146/annurev.neuro.22.1.567. PMID: 10202549.
[25] Daou A, Ross M, Johnson F, Hyson R, Bertram R. Electrophysiological characterization and
computational models of hvc neurons in the zebra finch. Journal of Neurophysiology 110 (2013)
1227 -- 1245.
[26] Bolhuis JJ, Okanoya K, Scarff C. Twitter evolution: Converging mechanisms in birdsong and human
speech. Nature Reviews Neuroscience 11 (2010) 747 -- 759.
[27] Fee MS, Scharff C. The songbird as a model for the generation and learning of complex sequential
behaviors. ILAR Journal 51 (2010) 362 -- 377. doi:10.1093/ilar.51.4.362.
[28] Mooney R. Neuronal mechanisms for learned birdsong. Learning and Memory 16 (2009) 655 -- 669.
doi:10.1101/lm.1065209.
[29] Simonyan K, Horwitz B, Jarvis E. Dopamine regulation of human speech and bird song: A critical
review. Brain and Language 122 (2012) 142 -- 150. doi:10.1016/j.bandl.2011.12.009.
[30] Teramitsu I, Kudo LC, London SE, Geschwind DH, White SA. Parallel foxp1 and foxp2 expression
in songbird and human brain predicts functional interaction. Journal of Neuroscience 24 (2004)
3152 -- 3163. doi:10.1523/JNEUROSCI.5589-03.2004.
[31] Jarvis ED, Gunturkun O, Bruce L, Csillag A, Karten H, Kuenzel W, et al. Avian brains and a
new understanding of vertebrate brain evolution. Nature Review Neuroscience 6 (2005) 151 -- 159.
doi:10.1038/nrnl1606.
[32] Doupe AJ, Perkel DJ, Reiner A, Stern EA. Birdbrains could teach basal ganglia research a new song.
Trends in Neuroscience 28 (2005) 353 -- 363. doi:10.1016/j.tins.2005.05.005.
[33] Nottebohm F, Stokes TM, Leonard CM. Central control of song in the canary, serinus canarius.
Journal of Comparative Neurology 165 (1976) 457 -- 486. doi:10.1002/cne.901650405.
[34] Fortune ES, Margoliash D. Parallel pathways and convergence onto hvc and adjacent neostriatum of
adult zebra finches (taeniopygia guttata). Journal of Comparative Neurology 360 (1995) 413 -- 441.
doi:10.1002/cne.903600305.
[35] Brainard MS, Doupe AJ. Auditory feedback in learning and maintenance of vocal behaviour. Nature
Reviews Neuroscience 1 (2000).
[36] Hamill OP, Marty A, Neher E, Sakmann B, Sigworth FJ. Improved patch-clamp techniques for
high-resolution current recording from cells and cell-free membrane patches. Pflugers Archiv 391
(1981) 85 -- 100. doi:10.1007/BF00656997.
[37] Hern´andez-Ochoa EO, Schneider MF. Voltage clamp methods for the study of membrane currents
and sr ca2+ release in adult skeletal muscle fibres. Progress in Biophysics and Molecular biology 108
(2012) 98 -- 118.
[38] Bagal SK, Marron BE, Owen RM, Storer RI, Swain NA. Voltaged gated sodium channels as drug
discovery targets. Channels (Austin) 9 (2015) 360 -- 366. doi:10.1080/19336950.2015.1079674.
[39] Frolov RV, Weckstrom M. Ion channels as therapeutic targets, part a. Advances in Protein Chemistry
and Structural Biology 103 (2016) 1 -- 386.
[40] Daou A, Margoliash D. Things to say about birds; brothers and not. Waiting for Godot 11901 (2018).
[41] Breen D. Characterizing Real World Neural Systems Using Variational Methods of Data Assimilation.
Ph.D. thesis, University of California San Diego (2017).
[42] Wang J, Breen D, Akinin A, Broccard F, Abarbanel HDI, Cauwenberghs G. Assimilation of biophysical
neuronal dynamics in neuromorphic vlsi. Biomedical Circuits and Systems 11 (2017) 1258 -- 1270.
[43] Kadakia N, Armstrong E, Breen D, Daou A, Margoliash D, Abarbanel H. Nonlinear statistical data
assimilation for hvcra neurons in the avian song system. Biological Cybernetics 110 (2016) 417 -- 434.
Frontiers
27
Miller et al.
[44] Breen D, Shirman S, Armstrong E, Kadakia N, Abarbanel H. Hvci neuron properties from statistical
data assimilation (2016).
[45] Nogaret A, Meliza CD, Margoiliash D, Abarbanel. Automatic construction of predictive neuron
models through large scale assimilation of electrophysiological data. Scientific Reports 6 (2016).
doi:10.1038/srep32749.
[46] Mooney R, Prather JF. The hvc microcircuit: The synaptic basis for interactions between song
motor and vocal plasticity pathways. Journal of Neuroscience 25 (2005) 1952 -- 1964. doi:10.1523/
JNEUROSCI.3726-04.2005.
[47] Kosche G, Vallentin D, Long MA. Interplay of inhibition and excitation shapes a premotor neural
sequence. J Neurosci. 35 (2015) 1217 -- 1227.
[48] Goldman DE. Potential, impedance, and rectification in membranes. The Journal of General
Physiology 27 (1943) 37 -- 60. doi:10.1085/jgp.27.1.37.
[49] Dutar P, Vu HM, Perkel DJ. Multiple cell types distinguished by physiological, pharmacological, and
anatomic properties in nucleus hvc of the adult zebra finch. Journal of Neurophysiology 80 (1998)
1828 -- 1838. doi:10.1152/jn.1998.80.4.1828. PMID: 9772242.
[50] Mooney R. Different subthreshold mechanisms underlie song selectivity in identified hvc neurons
of the zebra finch. Journal of Neuroscience 20 (2000) 5420 -- 5436. doi:10.1523/JNEUROSCI.
20-14-05420.2000.
[51] Kubota M, Taniguchi I. Electrophysiological characteristics of classes of neuron in the hvc of the
zebra finch. Journal of Neurophysiology 80 (1998) 914 -- 923. doi:10.1152/jn.1998.80.2.914. PMID:
9705478.
This is a provisional file, not the final typeset article
28
|
1611.00033 | 1 | 1611 | 2016-07-26T06:04:02 | Wavelet algorithm for the identification of P300 ERP component | [
"q-bio.NC"
] | Brain-computer interfaces have many algorithms based on the P300 component of ERP. Modern industry has started to produce consumer grade EEG equipment which is handy and not too expensive. This gives us an opportunity to use BCI in everyday practice. In order to improve the performance of these devices, we need effective algorithms for time series analysis and pattern recognition. We have tested Emotiv Insight headset in a real live environment and we have conducted several tests for series of standard wavelets in P300 pattern recognition task. | q-bio.NC | q-bio | Wavelet algorithm for the identification of P300 ERP component
© 2016, S.N. Agapov 1, V.А. Bulanov 2, А.V. Zaharov 3, М.S. Sergeeva 4.
1 [email protected]
2 [email protected]
IT Universe LLC, Department of BCI. Samara, Russia.
3 [email protected]
Samara State Medical University, Department of Neurology and Neurosurgery. Samara, Russia.
4 [email protected]
Samara State Medical University, Department of the Normal Physiology. Samara, Russia.
Abstract: Brain computer interfaces have many algorithms based on the P300 component of ERP.
Modern industry has started to produce consumer grade EEG equipment which is handy and not
too expensive. This gives us an opportunity to use BCI in everyday practice. In order to improve
the performance of these devices we need effective algorithms for time series analysis and pattern
recognition. We have tested Emotiv Insight headset in real live environment and we have conducted
several tests for series of standard wavelets in P300 pattern recognition task.
Copyright: 2016 IT Universe LLC. This is an open-access article distributed under the terms of the
Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original authors and source are credited.
Competing interests: The authors have declared that no competing interests exist.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 1 / 10
1. Introduction
Modern population is becoming older every year. In the 20th century the average life expectancy
rose significantly [1]. This resulted in bringing about several ageing-related blocking diseases such as
amiotrophic lateral sclerosis (ALS), brainstem stroke and others [2, 3]. Heart diseases have become the
main cause of death since the middle of the 20th century [4]. They lead to after-effects which can be the
cause of complete or partial paralysis. The rehabilitation of these people increases the expenses allocated
by a developed country's government on healthcare system [5, 6]. All these factors require fast development
of effective and cheap brain-computer interfaces. If BCI equipment is used together with appropriate
medical and rehab devices, it can considerably improve the abilities of movement-constrained patients.
Nowadays brain-computer interfaces (BCI) can use several different types of input signals from
brain: the P300 component of ERP-wave, imagery movements, slow cortical potentials (SCP), steady state
visually evoked potential (SSVEP) [7]. Modern industry has produced a lot of devices which use some of
these concepts. However, in everyday rehabilitation clinical practice their use is limited. One of the main
causes is the cost of the software and hardware parts of the brain-computer interfaces. Low-cost devices
often have poor SNR (signal-to-noise ratio) characteristics. The modern methods of digital signal
processing and machine learning can greatly improve the accuracy of consumer class equipment. This
improvement should give an opportunity to use EEG devices for medical and rehabilitation facilities.
The implementation of wavelet analysis for the recognition of P300 ERP component in EEG signal
registered with the help of Emotiv Insight (Emotiv Inc., http://emotiv.com/insight/) is shown in this paper.
2. Materials and methods
A group of 5 healthy men aged 29-44 (35.8 ± 7.2) participated in our experiments. We recorded
EEG signals with 5-channel Emotive Insight EEG headset (Fig. 1). This device uses channels AF3, AF4,
T7, T8, Pz according to the international 10-20 system (Fig. 2). The reference electrode was always placed
at the left ear. We used software that showed stimulus on the computer display («eSpeller», Java 1.8,
Windows 7, granted by CPR LLC, Samara, Russia). That software had been designed with "single
character paradigm" [11].
Fig. 1. Emotiv Insight 5-channel wireless EEG headset.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 2 / 10
Fig. 2. Electrodes disposition.
EEG was recorded in a small office room; all tests were performed when it was daylight. There was
some noise made by PC coolers. Datasets were recorded with the sampling rate 128 sec-1. The obtained
data was analyzed with MathWork® MATLAB R2015a (http://www.mathworks.com) in Department of
BCI of Samara State Medical University, and R statistical package (https://www.r-project.org).
2.1. Experimental design
We carried out our experiments within 9 days between 12:00 and 18:00. Each participant sat in
front of the PC display. The distance from a participant's eyes to the display was 0.5-0.7 m.
The participant wearing Emotiv Insight EEG headset was gazing at the display, when eSpeller
software showed him the square window with 250-mm-side size divided by 9 equal cells in the form of 3x3
matrix with a dot in each cell. The software flashed the cells in random order. 'Flashed' means changed the
image in the cell: it made the dot one and a half as large and it also altered the color from grey [RGB
179,179,179] to red [RGB 255, 0, 0] (Fig. 3). Each of the 9 cells was flashed only once in each cycle. One
cycle consists of 9 flashes. We set stimulus duration to 120 ms and pause duration to 180 ms. Therefore,
ISI (interstimulus interval) was set to 300 ms [22]. The 20 flashing cycles completed one experimental
session. There was a 10-sec break between the sessions.
Fig. 3. The window of eSpeller.
Cells 1 – 4, 6 – 9 are not flashing, the red dot is flashing in cell 5.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 3 / 10
Each participant had a task to concentrate on the only one cell at the moment. That cell was defined
as target cell and the red dot in this cell was called target visual stimulus. Therefore, the other 8 cells were
called non-target cells, and the red dots in them were named non-target visual stimulus. The cells were
flashed in random order.
In the course of the experiment each participant gazed at cells 1-9 in succession (Fig. 4).
Fig. 4. The succession in which cells are gazed at.
A participant gazes at the cell 1 first, after 20 flashing cycles they have a 10-sec break and then they
continue with cell 2. They repeat this procedure up to cell 9.
We presented visual stimuli and recorded EEG signals making time stamps simultaneously. As a
result, we got 9 files, one for each session, in accordance with the number of cells which were gazed at by
a participant. We conducted 3 trials for each participant. All the participants were healthy dexter men and
they were volunteers. In this paper the subjA, subjB, subjK, subjP, subjS represent the participants.
Therefore, we conducted 15 trials and recorded 135 EEG-signals – 9 files × 3 trials × 5 participants.
2.2. Continuous and discrete wavelet transformations
The Fourier transform requires stationarity of the signals and does not locate any events in the
signal in time domain. The wavelets give us this opportunity. The wavelets spread in many areas such as
time series analysis, image processing, variance stabilization and others. [12, 13]. One of the features of the
wavelets is scaling, which variates the width of the time window used for the analysis [14]. A wide class
of functions can be represented as wavelets, and then we can create an adapted wavelet from time series
data. Up until now many different wavelets have been created and they have a wide area of applications.
For example, they are Morlet, Daubechies, "mexican hat" wavelets. The Matlab wavelet toolbox presents
16 distinct wavelet families.
The function 𝜓(𝑡) which satisfies the next conditions is called wavelet:
𝜓-𝑢𝑑𝑢=1
𝜓𝑢𝑑𝑢=0
,
When these conditions are adhered to, there should be an interval [−𝑇,𝑇] of finite length and the
value of 𝜓(𝑡) close to 0 outside of that interval. As a result, we obtain the "small wave" or "wavelet" [15].
We used Daubechies wavelets (db4 and db6) for our work, which are shown in Fig. 5.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 4 / 10
)
*)
)
*)
Fig 5. Daubechies wavelets: db4 (left), db6 (right).
The continuous wavelet transform (CWT) of a signal 𝑥𝑡 ∈𝐿-𝑅 is defined as a dot product
С7,8
𝜓7,8𝑡
between the signal and the wavelet functions 𝜓7,8(𝑡):
С7,8 = <𝑥𝑡,𝜓7,8𝑡 >
scaled and translated wavelet function 𝜓(𝑡):
𝑎 𝜓((𝑡−𝑏)/𝑎)
𝜓7,8𝑡 =
where 𝑎− scale,
𝑏− translation.
the coefficients of the wavelet,
The CWT gives coefficients С7,8 which have maximum values at these scales and times where the
signal form is similar to that of the wavelet function 𝜓(𝑡). We make CWT with Matlab cwt() function.
𝑎?= 2?, 𝑏?,A= 2?𝑘 . These parameters give us a discrete family of wavelets:
𝜓?,A𝑡 = 2*? 𝜓2*?𝑡−𝑘, 𝑗,𝑘 ∈𝑍
The discrete wavelet transformation (DWT) is defined for discrete scale and translation parameters
The discrete wavelet transform adds some more constraints to the function 𝜓(𝑡) [16, 17]. However,
we can still perfectly reconstruct a signal 𝑥𝑡 from its discrete wavelet transform coefficients. In the case
of the orthogonal wavelets this is: 𝑥𝑡 =
𝐶?,A 𝜓?,A𝑡, 𝑗,𝑘 ∈𝑍
?,A
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 5 / 10
Therefore, we can define discrete (dyadic) wavelet transform as:
C?,A = <𝑥𝑡,𝜓?,A𝑡 >
3. Results
The CWT can be used for the recognition of the patterns which have the form of the wavelet
function [18]. When we transform a signal to the wavelet coefficients, the maximum coefficient gives time
and scale where the pattern appears. We have analyzed the signals which were recorded during experiment
using well-known wavelets such as Daubechies (db 4, 6), reverse biorthogonal wavelets (rbio) and symlets
(sym). These wavelets were used in many previous works for EEG filtering and analysis [19 - 21].
One epoch consisted of 128 samples (1 sec), which were counted from the start of visual stimulus.
Then we averaged 20 epochs and next we eliminated all linear trends and normalized the averaged signal.
For each wavelet we filtered a signal with the same wavelet. For the CWT we used scales 1-64, which
covered frequencies 0-30 Hz, and the time interval from 230 ms to 700 ms. It is known from the previous
works that P300 component normally appears in that time window [22]. We calculated the total of
coefficient's maximum at each time point; this statistic was used for pattern recognition:
𝑆=
K
8LM
max (С7,8)
where С7,8 CWT coefficients,
time window width.
𝑤 -
We obtained the accuracy of pattern recognition of P300 component for given wavelets. In Tables
1 - 4 we showed the accuracy of the algorithm for different numbers of the averaged epochs. The maximum
accuracy for each participant is marked yellow. The maximum accuracy among the participants is marked
green.
Table 1. P300 detection accuracy (%). The number of the averaged epochs is 5.
Participant
subjA
subjB
subjK
subjP
subjS
Db4
14.81
44.44
22.22
25.93
40.74
Db5
11.11
44.44
18.52
40.74
44.44
Wavelet
Db6
18.52
44.44
14.81
33.33
44.44
Sym6
14.81
40.74
18.52
25.93
44.44
Table 2. P300 detection accuracy (%). The number of the averaged epochs is 10.
Participant
subjA
subjB
subjK
subjP
subjS
Db4
29.63
37.04
37.04
33.33
51.85
Db5
22.22
33.33
37.04
44.44
55.56
Wavelet
Db6
18.52
22.22
25.93
40.74
55.56
Sym6
25.93
33.33
37.04
33.33
55.56
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
Rbio3.5
14.81
37.04
18.52
33.33
48.15
Rbio3.5
22.22
25.93
37.04
29.63
55.56
p. 6 / 10
Table 3. P300 detection accuracy (%). The number of the averaged epochs is 15.
Participant
subjA
subjB
subjK
subjP
subjS
Db4
29.63
44.44
44.44
22.22
59.26
Db5
29.63
51.85
40.74
37.04
70.37
Wavelet
Db6
33.33
40.74
37.04
37.04
66.67
Sym6
48.15
44.44
44.44
33.33
59.26
Table 4. P300 detection accuracy (%). The number of the averaged epochs is 20.
Particiapnt
subjA
subjB
subjK
subjP
subjS
4. Discussion
Db4
55.56
62.96
51.85
33.33
77.78
Db5
44.44
66.67
51.85
37.04
77.78
Wavelet
Db6
48.15
62.96
44.44
37.04
77.78
Sym6
51.85
59.26
51.85
37.04
74.07
Rbio3.5
22.22
40.74
37.04
29.63
70.37
Rbio3.5
40.74
62.96
33.33
37.04
81.48
We can sum up the results obtained from our experiments:
- classification accuracy depends strongly on the number of target stimuli. For example, the
maximum accuracy for participant subjS changed from 48,15% to 81,48%, when we increased the number
of target stimuli from 5 to 20;
- classification accuracy strongly depends on a participant. For example, the accuracy for
participant subjP was 37,04% and 81,48% for participant subjS (within 20 averaged epochs for both);
- classification accuracy depends weakly on the wavelet chosen. For example, the accuracy for
participant subjS was 74.07% with wavelet Sym6 and 81,48% with wavelet Rbio3.5;
- the average of the classification accuracy was 54,4±7,7% with 20 target stimuli. The minimum
was 33,33% for participant subjP with wavelet Db4; the maximum was 81,48% for participant subjS with
wavelet Rbio3.5.
We can conclude that the current realization of wavelet algorithm for patter recognition in EEG
signals has several drawbacks: a low level of generalization and low overall accuracy even with a large
quantity of target stimuli. We think there are several reasons for this:
- there are anthropomorphic differences between subjects. The shape and time localization of the
ERP-waves and their P300 component have high variance among humans. Our results confirm that well-
known fact;
- the wavelets chosen, in general, have a shape similar to ERP-wave, but this is not a sufficient
condition for detection. We can suppose that the adapted wavelet will show a better performance;
- we can presume that the Emotiv Insight device has hardware or software design features which
make an effect on the given algorithm and do not allow us to get a better result.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 7 / 10
5. Conclusion
We have presented an algorithm for the P300 component detection in an EEG signal using several
well-known wavelets. This algorithm can be used with Emotiv Insight device for a certain category of
people. Additional testing of the accuracy level for a certain participant is required before using the given
algorithm. Accuracy is likely to be improved if the adapted wavelet is used, and we are planning this
investigation in the future.
The results obtained need more investigation. For example, we can test the given algorithm on the
data acquired with other EEG equipment. Also, we could try to adapt a wavelet to the individual pattern of
the ERP wave of a participant.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 8 / 10
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
J. Oeppen, J. W. Vaupel, and others, "Broken limits to life expectancy," Science (80-. )., vol. 296,
no. 5570, pp. 1029–1031, 2002.
G. Logroscino, R. Tortelli, G. Rizzo, B. Marin, P. M. Preux, and A. Malaspina, "Amyotrophic lateral
sclerosis: An aging-related disease," Curr. Geriatr. Reports, vol. 4, no. 2, pp. 142–153, 2015.
B. a Yankner, T. Lu, and P. Loerch, "The aging brain.," Annu. Rev. Pathol., vol. 3, pp. 41–66, 2008.
[Online]. Available: http://www.who.int/mediacentre/factsheets/fs310/ru/.
K. McGregor and B. Pentland, "Head injury rehabilitation in the U.K.: An economic perspective,"
Soc. Sci. Med., vol. 45, no. 2, pp. 295–303, 1997.
C. D.N. and O. J., "A clinical and economic perspective on head injury rehabilitation," Journal of
Head Trauma Rehabilitation, vol. 8, no. 4. pp. 1–14, 1993.
G. Edlinger, B. Z. Allison, and C. Guger, "How many people can use a BCI system?," in Clinical
Systems Neuroscience, 2015, pp. 33–66.
R. Quian Quiroga, "Obtaining single stimulus evoked potentials with wavelet denoising," Phys. D
Nonlinear Phenom., vol. 145, no. 3–4, pp. 278–292, 2000.
R. Quian Quiroga, O. W. Sakowitz, E. Basar, and M. Schürmann, "Wavelet Transform in the
analysis of the frequency composition of evoked potentials," Brain Res. Protoc., vol. 8, no. 1, pp.
16–24, 2001.
[10] X. Li, X. Chen, Y. Yan, W. Wei, and Z. Wang, "Classification of EEG Signals Using a Multiple
Kernel Learning Support Vector Machine," Sensors, vol. 14, no. 7, pp. 12784–12802, 2014.
[11] R. Fazel-rezai, S. Gavett, W. Ahmad, A. Rabbi, and E. Schneider, "A Comparison among Several
P300 Brain-Computer Interface Speller Paradigms," Clin. EEG Neurosci., vol. 42, no. 4, pp. 209–
213, 2011.
[12] C. Heil, "Ten Lectures on Wavelets (Ingrid Daubechies)," SIAM Review, vol. 35, no. 4. pp. 666–
669, 1993.
[13] S. Mallat, "A Wavelet Tour of Signal Processing," A Wavelet Tour Signal Process., pp. 20–41,
1999.
[14] C. K. Chui, Wavelets : a tutorial in theory and applications, no. 2. 1992.
[15] D. B. Percival and A. T. Walden, "Wavelet Methods for Time Series Analysis," J. Nonparametr.
Stat., vol. 10, no. December 2014, p. 594, 2000.
[16] E. Erçelebi, "Electrocardiogram signals de-noising using lifting-based discrete wavelet transform,"
Comput. Biol. Med., vol. 34, no. 6, pp. 479–493, 2004.
[17] M. V. Wickerhauser, A. Jensen, A. la Cour-Harbo, A. Boggess, and F. J. Narcowich, Ripples in
Mathematics: The Discrete Wavelet Transform, vol. 110, no. 2. 2003.
[18] D. B. Percival and A. T. Walden, Wavelet Methods for Time Series Analysis. 2000.
[19] A. B. Geva and D. H. Kerem, "Forecasting generalized epileptic seizures from the EEG signal by
wavelet analysis and dynamic unsupervised fuzzy clustering," IEEE Trans. Biomed. Eng., vol. 45,
no. 10, pp. 1205–1216, 1998.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 9 / 10
[20] V. J. Samar, A. Bopardikar, R. Rao, and K. Swartz, "Wavelet Analysis of Neuroelectric Waveforms:
A Conceptual Tutorial," Brain Lang., vol. 66, no. 1, pp. 7–60, 1999.
[21] M. Murugappan, "Classification of human emotion from EEG using discrete wavelet transform," J.
Biomed. Sci. Eng., vol. 03, no. 04, pp. 390–396, 2010.
[22] S. J. Luck, An Introduction to the Event-Related Potential Technique, vol. 78, no. 3. 2005.
Wavelet algorithm for the identification of P300 ERP component, © 2016 Agapov et al.
p. 10 / 10
|
1709.03978 | 1 | 1709 | 2017-09-11T22:39:35 | A neuronal network model of interictal and recurrent ictal activity | [
"q-bio.NC"
] | We propose a neuronal network model which undergoes a saddle-node bifurcation on an invariant circle as the mechanism of the transition from the interictal to the ictal (seizure) state. In the vicinity of this transition, the model captures important dynamical features of both interictal and ictal states. We study the nature of interictal spikes and early warnings of the transition predicted by this model. We further demonstrate that recurrent seizures emerge due to the interaction between two networks. | q-bio.NC | q-bio |
A neuronal network model of interictal and recurrent ictal activity
M. A. Lopes,1, 2, 3, 4, ∗ K.-E. Lee,4, 5 and A. V. Goltsev4, 6
1College of Engineering, Mathematics and Physical Sciences,
University of Exeter, Devon EX4, United Kingdom
2Wellcome Trust Centre for Biomedical Modelling and Analysis,
University of Exeter, Devon EX4, United Kingdom
3EPSRC Centre for Predictive Modelling in Healthcare,
University of Exeter, Devon EX4, United Kingdom
4Department of Physics & I3N, University of Aveiro, 3810-193 Aveiro, Portugal
5Department of Anesthesiology and Center for Consciousness Science,
University of Michigan Medical School, Ann Arbor, Michigan, USA
6A.F. Ioffe Physico-Technical Institue, 194021 St. Petersburg, Russia
We propose a neuronal network model which undergoes a saddle-node on an invariant circle
bifurcation as the mechanism of the transition from the interictal to the ictal (seizure) state. In the
vicinity of this transition, the model captures important dynamical features of both interictal and
ictal states. We study the nature of interictal spikes and early warnings of the transition predicted by
this model. We further demonstrate that recurrent seizures emerge due to the interaction between
two networks.
PACS numbers: 87.19.lj,87.19.xm,05.70.Fh,05.10.-a,87.19.ln,87.18.Sn
I.
INTRODUCTION
Epilepsy affects nearly 1% of the population world-
wide [1]. This disabling neurological disorder is char-
acterized by spontaneous recurrent seizures, which cor-
relate to strongly synchronized neuronal activities, so-
called paroxysmal activity, revealed in electroencephalo-
grams (EEG). Anticonvulsant medications can prevent
seizures, but side effects are frequently reported [2]. For
about 30% of the patients, medications are not effec-
tive [3]. A minority undergoes surgery to remove the
epileptogenic brain tissue, but even in these cases the
patients may continue experiencing spontaneous seizures
[4, 5]. One of the main challenges has been to try to fore-
cast seizures [6 -- 8]. On one hand, the unpredictability of
seizure occurrence is a major burden of the condition [9]
and therefore being able to alert patients of impending
seizures could greatly improve their quality of life. On the
other hand, it would allow the design of closed-loop inter-
vention systems which could stop seizures [10]. Much of
the research in seizure prediction has been focused on al-
gorithms [8], however a better understanding of epilepsy
mechanisms is required.
Although epilepsy is an umbrella term for a range of
syndromes, the electrophysiological signatures are similar
between them [11]. For instance, different epileptogenic
lesions can produce similar electroencephalographic pat-
terns [12]. Also, it is noteworthy that it is possible to in-
duce seizures in non-epileptic brains across species both
in vivo and in vitro, which again present similar electro-
physiological features (see e.g.
[13, 14]). As argued by
Jirsa et al. [11], these facts suggest the existence of invari-
ant dynamical properties underlying seizure dynamics.
∗ [email protected]
Moreover, there are evidences that seizures self-terminate
via a critical transition [15]. Note that bifurcations are
the mechanisms of phase transitions in many-body in-
teracting systems [16, 17]. The two frameworks provide
complementary insights about the underlying transitions.
While there are extensive studies on what kind of bi-
furcations occur at the onset and offset of seizures (e.g.
[11, 18, 19]), there are fewer studies about the collective
nature of the phase transitions [20]. The fundamental
question is how does the spike interaction between large
populations of neurons may result in seizures.
Critical phenomena provide early-warnings of phase
transitions [17], which consequently open the possibility
to take action to prevent the occurrence of those transi-
tions. We have previously demonstrated that the inter-
action between neurons on a network give rise to collec-
tive phenomena and diverse phase transitions [21]. Note
that different phase transitions are associated with dif-
ferent precursors and different critical phenomena such
as (among other) bursts of neuronal activity, avalanches,
hysteresis, critical slowing-down, symmetry breaking,
and resonance phenomena [21, 22].
Here, we propose a neuronal network model consist-
ing of interacting excitatory and inhibitory neurons to
understand the nature and emergence of both interictal
and ictal activity. We start by presenting the model and
its dynamical states. We then study the properties of
interictal-like spikes, which are evoked in the vicinity of
the transition to the ictal state. As the dynamical state
moves towards the transition, early-warning phenomena
signal the impending transition. We demonstrate that
such phenomena is revealed through the stimulation of
interictal-like spikes, and the analysis of accompanying
low-fluctuating activity. Finally, we show that recurrent
ictal activity is an emergent collective phenomenon in a
system of two interacting neuronal networks.
II. NEURONAL NETWORK MODEL
Herein we consider a neuronal network model [21 -- 23],
which we will refer to as the stochastic cellular automata
neuronal network model (SCANNM). In the SCANNM,
neurons are modelled as stochastic integrate-and-fire neu-
rons: they integrate the inputs and fire a train of spikes
with a certain probability if the input is larger than an
activation threshold (see Appendix A for more details).
There are two populations of neurons, excitatory and in-
hibitory neurons. Excitatory neurons fire positive out-
puts, whereas inhibitory fire negative outputs to their
postsynaptic neighbours. In general, the two populations
have different response times to stimuli. Here we consider
the case in which excitatory neurons respond faster than
inhibitory neurons. Additionally, neurons are excited by
endogenous stimuli that account for random spikes com-
ing from other areas of the brain, as well as spontaneous
releases of neurotransmitters at the synapses. (cid:104)n(cid:105) is the
endogenous stimulation and we use it as control parame-
ter. The neurons form an uncorrelated random directed
network (properties of this kind of networks have been
studied, for example, in [24 -- 26]). The mean-field neu-
ronal dynamics are described by the fractions of active
excitatory (ρe) and inhibitory (ρi) neurons which follow
the rate equations [21, 23]
ρa
µa
= −ρa + Ψa(ρe, ρi,(cid:104)n(cid:105)),
(1)
where a = e, i, and ρ ≡ dρ/dt. Ψa(ρe, ρi,(cid:104)n(cid:105)) is the
probability that at time t a randomly chosen excitatory
(a = e) or inhibitory (a = i) neuron becomes active.
This function encodes the network structure, single neu-
ron stochastic firing rules, and endogenous stimulation
[21 -- 23].
The model is analytically solvable, but despite its sim-
plicity, it describes a rich repertoire of collective phenom-
ena, namely neuronal avalanches, bursty activity, hys-
teresis, bistability, different kinds of neuronal oscillations,
phase transitions, and stochastic resonance [21, 22]. Fur-
thermore, the SCANNM combines two usually distinct
modelling frameworks to describe mesoscopic brain dy-
namics, namely it allows the modelling of large-scale neu-
ronal networks like in Ref. [27], and it is simultaneously
described by a neural mass formulation (see for instance
[28]). It thus enables an analysis of both single neuron
dynamics within the network, and large-scale dynamics
of neuronal populations that can be treated numerically
and analytically.
Figure 1 depicts the different patterns of neuronal ac-
tivity in the SCANNM. In this paper we focus on the re-
gions corresponding to (i) low fluctuating activity around
a stable state, which we identify as a 'normal' state; (ii)
low fluctuating activity with sporadic single sharp os-
cillations, the 'interictal' state; and (iii) sustained net-
work oscillations, which is the model 'ictal' state. The
boundary between the interictal and ictal regions corre-
sponds to a saddle-node on an invariant circle (SNIC)
2
FIG. 1. Schematic diagram of the different dynamical states
of the SCANNM as function of the endogenous stimulation.
We find low neuronal activity at low stimulation ((cid:104)n(cid:105) < nc1).
At an intermediate stimulation level (nc1 < (cid:104)n(cid:105) < nc2), the
neuronal network exhibits transients of high activity, either
bursts or single sharp oscillations, which give place to damped
oscillations or sustained oscillations at (cid:104)n(cid:105) > nc2 respectively
depending on parameters. The shape of the sustained oscilla-
tions changes from slow, high-amplitude oscillations to fast,
low-amplitude oscillations as the stimulation increases. The
neuronal network produces damped oscillations to a high ac-
tivity state at (cid:104)n(cid:105) > nc3.
bifurcation (at (cid:104)n(cid:105) = nc2), which is the critical point of
a second-order phase transition from the interictal to the
ictal state [21]. At higher endogenous stimulations, the
ictal region is bounded by the critical point of a supercrit-
ical Hopf bifurcation that separates the ictal state from
a high activity state (at (cid:104)n(cid:105) = nc3). In the vicinity of the
Hopf bifurcation, the neuronal oscillations have high fre-
quency, and low amplitude, in contrast to the oscillations
close to the SNIC bifurcation, which are characterized by
low-frequency and high-amplitude [29]. In the interictal
state, interictal-like spikes (ILS) emerge at random, but
with a deterministic shape. Note that ILS are strongly
nonlinear events that comprise the synchronous activity
of almost 90% of the neurons in the network. An ILS
is described by a trajectory that goes around an unsta-
ble point in the (ρe, ρi)−phase plane (see Sec. III). Their
occurrence is deterministic if the activity overcomes a
threshold (a separatrix in the phase plane). This thresh-
old defines the number of excitatory neurons that must be
activated simultaneously in order to generate an ILS. As
an example, for a network of 104 neurons at (cid:104)n(cid:105) = 16 (i.e.,
at (nc2−(cid:104)n(cid:105))/nc2 ≈ 0.15), which is in a low activity state
where almost all neurons are inactive, the simultaneous
activation of just 75 excitatory neurons chosen at random
(i.e., about 1% of the excitatory neurons in a network
with 25% inhibitory neurons) generates an ILS formed
by the synchronized activity of about 9000 neurons [21].
One should note that the duration of these ILS is much
larger than the period of single neuron spikes. This im-
plies that an ILS is genuinely a collective phenomena in
low activitity(normal state)single sharposcillations(interictal spikes)bursting activity(bistable state)dampedoscillations(spindles)slow sharposcillations(ictal activity)fast oscillationsnc1c2c3nn spite of the fact that it can be elicited by a small num-
ber of neurons. For realistic parameters, namely mean
degree c = 1000, fraction of inhibitory neurons gi = 0.25,
synaptic efficacies ratio Ji/Je = −3, dimensionless fir-
ing threshold Ω = 30, ratio between excitatory and in-
hibitory response times α = µi/µe = 0.7, and 1/µe = 20
ms (see Appendix A and Refs. [21 -- 23]), the typical du-
ration of an ILS is about 100 ms which is comparable
to real interictal spikes [30]. Within the ictal region, the
frequency of sustained oscillations increases with endoge-
nous stimulation from very low frequencies up to 4 Hz
[21], which is comparable to the frequency of ictal activ-
ity [31]. These are natural features of the model, without
needing to calibrate parameters.
III. THE NATURE OF THE INTERICTAL-LIKE
SPIKES
In order to understand the nature of ILS, let us study
their phase trajectories in the plane ρe − ρi. ILS emerge
in the region nc1 < (cid:104)n(cid:105) < nc2 (see Fig. 1). In this re-
gion there are three fixed points: a stable fixed point
at low activity, a saddle point at an intermediate activ-
ity, and an unstable point at high activity [21]. For this
parameter region, the dynamics in the SCANNM is qual-
itatively equivalent to the dynamics of the Morris-Lecar
neuron near the SNIC bifurcation [32]. To the best of
our knowledge, a SNIC bifurcation has not been found in
any other neuronal network model. Thus, the SCANNM
gives an unique possibility to study the collective behav-
ior of neuronal populations near a SNIC bifurcation. We
follow the nomenclature used by Rinzel and Ermentrout
[32]: the stable point corresponds to a rest state (R), and
the saddle point is a threshold (T) on the separatrix that
divides the phase plane into two regions. There are two
heteroclinic orbits connecting T to R, one corresponding
to immediate exponential relaxation to R, and another
that goes around the unstable point (U), reaching high
activity before relaxing to R. The ILS follows the second
path (see Fig. 2) as spikes of single neurons do in the
Morris-Lecar model. Panels (a) and (b) of Fig. 2 also
portray the nullclines of the system,
ρa = 0 ⇔ ρa = Ψa(ρe, ρi,(cid:104)n(cid:105)).
(2)
The nullclines determine the maximums or minimums of
excitatory and inhibitory activity. In this case, the ac-
tivity of the population a increases ( ρa > 0) below the
respective nullcline, whereas it decreases above. Con-
sequently, the ILS move counterclockwise in the phase
plane. Thus, any excitatory activity perturbation that
drives the activity state below both nullclines results in
an ILS as the one displayed in Fig. 2(c). At the critical
point nc2, the points R and T merge, and there is a ho-
moclinic orbit around the unstable point connecting the
saddle-node to itself.
The distance between R and T defines an activation
threshold Ath((cid:104)n(cid:105)) for the generation of ILS. We demon-
3
(a) Schematic representation of the nullclines and
FIG. 2.
fixed points in the (ρe, ρi)−phase plane.
(b) Nullclines
and fixed points of the SCANNM (numerical integration of
Eqs. (2)). In panels (a) and (b) the blue and red lines corre-
spond to the ρe- and ρi-nullcline, respectively. The nullclines
intersect at the fixed points: the stable or rest state (R), the
saddle or threshold (T), and the unstable point (U). (c) Tra-
jectory of an ILS in the (ρe, ρi)−phase plane. The line is the
result of the numerical integration of Eqs. (1) and the dots
represent simulations of the model. Parameters: c = 1000,
Ω = 30, gi = 0.25, Ji = −3Je, σ2 = 10, (cid:104)n(cid:105) = 16, α = 0.7,
τ = 0.1 and N = 104 (see the Appendix A for details about
the parameters).
strate in Appendix B that Ath((cid:104)n(cid:105)) follows a square root
dependence with (cid:104)n(cid:105) in the vicinity of the SNIC bifurca-
tion,
Ath((cid:104)n(cid:105)) ∝(cid:112)nc2 − (cid:104)n(cid:105).
(3)
In a finite network, finite-size effects elicit activity fluctu-
ations which can overcome the activation threshold pro-
vided that the system is sufficiently close to the critical
point, which in turn results in the occasional generation
of ILS (see Fig. 3).
IV. EARLY WARNINGS OF THE TRANSITION
TO THE ICTAL STATE
If a control parameter such as the endogenous stimu-
lation (cid:104)n(cid:105) changes slowly from the normal state towards
4
FIG. 3. (a) Series of ILS generated at random due to finite-
size effects at (cid:104)n(cid:105) = 18.75 (nc2 = 18.8). (b) Zoom of an ILS
from panel (a). All ILS have the same shape. Parameters in
simulations are the same as in Fig. 2.
FIG. 4. Panels (a) and (b) depict the minimum signal am-
plitude Fmin to elicit an ILS and the minimum time between
two consecutive ILS Dmin as function of the endogenous stim-
ulation (cid:104)n(cid:105), respectively. Parameters in the numerical inte-
gration of the rate equations (1) are the same as in Fig. 2.
the ictal state through the interictal state, changes in
the neuronal dynamics can inform on how close is the
system to the critical point of the transition to the ictal
state. However, if the variation of the control param-
eter is too fast to observe its consequences in the dy-
namics, the transition cannot be anticipated [33]. In the
SCANNM, since the threshold decreases towards the crit-
ical point (Eq. (3)), finite-size effects are more likely to
generate ILS, and therefore the rate of ILS is expected to
increase. However, as the neuronal network approaches
the critical point nc2, the stable point R gets closer to
the saddle point T. The time D that the system spends
in this region of the phase plane increases as
D ∝ (nc2 − (cid:104)n(cid:105))−1/2.
(4)
This equation describes a general feature of a SNIC bi-
furcation [16]. Therefore, although it becomes easier to
elicit ILS from the rest state as the system approaches
the critical point, the refractory time also increases, i.e.,
it takes longer to invoke consecutive ILS. To study the
generation of ILS as function of the distance to nc2, we
stimulate the neuronal network with an excitatory force.
In Eqs. (1) we introduce an additional term (1 − ρe)F
that stimulates the inactive excitatory population (1−ρe)
with a delta-like field of amplitude F (duration equal to
a time step). We found the minimum amplitude Fmin
required to elicit ILS as function of endogenous stimula-
tion, and measured the minimum time between two con-
secutive ILS Dmin generated by two consecutive pulses
Fmin. Figure 4 shows that Dmin diverges as the system
approaches nc2, as Eq. (4) predicts, and Fmin tends to
zero. This is because Fmin is essentially a measure of the
activation threshold, and as such
Fmin ∝(cid:112)nc2 − (cid:104)n(cid:105).
(5)
For a sufficiently small fixed stimulation F , the neuronal
network generates ILS if F ≥ Fmin((cid:104)n(cid:105)). The stimulation
dependence of Dmin is a consequence of critical slowing
down near nc2. (Note that "critical slowing down" is a
phenomenon by which a system takes longer and longer
time to recover from small perturbations as it approaches
a critical point of a continuous phase transition [17].) At
the critical point, Fmin → 0 and D → ∞, meaning that
network oscillations emerge with zero frequency, which
corresponds to the homoclinic orbit mentioned above.
FIG. 5. Zero-frequency power enhancement as a precursor
of the transition to the ictal state. Panels (a) and (b) show
the power spectral density of activity fluctuations in decibel
(10 log10(PSD)) at (cid:104)n(cid:105) = 18.6 and (cid:104)n(cid:105) = 18.8, respectively.
(c) Maximum of the power spectral density at zero frequency
as function of the endogenous stimulation. Parameters in
simulations are the same as in Fig. 2.
Besides these dynamical changes involving ILS, the low
activity state is also affected by the critical slowing down.
This can be quantified by power spectral analysis of low
activity fluctuations near the critical point. Using the
Wiener-Khintchine theorem, we have demonstrated that
the power spectral density (PSD) of activity fluctuations
in the low activity state when (cid:104)n(cid:105) → nc2 has a sharp
zero-frequency peak which grows as Smax ∝ 1/(nc2−(cid:104)n(cid:105))
[21]. This behavior was demonstrated in a metastability
region in the vicinity of a first-order phase transition in
Ref. [21], and it also occurs near the second-order phase
transition under consideration. Figure 5 shows that the
zero-frequency peak of the PSD increases as the neuronal
network approaches the critical point nc2.
0 100000t 0 1 (a)89458970t01(b)1617181950 350(b)16171819 0 0.4(a) 0 100200f /Hz-150-100 -50PSD /dB(a) 0 100200f /Hz-150-100 -50PSD /dB(b) 18 18.518.902468max(PSD)10-3(c) 5
FIG. 6.
Ictal-like pattern evolution driven by endogenous
stimulation in the SCANNM. The blue line is time-dependent
stimulation (cid:104)n(cid:105) that varies in the range [16, 50] (right y-axis)
and drives the neuronal activity ρ (black line) from the low ac-
tivity state to fast low-amplitude oscillations with a DC shift,
which then evolves to high-amplitude, low-frequency oscilla-
tions, before returning to the low activity state. Parameters
in the numerical integration of the rate equations (1) are the
same as in Fig. 2.
V. MODEL OF RECURRENT TRANSITIONS
TO THE ICTAL STATE
If we relax the condition of a slowly changing control
parameter, the SCANNM is capable of mimicking a typ-
ical ictal pattern evolution [11, 31]. Figure 6 shows that
an abrupt increase of the endogenous stimulation can
bring the neuronal network from the low activity state to
the vicinity of the supercritical Hopf bifurcation, which
results in a DC shift of the neuronal activity accompa-
nied by fast low-activity oscillations. As then (cid:104)n(cid:105) slowly
decreases back to the 'normal' or the interictal state, the
frequency of the sustained network oscillations decreases,
whereas its amplitude increases.
At fixed parameters, the SCANNM can either be in the
normal, interictal, or ictal state. Recurrent transitions
between these states can either be achieved by a change
in parameters, or due to external stimuli. Another sce-
nario is to consider a network of networks, i.e., several
interacting neuronal networks. In general, a network of
networks can consist of multiple networks whose inter-
network connections can be both directed or undirected,
connecting different numbers of excitatory and inhibitory
neurons. We consider two interacting neuronal networks
as our minimal model of different interacting brain ar-
eas. If one of the networks is in the interictal state, then
a small additional excitatory input from another network
can induce a transition to the ictal state. Recurrent tran-
sitions will occur as a consequence of a recurrent input.
Such intermittent input can be generated by a network in
the bursting state (see Fig. 1). As previously described
in Ref. [21], the SCANNM produces recurrent irregular
bursts of neuronal activity when close to a first-order
phase transition. To illustrate the concept, we consider
two networks A and B of size N = 104. Network A
is in the bursting state [((cid:104)n(cid:105), α) = (18.7, 0.85), see the
Appendix A for the meaning of α] and network B is in
the interictal state [((cid:104)n(cid:105), α) = (18, 0.7)]. For simplicity,
network A sends axonal projections to network B, but
FIG. 7. Minimal model of recurrent ictal transitions.
(a)
Schematic representation of two networks, where the red net-
work (A) influences the blue (B). (b) Neuronal activity in net-
work A. (c) Neuronal activity in network B. The clusters of
ictal-like activity in network B are driven by the noisy bursts
of network A. Panels (d) and (e) are zooms of the activity
displayed in panels (b) and (c) respectively. Parameters in
simulations are the same as in Fig. 2, except for those re-
ferred in the text.
network B does not influence network A (see Fig. 7(a)).
We define a fraction gAB = 0.3 of excitatory neurons
chosen at random in each network, and we wire them
by directed connections from A to B (synaptic efficacies
JAB = 3Je, see the Appendix A). In Figs. 7(b) and (c),
we show the neuronal activity of the two networks. The
recurrent transitions to the ictal state in network B are
driven by the bursting activity in network A.
VI. DISCUSSION
In this paper we proposed a neuronal network model
to describe interictal spikes and recurrent ictal activity.
ILS are strongly nonlinear collective events that comprise
the synchronized activity of a large number of neurons.
ILS emerge from a low background activity when either
random fluctuations or stimuli force the neuronal activ-
ity to overcome a threshold (see Fig. 2). This thresh-
old becomes smaller as a control parameter (an endoge-
nous stimulation) moves the neuronal network towards
0 1000t010 1650 the transition to the ictal state. The transition corre-
sponds to a SNIC bifurcation at which sustained net-
work oscillations emerge with low frequency. This re-
gion of oscillations is also bounded at higher endogenous
stimulations by a supercritical Hopf bifurcation. Near
the Hopf bifurcation, oscillations have low-amplitude and
high-frequency.
It is conceivable that the transition to seizures may ei-
ther be a consequence of a gradual or an abrupt change
in the endogenous stimulation depending on the type of
epilepsy [33].
In the SCANNM, if we assume a grad-
ual increase of the endogenous stimulation, then we can
observe critical phenomena that signal the transition.
This transition can be compared to the high amplitude
slow (HAS) activity onset pattern observed in some focal
epilepsies (see [34] and references therein). We showed
that under this assumption, the ILS activation thresh-
old scales as the square root of the distance to the SNIC
bifurcation, whereas the minimum time between two con-
secutive spikes diverges at the critical point of the tran-
sition to the ictal state (see Fig. 4). Additionally, the
zero-frequency peak of the PSD of low activity fluctua-
tions reaches a maximum at the critical point (see Fig. 5).
Alternatively, if we instead assume that the control
parameter may change abruptly, then the SCANNM is
capable of mimicking the typical pattern evolution ob-
served in seizures (called low amplitude fast (LAF) ac-
tivity onset pattern in [34]). Seizures are often preceded
by a low-voltage, high-frequency discharge [35], and the
ictal pattern generally displays increasing amplitude and
decreasing frequency [31]. The SCANNM exhibits such
pattern evolution if we assume that at seizure onset the
endogenous stimulation abruptly increases, forcing the
neuronal network to jump from the interictal-like state
to the vicinity of the Hopf bifurcation. Then, as the stim-
ulation gradually decreases towards 'normal' levels, the
neuronal activity evolves from low-amplitude, fast oscil-
lations, to high-amplitude, slow oscillations (see Fig. 6).
Finally, we demonstrated the viability of modelling re-
current ictal transitions using two interacting neuronal
networks, where the intermittent output of one network
drives the other to recurrent seizures. This concept aligns
well with other modelling approaches which have also
explored the role of interacting populations to generate
ictal-like activity [36 -- 39]. Although for simplicity we
have considered here the mechanism to seizures as an
excitatory drive from another network, we would like to
note that this is equivalent to an interruption of inhibi-
tion. Such mechanism of triggering seizure-like activity
has been observed in genetically engineered mice, where
the shut-down of CA2 output leads to hyperexcitability
in the recurrent CA3 network [40]. Note however that
the networks involved in the generation of ictal activity
may be located in distant regions of the brain. This high-
lights the importance of studying large scale brain net-
works, rather than focal brain activity even in the case
of focal epilepsies [41].
Contrarily to previous computational models of
6
epilepsy [11, 42, 43], the SCANNM was not explicitly
designed to describe epileptiform activity [21, 23].
In-
stead, this model demonstrates that interictal and ictal-
like activity may be emergent phenomena of an inter-
acting neuronal network. The heterogeneous mean-field
equations of the SCANNM were derived from a minimal
set of fundamentals, namely neurons behave as stochastic
integrate-and-fire neurons, there are two types of neurons
(excitatory and inhibitory), and the neurons interact on
a complex network [21, 23, 44]. Our analytical and nu-
merical results were in good agreement with simulations
of the model for sufficiently large networks (typically for
N > 103 [23]). We conclude that the SCANNM thus al-
lows to study numerically and analytically neural mass-
like equations, and to measure and compare single neu-
ronal activity. It enables to examine how the activity of
single neurons can impact on the whole network. For the
case in point, ILS are a remarkable example of a collective
network phenomenon that can be evoked by the simul-
taneous activation of a few neurons. On the other hand,
once an ILS is excited, the SCANNM predicts that it is
very difficult to suppress it. The only way to make the
neuronal network return to the low activity state is by
applying a strong inhibitory stimulus to a considerable
macroscopic part of the network.
A recent
study has
reported apparently self-
contradictory evidence on the role of pre-ictal spikes for
the prediction of seizures [30], showing that different
seizures could be preceded by an increase or decrease of
the pre-ictal spike rate. The SCANNM provides a possi-
ble explanation for this observation. In the model, as a
transition to the ictal state is approached, two competing
mechanisms can influence the spike rate. On one hand,
the activation threshold of ILS decreases which leads to
a higher spike rate. On the other hand, critical slowing
down hinders the consecutive emergence of spikes (see
Fig. 4). It is then conceivable that the prevailing mech-
anism may vary from seizure to seizure, and as a result
the spike rate can increase or decrease before a seizure.
There is also conflicting evidence as to which it may or
may not be possible to predict seizures based on critical
phenomena [45] or using other data analysis [8]. Nev-
ertheless, based on the SCANNM we can propose two
measures to forecast seizures. First, Fig. 4(a) indicates
that the required external stimulation to evoke ILS be-
comes smaller as the neuronal network approaches the
transition to the ictal state. In the case of photosensitive
epilepsy, the stimulation can be intermittent photic stim-
ulation. In other epilepsies it may be necessary to use im-
planted electrodes to electrically stimulate the brain, like
it was proposed by Silva et al.
[46]. Thus, for a given
patient, and after a sufficient number of trials, it may
be possible to correlate the minimum required stimula-
tion to elicit ILS with the timing of impending seizures.
However, such method may be infeasible due to the risk
of inducing seizures due to this probing stimulation [47].
The second measure does not require stimulation, instead
it uses the analysis of ongoing EEG recordings. Figure 5
shows that the power of low frequencies should increase
in EEG recordings when approaching a transition to a
seizure. In fact, a gradual increase in power of low fre-
quencies has been observed as a precursor of spike-wave
discharges in absence epilepsy both in humans and rat
models [48 -- 51]. We acknowledge, however, that even if
this process takes place, it may often be unobservable be-
cause if the transition occurs in a faster time scale than
the scale of the low frequencies, then it is not possible to
find the gradual power increase.
VII. CONCLUSION
In this paper we proposed a neuronal network model
(the SCANNM) to describe interictal and ictal activ-
ity, as well as ictal-like pattern evolution, and spon-
taneous recurrent transitions to seizures. Additionally,
we found a set of precursors that signal the transition
to the ictal state. The neuronal activity state was de-
pendent of an endogenous stimulation which we used as
the control parameter. The interictal state was char-
acterized in the model by low fluctuating activity from
which interictal-like spikes could sporadically emerge.
We demonstrated that the required stimulation to elicit
interictal-like spikes tends to zero as the neuronal net-
work approaches the critical point of a saddle-node bi-
furcation. Furthermore, the transition is signaled by an
increase of the zero-frequency peak of the power spec-
trum of low activity fluctuations when the endogenous
stimulation varies slowly. On the other hand, for abrupt
changes in the control parameter, we showed that the
model can mimic a typical ictal pattern evolution: as
onset with low-voltage, high-frequency discharges, fol-
lowed by increasing amplitude, decreasing frequency os-
cillations. Finally, we showed that the model could also
reproduce recurrent transitions to the ictal state at fixed
parameters, as the result of the interaction between two
neuronal networks.
VIII. ACKNOWLEDGEMENTS
This work was partially supported by FET IP Project
MULTIPLEX 317532. A.V.G. is grateful to LA I3N for
Grant No. PEST UID/CTM/50025/2013. M.A.L. ac-
knowledges the financial support of the Medical Research
Council (MRC) via grant MR/K013998/01. K.E.L. was
supported by the Department of Anesthesiology at the
University of Michigan (Ann Arbor) and National In-
stitutes of Health (Bethesda, MD, USA) grant RO1
GM098578.
Appendix A: SCANNM
Here we describe the SCANNM [21 -- 23].
7
1. Network structure and stochastic dynamics
The neuronal network is composed of N stochastic neu-
rons, geN excitatory and giN inhibitory (ge + gi = 1).
We consider that the network has the structure of the
Erdos-R´enyi network. This is a random network with
small world properties, namely small mean shortest path
length like real neuronal networks in the brain [52]. The
neurons are connected by directed edges which repre-
sent synapses that allow active neurons to send spikes
to their postsynaptic neighbors.
In addition, neurons
also receive random spikes from endogenous stimulation
that represent spontaneous releases of neurotransmitters
in synapses and random spikes arriving from other areas
of the brain (this stimulation has properties of shot noise
[21]).
The dynamics of the stochastic neurons is determined
by the following rules. If during an integration time win-
dow τ the total input Vj(t) to an inactive neuron be-
comes larger than a threshold Ω, then with probability
τ µa the neuron becomes active and fires a spike train at
a constant frequency ν (the index a = e if the neuron is
excitatory and a = i if it is inhibitory). If the total input
Vj(t) to an active excitatory (inhibitory) neuron becomes
smaller than Ω, then the neuron stops to fire with proba-
bility τ µa. In this model, the rates µe and µi are the re-
ciprocal first-spike latencies of excitatory and inhibitory
neurons, respectively. We define a parameter α as the ra-
tio of the first-spike latency of excitatory neurons to the
first-spike latency of inhibitory neurons, α ≡ µi/µe. If
α < 1, then excitatory neurons respond faster to stimuli
than inhibitory neurons. This neuronal stochastic behav-
ior is meant to account for intrinsic noise within neurons
[53], namely, ion channel stochasticity [54].
2. Rate equations
The fractions ρe(t) and ρi(t) of active excitatory and
inhibitory neurons, respectively, characterize the neu-
ronal activity at time t. They are determined by the
rate equations (1), in which Ψa(ρe, ρi,(cid:104)n(cid:105)) is the prob-
ability that, at time t, the total input to a randomly
chosen excitatory (a = e) or inhibitory (a = i) neu-
ron is at least the threshold Ω at a given endogenous
stimulation (cid:104)n(cid:105). The functions Ψa(ρe, ρi,(cid:104)n(cid:105)) are deter-
mined by the network structure, the distribution func-
tion of endogenous stimulation (we consider the Gaus-
sian distribution), and the frequency-current relationship
of single neurons (a step function in this model [23]).
Note that the probability Ψa(ρe, ρi,(cid:104)n(cid:105)) is the same for
both excitatory and inhibitory neurons because, in the
network under consideration, excitatory and inhibitory
neurons occupy topologically equivalent positions. Thus,
Ψe(ρe, ρi,(cid:104)n(cid:105)) = Ψi(ρe, ρi,(cid:104)n(cid:105)) ≡ Ψ(ρe, ρi,(cid:104)n(cid:105)), where
Ψ(ρe, ρi,(cid:104)n(cid:105)) =(cid:80)
n,k,l≥0 Θ(nJn+kJe+lJi−Ω)G(n,(cid:104)n(cid:105)) ×
(A1)
Pk(geρec)Pl(giρic).
Here, c = cντ and c is the mean degree. Θ(x) is the
Heaviside step function. Pq(c) is the Poisson distribution
function,
Pq(c) = cqe−c/q!,
(A2)
and G(n,(cid:104)n(cid:105)) is the Gaussian distribution function,
.
G(n,(cid:104)n(cid:105)) = G0e−(n−(cid:104)n(cid:105))2/2σ2
and G0 is the normalization constant,(cid:80)∞
(A3)
G(n,(cid:104)n(cid:105)) defines the probability that a neuron receives n
spikes from endogenous stimulation per integration time
τ . (cid:104)n(cid:105) is the mean number of spikes, σ2 is the variance,
n=0 G(n,(cid:104)n(cid:105)) =
1. We use (cid:104)n(cid:105) as the control parameter characterizing the
endogenous stimulation. Note that Eqs. (1) and (A1) are
asymptotically exact in the thermodynamic limit, N →
∞ [21, 23].
In numerical simulations, we use the algorithm ex-
plained in [21]. We used the following model parameters
(except when mentioned otherwise): N = 104, c = 103,
Ω = 30, τ ν = 1, µeτ = 0.1, α = 0.7, and gi = 0.25.
Throughout this paper we used 1/µe ≡ 1 as time unit and
Je ≡ 1 as input unit. Following [55], we chose Ji = −3Je.
We also used Jn = Je and σ2 = 10 for the amplitude and
variance of the endogenous stimulation. These parameter
choices have been discussed in [21 -- 23].
Appendix B: The activation threshold of ILS
In this appendix we show how the activation thresh-
old Ath of ILS depends on the endogenous stimula-
tion (cid:104)n(cid:105) near the SNIC bifurcation. Since we consider
8
Ψe = Ψi ≡ Ψ, this implies that there is only one steady
state equation,
ρ = Ψ(ρ,(cid:104)n(cid:105)).
In this case, the SNIC bifurcation occurs when
(ρ,(cid:104)n(cid:105)) = 1,
∂Ψ
∂ρ
(B1)
(B2)
which determines the critical point nc2.
In the region
of ILS, close to the bifurcation, at 0 < nc2 − (cid:104)n(cid:105) (cid:28)
nc2, we can study low activity fluctuations δρ((cid:104)n(cid:105)) =
ρ((cid:104)n(cid:105)) − ρ(nc2) near the SNIC bifurcation. To find how
the activity fluctuations depend on the stimulation (cid:104)n(cid:105),
we apply the Taylor expansion to Ψ(ρ,(cid:104)n(cid:105)) in Eq. (B1)
over δρ((cid:104)n(cid:105)) and δn = (cid:104)n(cid:105) − nc2 up to the second order
in δρ((cid:104)n(cid:105)),
δρ((cid:104)n(cid:105)) ≈ ∂Ψ
∂(cid:104)n(cid:105) δn +
∂Ψ
∂ρ
+
∂2Ψ
∂ρ2 (δρ)2,
(B3)
where the derivatives of Ψ are taken at nc2. Using
Eq. (B2), we find
where
ρ((cid:104)n(cid:105)) − ρ(nc2) ≈ −C(cid:112)nc2 − (cid:104)n(cid:105),
(cid:17)−1
(cid:16) ∂2Ψ
(cid:115)
C =
−2
∂Ψ
∂(cid:104)n(cid:105)
.
∂ρ2
Consequently, near the critical point nc2, the activation
threshold Ath((cid:104)n(cid:105)) also follows
Ath((cid:104)n(cid:105)) ∝(cid:112)nc2 − (cid:104)n(cid:105).
(B4)
(B5)
(B6)
[1] P. N. Banerjee, D. Filippi, and W. A. Hauser, Epilepsy
[10] W. C. Stacey and B. Litt, Nat. Clin. Pract. Neurol. 4,
Res. 85, 31 (2009).
190 (2008).
[2] P. Ortinski and K. J. Meador, Epilepsy Behav. 5, 60
[11] V. K. Jirsa, W. C. Stacey, P. P. Quilichini, A. I. Ivanov,
(2004).
and C. Bernard, Brain 137, 2210 (2014).
[3] P. Kwan and M. J. Brodie, N. Engl. J. Med. 342, 314
[12] P. Perucca, F. Dubeau, and J. Gotman, Brain 137, 183
(2000).
(2014).
[4] J. de Tisi, G. S. Bell, J. L. Peacock, A. W. McEvoy, W. F.
Harkness, J. W. Sander, and J. S. Duncan, Lancet 378,
1388 (2011).
[5] I. Najm, L. Jehi, A. Palmini, J. Gonzalez-Martinez,
E. Paglioli, and W. Bingaman, Epilepsia 54, 772 (2013).
[6] M. Ch´avez, M. Le Van Quyen, V. Navarro, M. Baulac,
and J. Martinerie, IEEE Trans. Biomed. Eng 50, 571
(2003).
[7] M. Le Van Quyen, J. Soss, V. Navarro, R. Robertson,
M. Chavez, M. Baulac, and J. Martinerie, Clin. Neuro-
physiol. 116, 559 (2005).
[8] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehn-
ertz, Brain 130, 314 (2007).
[9] H. M. De Boer, M. Mula, and J. W. Sander, Epilepsy
Behav. 12, 540 (2008).
[13] Y. H. Raol and A. R. Brooks-Kayal, Prog. Mol. Biol.
Transl. Sci. 105, 57 (2011).
[14] G. Huberfeld, L. M. De La Prida, J. Pallud, I. Cohen,
M. Le Van Quyen, C. Adam, S. Clemenceau, M. Baulac,
and R. Miles, Nat. Neurosci. 14, 627 (2011).
[15] M. A. Kramer, W. Truccolo, U. T. Eden, K. Q. Lepage,
L. R. Hochberg, E. N. Eskandar, J. R. Madsen, J. W. Lee,
A. Maheshwari, E. Halgren, et al., Proc. Natl. Acad. Sci.
U.S.A. 109, 21116 (2012).
[16] S. H. Strogatz, Nonlinear Dynamics And Chaos: With
Applications To Physics, Biology, Chemistry, And Engi-
neering (Perseus Books Group, New York, 1994).
[17] M. Scheffer, J. Bascompte, W. A. Brock, V. Brovkin,
S. R. Carpenter, V. Dakos, H. Held, E. H. Van Nes,
M. Rietkerk, and G. Sugihara, Nature 461, 53 (2009).
9
[18] M. Breakspear, J. Roberts, J. R. Terry, S. Rodrigues,
N. Mahant, and P. Robinson, Cereb. Cortex 16, 1296
(2006).
[37] M. Goodfellow and P. Glendinning, J. Comput. Neurosci.
3, 1 (2013).
[38] A. Rothkegel and K. Lehnertz, Europhys. Lett. 95, 38001
[19] P. Suffczynski, S. Kalitzin, and F. Lopes Da Silva, Neu-
(2011).
roscience 126, 467 (2004).
[39] A. Rothkegel and K. Lehnertz, New J. Phys. 16, 055006
[20] D. Steyn-Ross and M. Steyn-Ross, Modeling Phase Tran-
(2014).
sitions in the Brain (Springer, New York, 2010).
[21] K.-E. Lee, M. A. Lopes, J. F. F. Mendes, and A. V.
Goltsev, Phys. Rev. E 89, 012701 (2014).
[22] M. Lopes, K.-E. Lee, A. Goltsev, and J. Mendes, Phys.
Rev. E 90, 052709 (2014).
[40] R. Boehringer, D. Polygalov, A. J. Huang, S. J. Mid-
dleton, V. Robert, M. E. Wintzer, R. A. Piskorowski,
V. Chevaleyre,
and T. J. McHugh, Neuron 94, 642
(2017).
[41] M. P. Richardson, J. Neurol. Neurosurg. Psychiatr. 83,
[23] A. V. Goltsev, F. V. de Abreu, S. N. Dorogovtsev, and
1238 (2012).
J. F. F. Mendes, Phys. Rev. E 81, 061921 (2010).
[42] F. Wendling, F. Bartolomei, J. Bellanger, and P. Chau-
[24] S. N. Dorogovtsev and J. F. F. Mendes, Adv. Phys. 51,
vel, Eur. J. Neurosci. 15, 1499 (2002).
1079 (2002).
[25] M. E. J. Newman, SIAM Rev. 45, 167 (2003).
[26] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and
[43] F. Wendling, P. Benquet, F. Bartolomei, and V. Jirsa,
J. Neurosci. Methods 260, 233 (2016).
[44] D. Holstein, A. Goltsev, and J. Mendes, Phys. Rev. E
D.-U. Hwang, Phys. Rep. 424, 175 (2006).
87, 032717 (2013).
[27] E. M. Izhikevich and G. M. Edelman, Proc. Natl. Acad.
[45] P. Milanowski and P. Suffczynski, Int. J. Neural Syst. 26,
Sci. U.S.A. 105, 3593 (2008).
1650053 (2016).
[28] R. A. Stefanescu, R. Shivakeshavan, and S. S. Talathi,
[46] F. L. da Silva and G. F. Harding, Epilepsy Res. 97, 278
Seizure 21, 748 (2012).
(2011).
[29] E. M. Izhikevich, Int. J. Bif. and Chaos 10, 1171 (2000).
[30] P. J. Karoly, D. R. Freestone, R. Boston, D. B. Gray-
den, D. Himes, K. Leyde, U. Seneviratne, S. Berkovic,
T. OBrien, and M. J. Cook, Brain , aww019 (2016).
[31] O. N. Markand, Semin. Neurol. 23, 7 (2003).
[32] J. Rinzel and G. B. Ermentrout (MIT Press, Cambridge,
USA, 1989) Chap. Analysis of neural excitability and os-
cillations, pp. 135 -- 169.
[47] S. N. Kalitzin, D. N. Velis, and F. H. L. da Silva, Epilepsy
Behav. 17, 310 (2010).
[48] T. Inouye, H. Sakamoto, K. Shinosaki, S. Toi,
and
S. Ukai, Electroencephalogr. Clin. Neurophysiol. 76, 205
(1990).
[49] D. Gupta, P. Ossenblok, and G. van Luijtelaar, Med.
Biol. Eng. Comput. 49, 555 (2011).
[50] E. Sitnikova and G. van Luijtelaar, Epilepsy Res. 84, 159
[33] F. L. da Silva, W. Blanes, S. N. Kalitzin, J. Parra, P. Suf-
(2009).
fczynski, and D. N. Velis, Epilepsia 44, 72 (2003).
[51] G. Van Luijtelaar, A. Hramov, E. Sitnikova, and A. Ko-
[34] Y. Wang, A. J. Trevelyan, A. Valentin, G. Alarcon,
P. N. Taylor, and M. Kaiser, PLOS Comput. Biol. 13,
e1005475 (2017).
ronovskii, Clin. Neurophysiol. 122, 687 (2011).
[52] O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilgetag,
Trends Cogn. Sci. 8, 418 (2004).
[35] F. Wendling, F. Bartolomei, J.-J. Bellanger, J. Bourien,
[53] Z. F. Mainen and T. J. Sejnowski, Science 268, 1503
and P. Chauvel, Brain 126, 1449 (2003).
(1995).
[36] M. Goodfellow, K. Schindler, and G. Baier, Neuroimage
[54] E. Schneidman, B. Freedman, and I. Segev, Neural Com-
55, 920 (2011).
put. 10, 1679 (1998).
[55] D. J. Amit and N. Brunel, Cereb. Cortex 7, 237 (1997).
|
1607.06251 | 2 | 1607 | 2016-11-07T14:46:30 | Next generation neural mass models | [
"q-bio.NC"
] | Neural mass models have been actively used since the 1970s to model the coarse grained activity of large populations of neurons and synapses. They have proven especially useful in understanding brain rhythms. However, although motivated by neurobiological considerations they are phenomenological in nature, and cannot hope to recreate some of the rich repertoire of responses seen in real neuronal tissue. In this chapter we consider the $\theta$-neuron model that has recently been shown to admit to an exact mean-field description for instantaneous pulsatile interactions. We show that the inclusion of a more realistic synapse model leads to a mean-field model that has many of the features of a neural mass model coupled to a further dynamical equation that describes the evolution of network synchrony. A bifurcation analysis is used to uncover the primary mechanism for generating oscillations at the single and two population level. Numerical simulations also show that the phenomena of event related synchronisation and desynchronisation are easily realised. Importantly unlike its phenomenological counterpart this \textit{next generation neural mass model} is an exact macroscopic description of an underlying microscopic spiking neurodynamics, and is a natural candidate for use in future large scale human brain simulations. | q-bio.NC | q-bio | Next generation neural mass models
Stephen Coombes and ´Aine Byrne
Centre for Mathematical Medicine and Biology, School of Mathematical Sciences,
University of Nottingham, University Park, Nottingham, NG7 2RD, UK.
Summary. Neural mass models have been actively used since the 1970s to model the coarse
grained activity of large populations of neurons and synapses. They have proven especially
useful in understanding brain rhythms. However, although motivated by neurobiological con-
siderations they are phenomenological in nature, and cannot hope to recreate some of the rich
repertoire of responses seen in real neuronal tissue. In this chapter we consider the θ-neuron
model that has recently been shown to admit to an exact mean-field description for instan-
taneous pulsatile interactions. We show that the inclusion of a more realistic synapse model
leads to a mean-field model that has many of the features of a neural mass model coupled to a
further dynamical equation that describes the evolution of network synchrony. A bifurcation
analysis is used to uncover the primary mechanism for generating oscillations at the sin-
gle and two population level. Numerical simulations also show that the phenomena of event
related synchronisation and desynchronisation are easily realised. Importantly unlike its phe-
nomenological counterpart this next generation neural mass model is an exact macroscopic
description of an underlying microscopic spiking neurodynamics, and is a natural candidate
for use in future large scale human brain simulations.1
1 Introduction
The term neural mass model is often used to refer to low dimensional models that aim to
describe the coarse grained activity of large populations of neurons and synapses. They are
typically cast as systems of ordinary differential equations (ODEs) and in their modern in-
carnations are exemplified by variants of the two dimensional Wilson-Cowan model [1]. This
model tracks the activity of an excitatory population of neurons coupled to an inhibitory
population. With the augmentation of such models by more realistic forms of synaptic and
network interaction they have proved especially successful in providing fits to neuroimaging
data. Historically one of the first examples in this area is the Zetterberg model [2] for the
electroencephalogram (EEG) rhythm. This is based on previous ideas developed by Lopes
da Silva and colleagues [3, 4] and is built from three interacting neural mass models, as a
minimal model of a cortical column. The first represents a population of pyramidal cells, the
second a population of excitatory interneurons, and the third a population of inhibitory in-
terneurons. Since its introduction the Zetterberg model has become more widely popularised
by the work of Jansen and Rit [5] and used to explain epileptic brain dynamics, particu-
larly by Wendling and colleagues, as recently reviewed in [6]. Another well known neural
mass model is that of Liley, which pays particular attention to the role of synaptic reversal
1Contribution to the Workshop "Nonlinear Dynamics in Computational Neuroscience:
from Physics and Biology to ICT" held in Turin (Italy) in September 2015.
6
1
0
2
v
o
N
7
]
.
C
N
o
i
b
-
q
[
2
v
1
5
2
6
0
.
7
0
6
1
:
v
i
X
r
a
2
Stephen Coombes and ´Aine Byrne
potentials, and see [7] for a discussion of this model within the context of Freeman's ideas
on the importance of chaos for cognition and perception [8]. As well as proving useful for
understanding EEG rhythms ranging from delta (1− 4 Hz) through to gamma (30− 70 Hz)
[9], neural mass models have been used to describe brain resonance phenomena [10], resting
brain state activity [11] and are very popular in the neuroimaging community. In this latter in-
stance they are often used for model driven fusion of multiple neuroimaging modalities, such
as EEG and functional magnetic resonance imaging (fMRI) [12], as well as to augment the
dynamic causal modelling framework for understanding how event-related responses result
from the dynamics of coupled neural populations [13]. Moreover, they are now an integral
part of the Virtual Brain project that aims to deliver the first open simulation of the human
brain based on individual large-scale connectivity [14], as well as play a key role in the neuro-
computational modelling of neurological and psychiatric disorders [15]. This latter work is
especially viable since neural mass models can incorporate accurate descriptions of synaptic
processing, typically in the form of a synaptic response function that is driven by firing rate
rather than by the arrival times of individual action potentials. However, it is important to
remember that at heart all neural mass models to date are essentially phenomenological, with
state variables that track coarse grained measures of the average membrane potential, popu-
lation firing rate or synaptic activity. At best they are expected to provide appropriate levels
of description for many thousands of near identical interconnected neurons with a preference
to operate in synchrony. This latter assumption is especially important for the generation of
a sufficiently strong physiological signal that can be detected non-invasively. The variation
of synchrony within a neuronal population is believed to underly the decrease or increase of
power seen in given EEG frequency bands. The former phenomenon is called event-related
desynchronisation (ERD), and the latter event-related synchronisation (ERS) [16]. Unfortu-
nately the assumption of synchrony within neural mass models means that they cannot hope
to describe ERD and ERS, at least not at the single population level. Rather this sets a natural
challenge for the next generation of neural mass models. It is precisely this issue that we take
up in this chapter.
As a starting point to move beyond the current neural mass models we draw inspiration
from the physics of self-organised networks. In particular the observation of macroscopic co-
herent states in large networks of coupled spiking neuron models has inspired a search for
equivalent low-dimensional dynamical descriptions, and see [17] for a recent review of os-
cillatory network dynamics in neuroscience. However, although the mathematical step from
microscopic to macroscopic dynamics has proved elusive for the majority of spiking mod-
els the θ-neuron model has proven amenable to such a reduction for pulsatile coupling by
Luke et al. [18]. A similar approach by Montbri´o et al. [19] has been used to reduce net-
works of quadratic integrate-and-fire neurons. Here we show how to naturally augment these
approaches to incorporate the biologically realistic forms of synaptic coupling that are com-
monly adopted within current neural mass models. In this way we arrive at the first instance
of a next generation neural mass model, with a derived (as opposed to postulated) population
firing rate that couples directly to a dynamical variable describing population synchrony. In
§2 we discuss the main elements of synaptic processing that are incorporated within standard
neural mass models, and give a heuristic description of how to close the equations of motion
in terms of population firing rates. The same model of a synapse is used in §3, though this
time driven by the spike times arising in a network of θ-neurons. For a large globally coupled
network an exact mean field description is derived, and the form of the equations compared
and contrasted with standard phenomenological neural mass models. In §4 we present a bifur-
cation analysis for the single and two population mean-field models, and use this to highlight
Next generation neural mass models
3
the primary mechanisms for generating population oscillations. Importantly we also show,
through direct numerical simulations, that the model supports ERD and ERS. Finally in §5
we reflect upon the use of such models in future large scale human brain simulations, as well
as their subsequent mathematical analysis.
2 Neural mass modelling
Neural mass models generate brain rhythms using the notion of population firing rates, aiming
to side-step the need for large scale simulations of more realistic networks of spiking neu-
rons. However, both approaches often make use of the same level of description for synaptic
processing, in a manner that we shall now clarify.
At a synapse, presynaptic firing results in the release of neurotransmitters that causes a
change in the membrane conductance of the post-synaptic neuron. This post-synaptic current
may be written I = g(vsyn − v), where v is the voltage of the post-synaptic neuron, vsyn is its
membrane reversal potential and g is a conductance. This conductance is proportional to the
probability that a synaptic receptor channel is in an open conducting state. This probability
depends on the presence and concentration of neurotransmitter released by the presynaptic
neuron. The sign of vsyn relative to the resting potential (assumed to be zero) determines
whether the synapse is excitatory (vsyn > 0) or inhibitory (vsyn < 0). The effect of some
synapses can be described with a function that fits the shape of the post-synaptic response due
to the arrival of action potential at the pre-synaptic release site. A post-synaptic conductance
change g(t) would then be given by g(t) = ks(t − T ) for t ≥ T , where T is the arrival time
of a pre-synaptic action potential, s(t) fits the shape of a realistic post-synaptic conductance,
and k is a constant. A common (normalised) choice for s(t) is the α-function:
s(t) = α2te−αtΘ (t),
(1)
where Θ is a Heaviside step function. The conductance change arising from a train of action
potentials, with firing times T m, is given by
g(t) = k∑
m
s(t − T m).
(2)
If s is the Green's function of a linear differential operator, so that Qs = δ , then we may write
(2) in the equivalent form
This is indeed the case for the choice (1) for which
δ (t − T m).
m
Qg = k∑
(cid:18)
Q =
1 +
(cid:19)2
.
1
α
d
dt
(3)
(4)
In many neural population models it is assumed that the interactions are mediated by fir-
ing rates rather than action potentials (spikes) per se. To see how this might arise we perform
a short-time average of (3) over some time-scale τ and assume that s is sufficiently slow so
that (cid:104)Qg(cid:105)t is approximately constant, where
1
(cid:104)x(cid:105)t =
τ
x(t(cid:48))dt(cid:48),
(cid:90) t
(5)
t−τ
4
Stephen Coombes and ´Aine Byrne
then we have that Qg = k f , where f is the instantaneous firing rate (number of spikes per
time ∆). For a single neuron (real or synthetic) experiencing a constant drive it is natural to
assume that this firing rate is a function of the drive alone. If for the moment we assume that
a neuron spends most of its time close to rest such that vsyn − v ≈ vsyn, and absorb a factor
vsyn into k, then for synaptically interacting neurons this drive is directly proportional to the
conductance state of the presynaptic neuron. Thus for a single population of identically and
globally coupled neurons operating synchronously we are led naturally to equations like:
Qg = κ f (g),
(6)
for some strength of coupling κ. A common choice for the population firing rate function is
the sigmoid
f (g) =
,
(7)
f0
1 + e−r(g−g0)
which saturates to f0 for large g. This functional form, with threshold g0 and steepness param-
eter r, is not derived from a biophysical model, rather it is seen as a physiologically consistent
choice. The extension to multiple interacting populations is straight forward, and the popular
Jansen-Rit model [5], provides a classic example of such a generalisation. This can be written
in the form
QEgP = κP f (gE − gI), QEgE = κE f (w1gP) + A, QIgI = κI f (w2gP),
(8)
which describes a network of interacting pyramidal neurons (P), inhibitory interneurons (I)
and excitatory interneurons (E). Here, Qa is given by (4) under the replacement α → αa
for a ∈ {E,I}, w1,2, κE,I,P are constants, and A is an external input. It has been used to
model both normal and epileptic patterns of cortical activity and its bifurcation structure has
been systematically analysed in [20, 21]. Despite its usefulness in describing certain large
scale brain rhythms, and especially alpha (8− 13 Hz), it suffers the same deficiencies as all
other neural mass models, namely it cannot track the level of synchrony within a neuronal
population.
3 θ-neuron network and reduction
The θ-neuron model or Ermentrout-Kopell canonical model is now widely known throughout
computational neuroscience as a parsimonious model for capturing the firing and response
properties of a cortical cell [22]. It is described by a purely one dimensional dynamical system
evolving on a circle according to
θ = (1− cosθ ) + (1 + cosθ )η,
θ ∈ [−π,π),
(9)
d
dt
where η represents a constant drive. For η < 0 the θ-neuron supports a pair of equilibria θ±,
with θ+ < 0 and θ− > 0, and no equilibria for η > 0. In the former case the equilibria at
θ+ is stable and the one at θ− unstable. In neurophysiological terms, the unstable fixed point
at θ− is a threshold for the neuron model. Any initial conditions with θ ∈ (θ+,θ−) will be
attracted to the stable equilibrium, while initial data with θ > θ− will make a large excursion
around the circle before returning to the rest state. For η > 0 the θ-neuron oscillates with
√
η. When η = 0 the θ-neuron is poised at a saddle-node on an invariant circle
frequency 2
(SNIC) bifurcation.
Next generation neural mass models
5
A network of θ-neurons can be described with the introduction of an index i = 1, . . . ,N
and the replacement η → ηi + Ii, where Ii describes the synaptic input current to neuron i.
For a globally coupled network this can be written in the form Ii = g(t)(vsyn − vi) for some
global conductance g and local voltage vi. As a model for the conductance we take the form
used in (3) and write
Qg(t) =
k
N
δ (t − T m
j ),
∑
m∈Z
(10)
N
∑
j=1
where T m
is the mth firing time of the jth neuron. These are defined to happen every time
j
θ j increases through π. It is well known that the θ-neuron model is formally equivalent
to a quadratic integrate-and-fire model for voltage dynamics [23] under the transformation
vi = tan(θi/2) (so that cosθi = (1 − v2
i )). This voltage
relationship allows us to write the network dynamics as
i ) and sinθi = 2vi/(1 + v2
i )/(1 + v2
θi = (1− cosθi) + (1 + cosθi)(ηi + g(t)vsyn)− g(t)sinθi,
d
dt
Qg = 2
k
N
N
∑
j=1
P(θ j).
(11)
(12)
j ) = δ (θ j(t)− π) θ j(T m
Here P(θ ) = δ (θ −π) and is periodically extended such that P(θ ) = P(θ +2π), and we have
used the result that δ (t − T m
j ). For the case that vsyn (cid:29) vi, Q = 1,
and P(θ ) has a shape of the form (1− cos(θ ))n for some positive integer n we recover the
model of Luke et al. [18]. In this case these authors have shown how to obtain an exact mean-
field reduction making use of the Ott-Antonsen (OA) ansatz. The same style of reduction has
also been used by Paz´o and Montbri´o to study pulse-coupled Winfree networks [24]. The
OA anstaz was originally used to find solutions on a reduced invariant manifold of the Ku-
ramoto model [25], and essentially assumes that the distribution of phases as N → ∞ has a
simple unimodal shape, capable of describing synchronous (peaked) and asynchronous (flat)
distributions. In the following we show how their reduction approach extends to the more bio-
logically realistic model described by (11)-(12) that includes synaptic reversal potentials and
causal non-instantaneous synaptic responses. We note that even in the limit of fast synaptic
interactions we do not recover models of the type described in [18, 24, 19] due to our focus
on conductance changes and the inclusion of voltage shunts.
3.1 Mean field reduction
In the following we shall choose the background drives ηi to be random variables drawn from
a Lorentzian distribution L(η) with
L(η) =
1
π
∆
(η − η0)2 + ∆ 2 ,
(13)
where η0 is the centre of the distribution and ∆ the width at half maximum. For the choice of
Q we shall take equation (4). In the coupled network, and if the frequencies of the individual
neurons are similar enough, then one may expect some degree of phase locking (ranging
from synchrony to asynchrony), itself controlled in part by the the time-to-peak, 1/α, of the
synaptic filter. In the limit N → ∞ the state of the network at time t can be described by a
continuous probability distribution function ρ(η,θ ,t), which satisfies the continuity equation
(arising from the conservation of oscillators):
6
Stephen Coombes and ´Aine Byrne
∂
∂t
ρ +
∂
∂θ
ρc = 0,
c = lim
N→∞
1
N
N
∑
j=1
d
dt
θ j.
(14)
The global drive to the network given by the right hand side of (12) can be constructed as
lim
N→∞
1
N
N
∑
j=1
P(θ j) =
Hence,
(cid:90) 2π
0
(cid:90) ∞
−∞
dθ
dρ(η,θ ,t)P(θ ).
c = (1− cosθ ) + (1 + cosθ )(η + gvsyn)− gsinθ ,
(cid:90) 2π
0
(cid:90) ∞
−∞
Qg =
k
π ∑
m∈Z
dθ
dρ(η,θ ,t)eim(θ−π),
(15)
(16)
(17)
(20)
(21)
where we have used the result that 2πP(θ ) = ∑m∈Z eim(θ−π). The formula for c above may
be written conveniently in terms of e±iθ as
c = f eiθ + h + f e−iθ ,
(18)
where f = ((η − 1) + vsyng + ig)/2 and h = (η + 1) + vsyng, and f denotes the complex
conjugate of f .
The OA ansatz assumes that ρ(η,θ ,t) has the product structure ρ(η,θ ,t) = L(η)F(η,θ ,t).
Since F(η,θ ,t) should be 2π periodic in θ it can be written as a Fourier series:
(cid:41)
F(η,θ ,t) =
1
2π
∞
∑
n=1
1 +
Fn(η,t)einθ + cc
,
(19)
where cc denotes complex conjugate. The insight in [26] was to restrict the Fourier coef-
ficients such that Fn(η,t) = a(η,t)n, where a(η,t) ≤ 1 to avoid divergence of the series.
There is also a further requirement that a(η,t) can be analytically continued from real η into
the complex η-plane, and that this continuation has no singularities in the lower half η-plane,
and that a(η,t) → 0 as Imη → −∞. If we now substitute (18) into the continuity equation
(14), use the OA ansatz, and balance terms in eiθ we obtain an evolution equation for a(η,t)
as
It is now convenient to introduce the Kuramoto order parameter
∂
∂t
a− ia2 f − iah− i f = 0.
(cid:90) 2π
(cid:90) ∞
Z(t) =
dθ
−∞
0
dηρ(η,θ ,t)eiθ ,
(cid:82) 2π
where Z ≤ 1. Using the OA ansatz (and using the orthogonality properties of eiθ , namely
0 eipθ eiqθ dθ = 2πδp+q,0) we then find that
Z(t) =
dηL(η)a(η,t).
(22)
By noting that the Lorentzian (13) has simple poles at η± = η0 ± i∆ the integral in (22)
may be performed by choosing a large semi-circle contour in the lower half η-plane. This
yields Z(t) = a(η−,t), giving Z(t) = a(η+,t). Hence, the dynamics for g given by (17) can
be written as Qg = κ f (Z), where κ = k/π and
−∞
(cid:40)
(cid:90) ∞
f (Z) = ∑
m∈Z
(−Z)m =
Next generation neural mass models
1−Z2
Z < 1.
1 + Z + Z +Z2 ,
7
(23)
The dynamics for Z is obtained from (20) as dZ/dt = F (Z;η0,∆ ) + G (Z,g;vsyn), where
F (Z;η0,∆ ) = −i (Z − 1)2
2
G (Z,g;vsyn) = i (Z + 1)2
+
(Z + 1)2
vsyng− Z2 − 1
2
[−∆ + iη0] ,
2
g.
2
(24)
(25)
Here we may view (24) as describing the intrinsic population dynamics and (25) the dynamics
induced by synaptic coupling. Thus the form of the mean field model is precisely that of a
neural mass model as given by equation (6). Importantly the firing rate f is a derived quantity
that is a real function of the complex Kuramoto order parameter for synchrony. This in turn
is described by a complex ODE with parameters from the underlying microscopic model.
Fig. 1. Comparison between the reduced mean field model (blue) and a simulation of a net-
work of 500 θ-neurons (red). Left: Phase plane projection for the Kuramoto order parame-
ter Z = ReiΨ . Right: Phase plane projection for the synaptic conductance g. Here η0 = 20,
∆ = 0.5, vsyn = −10 , κ = 1, α = 0.95.
To illustrate the validity of the reduction presented above we show a simulation of a
network with N = 500 θ-neurons and the mean field model in Fig. 1. Here we plot the
real and imaginary parts of Z which we write in the form ReiΨ for the mean field reduction
and calculate as Z = N−1 ∑N
j=1 eiθ j for the finite size network simulation. It is abundantly
clear that the two realisations agree very well. If the size of the population in the large scale
simulations is reduced then one can more easily see finite size fluctuations as expected.
4 Next generation neural mass model: analysis
The mean field model derived in §3.1 is a natural candidate for a next generation neural mass
model. It generalises the form of the phenomenological neural mass model whilst maintaining
8
Stephen Coombes and ´Aine Byrne
contact with biological reality in that it preserves the notion of both population rate and syn-
chrony. An almost identical model has recently been discussed by Laing [27], although here
the focus was on a first order synapse model (namely Q = (1 + α−1d/dt)) with no provision
for synaptic reversal potentials. In mathematical terms we are now faced with understanding
the dynamics of a coupled system of ODEs given by
Qg = κ f (Z),
d
dt
Z = F (Z;η0,∆ ) + G (Z,g;vsyn),
(26)
with f , F and G given by (23), (24) and (25) respectively, and Q a linear differential operator
such as given by (4). One practical way to assess the emergent behaviour of the model (26)
under parameter variation is through numerical bifurcation analysis. We now pursue this for
(26), as well as for its natural extension to cover two interacting populations.
4.1 Bifurcation diagrams
We first consider the case of a purely inhibitory population. Using XPPAUT [28] we find
that for a wide range of system parameters it is possible to find a Hopf bifurcation of a
steady state to a periodic orbit under parameter variation of η0 (controlling the mean value
of the background drive). To illustrate the relatively large volume of parameter space that can
support oscillations by this mechanism we show a two parameter continuation of the Hopf
bifurcation in (∆ ,η0) (controlling the shape of the distribution (13)) in Fig. 2, for several
values of the coupling strength κ and reversal potential vsyn. Since the OA ansatz does not
hold for ∆ = 0 (so that some degree of heterogeneity must be present) and oscillations are
only possible for η0 > 0, the bifurcation diagrams are only presented for the scenario ∆ ,η0 >
0.
Fig. 2. Two parameter continuation of a Hopf bifurcation in the single population model de-
scribed by (26) using (4) with α = 1. Left: Curves obtained for various values of κ with
vsyn = −5. Right: Curves obtained under for various values of vsyn with κ = 1. In both dia-
grams the area under the curves represents the parameter window for oscillatory behaviour.
Next generation neural mass models
9
Suppose now that we have two populations, one excitatory and one inhibitory, with re-
ciprocal connections. Introducing the labels E and I for each population then the natural
generalisation of (26) is
Qabgab = κab f (Zb),
d
dt
Za = Fa(Za) +∑
b
Gb(Za,gab),
(27)
where a,b ∈ {E,I}. Here, Qab is obtained from (4) under the replacement α → αab,
Fa(Za) = F (Za;ηa
syn). The system of equations (27)
0 ,∆ a) and Gb(Za,gab) = G (Za,gab;vab
Fig. 3. Bifurcation diagram for the reciprocally connected PING network defined by (27)
under variation of κIE, for both the excitatory (blue) and inhibitory (red) populations. Solid
lines: stable; dashed lines: unstable. Circles show maximum and minimum values of f (ZE )
and f (ZI) over one period of oscillation when no steady states are stable. The inset shows
syn = −10,
a PING rhythm for κIE = 0.65. Parameters: αEI = 0.8, αIE = 10, κEI = 0.5, vEI
vIE
syn = 10, ηE
0 = 0, ∆ E = ∆ I = 0.5, κEE = κII = 0.
0 = 10, ηI
generalises those recently presented by Laing [27] (to include reversal potentials, higher or-
der synapses and self coupling), who highlighted the ability of such networks to produce a
so-called pyramidal-interneuronal network gamma (PING) rhythm [29], as shown in the inset
of Fig. 3. Here we also show a bifurcation digram as a function of κIE, when κEE = κII = 0,
which shows that periodic behaviour can be destroyed in a supercritical Hopf bifurcation as
κIE is decreased. In Fig. 4 we show bifurcation diagrams under the variation of κEI and ηI
0.
We see that when κEI is decreased periodic behaviour can be destroyed in a supercritical
Hopf bifurcation. With an increase in ηI
0 it is also possible to generate a supercritical Hopf
bifurcation to terminate the PING rhythm.
The inclusion of self coupling leads to a wide variety of bifurcations, as seen in Fig. 5.
As ηI
0 is increased we observe the appearance of oscillatory behaviour through a supercritical
Hopf bifurcation, which is destroyed at a second supercritical Hopf bifurcation when ηI
0 is
increased further. Note the appearance/disappearance of period doubling through two period
10
Stephen Coombes and ´Aine Byrne
Fig. 4. Corresponding bifurcation diagrams to Fig. 3 under variation in κEI (left) and ηI
0
(right), for both the excitatory (blue) and inhibitory (red) populations. Note that PING
rhythms can be terminated by either decreasing the strength of coupling to the excitatory pop-
ulation from the inhibitory population or increasing the natural frequency of the inhibitory
population. Parameters as in Fig. 3 with κIE = 0.9 for both.
doubling bifurcations on this branch of periodic orbits. We also observe the appearance and
disappearance of an isola of periodic orbits through two saddle node bifurcations of peri-
odic orbits. The first saddle node occurs before the second Hopf bifurcation, i.e. there exists
two stable periodic orbits for this window of parameter space. Further increasing ηI
0 leads
to another saddle node bifurcation of periodic orbits, shortly followed by a torus bifurcation
and then a saddle-node on invariant circle bifurcation which destroys the unstable branch
of the periodic orbit. The stable branch of the periodic orbit is destroyed at a supercritical
Hopf bifurcation, as ηI
0 is increased further. Along the unstable fixed point branch there are
four Hopf bifurcations all of which either create or destroy unstable periodic behaviour. Be-
tween the second and third of these bifurcations there are two torus bifurcations, one on each
periodic orbit.
0 = −20 and ηI
The inset in Fig. 5 shows the behaviour of the system for ηI
0 = 25 respec-
tively. In both cases the excitatory population has two frequencies and the lower frequency is
synchronised to the inhibitory population. For ηI
0 = 25, the system follows the orbit created
by the isola, and there are two peaks in the (inhibitory) firing rate per period.
4.2 Event related synchronisation and desynchronisation
system), where J(t) is a smoothed rectangular pulse of the form J(t) =(cid:82) t−∞ ηD(t − s)A(s)ds,
Here we show that a time varying input to a single population model given by (26) can both
disrupt and enhance the degree of synchrony within the population. We include such a drive
under the replacement η0 → η0 + J(t) (describing a homogeneous drive to the microscopic
for A(t) = σΘ (t − T )Θ (T + τ − t) and ηD(t) is an α-function of the form (1) under the re-
placement α → αD. In Fig. 6 we show that for the inhibitory population considered in §3.1
operating in its oscillatory regime that such a drive can initially cause a population to desyn-
chronise (during the pulse) though that upon termination of the pulse the system can rebound
and generate a stronger spectral power peak than seen before the presentation of the pulse
(before relaxing back to the undriven periodic orbit). Thus, in contrast to a standard neural
mass model, the mean field model (26) is mechanistically able to support the phenomena
Next generation neural mass models
11
Fig. 5. Bifurcation diagram for ηI
0 showing the exotic behaviour of the two population model
when self coupling is reintroduced. Solid lines: stable; dashed lines: unstable; green (blue)
dotted line: stable (unstable) oscillations; orange crosses: period doubling bifurcations; red
stars: torus bifurcations. Of particular interest is the appearance/disappearance of an isola at
0 (cid:39) 15− 50. The insets shows the firing rate for the excitatory (blue) and inhibitory (red)
ηI
populations. In the upper inset ηI
0 = 25. Other parameters:
αEE = 1, αEI = 0.7, αIE = 1.4, αII = 0.4, κEE = 1.5, κEI = 2, κIE = 1, κII = 3, vEE
syn = 10,
syn = −vEI
syn = −12, ηE
vIE
0 = −20 and in the lower inset ηI
0 = 20, ∆ E = ∆ I = 0.5.
syn = 8, vII
of ERD and ERS. A more thorough discussion of this observation can be found in Byrne
et al. [30], where the model is used to describe magnetoencephalography (MEG) data for
movement related beta decrease and post-movement beta rebound [16].
5 Discussion
The desire to understand large scale brain dynamics as observed using EEG, MEG and fMRI
has prompted the increasing use of computational models [31]. Many of these approaches,
such as exemplified by the Virtual Brain project [14], make use of networks of interconnected
neural mass models. However, the inability of a single neural mass model to support the well
documented phenomena of ERS and ERD reminds us that these phenomenological models
could be improved upon. Of course, building more detailed biophysically realistic models of
neurons and their interactions would improve this state of affairs, though at a price. This being
not only computational complexity but our ability to interpret the behaviour of very high di-
mensional models in a meaningful way. The model that we have presented here is very much
in the original spirit of neural mass modelling, yet importantly it can be interpreted directly
in terms of an underlying spiking model. Moreover the derived structure of the macroscopic
equations can be viewed as a modification of the standard neural mass framework whereby
the firing rate of the system is now coupled to the degree of synchrony within a population.
Given the recent success of this model in explaining beta rebound [30], we advocate strongly
for its subsequent use in future population-level modelling approaches for understanding in
vivo brain activity states. Indeed we would like to think that this is a first example of a next
12
Stephen Coombes and ´Aine Byrne
Fig. 6. Left: Phase plane for Z = ReiΨ , demonstrating the behaviour of the system (26) in
response to a drive in the form of a smoothed rectangular pulse. The blue curve represents
the system before the pulse arrives, as it settles to its non-perturbed dynamics (t < T ),the
red curve demonstrates how the system behaves when the pulse is switched on (at t = T )
and the green how the system reacts once the drive is switched off (at t = T + τ). Right:
The corresponding spectrogram of the synaptic current demonstrating rebound (an enhanced
spectral peak on cessation of the applied pulse). Parameter values as in Fig. 1, with T = 40,
τ = 12, η0 = 21.5, αD = 6 and σ = 15.
generation neural mass model and that there will be others to follow. For phase oscillator sin-
gle neuron models, this is intrinsically linked to the mathematical challenge of generalising
the OA approach, whilst for more general conductance based neurons one might appeal to the
McKean-Vlasov-Fokker-Planck approach of Baladron et al. [32]. However, even before ex-
tending the model we have presented here to include more biological features (such as action
potential generation, dendritic processing, and stochasticity) there is still much to be done in
understanding the response of the model to input. Of particular utility would be an under-
standing of the response to periodic forcing, as this would be a precursor to understanding
patterns of phase-locking, clustering, chaos, and the multiplicity of attractors expected at the
network level. Moreover, given that neural mass models are themselves the building blocks
of neural field models [33] it would be interesting to pursue the analysis of bumps, waves and
patterns in continuum versions of the model presented here, along the lines recently devel-
oped by Laing for both synaptic and gap junction coupled systems [34, 27].
Acknowledgement. SC was supported by the European Commission through the FP7 Marie
Curie Initial Training Network 289146, NETT: Neural Engineering Transformative Tech-
nologies.
References
1. H R Wilson and J D Cowan. Excitatory and inhibitory interactions in localized popula-
tions of model neurons. Biophysical Journal, 12:1–24, 1972.
Next generation neural mass models
13
2. L H Zetterberg, L Kristiansson, and K Mossberg. Performance of a model for a local
neuron population. Biological Cybernetics, 31:15–26, 1978.
3. F H Lopes da Silva, A Hoeks, H Smits, and L H Zetterberg. Model of brain rhythmic
activity: The alpha-rhythm of the thalamus. Kybernetik, 15:27–37, 1974.
4. F H Lopes da Silva, A van Rotterdam, P Barts, E van Heusden, and W Burr. Models of
neuronal populations: the basic mechanisms of rhythmicity. Progress in Brain Research,
45:281–308, 1976.
5. B H Jansen and V G Rit. Electroencephalogram and visual evoked potential generation in
a mathematical model of coupled cortical columns. Biological Cybernetics, 73:357–366,
1995.
6. F Wendling, P Benquet, F Bartolomei, and V Jirsa. Computational models of epilepti-
form activity. Journal of Neuroscience Methods, 260:233–251, 2016.
7. M P Dafilis, F Frascoli, P J Cadusch, and D T J Liley. Chaos and generalised multista-
bility in a mesoscopic model of the electroencephalogram. Physica D, 238:1056–1060,
2009.
8. W J Freeman. Tutorial on neurobiology: From single neurons to brain chaos. Interna-
tional Journal of Bifurcation and Chaos, 2:451–482, 1992.
9. R C Sotero, N J Trujillo-Barreto, Y Iturria-Medina, F Carbonell, and J C Jimenez. Real-
istically coupled neural mass models can generate EEG rhythms. Neural Computation,
19:478–512, 2007.
10. A Spiegler, T R Knosche, K Schwab, J Haueisen, and F M Atay. Modeling brain
resonance phenomena using a neural mass model. PLoS Computational Biology,
7(12):e1002298, 2011.
11. G Deco, V K Jirsa, and A R McIntosh. Emerging concepts for the dynamical organization
of resting-state activity in the brain. Nature Reviews Neuroscience, 12:43–56, 2011.
12. P Valdes-Sosa, J M Sanchez-Bornot, R C Sotero, Y Iturria-Medina, Y Aleman-Gomez,
J Bosch-Bayard, F Carbonell, and T Ozaki. Model driven EEG/fMRI fusion of brain
oscillations. Human Brain Mapping, 30:2701–21, 2009.
13. R Moran, D A Pinotsis, and K Friston. Neural masses and fields in dynamic causal
modeling. Frontiers in Computational Neuroscience, 7(57):1–12, 2013.
14. P Sanz-Leon, S A Knock, A Spiegler, and V K Jirsa. Mathematical framework for large-
scale brain network modeling in The Virtual Brain. NeuroImage, 111:385–430, 2015.
15. B S Bhattacharya and F N Chowdhury, editors. Validating Neuro-Computational Models
of Neurological and Psychiatric Disorders. Springer, 2015.
16. G Pfurtscheller and F H Lopes da Silva. Event-related EEG/MEG synchronization and
desynchronization: basic principles. Clinical Neurophysiology, 110:1842–1857, 1999.
17. P Ashwin, S Coombes, and R Nicks. Mathematical frameworks for oscillatory network
dynamics in neuroscience. Journal of Mathematical Neuroscience, 6(2), 2016.
18. T B Luke, E Barreto, and P So. Complete classification of the macroscopic behaviour of
a heterogeneous network of theta neurons. Neural Computation, 25:3207–3234, 2013.
19. E Montbri´o, D Paz´o, and A Roxin. Macroscopic description for networks of spiking
neurons. Physical Review X, 5:021028, 2015.
20. A Spiegler, S J Kiebel, F M Atay, and T R Knosche. Bifurcation analysis of neural mass
models: Impact of extrinsic inputs and dendritic time constants. NeuroImage, 52:1041–
1058, 2010.
21. J Touboul, F Wendling, P Chauvel, and O Faugeras. Neural mass activity, bifurcations,
and epilepsy. Neural Computation, 23:3232–3286, 2011.
22. G B Ermentrout and N Kopell. Parabolic bursting in an excitable system coupled with a
slow oscillation. SIAM Journal on Applied Mathematics, 46:233–253, 1986.
14
Stephen Coombes and ´Aine Byrne
23. P E Latham, B J Richmond, P G Nelson, and S Nirenberg. Intrinsic Dynamics in Neu-
ronal Networks . I . Theory. Journal of Neurophysiology, 83:808–827, 2000.
24. D Paz´o and E Montbri´o. Low-dimensional dynamics of populations of pulse-coupled
oscillators. Physical Review X, 4:011009, 2014.
25. Y Kuramoto. Collective synchronization of pulse-coupled oscillators and excitable units.
Physica D, 50:15–30, 1991.
26. E Ott and T M Antonsen. Low dimensional behavior of large systems of globally coupled
oscillators. Chaos, 18:037113, 2008.
27. C R Laing. Computational Models of Brain and Behavior, chapter Phase oscillator net-
work models of brain dynamics. Wiley-Blackwell, 2016.
28. G B Ermentrout. Simulating, analyzing, and animating dynamical systems: A guide to
XPPAUT for researchers and students. SIAM Books, Philadelphia, 2002.
29. C Borgers and N Kopell. Synchronization in networks of excitatory and inhibitory neu-
30.
rons with sparse, random connectivity. Neural Computation, 15:509–538, 2003.
´A Byrne, M J Brookes, and S Coombes. A mean field model for movement induced
changes in the β rhythm. NeuroImage, submitted, 2016.
31. I Bojak and M Breakspear. Neuroimaging, neural population models for. In Encyclope-
dia of Computational Neuroscience, pages 1–29. Springer, 2014.
32. J Baladron, D Fasoli, O Faugeras, and J Touboul. Mean field description of and propaga-
tion of chaos in networks of Hodgkin-Huxley and FitzHugh-Nagumo neurons. Journal
of Mathematical Neuroscience, 2(10), 2012.
33. S Coombes, P beim Graben, R Potthast, and J J Wright, editors. Neural Field Theory.
Springer Verlag, 2014.
34. C R Laing. Exact neural fields incorporating gap junctions. SIAM Journal on Applied
Dynamical Systems, 14:1899–1929, 2015.
|
1610.01217 | 2 | 1610 | 2016-10-08T05:11:51 | New hallmarks of criticality in recurrent neural networks | [
"q-bio.NC"
] | A rigorous understanding of brain dynamics and function requires a conceptual bridge between multiple levels of organization, including neural spiking and network-level population activity. Mounting evidence suggests that neural networks of cerebral cortex operate at criticality. How operating near this network state impacts the variability of neuronal spiking is largely unknown. Here we show in a computational model that two prevalent features of cortical single-neuron activity, irregular spiking and the decline of response variability at stimulus onset, are both emergent properties of a recurrent network operating near criticality. Importantly, our work reveals that the relation between the irregularity of spiking and the number of input connections to a neuron, i.e., the in-degree, is maximized at criticality. Our findings establish criticality as a unifying principle for the variability of single-neuron spiking and the collective behavior of recurrent circuits in cerebral cortex. | q-bio.NC | q-bio |
New hallmarks of criticality in recurrent neural networks
Yahya Karimipanah, Zhengyu Ma and Ralf Wessel
Department of Physics, Washington University, St. Louis, Missouri 63130
A rigorous understanding of brain dynamics and function requires a conceptual bridge between
including neural spiking and network-level population activity.
multiple levels of organization,
Mounting evidence suggests that neural networks of cerebral cortex operate at criticality. How
operating near this network state impacts the variability of neuronal spiking is largely unknown.
Here we show in a computational model that two prevalent features of cortical single-neuron activ-
ity, irregular spiking and the decline of response variability at stimulus onset, are both emergent
properties of a recurrent network operating near criticality. Importantly, our work reveals that the
relation between the irregularity of spiking and the number of input connections to a neuron, i.e.,
the in-degree, is maximized at criticality. Our findings establish criticality as a unifying principle for
the variability of single-neuron spiking and the collective behavior of recurrent circuits in cerebral
cortex.
I.
Introduction
II. Results
Linking the evolutionary-derived multi-scale organiza-
tion of the brain to neural dynamics and computation
represents a major challenge in systems neuroscience [1].
Among the vast spectrum of spatial and temporal scales
of brain activity, two experimentally accessible levels of
brain organization are (i) the single-neuron spiking and
(ii) the population activity of the network in which the
neurons are embedded to various degrees. Single-neuron
spiking in cerebral cortex is characterized by statistical
properties, such as irregular spiking [2] and reduced vari-
ability during sensory stimulation [3]. Population activ-
ity is characterized by complex spatiotemporal activity,
including scale-free activity [4], which is predicted to oc-
cur for a network state near criticality [5]. These obser-
vations at two adjacent levels of brain organization raise
the question, to what extent the network state controls
the variability of single-neuron spiking?
Irregular spiking, defined as the mean coefficient of
variation (CV) being larger than one, is known as one of
the most widespread features of cortical activity in vivo
[2, 6]. Theoretical studies have linked the prevalence of
irregular spiking to the fluctuations of synaptic inputs at
the sub-threshold regime [7, 8]. Despite the popularity of
this hypothesis, here we provide an alternative scenario,
by which the onset of irregular spiking can simply emerge
at the transition between two phases of order and chaos.
Further, using a computational model we show that even
at the presence of other mechanisms for irregular spiking
(CV > 1), neuronal activity shows maximum irregular-
ity when the network resides at criticality. Moreover, in
addition to irregular spiking, we show that criticality also
gives rise to pronounced decline in neural variability after
the onset of stimulus, as well as maximum correlation be-
tween a number of single-neuron properties such as their
CV's and firing rates. Our findings provide us with ro-
bust measures of critical dynamics that at the same time
could further our understanding about the implications
of criticality for brain dynamics and function.
To explore the impact of critical dynamics on the
statistics of single-neuron spiking, we used a model net-
work consisting of excitatory binary probabilistic neurons
with sparse connectivity and external inputs (see fig.1a
and Methods). This model permits the use of the maxi-
mum eigenvalue λ of the transition probability matrix Pij
as a control parameter to tune the network near the crit-
ical point [9, 10]. Such tuning results in characteristic
avalanche size distributions for the subcritical (λ < 1),
critical (λ = 1), and supercritical (λ > 1) regime (fig.1
b-d).
To investigate the hypothesized impact of the network
state on the statistics of neuronal spiking, we simulated
the network activity for different values of λ and quanti-
fied the single-neuron spiking statistics using the CV of
the inter-spike-interval distributions (fig.1 e). The CV
is defined as the ratio of the standard deviation and the
mean of the inter-spike-interval distribution for a given
neuron. The irregular spiking is basically characterized
by CV > 1, whereas CV = 1 is considered as Poisson
spiking.
It turned out that at the presence of a con-
stant slow drive (see Methods) the CV values distributed
around a mean greater than one, thus indicating irregular
spiking. In contrast, small deviations of the network state
towards either the subcritical or the supercritical regime
resulted in CV values distributed around a mean of 1 or
less. In summary, when tuning the network from the sub-
critical to the supercritical state, the mean CV peaked
near λ = 1 (fig.1f); with increasing network size, the
peak moved closer towards λ = 1 and becomes narrower
as well. This observation suggests that, at the large-size
limit, the irregular spiking is an emergent property of re-
current networks operating at the critical point (λ = 1).
In other words, the scale-free fluctuations in network ac-
tivity at criticality translate into irregular single-neuron
spiking.
The above results could be confounded by the finite
network size.
Importantly, the width of the mean CV
as a function of λ decreased with increasing network size
2
Fig. 1: Irregular spiking emerges in a recurrent network operating at the critical point: (a) The model network consists of
binary probabilistic model neurons with sparse connectivity (black) and weak external inputs (gray) to a fraction of the neurons. (b-d)
Simulated spike trains (black raster), neuronal avalanches (gray), and corresponding avalanche size distributions (for 5 × 105 simulation
time-steps) for a network of N = 500 neurons with 10% connectivity and external inputs η = 1/(10N ) to all neurons. Simulations were
conducted for three different network states: subcritical ((b) λ = 0.9, blue), critical ((c) λ = 1.0, red), and supercritical ((d) λ = 1.1,
green). (e) Inter-spike-interval CV distributions of simulated spike trains from a network of N = 5000 neurons with 3% connectivity and
external inputs η = 1/(5N ) to all neurons. Simulations were conducted for the subcritical (λ = 0.9, blue), critical (λ = 1.02, red), and
supercritical (λ = 1.06, green) network states. (f ) The average CV as a function of the maximum eigenvalue λ of the transition probability
matrix Pij for three network sizes. Connectivity was 3% and external input was 1/(5N ) to all neurons. The curves were based on 13
values of λ within the range shown and were smoothed using Matlab spline. (g) The average population coupling and the change in mean
response as a function of λ for a network of N = 5000 neurons with 3% connectivity and external inputs of strength η = 1/(5N ) applied
to all neurons. To compute the change in mean response, we increased the external input strength by a factor of 10 half-way through the
simulation, i.e., from η = 1/(5N ) to η = 2/N .
and the peak location moved closer to λ = 1 (fig.1f).
To further test whether the deviation of the peak from
λ = 1 is indeed due to the finite size effect, we looked at
two major characteristics of criticality, namely maximum
correlations and mean response. In order to see how the
average correlation among neurons behaves in terms of
the control parameter λ, we computed a commonly-used
measures of coordinated network activity that is known
to be maximized at criticality. One such quantity is the
average population coupling, which represents a measure
of the overall level of correlated fluctuations within the
network [11] (see Methods).
Comparing the CV and the average population cou-
pling reveals that they both peak at the same point (close
to λ = 1), also similar to the change in mean response
to an increase in external input. This demonstrates that
the peak for CV coincides with the critical point, which is
characterized by maximum average population coupling
and change in mean response. Therefore, maximum CV
could be regarded as a hallmark of critical dynamics for
recurrent neural networks.
The observed onset of irregular spiking near criticality
(fig.1f) has an intuitive explanation in the extreme limits
of connection strength. In the limit of weak connections
(λ < 1), spiking is largely driven by the Poisson external
input alone, thus resulting in Poisson like spiking. On
the other hand, neurons become mostly active leading to
more regular spiking in the limit of very strong connec-
tions (λ > 1). However, it is at criticality (λ ≈ 1) that
the scale-free fluctuations in network activity translate
to single-neuron irregular spiking (hCV i > 1). Further-
more, although there can be alternative mechanisms for
irregular spiking, such as synchronous inputs (see discus-
sion), we show that the impact of the network state on
the statistics of single-neuron spiking is largely robust
with respect to structured external inputs; the irregu-
larity of spiking is always maximized near network crit-
icality even at the presence of other sources of irregular
spiking (fig.S1,fig.S2,fig.S3). The observed impact of the
network state on the statistics of single-neuron spiking (
3
Fig. 2: Solely for a recurrent network operating near criticality does the irregularity of spiking increase with a neuron's
in-degree: (a, b) The inter-spike-interval CVs from simulated spike trains versus the neuron's in degree for a model network in three
different states: (a) critical (λ = 1.02), (b) subcritical (λ = 0.9) and supercritical (λ = 1.1). Network parameters were N = 5000, 3%
connectivity, and η = 1/(5N ) applied to all neurons. (c) Correlation between CV and in-degree as a function of λ. Other network
parameters as in (a, b). The curves were smoothed using Matlab spline. The colored dots correspond to the distributions in (a, b). (d, e)
CV versus rate for three network states. Network parameters as in (a, b). (f) Correlation between CV and rate as a function of λ. Other
network parameters as in (a, b). (g, h) Population coupling versus in-degree for three network states. Network parameters as in (a, b).
(i) Correlation between population coupling and in-degree as a function of λ. Other network parameters as in (a, b).
fig.1e,f) raised the question to what extent this network-
to-neuron impact is regulated by a neuron's in-degree,
i.e., the number of input connections to a neuron. To
address this question, we took advantage of the distri-
bution of in-degrees provided by a model network with
sparse and random connectivity.
We found that near criticality (λ ≈ 1), a neuron's CV
tended to increase with increasing in-degree (fig.2a). Im-
portantly, however, this correlation between a neuron's
activity statistics (CV) and its connectivity (in-degree)
changed drastically when tuning the network state away
from criticality. In the subcritical regime only weak cor-
relation was found and in the supercritical regime the
relation reversed, i.e., a neuron's CV tended to decrease
with increasing in-degree (fig.2b). In general, when tun-
ing the network state through the critical regime, the
relation between a neuron's CV and its in-degree trans-
forms from small correlation in the subcritical regime
(λ < 1), to maximized correlation at the critical point
(λ ≈ 1), and to anti-correlation in the supercritical
regime (λ > 1), (fig.2c).
The relationship that solely near criticality is the ir-
regular spiking maximally reflective of the neuron's in-
degree, suggests a novel measure to test the criticality
hypothesis in electrophysiological experiments.
It is of
practical importance that in a network dominated by ex-
citatory neurons, a neuron's firing rate scales with its
in-degree and that this relation is independent of the
network state. Thus, spike train recordings from a pop-
ulation of neurons can be informative about the network
state: a maximum correlation between a neuron's CV
and its firing rate is indicative of a critical network state,
whereas weaker correlation or anti-correlation is indica-
tive of the subcritical or supercritical network state, re-
spectively (fig.2d-f).
In comparison, the relation between a neuron's "pop-
ulation coupling" (see Methods) and its in-degree is less
decisive about the network state, as population coupling
increases with a neuron's in degree for all three network
states (fig.2g, h). However, this relation is also dominant
4
Fig. 3: The network state determines the change in response variability after stimulus onset.
(a) Average Fano factor (solid
lines) from repeated step increases in the external input for three different network states: subcritical (λ = 0.95, blue), critical (λ = 1.02,
red), and supercritical (λ = 1.07, green). Spike trains were simulated for a network with N = 5000 neurons and 3% connectivity. External
input was applied to 10% of the neurons and external input strength switched from η = 1/N to η = 5/N as indicated. The Fano factor is
the ratio of the variance and mean in the number of spikes within a given time window and repeated trials. The sliding window was 200
time steps and the sliding increment was 20 time steps. We calculated the average Fano factor from 60 randomly sampled model neurons
and 2000 repeated trials. (b) The mean and variance of the spike rate when the average external input strength switched from 1/N (first
half) to η = 5/N (second half) for three different network states. All network parameters as in (a).
for the critical network state (fig.2i). Consistent with
these model results, recent experimental data from in
vivo recordings showed increased population coupling
with increasing synaptic inputs [11]. However, unless
the network state is manipulated and a maximum in
the correlation between the population coupling and the
in-degree is determined, that data set is not informative
about the network state.
In addition to the irregular spiking evaluated above,
a decline of response variability at stimulus onset is a
prominent cortical phenomenon [3]. The important ques-
tion that how an external input suppresses the variability
of ongoing activity has remained intriguing [12, 13]. To
address this question in the context of the network state,
we simulated spiking activity of the model network near
criticality (λ ≈ 1) in response to repeated increases in the
external input. We quantified the across-trial firing rate
variability in terms of the Fano factor, which is a mea-
sure of the across-trial variability in the number of spikes
in relation to the trial-averaged mean number of spikes
during a given window of time (see Methods). The Fano
factor has a value of approximately 1 for repeated Pois-
son spike trains. Recordings from cortical neurons show a
Fano factor above 1 for ongoing activity and a significant
drop during external stimulation [3]. Our simulations of
a model network in the critical state reproduced this drop
in Fano factor from a high value during ongoing activity
to a lower value (still above 1) during external stimulation
(fig.3a). In contrast, in the subcritical and the supercrit-
ical network state, the Fano factor remained unchanged
when switching from ongoing to evoked activity. This re-
sult demonstrates that the experimentally observed drop
in Fano factor could naturally emerge as a characteristic
feature of the critical network state. The result is quali-
tatively robust with respect to the details of the external
inputs (fig.S4).
Since the Fano factor is defined as a ratio of the vari-
ance divided by the mean, the declining Fano factor at
criticality could result trivially from an increased mean
spike count accompanied by a weak dependence of the
variance on the network state during external stimula-
tion. Evaluating the changes in the spike count and
the variance separately ruled out this possibility (fig.3b).
The "mean spike count" and the variance increased with
external stimulation for all three network states. How-
ever, only near criticality, when the network is maximally
sensitive to external stimuli, did the mean spike count in-
crease more drastically, thus yielding a reduced Fano fac-
tor.
In conclusion, our model simulation suggests that
the experimentally observed drop in Fano factor after
stimulus onset [3] can be the result of the cortical circuit
operating at criticality.
III. Discussion
A wealth of evidence indicates that neural networks
of cerebral cortex operate at criticality [4, 14, 15]. Here,
we showed that (i) irregular spiking [2, 6, 16], (ii) its
relation to the neuron's in-degree, and (iii) the decline
of response variability at stimulus onset [3], are all
emergent properties of a recurrent network operating at
criticality. A number of separate dynamical, biophysical,
and structural mechanisms have been proposed to
generate the observed irregular spiking [12, 17 -- 21]. The
significance of our work resides in part in establishing
criticality as one unifying principle for both the collective
behavior of cortical circuits and the statistics of single-
neuron spiking.
Indeed, experimental evidence for the
predicted coexistence of irregular spiking and criticality
has recently been provided in recordings of ongoing
cortical activity in vivo [6]. While the coexistence of
irregular spiking and power-law avalanche size distri-
butions has been demonstrated in more complex model
networks[22 -- 26], our work extends qualitatively beyond
these important earlier studies in three fundamental
dimensions. First, the choice of a network of excitatory
probabilistic integrate-and-fire model neurons allows the
precise analytic evaluation of the network state via a
single control parameter λ, i.e., the maximum eigenvalue
of the transfer probability matrix.
This approach
avoids the need to rely on the cumbersome and less
precise avalanche analysis to evaluate the network state.
Second, our discovery that the relation between the
irregularity of spiking and the neuron's in-degree/firing
rate is maximized at criticality provides an important
new measure of criticality. Furthermore, this relation
establishes a valuable conceptual link between criticality
and the important field of network theory, where a
node's in-degree is a basic system parameter. Third, as
the observed decline in response variability is regarded
as an essential mechanism to enhance response fidelity
to stimuli [3], our discovery of its relation to network
criticality offers a starting point toward unraveling the
possible roles of critical dynamics in neural coding. In
conclusion, it will be interesting to see to what extent the
presented findings will generate a paradigm shift in the
study of criticality of neural systems: our results build
a much-needed bridge between critical dynamics and
neural coding, and provide novel and robust measures
to test the criticality hypothesis itself.
IV. Methods
We simulated a model network consisting of excitatory bi-
nary probabilistic model neurons with sparse connectivity and
external inputs. Network size ranged from 5000 to 20000
model neurons. The strength of the connection from neuron j
to neuron i is quantified in terms of the transition probability
Pij , which is the probability that a spike in neuron j causes
a spike in neuron i in the next simulation time step. For a
network of N neurons and an average connectivity K, each
neuron is connected to N − 1 other neurons with probability
K/N . For each (on average) K(N − 1) connections among
neurons a Pij is assigned by drawing a random number from
uniform distribution in the interval [0 2
K ]. With sufficiently
5
large this yields a network with a normally distributed con-
nectivity with average K and a transition matrix Pij with
maximum (absolute value) eigenvalue of 1. In order to devi-
ate the network from the critical point we can simply multiply
Pij by a factor smaller(greater) than one. The binary state
Xi(t) of neuron i denotes whether the model neuron spikes
(Xi(t) = 1 or does not spike (Xi(t) = 0 at time t. At each
time step, the state of all neurons were updated synchronously
according to the following update rule:
Xi(t + 1) = Θ(cid:20)(cid:16)1 − ηi(t)(cid:17)Xj
Pij Xj(t) + ηi(t) − ξi(t)(cid:21) (1)
where ηi(t) of neuron i quantifies the probability of that
neuron spiking only due to external input, ξ(t) is a random
number in [01] drawn from a uniform distribution, and Θ is
the step function. In addition to the update rule, a refractory
period of two time-steps was implemented. The external input
ηi(t) was chosen to be smaller than the transition probabil-
ity Pij , which itself was small for large networks, Pij ∼ 1/N .
Because of the weak external inputs, we employed the approx-
imation 1 − η ≈ 1 in the above update rule. The maximum
eigenvalue λ of the transition probability matrix Pij describes
the network state at the thermodynamic limit: λ < 1 denotes
subcritical regime, λ ≈ 1 denotes the near critical regime and
λ > 1 denotes the supercritical regime. However, for finite-
sized networks we evaluate the exact critical point based on
the peak of average population coupling (see below). The
Pij values were drawn from a uniform distribution and then
scaled by constant to reach the desired maximum eigenvalue
λ. The constant external input ηi(t) was either constant.
The external input ηi(t) was either constant or was mod-
eled as a binary Poisson process followed by smoothing with
a Gaussian filter with a width of 20 time steps and multiplied
by an amplitude factor η0 between 0 and 1. The synchronous
input (fig.S1) was simulated with replicating a single binary
Poisson process smoothed with a Gaussian filter with a width
of 100 time-steps. In order to apply some variability among
the stimuli received by different neurons each smoothed Pois-
son process was multiplied by a factor of η0 + 0.2ǫ where ǫ
was drawn from a normal distribution.
We analyzed the simulated spike trains with respect to five
complementary statistical measures: neuronal avalanches,
coefficient of variation, population coupling, change in mean
response, and Fano factor. Following commonly-used proce-
dures [27], a neuronal avalanche was defined as an episode of
continuous (each time step) network spiking, framed by time
steps of no spikes. Avalanche size was taken as the number of
spiking neurons. The single-neuron spike train variability was
quantified using the coefficient of variation CV ≡ σisi/µisi,
defined as the ratio of the standard deviation σisi and the
mean µisi of the inter-spike-interval (ISI) distribution for a
given neuron. We managed simulation times to be sufficiently
long to ensure stable CV values. The coordination within
the network was quantified using the population coupling
, which is defined as the zero-lag cross-correlation between
the spike train Xi(t) of neuron i and the remaining network
activity Ni(t) = Pj6=i Xj (t) from all other spike trains, i.e.,
ci = D(cid:0)Xi(t)−hXi(t)i(cid:1)(cid:0)Ni(t)−hNi(t)i(cid:1)E
, where the angled brack-
ets indicate a time average [11]. Averaging the population
coupling over many neurons within a large network yields
the average population coupling , which represents a measure
of the overall level of coordination within the network. The
σX σN
change in mean response was computed as the difference in
the mean spike counts for external inputs of η = 1/(5N ) and
η = 2/N . When quantifying the spike train variability in the
context of repeated experimental situations, it is convenient
to use the Fano factor, which,
for a given time window
and repeated trials, is defined as the ratio of the variance
and mean in the number of spikes. We chose a window of
200 time-steps (sliding in increments of 20 time steps) and
repeated trials 2000 times. We calculated the average Fano
factor from 60 randomly sampled model neurons.
Acknowledgment We thank Anders Carlsson, John
Clark, and Woodrow Shew for comments on previous ver-
sions of the manuscript. This research was supported by a
Whitehall Foundation grant #20121221 (R.W.) and a NSF
CRCNS grant #1308159 (R.W.).
6
[1] H. Sompolinsky, "Computational neuroscience: beyond
the local circuit," Current opinion in neurobiology,
vol. 25, pp. xiii -- xviii, 2014.
[2] W. R. Softky and C. Koch, "The highly irregular firing
of cortical cells is inconsistent with temporal integration
of random epsps," The Journal of Neuroscience, vol. 13,
no. 1, pp. 334 -- 350, 1993.
[3] M. M. Churchland, M. Y. Byron, J. P. Cunningham, L. P.
Sugrue, M. R. Cohen, G. S. Corrado, W. T. Newsome,
A. M. Clark, P. Hosseini, B. B. Scott, et al., "Stimu-
lus onset quenches neural variability: a widespread cor-
tical phenomenon," Nature neuroscience, vol. 13, no. 3,
pp. 369 -- 378, 2010.
[4] H. G. Schuster, D. Plenz, and E. Niebur, Criticality in
neural systems. John Wiley & Sons, 2014.
[5] P. Bak, C. Tang, and K. Wiesenfeld, "Self-organized crit-
icality: An explanation of the 1/f noise," Physical review
letters, vol. 59, no. 4, p. 381, 1987.
[6] T. Bellay, A. Klaus, S. Seshadri, and D. Plenz, "Irregular
spiking of pyramidal neurons organizes as scale-invariant
neuronal avalanches in the awake state," eLife, vol. 4,
p. e07224, 2015.
[7] S. Denève and C. K. Machens, "Efficient codes and bal-
anced networks," Nature neuroscience, vol. 19, no. 3,
pp. 375 -- 382, 2016.
[8] W. Gerstner, W. M. Kistler, R. Naud, and L. Panin-
ski, Neuronal dynamics: From single neurons to networks
and models of cognition. Cambridge University Press,
2014.
[9] D. B. Larremore, W. L. Shew, and J. G. Restrepo,
"Predicting criticality and dynamic range in complex
networks: effects of topology," Physical review letters,
vol. 106, no. 5, p. 058101, 2011.
[10] D. B. Larremore, M. Y. Carpenter, E. Ott, and J. G. Re-
strepo, "Statistical properties of avalanches in networks,"
Physical Review E, vol. 85, no. 6, p. 066131, 2012.
[11] M. Okun, N. A. Steinmetz, L. Cossell, M. F. Iacaruso,
H. Ko, P. Barthó, T. Moore, S. B. Hofer, T. D. Mrsic-
Flogel, M. Carandini, et al., "Diverse coupling of neu-
rons to populations in sensory cortex," Nature, vol. 521,
no. 7553, pp. 511 -- 515, 2015.
[12] D. Sussillo and L. F. Abbott, "Generating coherent pat-
terns of activity from chaotic neural networks," Neuron,
vol. 63, no. 4, pp. 544 -- 557, 2009.
[13] K. Rajan, L. Abbott, and H. Sompolinsky, "Stimulus-
dependent suppression of chaos in recurrent neural net-
works," Physical Review E, vol. 82, no. 1, p. 011903, 2010.
[14] O. Arviv, A. Goldstein, and O. Shriki, "Near-critical dy-
namics in stimulus-evoked activity of the human brain
and its relation to spontaneous resting-state activity,"
The Journal of Neuroscience, vol. 35, no. 41, pp. 13927 --
13942, 2015.
[15] W. L. Shew, W. P. Clawson, J. Pobst, Y. Karimipanah,
N. C. Wright, and R. Wessel, "Adaptation to sensory in-
put tunes visual cortex to criticality," Nature Physics,
vol. 11, no. 8, pp. 659 -- 663, 2015.
[16] M. N. Shadlen and W. T. Newsome, "The variable dis-
charge of cortical neurons: implications for connectivity,
computation, and information coding," The Journal of
neuroscience, vol. 18, no. 10, pp. 3870 -- 3896, 1998.
[17] C. van Vreeswijk and H. Sompolinsky, "Chaos in neuronal
networks with balanced excitatory and inhibitory activ-
ity," Science, vol. 274, no. 5293, pp. 1724 -- 1726, 1996.
[18] C. F. Stevens and A. M. Zador, "Input synchrony and the
irregular firing of cortical neurons," Nature neuroscience,
vol. 1, no. 3, pp. 210 -- 217, 1998.
[19] A. Kumar, S. Rotter, and A. Aertsen, "Conditions for
propagating synchronous spiking and asynchronous fir-
ing rates in a cortical network model," The Journal of
neuroscience, vol. 28, no. 20, pp. 5268 -- 5280, 2008.
[20] A. Litwin-Kumar and B. Doiron, "Slow dynamics and
high variability in balanced cortical networks with clus-
tered connections," Nature neuroscience, vol. 15, no. 11,
pp. 1498 -- 1505, 2012.
[21] S. Ostojic, "Two types of asynchronous activity in net-
works of excitatory and inhibitory spiking neurons," Na-
ture neuroscience, vol. 17, no. 4, pp. 594 -- 600, 2014.
[22] D.-M. Chen, S. Wu, A. Guo, and Z. Yang, "Self-organized
criticality in a cellular automaton model of pulse-coupled
integrate-and-fire neurons," Journal of physics A: math-
ematical and general, vol. 28, no. 18, p. 5177, 1995.
[23] C. W. Eurich, J. M. Herrmann, and U. A. Ernst, "Finite-
size effects of avalanche dynamics," Physical review E,
vol. 66, no. 6, p. 066137, 2002.
[24] M. Benayoun, J. D. Cowan, W. van Drongelen, and
E. Wallace, "Avalanches in a stochastic model of spiking
neurons," PLoS Comput Biol, vol. 6, no. 7, p. e1000846,
2010.
[25] D. Millman, S. Mihalas, A. Kirkwood, and E. Niebur,
"Self-organized criticality occurs in non-conservative neu-
ronal networks during/up/'states," Nature physics, vol. 6,
no. 10, pp. 801 -- 805, 2010.
[26] N. Stepp, D. Plenz, and N. Srinivasa, "Synaptic plastic-
ity enables adaptive self-tuning critical networks," PLoS
Comput Biol, vol. 11, no. 1, p. e1004043, 2015.
[27] J. M. Beggs and D. Plenz, "Neuronal avalanches in neo-
cortical circuits," The Journal of neuroscience, vol. 23,
no. 35, pp. 11167 -- 11177, 2003.
[28] J. M. Beggs and N. Timme, "Being critical of criticality
in the brain," Front Physiol, vol. 3, p. 163, 2012.
Supplementary Figures
7
Fig. S1: Irregular spiking at criticality in a recurrent network with synchronous external inputs.
(a) The temporal structure
and strength of the external input η(t) to 10% of the neurons in a recurrent model network of 5000 neurons and 1% connectivity. The
external input η(t) was generated from Poisson pulses of rate 10/N , smoothed by a Gaussian filter of width 100 time-steps and amplitude
of 0.2(1 + ǫ), where ǫ is drawn from a normal distribution (see Methods). This synchronous input was added to a background constant
external input of 1/(10N ). (b) Inter-spike-interval CV distributions of simulated spike trains for the subcritical (λ = 0.95, blue), critical
(λ = 1.09, red), and supercritical (λ = 1.09, green) network state. At the critical regime the spike trains show highest irregularity, which
is indicated by the peak of the CV distribution located near 1.3. (c, d) The inter-spike-interval CVs from simulated spike trains versus
the neuron's in degree (c) and its normalized rate (d) for the three network states. (e, f ) The population coupling from simulated spike
trains versus the neuron's in degree (e) and its normalized rate (f) for the three network states. (g) The population coupling vs a neuron's
CV for the three network states. (h) The Spearman correlation coefficients between CV and in-degree (rate), population coupling and
in-degree (rate), population coupling and CV are all maximized near criticality.
8
Fig. S2:
Irregular spiking at criticality in a recurrent network with asynchronous external inputs. (a) The temporal
structure and strength of the external input ǫ to 10% of the neurons in a recurrent model network of 5000 neurons and 1% connectivity.
The external input ǫ was generated by independent Poisson pulses of rate 5/N , smoothed by a Gaussian filter of width 20 time-steps and
amplitude η0 = 0.5 (see Methods). (b) Inter-spike-interval CV distributions of simulated spike trains for the subcritical (λ = 0.95, blue),
critical (λ = 1.02, red), and supercritical (λ = 1.09, green) network state. (c, d) The inter-spike-interval CVs from simulated spike trains
versus the neuron's in degree (c) and its normalized rate (d) for the three network states. (e, f ) The population coupling from simulated
spike trains versus the neuron's in degree (e) and its normalized rate (f) for the three network states. (g) The population coupling vs a
neuron's CV for the three network states. (h) The Spearman correlation coefficients between CV and in-degree (rate), population coupling
and in-degree (rate), population coupling and CV are all maximized near criticality.
9
synchronized stimuli
synchronized stimuli
synchronized stimuli
synchronized stimuli
a
c
e
b
d
f
random stimuli
random stimuli
Fig. S3: The average CV and average population coupling are maximized near network criticality for external inputs
of different spatiotemporal structure.
(a, b) Average CV (a) and average population coupling (b) vs the control parameter ??
for synchronous external inputs (see Fig. S1a, but with different stimulation amplitudes (see fig.S1a, but with stimulation amplitude
η0 = 0.1, 0.2, 0.5; see Methods) for three different network sizes. (c, d) Average CV (c) and average population coupling (d) vs the control
parameter λ for synchronous external inputs (see fig.S1a) for a network size of N = 5000 and for three different stimulus amplitudes. (e,
f ) Average CV (e) and average population coupling (f) vs the control parameter λ for asynchronous external inputs (see fig.S2a) for three
different network sizes (see legend in (a)).
10
Fig. S4: The change in response variability for external inputs of different temporal structure. (a) Average external input
for the case of synchronous external inputs applied to 10% of the neurons as described in fig.S1a. (b) The average Fano factor computed
for 60 neurons over 2000 trials of different stimuli, but of the same type. In the presence of temporally structured stimuli, the Fano factor
can increase at sub-criticality, due to the across-trial variability in the stimuli. (c) The average Fano factor computed for the same neurons
over 2000 trials of the exactly the same stimulus (as is shown in (a). (d-f ) Same as (a-b) but for asynchronous external inputs as described
in fig.S2a.
|
1809.08045 | 3 | 1809 | 2019-08-24T12:58:19 | Stochasticity from function -- why the Bayesian brain may need no noise | [
"q-bio.NC",
"cond-mat.dis-nn",
"cs.NE",
"physics.bio-ph",
"stat.ML"
] | An increasing body of evidence suggests that the trial-to-trial variability of spiking activity in the brain is not mere noise, but rather the reflection of a sampling-based encoding scheme for probabilistic computing. Since the precise statistical properties of neural activity are important in this context, many models assume an ad-hoc source of well-behaved, explicit noise, either on the input or on the output side of single neuron dynamics, most often assuming an independent Poisson process in either case. However, these assumptions are somewhat problematic: neighboring neurons tend to share receptive fields, rendering both their input and their output correlated; at the same time, neurons are known to behave largely deterministically, as a function of their membrane potential and conductance. We suggest that spiking neural networks may, in fact, have no need for noise to perform sampling-based Bayesian inference. We study analytically the effect of auto- and cross-correlations in functionally Bayesian spiking networks and demonstrate how their effect translates to synaptic interaction strengths, rendering them controllable through synaptic plasticity. This allows even small ensembles of interconnected deterministic spiking networks to simultaneously and co-dependently shape their output activity through learning, enabling them to perform complex Bayesian computation without any need for noise, which we demonstrate in silico, both in classical simulation and in neuromorphic emulation. These results close a gap between the abstract models and the biology of functionally Bayesian spiking networks, effectively reducing the architectural constraints imposed on physical neural substrates required to perform probabilistic computing, be they biological or artificial. | q-bio.NC | q-bio | Stochasticityfromfunction --
whytheBayesianbrainmayneednonoise
DominikDolda,b,1,2,IljaBytschoka,1,AkosF.Kungla,b,AndreasBaumbacha,
OliverBreitwiesera,WalterSennb,JohannesSchemmela,KarlheinzMeiera,MihaiA.Petrovicia,b,1,2
aKirchhoff-InstituteforPhysics,HeidelbergUniversity.
bDepartmentofPhysiology,UniversityofBern.
1Authorswithequalcontributions.
2Correspondingauthors:[email protected],[email protected].
August27,2019
Abstract
Anincreasingbodyofevidencesuggeststhatthetrial-to-trialvariabilityofspikingactivityinthebrainisnotmerenoise,but
ratherthereflectionofasampling-basedencodingschemeforprobabilisticcomputing.Sincetheprecisestatisticalproperties
ofneuralactivityareimportantinthiscontext,manymodelsassumeanad-hocsourceofwell-behaved,explicitnoise,either
ontheinputorontheoutputsideofsingleneurondynamics,mostoftenassuminganindependentPoissonprocessineither
case.However,theseassumptionsaresomewhatproblematic:neighboringneuronstendtosharereceptivefields,rendering
boththeirinputandtheiroutputcorrelated;atthesametime,neuronsareknowntobehavelargelydeterministically,asa
functionoftheirmembranepotentialandconductance.Wesuggestthatspikingneuralnetworksmayhavenoneedfornoise
toperformsampling-basedBayesianinference.Westudyanalyticallytheeffectofauto-andcross-correlationsinfunctional
Bayesianspikingnetworksanddemonstratehowtheireffecttranslatestosynapticinteractionstrengths,renderingthem
controllablethroughsynapticplasticity.Thisallowsevensmallensemblesofinterconnecteddeterministicspikingnetworksto
simultaneouslyandco-dependentlyshapetheiroutputactivitythroughlearning,enablingthemtoperformcomplexBayesian
computationwithoutanyneedfornoise,whichwedemonstrateinsilico,bothinclassicalsimulationandinneuromorphic
emulation.TheseresultscloseagapbetweentheabstractmodelsandthebiologyoffunctionallyBayesianspikingnetworks,
effectively reducing the architectural constraints imposed on physical neural substrates required to perform probabilistic
computing,betheybiologicalorartificial.
1
9
1
0
2
g
u
A
4
2
]
.
C
N
o
i
b
-
q
[
3
v
5
4
0
8
0
.
9
0
8
1
:
v
i
X
r
a
1 Introduction
An ubiquitous feature of in-vivo neural responses is their
stochasticnature[1 -- 6].Theclearpresenceofthisvariabil-
ityhasspawnedmanyfunctionalinterpretations,withthe
Bayesian-brainhypothesisarguablybeingthemostnotable
example[7 -- 12].Underthisassumption,theactivityofaneu-
ral network is interpreted as representing an underlying
(prior)probabilitydistribution,withsensorydataproviding
theevidenceneededtoconstrainthisdistributiontoa(pos-
terior)shapethatmostaccuratelyrepresentsthepossible
statesoftheenvironmentgiventhelimitedavailableknowl-
edgeaboutit.
Neural network models have evolved to reproduce this
kindofneuronalresponsevariabilitybyintroducingnoise-
generatingmechanisms,betheyextrinsic,suchasPoisson
input [13 -- 16] or fluctuating currents [17 -- 22], or intrinsic,
suchasstochasticfiring[23 -- 28]ormembranefluctuations
[29,30].However, while representing, to some degree, reason-
ableapproximations,noneofthecommonlyusedsourcesof
stochasticityisfullycompatiblewithbiologicalconstraints.
Contrarytotheindependentwhitenoiseassumption,neu-
ronal inputsare bothauto- and cross-correlatedto asig-
nificant degree [31 -- 37], with obvious consequences for a
network's output statistics [38]. At the same time, the as-
sumptionofintrinsicneuronalstochasticityisatoddswith
experimentalevidenceofneuronsbeinglargelydeterminis-
ticunits[39 -- 41].Althoughsynaptictransmissionsfromin-
dividualreleasesitesarestochastic,averagedacrossmul-
tiplesites,contactsandconnections,theylargelyaverage
out[42].Therefore,itremainsaninterestingquestionhow
corticalnetworksthatusestochasticactivityasameansto
performprobabilisticinferencecanrealisticallyattainsuch
apparentrandomnessinthefirstplace.
We address this question within the normative frame-
workofsampling-basedBayesiancomputation[29,43 -- 47],
in which the spiking activity of neurons is interpreted as
Markov Chain Monte Carlo sampling from an underlying
distributionoverahigh-dimensionalbinarystatespace.In
contrasttootherworkondeterministicchaosinfunctional
spiking networks, done mostly in the context of reservoir
computing (e.g.,
[48, 49]), we provide a stringent connec-
tiontothespike-basedrepresentationandcomputationof
probabilities,aswellasthesynapticplasticityrequiredfor
learning the above. We demonstrate how an ensemble of
dynamicallyfullydeterministic,butfunctionallyprobabilis-
ticnetworks,canlearnaconnectivitypatternthatenables
probabilistic computation with a degree of precision that
matchestheoneattainablewithidealized,perfectlystochas-
ticcomponents.Thekeyelementofthisconstructionisself-
consistency,inthatallinputactivityseenbyaneuronisthe
resultofoutputactivityofotherneuronsthatfulfillafunc-
tionalroleintheirrespectivesubnetworks.Thepresentwork
supportsprobabilisticcomputationinlightofexperimental
2
(cid:88)
gsyn
k,x (t) =
synapses j
spikes s
Erev
(cid:88)
x∈{e,i}
Cm
duk
dt
= gl (El − uk) +
gsyn
k,x (Erev
wkj θ(t − ts) exp (−
x − uk) ,
t − ts
τ syn ) ,
uk(ts) ≥ ϑ ⇒ uk(t ∈ (ts, ts + τref ]) = ,
evidencefrombiologyandsuggestsaresource-efficientim-
plementationofstochasticcomputingbycompletelyremov-
ingtheneedforanyformofexplicitnoise.
2 Methods
2.1 Neuronmodelandsimulationdetails
We consider deterministic Leaky Integrate-and-Fire (LIF)
neurons with conductance-based synapses and dynamics
describedby
(1)
(cid:88)
(2)
(3)
withmembranecapacitance Cm,leakconductance gl,leak
potential El, excitatory and inhibitory reversal potentials
e/i andconductances gsyn
k,e/i,synapticstrength wkj,synap-
tictimeconstant τ synand θ(t)theHeavisidestepfunction.
For gsyn
k,e/i,thefirstsumcoversallsynapticconnectionspro-
jecting to neuron k. A neuron spikes at time ts when its
membranepotentialcrossesthethreshold ϑ,afterwhichit
becomes refractory. During the refractory period τref, the
membranepotentialisclampedtotheresetpotential .We
havechosentheabovemodelbecauseitprovidesacompu-
tationallytractableabstractionofneurosynapticdynamics
[41],butourgeneralconclusionsarenotrestrictedtothese
specificdynamics.
We further use the short-term plasticity mechanism de-
scribed in [50] to modulate synaptic interaction strengths
with an adaptive factor USE × R(t), where the time-
dependenceisgivenby1
(4)
where δ(t)istheDiracdeltafunction, ts denotesthetime
ofapresynapticspike,whichdepletesthereservoir Rbya
fraction USE,and τrecisthetimescaleonwhichthereser-
voir Rrecovers.Thisenablesabettercontrolovertheinter-
neuroninteraction,aswellasoverthemixingpropertiesof
ournetworks[47].
Backgroundinput,suchasspikesfromaPoissonsource,
entersEq.(1)assynapticinput,butwithoutshort-termplas-
ticity(asin[44])tofacilitatethemathematicalanalysis(see
Supportinginformationformoredetails).
All simulations were performed with the network spec-
ification language PyNN 0.8 [51] and the spiking neural
networksimulatorNEST2.4.2[52].
1In[50]thepostsynapticresponseonlyscaleswith R(t),whereashere
wescaleitwith USE × R(t).
1 − R
τrec − USERδ(t − ts) , USE , R ∈ [0, 1] ,
dR
dt
=
ptarget
z
τ syn
e−
τref
τ syn
τref
αWkjCm
τ syn − 1
(cid:17)−1
(cid:17)(cid:105) ,
(cid:17)
1 − τ syn
(cid:88)
− τeff
= p(z) ∝ exp(cid:0) 1
(cid:16)
(cid:104)
2.2 Samplingframework
As a model of probabilistic inference in networks of spik-
ingneurons,weadopttheframeworkintroducedin[44,46].
There,theneuronaloutputbecomesstochasticduetoahigh-
frequencybombardmentofexcitatoryandinhibitoryPoisson
stimuli(Fig.1A),elevatingneuronsintoahigh-conductance
state(HCS)[53,54],wheretheyattainahighreactionspeed
duetoareducedeffectivemembranetimeconstant.Under
theseconditions,aneuron'sresponse(oractivation)function
becomesapproximatelylogisticandcanberepresentedas
ϕ(µ) = (1 + exp (−(µ − u0)/α))−1 with inverse slope α
andinflectionpoint u0.Togetherwiththemeanfreemem-
brane potential µ and the mean effective membrane time
constant τeff (seeEqs.(17b)and(20c)),thescalingparam-
eters α and u0 are used to translate the weight matrix
W and bias vector b of a target Boltzmann distribution
2 zT W z + zT b(cid:1)withbinaryrandom
variables z ∈ {0, 1}n tosynapticweightsandleakpoten-
tialsinasamplingspikingnetwork(SSN):
(cid:16)
(cid:17)(cid:16)
(5)
(cid:16)
Erev
kj − µ
(6)
e−
τeff − 1
(cid:104)gsyn
x (cid:105)
Erev
where wkjisthesynapticweightfromneuron jtoneuron k,
gl
Elavectorcontainingtheleakpotentialsofallneurons, b
i },depend-
thecorrespondingbiasvector, Erev
ingonthenatureoftherespectivesynapse,and τm = Cm
(seeEq.(68)toEq.(73)foraderivation).Thistranslation
effectivelyenablessamplingfrom ptarget
,wherearefractory
neuron is considered to represent the state zk = 1 (see
Fig.1B,C).
2.3 Measuresofnetworkperformance
Toassesshowwellasamplingspikingnetwork(SSN)sam-
plesfromitstargetdistribution,weusetheKullback-Leibler
divergence[55]
(7)
whichisameasureforthesimilaritybetweenthesampled
distribution pnet andthetargetdistribution ptarget.Forin-
ferencetasks,wedeterminethenetwork'sclassificationrate
onasubsetoftheuseddatasetwhichwasputasidedur-
ing training. Furthermore, generative properties of SSNs
areinvestigatedeitherbylettingthenetworkcompletepar-
tiallyoccludedexamplesfromthedatasetorbylettingit
generatenewexamples.
(cid:0)pnet (cid:107) ptarget(cid:1) =
(cid:18) pnet
(αb + u0) −
kj ∈ {Erev
(cid:88)
pnet
z ln
z
ptarget
z
, Erev
e
(cid:19)
,
El =
τm
τeff
τref
,
x
τeff
x∈{e,i}
DKL
z
wkj =
gl
z
,
(cid:105)
(cid:104)
zi=1,zj =1
(cid:3) ,
∆W ij = η
ptarget
zi=1,zj =1 − pnet
zi=1 − pnet
∆bi = η(cid:2)ptarget
Figure 1: Sampling spiking networks (SSNs) with and without explicit
noise. (A) Schematic of a sampling spiking network, where each neuron
(circles) encodes a binary random variable zi ∈ {0, 1}. In the original
model, neurons were rendered effectively stochastic by adding external
Poissonsourcesofhigh-frequencybalancednoise(redboxes).(B)Aneuron
representsthestate zk = 1whenrefractoryand zk = 0otherwise.(C)
ThedynamicsofneuronsinanSSNcanbedescribedassampling(red
bars)fromatargetdistribution(bluebars).(D)InsteadofusingPoisson
processesasasourceofexplicitnoise,wereplacethePoissoninputwith
spikes coming from other networks performing spike-based probabilistic
inference by creating a sparse, asymmetric connectivity matrix between
severalSSNs.Forinstance,theredneuronreceivesnotonlyinformation-
carryingspikesfromitshomenetwork(blacklines),butalsospikesfrom
theothertwoSSNsasbackground(redarrows),andinturnprojectsback
towardsthesenetworks.Othersuchbackgroundconnectionsareindicated
inlightgray.
2.4 Learningalgorithm
Networks were trained with a Hebbian wake-sleep algo-
rithm
(8)
(9)
whichminimizesthe DKL (pnet (cid:107) ptarget)[56]. η isalearn-
ingrate(seeSupportinginformationforusedhyperparam-
eters).Forhigh-dimensionaldatasets(e.g.handwrittenlet-
tersanddigits),Boltzmannmachinesweretrainedwiththe
CASTalgorithm[57],avariantofwake-sleepwithatemper-
ing scheme, and then translated to SSN parameters with
Eqs. (5) and (6) instead of training the SSNs directly to
reducesimulationtime.
2.5 Experimentsandcalculations
Detailstoallexperimentsaswellasadditionalfiguresand
captionstovideoscanbefoundintheSupportinginforma-
tion.Detailedcalculationsarepresentedattheendofthe
maintext.
3 Results
We approach the problem of externally-induced stochas-
ticity incrementally. Throughout
the
manuscript, we discern between background input, which
isprovidedbyotherfunctionalnetworks,andexplicitnoise,
for which we use the conventional assumption of Poisson
the remainder of
zi=1
3
t [a.u.]u2u1u0101100zB000001010011100101110111z102101p(z)CADz0z1z2 spiketrains.Westartbyanalyzingtheeffectofcorrelated
backgroundontheperformanceofSSNs.Wethendemon-
strate how the effects of both auto- and cross-correlated
background can be mitigated by Hebbian plasticity. This
ultimately enables us to train a fully deterministic net-
workofnetworkstoperformdifferentinferencetaskswith-
outrequiringanyformofexplicitnoise.Thisisfirstshown
for larger ensembles of small networks, each of which re-
ceivesitsowntargetdistribution,whichallowsastraightfor-
wardquantitativeassessmentoftheirsamplingperformance
DKL (pnet (cid:107) ptarget).Westudythebehaviorofsuchensem-
blesbothincomputersimulationsandonmixed-signalneu-
romorphichardware.Finally,wedemonstratethecapability
ofourapproachfortrulyfunctional,larger-scalenetworks,
trainedonhigher-dimensionalvisualdata.
3.1 Backgroundautocorrelations
Unlike ideal Poisson sources, single spiking neurons pro-
duceautocorrelatedspiketrains,withtheshapeoftheau-
tocorrelationfunction(ACF)dependingontheirrefractory
time τref and mean spike frequency ¯r = p(z = 1)τref−1.
For higher output rates, spike trains become increasingly
dominated by bursts, i.e., sequences of equidistant spikes
with an interspike interval (ISI) of ISI ≈ τref. These fixed
structures also remain in a population, since the popula-
tionautocorrelationisequaltotheaveragedACFsofthe
individualspiketrains.
We investigated the effect of such autocorrelations on
the output statistics of SSNs by replacing the Poisson
input in the ideal model with spikes coming from other
SSNs. As opposed to Poisson noise, the autocorrelation
oftheSSN-generated
(excitatoryorinhibitory)background Sx, x ∈ {e, i}(Fig.2B)
isnon-singularandinfluencesthefreemembranepotential
C(Sx, Sx, ∆) = (cid:104)Sx(t)Sx(t+∆)(cid:105)−(cid:104)Sx(cid:105)2
(FMP)distribution(Fig.2C)andtherebyactivationfunction
(Fig. 2D) of individual sampling neurons. With increasing
firing rates (controlled by the bias of the neurons in the
backgroundSSNs),thenumberofsignificantpeaksinthe
ACFincreasesaswell(seeEq.(54)):
(10)
where ¯pistheprobabilityforabursttostart.Thisregularity
inthebackgroundinputmanifestsitselfinareducedwidth
σ(cid:48)oftheFMPdistribution(seeEq.(30))
(cid:112)
(11)
withascalingfactor √βthatdependsontheACF,whichin
turntranslatestoasteeperactivationfunction(seeEqs.(36)
and(37))
(12)
ek ln ¯pδ(cid:0)[n − k]τref
) ∼ N (µ(cid:48) = µ, σ(cid:48) =
C(Sx, Sx, nτref ) ≈
(cid:90) ∞
∞(cid:88)
(cid:1) ,
f (ufree
Var(Sx)
βσ)
k=1
i
,
f (u)du ≈ ϕ(µ)
p(zi = 1) ≈
ϑ
(cid:12)(cid:12)(cid:12)(cid:12)u(cid:48)
0=u0,α(cid:48)=√βα
withinflectionpoint u(cid:48)0andinverseslope α(cid:48).Thus,autocor-
relationsinthebackgroundinputleadtoareducedwidth
oftheFMPdistributionandhencetoasteeperactivation
functioncomparedtotheoneobtainedusinguncorrelated
Poisson input. For a better intuition, we used an approx-
imation of the activation function of LIF neurons, but the
argumentalsoholdsfortheexactexpressionderivedin[44],
asverifiedbysimulations(Fig.2D).
Apartfromtheaboveeffect,thebackgroundautocorrela-
tionsdonotaffectneuronpropertiesthatdependlinearly
on the synaptic noise input, such as the mean FMP and
the inflection point of the activation function (equivalent
tozerobias).Therefore,theeffectofthebackgroundauto-
correlations can be functionally reversed by rescaling the
functional(fromotherneuronsintheprincipalSSN)affer-
entsynapticweightsbyafactorequaltotheratiobetween
thenewandtheoriginalslope α(cid:48)/α(Eqs.(5)and(6)),as
showninFig.2E.
3.2 Backgroundcross-correlations
In addition to being autocorrelated, background input to
pairs of neurons can be cross-correlated as well, due to
either shared inputs or synaptic connections between the
neuronsthatgeneratesaidbackground.Thesebackground
cross-correlations can manifest themselves in a modified
cross-correlation between the outputs of neurons, thereby
distortingthedistributionsampledbyanSSN.
However,dependingonthenumberandnatureofpresy-
naptic background sources, background cross-correlations
maycancelouttoasignificantdegree.Thecorrelationco-
efficient(CC)oftheFMPsoftwoneuronsfedbycorrelated
noiseamountsto(seeEq.(59))
(cid:1) (13)
(cid:0)Erev
(cid:1)(cid:0)Erev
d∆ λli,mj C (Sl,i, Sm,j, ∆)(cid:101)C (κ, κ, ∆) ,
where l sums over all background spike trains Sl,i pro-
jecting to neuron i and m sums over all background
spike trains Sm,j projecting to neuron j. (cid:101)C (κ, κ, ∆) is
the unnormalized autocorrelation function of
the post-
synaptic potential
i.e., (cid:101)C (κ, κ, ∆) =
(cid:104)κ(t)κ(t + ∆)(cid:105), and C (Sl,i, Sm,j, ∆) the cross-correlation
function of
is given by
the background inputs. λli,mj
λli,mj = (cid:112)Var (Sl,i) Var (Sm,j). The background cross-
correlationisgatedintothecross-correlationofFMPsby
thenatureoftherespectivesynapticconnections:ifthetwo
neuronsconnecttothecross-correlatedinputsbysynapses
of different type (one excitatory, one inhibitory), the sign
of the CC is switched (Fig. 2F). However, individual con-
tributions to the FMP CC also depend on the difference
ofthemeanfreemembranepotentialandthereversalpo-
tentials,sothegatingofcross-correlationsisnotsymmet-
ricforexcitatoryandinhibitorysynapses.Nevertheless,if
(PSP) kernel κ,
jm − µj
il − µi
(cid:88)
ρ(ufree
wilwjm
, ufree
) ∝
(cid:90)
l,m
·
j
i
4
τref
Figure2:EffectofcorrelatedbackgroundonSSNdynamicsandcompensationthroughreparametrization.(A)FeedforwardreplacementofPoissonnoise
byspikingactivityfromotherSSNs.Inthisillustration,theprincipalSSNconsistsofthreeneuronsreceivingbackgroundinputonlyfromotherfunctional
SSNsthatsamplefromtheirownpredeterminedtargetdistribution.Forclarity,onlytwooutofatotalof[260,50,34](toptobottomin(B))background
SSNsperneuronareshownhere.Bymodifyingthebackgroundconnectivity(grayandbluearrows)theamountofcross-correlationinthebackground
inputcanbecontrolled.Atthisstage,thebackgroundSSNsstillreceivePoissoninput(redboxes).(B)Byappropriateparametrizationofthebackground
SSNs,weadjustthemeanspikefrequencyofthebackgroundneurons(blue)tostudytheeffectofbackgroundautocorrelations C(Sx, Sx, ∆).Higherfiring
probabilitiesincreasethechanceofevokingbursts,whichinducebackgroundautocorrelationsfortheneuronsintheprincipalSSNatmultiplesof τref
(darkblue:simulationresults;lightblue: ek ln ¯p with k = ∆
,seeEq.(10)).(C)BackgroundautocorrelationnarrowstheFMPdistributionofneurons
in the principal SSN: simulation (blue bars) and the theoretical prediction (Eq. (11), blue line) vs. background Poisson noise of the same rate (gray).
Background intensities correspond to (B). (D) Single-neuron activation functions corresponding to (B,C) and the theoretical prediction (Eq. (12), blue
line).Forautocorrelatednoise,theslopeoftheresponsecurvechanges,buttheinflectionpoint(with p(z = 1) = 0.5)isconserved.(E)Kullback-Leibler
(cid:0)pnet (cid:107) ptarget(cid:1)(medianandrangebetweenthefirstandthirdquartile)forthethreecasesshownin(B,C,D)aftersamplingfrom50
divergence DKL
differenttargetdistributionswith10differentrandomseedsforthe3-neuronnetworkdepictedin(A).Appropriatereparametrizationcanfullycancelout
theeffectofbackgroundautocorrelations(blue).Theaccordingresultswithoutreparametrization(gray)andwithPoissoninput(red)arealsoshown.(F)A
pairofinterconnectedneuronsinabackgroundSSNgeneratescorrelatednoise,asgivenbyEq.(13).Theeffectofcross-correlatedbackgroundonapair
oftargetneuronsdependsonthenatureofsynapticprojectionsfromthebackgroundtotheprincipalSSN.Here,wedepictthecasewheretheirinteraction
isexcitatory;theinhibitorycaseisamirrorimagethereof.Left:Ifforwardprojectionsareofthesametype,postsynapticpotentialswillbepositively
correlated.Middle:Differentsynapsetypesintheforwardprojectiononlychangethesignofthepostsynaptic potentialcorrelations.Right:Formany
backgroundinputswithmixedconnectivitypatterns,correlationscanaverageouttozeroevenwhenallinputcorrelationshavethesamesign.(G)Same
experimentasin(E),withbackgroundconnectionstatisticsadjustedtocompensateforinputcross-correlations.Theuncompensatedcasesfrom(F,left)
and(F,middle)areshowningray.(H)Correlation-cancellingreparametrizationintheprincipalSSN.Bytransformingthestatespacefrom z ∈ {0, 1}nto
z(cid:48) ∈ {−1, 1},inputcorrelationsattainthesamefunctionaleffectassynapticweights(Eq.(15));simulationresultsgivenasreddots,linearfitasredline.
Weightrescalingfollowedbyatransformationbackintothe z ∈ {0, 1}nstatespace,showningreen(whichaffectsbothweightsandbiases)cantherefore
alleviatetheeffectsofcorrelatedbackground.(I)Similarexperimentasin(E)foranetworkwithtenneurons,withparametersadjustedtocompensatefor
inputcross-correlations.Asinthecaseofautocorrelatedbackground,cross-correlationscanbecancelledoutbyappropriatereparametrization.
cialsetting,furtheranalysisisneededtoassessitscompat-
theconnectivitystatistics(in-degreeandsynapticweights)
from the background sources to an SSN are chosen ap-
ibilitywiththecorticalconnectomewithrespecttoconnec-
tivity statistics or synaptic weight distributions. However,
propriatelyandenoughpresynapticpartnersareavailable,
even if cortical architecture prevents a clean implementa-
thetotalpairwisecross-correlationbetweenneuronsinan
tionofthisdecorrelationmechanism,SSNscanthemselves
SSNcancelsoutonaverage,leavingthesamplingperfor-
manceunimpaired(Fig.2G).Notethatthiswayofreducing
compensate for residual background cross-correlations by
modifying their parameters, similar to the autocorrelation
cross-correlationsisindependentoftheunderlyingweight
distributionofthenetworksprovidingthebackground;the
compensationdiscussedabove.
required cross-wiring of functional networks could there-
Todemonstratethisability,weneedtoswitchfromthe
natural state space of neurons z ∈ {0, 1}N to the more
fore, in principle, be encoded genetically and does not
needtobelearned.Furthermore,averysimplecross-wiring
symmetricspace z(cid:48) ∈ {−1, 1}N.2Byrequiring p(z(cid:48))
rule, i.e., independently and randomly determined connec-
to conserve state probabilities (and thereby also correla-
tions,alreadysufficestoaccomplishlowbackgroundcross-
correlations and therefore reach a good sampling perfor-
2The z = 0stateforasilentneuronisarguablymorenatural,because
ithasnoeffectonitspostsynapticpartnersduringthisstate.Incontrast,
mance.Whereasthismethodisguaranteedtoworkinanartifi-
z ∈ {−1, 1}would,forexample,implyefferentexcitationuponspikingand
constantefferentinhibitionotherwise.
!
= p(z)
wpre
ij
5
exc.inh.randomAF0200400600800simulation time [ms]0.00.51.0v0p(z=1)=0.10.00.51.0(Sx,Sx,)p(z=1)=0.605101520 [ref]0.00.51.0p(z=1)=0.902004006008000200400600800656361u [mV]103102101100101100DKL(pnetptarget)wpreij>0100103105simulation time [ms]103102101100101100wpreij<00.00.10.20.3CC102101IGD0.00.20.4W000.20.40.6CCg1CC=g(W0)H103102101100101100DKL(pnetptarget)103102101100101100100103105simulation time [ms]103102101100101100C-60-56-52-48u [mV]0.00.20.40.00.20.40.00.20.4p(u)EB0.00.51.00.00.51.0p(z=1)-56-54-52-50 [mV]0.00.51.0wpreij>0ρ(ui,uj)>0wpreij>0ρ(ui,uj)<0wpreij>0ρ(ui,uj)≈0z1z2z3 1
4
1
2
1
4
i
g1
coliW .
W (cid:48) =
W and b(cid:48) =
b +
(cid:88)
ρ(Si, Sj) − g0
,
w(cid:48)ij = g−1[ρ(Si, Sj)] ≈
tions), the desired change of state variables z(cid:48) = 2z − 1
canbeachievedwithalinearparametertransformation(see
Eqs.(66)and(67)):
(14)
Inthe {−1, 1}N statespace,bothsynapticconnections w(cid:48)ij
andbackgroundcross-correlations ρ(Si, Sj)shiftprobabil-
ity mass between mixed states (zi, zj) = ±(1,−1) and
aligned states (zi, zj) = ±(1, 1) (see Supporting informa-
tion,Fig.S1).Therefore,byadjusting band W,itispos-
sible to find a W (cid:48) (Fig. 2H) that precisely conserves the
desiredcorrelationstructurebetweenneurons:
(15)
withconstants g0and g1(Fig.2I).Therefore,whenanSSN
learns a target distribution from data, background cross-
correlationsareequivalenttoanoffsetintheinitialnetwork
parametersandareautomaticallycompensatedduringtrain-
ing.Fornow,wecanconcludethattheactivityofSSNscon-
stitutesasufficientsourceofstochasticityforotherSSNs,
since all effects that follow from replacing Poisson noise
inanSSNwithfunctionaloutputfromotherSSNs(which
at this point still receive explicit noise) can be compen-
satedbyappropriateparameteradjustments.Theseareim-
portantpreliminaryconclusionsforthenextsections,where
we show how all noise can be eliminated in an ensemble
of interconnected SSNs endowed with synaptic plasticity
withoutsignificantpenaltytotheirrespectivefunctionalper-
formance.
3.3 Samplingwithoutexplicitnoiseinlargeen-
Weinitializedanensembleof1006-neuronSSNswithan
inter-networkconnectivityof = 0.1andrandomsynaptic
weights.Asopposedtothepreviousexperiments,noneofthe
neuronsintheensemblereceiveexplicitPoissoninputand
theactivityoftheensembleitselfactsasasourceofstochas-
ticityinstead,asdepictedinFig.1D.Noexternalinputis
neededtokick-startnetworkactivity,assomeneuronsspike
spontaneously,duetotherandominitializationofparame-
ters(seeFig.3A).Theexistenceofinhibitoryweightsdis-
ruptstheinitialregularity,initiatingthesamplingprocess.
Ongoinglearning(Equations(8)and(9))shapesthesam-
pleddistributionstowardstheirrespectivetargets(Fig.3B),
theparametersofwhichweredrawnrandomly(seeSupport-
inginformation).Ourensembleachievedasamplingperfor-
mance(median DKL)of 1.06+0.27
−0.40×10−3,whichissimilarto
themedianperformanceofanidealizedsetup(independent,
Poisson-drivenSSNsasin[44])of 1.05+0.15
−0.35 × 10−3(errors
aregivenbythefirstandthirdquartile).Toputtheabove
sembles
strate
DKL values in perspective, we compare the sampled and
targetdistributionsofoneoftheSSNsintheensembleat
variousstagesoflearning(Fig.3C).Thus,despitethefully
deterministic nature of the system, the network dynamics
and achieved performance after training is essentially in-
distinguishable from that of networks harnessing explicit
noisefortherepresentationofprobability.Insteadoftrain-
ingensembles,theycanalsobesetupbytranslatingthe
parametersofthetargetdistributionstoneurosynapticpa-
rametersdirectly,asdiscussedintheprevioussection(see
Supportinginformation,Fig.S2).
3.4 Implementation on a neuromorphic sub-
To test the robustness of our results, we studied an im-
plementationofnoise-freesamplingonanartificialneural
substrate,whichincorporatesunreliablecomponentsandis
thereforesignificantlymoredifficulttocontrol.Forthis,we
used the BrainScaleS system [58], a mixed-signal neuro-
morphicplatformwithanalogneurosynapticdynamicsand
digitalinter-neuroncommunication(Fig.3D,seealsoSup-
portinginformation,Fig.S3).Amajoradvantageofthisim-
plementationistheemulationspeedupof 104withrespectto
biologicalreal-time;however,forclarity,weshallcontinue
usingbiologicaltimeunitsinsteadofactualemulationtime.
The additional challenge for our neuronal ensemble is
tocopewiththenaturalvariabilityofthesubstrate,caused
mainlybyfixed-patternnoise,orwithotherlimitationssuch
as a finite weight resolution (4 bits) or spike loss, which
canallbesubstantial[59,60].Itisimportanttonotethat
theabilitytofunctionwhenembeddedinanimperfectsub-
strate with significant deviations from an idealized model
represents a necessary prerequisite for viable theories of
biologicalneuralfunction.
We emulated an ensemble of 15 4-neuron SSNs, with
an inter-SSN connectivity of = 0.2 and with randomly
drawntargetdistributions(seeSupportinginformation).The
biases were provided by additional bias neurons and ad-
justed during learning via the synaptic weights between
biasandsamplingneurons,alongwiththesynapseswithin
the SSNs, using the same learning rule as before (Equa-
tions (8) and (9)). After 200 training steps, the ensemble
−1.15 · 10−2 (errors given
reached a median DKL of 3.99+1.27
by the distance to the first and third quartile) compared
−0.55 before training (Fig. 3E). As a point of refer-
to 1.18+0.47
ence,wealsoconsideredtheidealizedcasebytrainingthe
samesetofSSNswithoutinterconnectionsandwithevery
neuronreceivingexternalPoissonnoisegeneratedfromthe
hostcomputer,reachinga DKLof 2.49+3.18
This relatively small performance loss of the noise-free
ensemblecomparedtotheidealcaseconfirmsthetheoreti-
calpredictionsandsimulationresults.Importantly,thiswas
achievedwithonlyarathersmallensemble,demonstrating
thatlargenumbersofneuronsarenotneededforrealizing
−0.71 · 10−2.
6
Figure3:Samplingwithoutexplicitnoisefromasetofpredefinedtargetdistributionsinsoftware(A-C)andonaneuromorphicsubstrate(D-G).(A)(top)
Temporalevolutionofspikingactivityinanensembleof100interconnected6-neuronSSNswithnosourceofexplicitnoise.Aninitialburstofregular
activitycausedbyneuronswithastrongenoughpositivebiasquicklytransitionstoasynchronousirregularactivityduetoinhibitorysynapses.(bottom)
Distribution of mean neuronal firing rates of the ensembles shown in (B,C) after training. (B) Median sampling quality of the above ensemble during
learning,foratestsamplingperiodof 106ms.Attheendofthelearningphase,thesamplingqualityofindividualnetworksintheensemble(blue)ison
parwiththeoneobtainedinthetheoreticallyidealcaseofindependentnetworkswithPoissonbackground(black).Errorbarsgivenover5simulationruns
withdifferentrandomseeds.(C)Illustrationofasingletargetdistribution(magenta)andcorrespondingsampleddistribution(blue)ofanetworkinthe
ensembleatseveralstagesofthelearningprocess(dashedgreenlinesin(B)).(D)PhotographofawaferfromtheBrainScaleSneuromorphicsystemused
in(E),(F)and(G)beforepost-processing(i.e.,addingadditionalstructureslikebusesontop),whichwouldmasktheunderlyingmodularstructure.Blue:
exemplarymembranetraceofananalogneuronreceivingPoissonnoise.(E)Performanceofanensembleconsistingof154-neuronSSNswithnoexternal
noiseduringlearningontheneuromorphicsubstrate,showninlightblueforeachSSNandwiththemedianshownindarkblue.Thelargefluctuations
comparedto(B)areasignatureofthenaturalvariabilityofthesubstrate'sanalogcomponents.Thedashedbluelinerepresentsthebestachievedmedian
(cid:0)pnet (cid:107) ptarget(cid:1) = 3.99 × 10−2.Forcomparison,wealsoplottheoptimalmedianperformanceforthetheoreticallyidealcaseof
performanceat DKL
(cid:0)pnet (cid:107) ptarget(cid:1) = 2.49 × 10−2 (dashedblackline).(F)Left:
independent,Poisson-drivenSSNsemulatedonthesamesubstrate,whichliesat DKL
DemonstrationofsamplingintheneuromorphicensembleofSSNsafter200trainingsteps.Individualnetworksinlightblue,medianperformanceindark
blue.Dashedblueline:medianperformancebeforetraining.Dashedblackline:medianperformanceofidealnetworks,asin(E).Right:Bestachieved
performance,after 100 sofbiotime(10 msofhardwaretime)forallSSNsintheensembledepictedasbluedots(sortedfromlowesttohighest DKL).
Forcomparison,thesameisplottedasblackcrossesfortheiridealcounterparts.(G)Sampled(blue)andtarget(magenta)distributionsoffourofthe15
SSNs.Theselectionismarkedin(F)withgreentriangles(left)andverticalgreendashedlines(right).Sincewemadenoparticularselectionofhardware
neuronsaccordingtotheirbehavior,hardwaredefectshaveasignificantimpactonasmallsubsetoftheSSNs.Despitetheseimperfections,amajorityof
SSNsperformclosetothebestvaluepermittedbythelimitedweightresolution(4bits)ofthesubstrate.
thiscomputationalparadigm.
InFig.3FandVideoS1,weshowthesamplingdynamics
ofallemulatedSSNsafterlearning.WhilemostSSNsare
able to approximate their target distributions well, some
sampled distributions are significantly skewed (Fig. 3G).
Thisiscausedbyasmallsubsetofdysfunctionalneurons,
whichwehavenotdiscardedbeforehand,inordertoavoid
animplausiblyfine-tuneduse-caseoftheneuromorphicsub-
strate. These effects become less significant in larger net-
works trained on data instead of predefined distributions,
wherelearningcannaturallycopewithsuchoutliersbyas-
signingthemsmalleroutputweights.Nevertheless,thesere-
sultsdemonstratethefeasibilityofself-sustainedBayesian
computationthroughsamplinginphysicalneuralsubstrates,
without the need for any source of explicit noise. Impor-
tantly,andincontrasttootherapproaches[61],everyneu-
ronintheensembleplaysafunctionalrole,withnoneuronal
real-estate being dedicated to the production of (pseudo-
)randomness.
3.5 EnsemblesofhierarchicalSSNs
When endowed with appropriate learning rules, hierarchi-
cal spiking networks can be efficiently trained on high-
dimensionalvisualdata[62,60,47,63 -- 65].Suchhierarchi-
calnetworksarecharacterizedbythepresenceofseveral
layers,withconnectionsbetweenconsecutivelayers,butno
lateralconnectionswithinthelayersthemselves.Whenboth
feedforwardandfeedbackconnectionsarepresent,suchnet-
worksareabletobothclassifyandgenerateimagesthatare
similartothoseusedduringtraining.
Inthesenetworks,informationprocessinginbothdirec-
tions is Bayesian in nature. Bottom-up propagation of in-
formationenablesanestimationoftheconditionalprobabil-
ityofaparticularlabeltofittheinputdata.Additionally,
top-downpropagationofneuralactivityallowsgeneratinga
subsetofpatternsinthevisiblelayerconditionedonincom-
pleteorpartiallyoccludedvisualstimulus.Whennoinput
is presented, such networks will produce patterns similar
7
100101102103training step103102101100101100DKL(pnetptarget)Bz101102103104p(z)2000 steps, DKL9.31104100101102training steps102101100DKL(pnetptarget)EA101103105biol. time [ms]102101100DKL(pnetptarget)FC00.20.40.60.81.0mean rate r [1ref]012p(r)z101102103104p(z)100 steps, DKL0.12z101102103104p(z)5 steps, DKL2.33Dz103102101p(z)DKL 0.017GzDKL 0.0360100200sim. time [ms]neuronszDKL 0.419151015networkszDKL 0.048 Figure4:Bayesianinferenceonvisualinput.(A)IllustrationoftheconnectivitybetweentwohierarchicalSSNsinthesimulatedensemble.EachSSN
hadavisiblelayerv,ahiddenhandalabellayerl.NeuronsinthesamelayerofanSSNwerenotinterconnected.EachneuroninanSSNreceived
onlyactivityfromthehiddenlayersofotherSSNsasbackground(nosourcesofexplicitnoise).(B)AnensembleoffoursuchSSNs(red)wastrainedto
performgenerativeanddiscriminativetasksonvisualdatafromtheEMNISTdataset.WeusedtheclassificationrateofrestrictedBoltzmannmachines
trainedwiththesamehyperparametersasabenchmark(blue).Errorbarsaregiven(onblue)over10testrunsand(onred)over10ensemblerealizations
withdifferentrandomseeds.(C)IllustrationofascenariowhereoneofthefourSSNs(redboxes)receivedvisualinputforclassification(B).Atthesame
time,theotherSSNscontinuouslygeneratedimagesfromtheirrespectivelearneddistributions.(D)Patterngenerationandmixingduringunconstrained
dreaming. Here, we show the activity of the visible layer of all four networks from (B), each spanning three rows. Time evolves from left to right. For
furtherillustrationsofthesamplingprocessintheensembleofhierarchicalSSNs,seeSupportinginformation,Figs.S4andS5 andVideoS2.(E)Pattern
completionandrivalryfortwoinstancesofincompletevisualstimulus.Thestimulusconsistedofthetoprightandbottomrightquadrantofthevisible
layer,respectively.Inthefirstrun,weclampedthetoparcofa"B"compatiblewitheithera"B"oran"R"(topthreerows,red),inthesecondrunwe
chosethebottomlineofan"L"compatiblewithan"L",an"E",a"Z"ora"C"(bottomthreerows,red).AnensembleofSSNsperformsBayesianinference
byimplicitlyevaluatingtheconditionaldistributionoftheunstimulatedvisibleneurons,whichmanifestsitselfhereassamplingfromallimageclasses
compatiblewiththeambiguousstimulus(seealsoSupportinginformation,Fig.S6).
−3.0%witherrorsgivenbythedistancetothefirst
of 91.5+3.6
tothoseenforcedduringtraining("dreaming").Ingeneral,
andthirdquartile,whichisclosetothe 94.0+2.1
−1.5%achieved
the exploration of a multimodal solution space in genera-
by the idealized reference setup provided by the abstract
tive models is facilitated by some noise-generating mech-
Boltzmannmachines(Fig.4B).Atthesametime,allother
anism. We demonstrate how even a small interconnected
set of hierarchical SSNs can perform these computations
SSNsremainedcapableofgeneratingrecognizableimages
(Fig. 4C). It is expected that direct training and a larger
self-sufficiently,withoutanysourceofexplicitnoise.
numberofSSNsintheensemblewouldfurtherimprovethe
Weusedanensembleoffour3-layerhierarchicalSSNs
results, but a functioning translation from the abstract to
trained on a subset of the EMNIST dataset [66], an ex-
thebiologicaldomainalreadyunderpinsthesoundnessof
tendedversionofthewidelyusedMNISTdataset[67]that
theunderlyingtheory.
includesdigitsaswellascapitalandlower-caseletters.All
Withoutvisualstimulus,allSSNssampledfreely,gener-
SSNshadthesamestructure,with784visibleunits,200
ating images similar to those on which they were trained
hidden units and 5 label units (Fig. 4A). To emulate the
(Fig.4D).Withoutanysourceofexplicitnoise,theSSNs
presenceofnetworkswithdifferentfunctionality,wetrained
were capable to mix between the relevant modes (images
eachofthemonaseparatesubsetofthedata.(Tocombine
belongingtoallclasses)oftheirrespectiveunderlyingdis-
samplinginspacewithsamplingintime,multiplenetworks
tributions,whichisahallmarkofagoodgenerativemodel.
can also be trained on the same data, see Supporting in-
Wefurtherextendedtheseresultstoanensembletrained
formationFig.S5 andVideoS2.)Sincetrainingthespik-
on the full MNIST dataset, reaching a similar generative
ingensembledirectlywascomputationallyprohibitive,we
performance for all networks (see Supporting information
trainedfourBoltzmannmachinesontherespectivedatasets
Fig.S5 andVideoS2).
andthentranslatedtheresultingparameterstoneurosynap-
TotestthepatterncompletioncapabilitiesoftheSSNs
ticparametersoftheensembleusingthemethodsdescribed
intheensemble,westimulatedthemwithincompleteand
earlierinthemanuscript(seealsoSupportinginformation,
ambiguous visual data (Fig. 4E). Under these conditions,
Fig.S2).
SSNsonlyproducedimagescompatiblewiththestimulus,
TotestthediscriminativepropertiesoftheSSNsinthe
alternating between different image classes, in a display
ensemble,onewasstimulatedwithvisualinput,whilethe
ofpatternrivalry.Asinthecaseoffreedreaming,thekey
remainingthreewerelefttofreelysamplefromtheirunder-
mechanismfacilitatingthisformofexplorationwasprovided
lyingdistribution.Wemeasuredamedianclassificationrate
8
CBEACELZ13579Aa15YW0.40.60.81.0BRTXVDvhl(visible)(hidden)(label) bythefunctionalactivityofotherneuronsintheensemble.
4 Discussion
Based on our findings, we argue that sampling-based
Bayesian computation can be implemented in fully deter-
ministic ensembles of spiking networks without requiring
anyexplicitnoise-generatingmechanism.Ourapproachhas
afirmtheoreticalfoundationinthetheoryofsamplingspik-
ing neural networks, upon which we formulate a rigorous
analysisofnetworkdynamicsandlearninginthepresence
orabsenceofnoise.
While in biology various explicit sources of noise exist
[68 -- 70],theseformsofstochasticityareeithertooweak(in
caseofionchannels)ortoohigh-dimensionalforefficient
exploration(inthecaseofstochasticsynaptictransmission,
asusedfor,e.g.,reinforcementlearning[71]).Furthermore,
arigorousmathematicalframeworkforneuralsamplingwith
stochasticsynapsesisstilllacking.Ontheotherhand,in
thecaseofpopulationcodes,neuronalpopulationnoisecan
be highly correlated, affecting information processing by,
e.g.,inducingsystematicsamplingbiases[33].
Inourproposedframework,eachnetworkinanensemble
playsadualrole:whilefulfillingitsassignedfunctionwithin
itshomesubnetwork,italsoprovidesitspeerswiththespik-
ingbackgroundnecessaryforstochasticsearchwithintheir
respective solution spaces. This enables a self-consistent
and parsimonious implementation of neural sampling, by
allowing all neurons to take on a functional role and not
dedicatinganyresourcespurelytotheproductionofback-
groundstochasticity.Theunderlyingidealiesinadapting
neuro-synapticparametersby(contrastive)Hebbianlearn-
ingtocompensateforauto-andcross-correlationsinduced
byinteractionsbetweenthefunctionalnetworksintheen-
semble. Importantly, we show that this does not rely on
thepresenceofalargenumberofindependentpresynaptic
partners for each neuron, as often assumed by models of
corticalcomputationthatusePoissonnoise(see,e.g.,[72]).
Instead,onlyasmallnumberofensemblesisnecessaryto
implementnoise-freeBayesiansampling.Thisbecomespar-
ticularlyrelevantforthedevelopmentofneuromorphicplat-
forms by eliminating the computational footprint imposed
bythegenerationanddistributionofexplicitnoise,thereby
reducingpowerconsumptionandbandwidthconstraints.
Forsimplicity,wechosenetworksofsimilarsizeinour
simulations.However,thepresentedresultsarenotcontin-
gentonnetworksizesintheensembleandlargelyindepen-
dentoftheparticularfunctionality(underlyingdistribution)
ofeachSSN.Theirapplicabilitytoscenarioswherediffer-
ent SSNs learn to represent different data is particularly
relevant for cortical computation, where weakly intercon-
nected areas or modules are responsible for distinct func-
tions [73 -- 77]. Importantly, these ideas scale naturally to
largerensemblesandlargerSSNs.Sinceeachneurononly
needsasmallnumberofpresynapticpartnersfromtheen-
9
semble,largernetworksleadtoasparserinterconnectivity
betweenSSNsintheensembleandhencesoftenstructural
constraints.Preliminarysimulationsshowthattheprinciple
ofusingfunctionaloutputasnoisecanevenbeappliedto
connectionswithinasingleSSN,eliminatingtheartificial
separationbetweennetworkandensembleconnections(see
Fig.S7andVideoS3intheSupportinginformation).
Eventhoughwehaveusedasimplifiedneuronmodelin
our simulations to reduce computation time and facilitate
the mathematical analysis, we expect the core underlying
principlestogeneralize.Thisisevidencedbyourresultson
neuromorphic hardware, where the dynamics of individual
neurons and synapses differ significantly from the mathe-
maticalmodel.Suchanabilitytocomputewithunreliable
componentsrepresentsaparticularlyappealingfeaturein
the context of both biology and emerging nanoscale tech-
nologies.
Finally,thesuggestednoise-freeBayesianbrainrecon-
cilesthedebateonspatialversustemporalsampling[78,29].
Infact,thenetworksofspikingneuronsthatprovideeach
other with virtual noise may be arranged in parallel sen-
sorystreams.Anambiguousstimuluswilltriggerdifferent
representationsoneachlevelofthesestreams,formingahi-
erarchyofprobabilisticpopulationcodes.Whilethesepop-
ulationcodeslearntocoverthefullsensorydistributionin
space,theywillalsogeneratesamplesofthesensorydistri-
butionintime(seeFig.S5intheSupportinginformation).
Attentionmayselectthemostlikelyrepresentation,while
suppressingtherepresentationsintheotherstreams.Anal-
ogously, possible actions may be represented in parallel
motor streams during planning and a motor decision may
selecttheonetobeperformed.Whenrecordinginpremotor
cortex,suchaselectioncausesanoisereduction[79],that
wesuggestiseffectivelythesignatureofchoosingthemost
probableactioninaBayesiansense.
5 Conclusion
From a generic Bayesian perspective, cortical networks
canbeviewedasgeneratorsoftargetdistributions.Toen-
ablesuchcomputation,modelsassumeneuronstopossess
sourcesofperfect,well-behavednoise -- anassumptionthat
is both impractical and at odds with biology. We showed
howlocalplasticityinanensembleofspikingnetworksal-
lowsthemtoco-shapetheiractivitytowardsasetofwell-
definedtargets,whilereciprocallyusingtheverysameac-
tivity as a source of (pseudo-)stochasticity. This enables
purelydeterministicnetworkstosimultaneouslylearnava-
rietyoftasks,completelyremovingtheneedfortruerandom-
ness. While reconciling the sampling hypothesis with the
deterministic nature of single neurons, this also offers an
efficientblueprintforin-silicoimplementationsofsampling-
basedinference.
(cid:115)
(cid:18)
6 Calculations
6.1 Freemembranepotentialdistributionwith
colorednoise
In the high-conductance state (HCS),
it can be shown
thatthetemporalevolutionofthefreemembranepotential
(FMP)ofanLIFneuronstimulatedbybalancedPoissonin-
putsisequivalenttoanOrnstein-Uhlenbeck(OU)process
withthefollowingGreen'sfunction[80]:
1
(cid:19)
.
,
,
,
k
−
θ =
µ =
νkw2
σ2 =
k∈{e,i}
k∈{e,i}
k∈{e,i}
· exp
k τ syn
νkwkErev
1
τ syn ,
1 − e−2θt
f (u, tu0) =
with
k − µ)2τ syn
(cid:104)gtot(cid:105) = gl +
(u − µ + (µ − u0)e−θt)2
(cid:104)gtot(cid:105)
k(Erev
(cid:104)gtot(cid:105)2
wkνkτ syn
2πσ2(1 − e−2θt)
1
2σ2
glEl +(cid:80)
(cid:80)
(cid:88)
(16)
(17a)
(17b)
(17c)
(17d)
where νk arethenoisefrequencies, wk thenoiseweights
usedinprevious
andwedroppedtheindexnotation ufree
sectionsforconvenience.ThestationaryFMPdistribution
isthengivenbyaGaussian[26,80]:
(cid:19)
(cid:18)
(18)
Replacingthewhitenoise η(t)intheOUprocess,defined
by (cid:104)η(cid:105) = const.and (cid:104)η(t)η(t(cid:48))(cid:105) = νδ(t − t(cid:48)) + ν2[26],with
(Gaussian)colorednoise ηc,definedby (cid:104)ηc(cid:105) = const.and
(cid:104)ηc(t)ηc(t(cid:48))(cid:105) = γ(t − t(cid:48))[81]where γ(t − t(cid:48))isafunction
thatdoesnotvanishfor t− t(cid:48) (cid:54)= 0,thestationarysolutionof
theFMPdistributionisstillgivenbyaGaussianwithmean
µ(cid:48) andwidth σ(cid:48) [81,82].Sincethenoisecorrelationsonly
appearwhencalculatinghigher-ordermomentsoftheFMP,
themeanvalueoftheFMPdistributionremainsunchanged
(cid:1)2
µ(cid:48) = µ.However,thevariance σ(cid:48)2 = (cid:104)
(cid:105)of
thestationaryFMPdistributionchangesduetothecorre-
lations,asdiscussedinthenextsection.
(cid:0)u(t) − (cid:104)u(t)(cid:105)
1
2πσ2 exp
(u − µ)2
(cid:114)
f (u) =
2σ2
−
k
.
10
tion
6.2 Widthoffreemembranepotentialdistribu-
IntheHCS,theFMPcanbeapproximatedanalyticallyas
[80]
(cid:88)
(cid:88)
(cid:19)
i
,
·
(cid:18)
(cid:19)(cid:21)
,
(cid:18)
(cid:20)
·
Λk
−
−
u(t) = u0 +
ΛkΘ(t − ts)
exp
exp
spikes s
u0 =
Λk =
(cid:104)gtot(cid:105)
− exp
− exp
u(t) = u0 +
(cid:104)τeff(cid:105) =
t − t(cid:48)
(cid:104)τeff(cid:105)
t − ts
(cid:104)τeff(cid:105)
with
k − (cid:104)τeff(cid:105)
k∈{e,i}
t − ts
−
τ syn
k
k∈{e,i}
t − t(cid:48)
−
τ syn
k
= u0 + Λe
(cid:88)
(cid:19)
(cid:0)Se ∗ κe
(cid:18)
glEl + ((cid:104)gtot(cid:105) − gl)µ
(cid:1) ,
τ syn
k wk
(cid:104)gtot(cid:105)
Cm
(cid:104)gtot(cid:105)
(19)
(20a)
(cid:0)Erev
k − µ(cid:1)
(20b)
(cid:0)τ syn
(20c)
By explicitly writing the excitatory and inhibitory noise
spike trains as Se/i(t(cid:48)) = (cid:80)
spikes s δ(t(cid:48) − ts), this can be
rewrittento
(cid:90)
(cid:18)
(cid:19)(cid:21)
(cid:20)
(cid:18)
dt(cid:48)Sk(t(cid:48))Θ(t − t(cid:48))
(21a)
(cid:0)Si ∗ κi
(cid:1)(t)
(cid:1)(t) + Λi
(21b)
= u0 +(cid:2)(ΛeSe + ΛiSi) ∗ κ(cid:3)(t) ,
(21c)
where ∗denotestheconvolutionoperatorandwith
(cid:20)
(cid:19)
(cid:18)
(cid:19)(cid:21)
. (22)
. The width of the
For simplicity, we assume τ syn
FMPdistributioncannowbecalculatedas
(cid:1)2
(cid:0)u(t) − (cid:104)u(t)(cid:105)
(23a)
(cid:2)u0 + (Stot ∗ κ)(t)(cid:3)2
(23b)
(cid:2)(Stot ∗ κ)(t)(cid:3)2
(23c)
(cid:2)(Stot ∗ κ)(t) − (cid:104)(Stot ∗ κ)(t)(cid:105)
(23d)
where the average is calculated over t and Stot(t) =
ΛeSe(t) + ΛiSi(t). Since the average is an integral over
(cid:82) T /2
−T /2(·) dt,wecanusetheidentity
t,i.e. (cid:104)(·)(cid:105) → lim
(cid:82) (f ∗ g)(t) dt = ((cid:82) f (t) dt)((cid:82) g(t) dt), that is (cid:104)f ∗ g(cid:105) =
(cid:82) g(t) dt = (cid:104)f(cid:105) ∗ ginthelimitof T → ∞,toarriveat
thefollowingsolution:
(cid:0)u(t) − (cid:104)u(t)(cid:105)
(cid:1)2
(cid:2)(Stot(t) − (cid:104)Stot(t)(cid:105)) ∗ κ(t)(cid:3)2
(24)
(cid:105) = (cid:104)u(t)2(cid:105) − (cid:104)u(t)(cid:105)2
(cid:3)2
(cid:105) − (cid:104)(Stot ∗ κ)(t)(cid:105)2
(cid:105) ,
(cid:105) − (cid:104)u0 + (Stot ∗ κ)(t)(cid:105)2
(cid:104)
= (cid:104)
= (cid:104)
= (cid:104)
e = τ syn
κe/i(t) = Θ(t)
(cid:104)
= (cid:104)
− exp
(cid:104)τeff(cid:105)
T→∞
τ syn
e/i
(cid:104)f(cid:105)
(cid:105) .
exp
−
−
t
t
1
T
.
(cid:105)
1
1
(cid:105)
lim
(cid:105) ,
(cid:104)¯u(t)¯u(t + ∆)(cid:105)
= (cid:104)
(cid:104)
= lim
T→∞
= lim
T→∞
= lim
T→∞
More generally, we obtain with a similar calculation the
autocorrelationfunction(ACF)oftheFMP:
(cid:0)( ¯Stot ∗ κ)(t)(cid:1)(cid:0)( ¯Stot ∗ κ)(t + ∆)(cid:1)
(25)
with ¯x(t) = x(t) − (cid:104)x(t)(cid:105) and by using (cid:104)¯u(t)¯u(t + ∆)(cid:105) =
(cid:104)u(t)u(t + ∆)(cid:105) − (cid:104)u(t)(cid:105)2.Thiscanbefurthersimplifiedby
applying the Wiener -- Khintchine theorem [83, 84], which
T→∞(cid:104)x(t)x(t + ∆)(cid:105)T = F−1(cid:0)
F(x)2)(cid:1)(∆)
states that
−T /2(·) dt (due to (cid:82) x(t)x(t + ∆) dt =
(cid:82) T /2
with (cid:104)(·)(cid:105)T →
(cid:0)x(t) ∗ x(−t)(cid:1)(∆)). Thus, for the limit T → ∞, we can
rewritethisas
(cid:0)( ¯Stot ∗ κ)(t)(cid:1)(cid:0)( ¯Stot ∗ κ)(t + ∆)(cid:1)
T F−1(cid:0)
F( ¯Stot ∗ κ)2(cid:1)(∆)
(26a)
F( ¯Stot)F(κ)2(cid:1)(∆)
T F−1(cid:0)
(26b)
(cid:18)
F−1(cid:0)
F( ¯Stot)2(cid:1)
F(κ)2(cid:1)(cid:19)
∗ F−1(cid:0)
(26c)
andbyapplyingtheWiener -- Khintchinetheoremagainin
reverse
(cid:104)¯u(t)¯u(t + ∆)(cid:105) =(cid:0) lim
(27)
where the variance of the FMP distribution is given for
∆ = 0. Thus, the unnormalized ACF of the FMP can be
calculatedbyconvolvingtheunnormalizedACFoftheback-
groundspiketrains(Stot)andthePSPshape(κ).Incase
ofindependentexcitatoryandinhibitoryPoissonnoise(i.e.,
(cid:104) ¯S(t) ¯S(t(cid:48))(cid:105) = νδ(t − t(cid:48))),weget
(cid:2) ¯Stot(t)(cid:3)(cid:2) ¯Stot(t + ∆(cid:48))(cid:3)
(28a)
e (cid:104) ¯Se(t) ¯Se(t + ∆(cid:48))(cid:105)
(cid:105) = Λ2
(cid:88)
(28b)
i (cid:104) ¯Si(t) ¯Si(t + ∆(cid:48))(cid:105)
+ Λ2
(cid:2) ¯Stot(t)(cid:3)(cid:2) ¯Stot(t + ∆(cid:48))(cid:3)
(cid:2)κ(t)(cid:3)(cid:2)κ(t + ∆(cid:48))(cid:3)
(cid:1)(∆) ,
1
T (cid:104)
T→∞
(cid:105)∞
(∆) ,
∗ (cid:104)
(cid:105)T
1
T
kνkδ(∆(cid:48))
Λ2
=
(cid:104)
andtherefore
k∈{e,i}
Λ2
kνkδ(∆(cid:48))
Var(u) =(cid:0) (cid:88)
(cid:2)κ(t)(cid:3)(cid:2)κ(t + ∆(cid:48))(cid:3)
(cid:88)
(cid:90) ∞
(cid:88)
Λ2
kνk (cid:104)κ2(t)(cid:105)
∗ (cid:104)
=
k∈{e,i}
k∈{e,i}
=
Λ2
kνk
κ2(t) dt ,
0
k∈{e,i}
(cid:1)(∆ = 0)
(cid:105)∞
(29a)
(29b)
(29c)
11
=
Λ2
β =
kνk
k∈{e,i}
(cid:80)
0 κ2(t) dt
σ2
colored
σ2
Poisson
(cid:82) dt κ(t)κ(t + ∆)
(cid:82) d∆ (cid:104) ¯Stot(t) ¯Stot(t + ∆)(cid:105) ·
(cid:82) ∞
whichagreeswiththeresultgivenin[80].Ifthenoisespike
trains are generated by processes with refractory periods,
theabsenceofspikesbetweenrefractoryperiodsleadsto
negativecontributionsintheACFofthenoisespiketrains.
ThisleadstoareducedvalueofthevarianceoftheFMP
andhence,alsotoareducedwidthoftheFMPdistribution.
Thefactor √βbywhichthewidthoftheFMPdistribution
(Eqn.(11))changesduetotheintroductionofcoloredback-
groundnoiseisgivenby
(30)
. (31)
ForthesimplifiedcaseofaPoissonprocesswithrefractory
period, one can show that (cid:82) d∆ (cid:104) ¯Stot(t) ¯Stot(t + ∆)(cid:105) has
areducedvaluecomparedtoaPoissonprocesswithoutre-
fractoryperiod[26],leadingto β ≤ 1.Eventhoughwedo
not show this here for neuron-generated spike trains, the
twocasesaresimilarenoughthat β ≤ 1canbeassumed
toapplyinthiscaseaswell.
Inthenextsection,wewillshowthatthefactor βcanbe
usedtorescaletheinverseslopeoftheactivationfunction
to transform the activation function of a neuron receiving
whitenoisetotheactivationfunctionofaneuronreceiving
equivalent (in frequency and weights), but colored noise.
That is, the rescaling of the FMP distribution width due
to the autocorrelated background noise translates into a
rescaledinverseslopeoftheactivationfunction.
6.3 ApproximateinverseslopeofLIFactivation
As stated earlier, the FMP of an LIF neuron in the HCS
isdescribedbyanOUprocesswithaGaussianstationary
FMP distribution (both for white and colored background
noise).Asafirstapproximation,wecandefinetheactivation
function as the probability of the neuron having a FMP
abovethreshold(seeEq.(18))
(32)
(33)
(34)
Eventhoughthisisonlyanapproximation(aswearene-
glectingtheeffectofthereset),theerrorfunctionisalready
similartothelogisticactivationfunctionobservedinsimu-
lations[44].
Theinverseslopeofalogisticactivationfunctionisde-
(cid:90) ∞
(cid:114)
(cid:18)
(cid:90) ∞
(cid:1)(cid:19)
(cid:18)
1 − erf(cid:0) ϑ − µ
function
1
2πσ2 exp
p(zi = 1) ≈
(u − µ)2
√2σ
ϑ
1
2
f (u)du
(cid:19)
2σ2
−
du
=
=
ϑ
.
,
.
α
ϕ
=
d
dµ
d
dµ
α−1 =
α−1 =
(cid:114)
p(zi = 1)
1
2πσ2 ,
finedattheinflectionpoint,i.e.,
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)µ=u0
(cid:18) µ − u0
(35)
Bycalculatingtheinverseslopeviatheactivationfunction
derivedfromtheFMPdistribution,weget
(cid:12)(cid:12)(cid:12)(cid:12)µ=ϑ
(36)
(37)
fromwhichitfollowsthattheinverseslope αisproportional
to the width of the FMP distribution σ. Thus, rescaling
the variance of the FMP distribution by a factor β leads,
approximately, to a rescaling of the inverse slope of the
activationfunction α(cid:48) = √βα.
6.4 Originofside-peaksinthenoiseautocor-
relationfunction
Forhighrates,thespiketraingeneratedbyanLIFneuron
intheHCSshowsregularpatternsofinterspikeintervals
whichareroughlyequaltotheabsoluterefractoryperiod.
Thisoccurs(i)duetotherefractoryperiodintroducingreg-
ularityforhigherrates,sinceISI's < τref arenotallowed
andthemaximumfiringrateoftheLIFneuronisbounded
by 1
,and(ii)duetoanLIFneurons'stendencytospike
consecutivelywhentheeffectivemembranepotential
(38)
(39)
issuprathresholdaftertherefractoryperiod[80].Theprob-
abilityofaconsecutivespikeaftertheneuronhasspiked
onceattime tisgivenby(undertheassumptionoftheHCS)
(40)
following an
due to the effective membrane potential
Ornstein-Uhlenbeck process (whereas the FMP is a low-
pass filter thereof, however with a very low time constant
τeff),seeEq.(16).Theprobabilitytospikeagainafterthe
secondrefractoryperiodisthengivenby
(41)
p1 = p(spike at t + τref first spike at t)
dut+τref f (ut+τref , τref ut = ϑ) ,
glEl +(cid:80)
p2 = p(spike at t + 2τref spike at t + τref , t)
τeff u = ueff − u ,
(cid:90) ∞
gsyn
k (t)Erev
k∈{e,i}
gtot(t)
ueff (t) =
τref
=
ϑ
k
,
=
ϑ du2du1 f (u2, τref u1)f (u1, τref u0 = ϑ)
ϑ du1 f (u1, τref u0 = ϑ)
(cid:82) ∞
ϑ
(cid:82) ∞
(cid:82) ∞
,
12
pn = p(spike at t + nτref spike at t + (n − 1)τref , . . . , t)
with un = ut+nτref,oringeneralafter n − 1spikes
(cid:90) ∞
(cid:82) ∞
ϑ
=
i=1
pi ,
Pn =
f n(un) =
n−1(cid:89)
(cid:82) ∞
dun−1 f (n−1)(un−1) ,
ϑ dun−1 f (n−1)(un−1)
p(burst of length n) = Pn · (1 − pn) .
ϑ dun−1 f (un, τref un−1)f (n−1)(un−1)
(42)
,(43)
for n > 1and f 1(u1) = f (u1, τref u0 = ϑ).Theprobability
toobserve nspikesinsuchasequenceisthengivenby
(44)
andtheprobabilitytofindaburstoflength n(i.e.,theburst
ends)
(45)
Withthis,onecancalculatetheaveragelengthoftheoccur-
ringbursts(cid:80)∞i=1 i· p(burst of length i),fromwhichwecan
already see how the occurrence of bursts depends on the
meanactivityoftheneuron.Asimplesolutioncanbefound
forthespecialcaseof τ syn (cid:28) τref,sincethentheeffective
membranepotentialdistributionhasalreadyconvergedto
thestationarydistributionaftereveryrefractoryperiod,i.e.,
f (un, τref un−1) = f (un)andhence
(46)
forall n.Thus,forthisspecialcasetheaverageburstlength
canbeexpressedas
∞(cid:88)
(47)
(48)
Bychangingthemeanmembranepotential(e.g.,byadjust-
ingtheleakpotentialoraddinganexternal(bias)current),
theprobabilityofconsecutivespikes ¯pcanbedirectlyad-
justedandhence,alsotheaveragelengthofbursts.Since
these bursts are fixed structures with interspike intervals
equaltotherefractoryperiod,theytranslateintoside-peaks
atmultiplesoftherefractoryperiodinthespiketrainACF,
aswedemonstratebelow.
TheACFofthespiketrain Sisgivenby
(49)
C(S, S, ∆) = (cid:104)StSt+∆(cid:105) − (cid:104)S(cid:105)2
where the first term of the numerator is (cid:104)StSt+∆(cid:105) =
(21a)
p(spike at t + ∆, spike at t) (notation as in Eqs.
pn = p(spike at t + nτref spike at t + (n − 1)τref , . . . , t)
i · p(burst of length i) =
i · ¯pi−1(1 − ¯p) ,
i=1
1
1 − ¯p
(cid:90) ∞
∞(cid:88)
duf (u) = ¯p
Var(S)
i=1
=
=
ϑ
.
,
and(23a)).Thistermcanbeexpressedas
Var(S)
a
can
with
spike
bursts
for ∆ < τref,
p(spike at t + ∆, spike at t)
= p(spike at t + ∆ spike at t) · p(spike at t) ,
= p(spike at t + ∆ spike at t) · (cid:104)S(cid:105) ,
(50)
(51)
where we assumed that the first spike starts the burst at
arandomtime t.Therefore,inordertocalculatetheACF,
wehavetocalculatetheprobabilitythataspikeoccursat
time t + ∆giventhattheneuronspikesattime t.Thishas
toincludeeverypossiblecombinationofspikesduringthis
interval.Inthefollowing,wearguethatatmultiplesofthe
refractoryperiod,themaincontributiontotheACFcomes
frombursts.
• First,
the term p(spike at t +
∆ spike at t)inEq.(50)vanishessincetheneuronis
refractoryandcannotspikeduringthisinterval.Thus,
the ACF becomes negative as only the term − (cid:104)S(cid:105)2
inEq.(49)remains,wherebothnumeratoranddenom-
inatorarepositive.
only
τref,
• For ∆ =
occur
probability
when
neuron
the
p1 = (cid:82) ∞
ϑ dut+τref f (ut+τref , τref ut = ϑ),wherewe
assumed for simplicity that the first spike starts the
burstspontaneously.
• Sincefor τref < ∆ < 2τref,theneurondidnotburst
withprobability 1 − p1,itispossibletofindaspikein
thisinterval,leadingagaintonegative,butdiminished,
valuesintheACF.
• For ∆ = 2τref,wenowhavetwowaystoobservespikes
at tand t + 2τref:(i)Thespikesarepartofaburstof
length 2 or (ii) there was no intermediate spike and
thespikeshaveanISIof 2τref.Sinceforlargerrates,
having large ISIs that are exact multiples of τref
is
unlikely,wecanneglectthecontributionof(ii).
• Ifwegofurtherto ∆ = nτref,wegetevenmoreaddi-
tionaltermsincludingburstsoflength < n.However,
these terms can again be neglected as, compared to
havingaburstoflength n,itisratherunlikelytoget
a burst pattern with missing intermediate spikes, i.e.,
havingpartialburstswhichareamultipleof τrefapart.
for ∆ → ∞, we have (cid:104)StSt+∆(cid:105) − (cid:104)S(cid:105)2 =
• Finally,
(cid:104)St(cid:105)(cid:104)St+∆(cid:105) − (cid:104)S(cid:105)2 = (cid:104)S(cid:105)(cid:104)S(cid:105) − (cid:104)S(cid:105)2 = 0andtheACF
(Eq.(49))vanishes.
Consequently,wecanapproximatetheACFatmultiples
of the refractory period by calculating the probability of
findingaburstof nspikes(Eq.(43)):
(52)
Pk+1δ(cid:0)[n − k]τref
C(S, S, nτref ) ≈
∞(cid:88)
(cid:1) ,
k=1
k=1
=
k=1
∞(cid:88)
∞(cid:88)
C(S, S, nτref ) ≈
andforthespecialcaseof τ syn (cid:28) τref
(cid:1)
¯pkδ(cid:0)[n − k]τref
(53)
ek ln ¯pδ(cid:0)[n − k]τref
(cid:1) .
(54)
Hence,sinceincreasingthemeanrate(orbias)oftheneu-
ronleadstoanincreasein ¯pandthustoareduceddecay
constant ln ¯p,moresignificantside-peaksemerge.
For τ syn ≈ τref, the effective membrane distribution is
not yet stationary and therefore, this approximation does
not hold. To arrive at the exact solution, one would have
torepeattheabovecalculationforallpossiblespiketime
combinations,leadingtoarecursiveintegral[26].Further-
more, one would also need to take into account the sit-
uation where the first spike is itself part of a burst, i.e.,
is not the first spike in the burst. To circumvent a more
tedious calculation, we use an approximation which is in
between the two cases τ syn (cid:28) τref and τ syn ≈ τref: we
use ¯p =(cid:82) ∞
ϑ duf (u, τref ϑ),whichprovidesareasonable
approximationforshortbursts.
6.5 Cross-correlationoffreemembranepoten-
tialsreceivingcorrelatedinput
Similarly to the ACF of the membrane potential, one can
calculatethecrosscorrelationfunctionoftheFMPsoftwo
neurons receiving correlated noise input. First, the mem-
branepotentialsaregivenby
(55)
(56)
Thecovariancefunctioncanbewrittenas
0 + Stot,1 ∗ κ ,
0 + Stot,2 ∗ κ .
u1 = u1
u2 = u2
(cid:104)¯u1(t)¯u2(t + ∆)(cid:105)
= (cid:104)u1(t)u2(t + ∆)(cid:105) − (cid:104)u1(t)(cid:105)(cid:104)u2(t)(cid:105)
= (cid:104)
− (cid:104)
= ...
(cid:0)Stot,1 ∗ κ(cid:1)(t)(cid:0)Stot,2 ∗ κ(cid:1)(t + ∆)(cid:105)
(cid:0)Stot,1 ∗ κ(cid:1)(t)(cid:105)(cid:104)
(cid:0)Stot,2 ∗ κ(cid:1)(t)(cid:105)
=(cid:0) lim
(cid:1)(∆) ,
(57a)
(57b)
(57c)
with ¯u = u − (cid:104)u(cid:105), from which we obtain the crosscorre-
lation function by normalizing with the product of stan-
dard deviations of u1 and u2 (for notation, see Eq. (27)).
The term containing the input correlation coefficient is
(cid:104) ¯Stot,1(t) ¯Stot,2(t + ∆(cid:48))(cid:105). Plugging in the spike trains, we
1
T (cid:104) ¯Stot,1(t) ¯Stot,2(t + ∆(cid:48))(cid:105)T
∗ (cid:104)κ(t)κ(t + ∆(cid:48))(cid:105)∞
T→∞
13
getfourcrosscorrelationterms
(cid:88)
·
i(cid:105)
i(cid:105)
(cid:1)
(cid:20)
(∆) ,
(cid:104)gtot
wlwm
l,m∈{e,i}
l,m∈{e,i}
m − µ2
l − µ1
(cid:104)¯u1(t)¯u2(t + ∆)(cid:105)
= ζ1ζ2
(cid:0)Erev
(cid:1)(cid:0)Erev
(cid:0)τ syn − (cid:104)τeff
Λl,1Λm,2 (cid:104) ¯Sl,1(t) ¯Sm,2(t + ∆(cid:48))(cid:105) .
(cid:104) ¯Stot,1(t) ¯Stot,2(t + ∆(cid:48))(cid:105)
=
(58)
Sinceexcitatoryaswellasinhibitorynoiseinputsareran-
domly drawn from the same pool of neurons, we can as-
sumethat (cid:104) ¯Sl,1(t) ¯Sm,2(t + ∆(cid:48))(cid:105)isapproximatelyequalfor
all combinations of synapse types when averaging over
enough inputs, regardless of the underlying correlation
structure/distributionofthenoisepool.Thefirstterm,how-
ever, depends on the synapse types since the Λ-terms
(Eq.(20b))containthedistancebetweenreversalpotentials
andmeanFMP:
(cid:88)
(cid:104) ¯Sl,1(t) ¯Sm,2(t + ∆(cid:48))(cid:105) ∗ (cid:104)κ(t)κ(t + ∆(cid:48))(cid:105)
(cid:104)Λe,1Λe,2(cid:105)inputs = −(cid:104)Λe,1Λi,2(cid:105)inputs ,
(cid:104)Λi,1Λi,2(cid:105)inputs = −(cid:104)Λi,1Λe,2(cid:105)inputs ,
(cid:21)
(59)
(cid:1). The cross-
with constants ζi = τ syn
correlationvanisheswhen,aftersummingovermanyinputs,
thefollowingidentitieshold:
(60a)
(60b)
where (cid:104)(·)(cid:105) is an average over all inputs, i.e., all neurons
thatprovidenoise.
Whilenotrelevantforoursimulations,itisworthnoting
thattheexcitatoryandinhibitoryweightswithwhicheach
neuroncontributesitsspiketrainscanberandomlydrawn
fromnon-identicaldistributions.Byenforcingthefollowing
correlationbetweenthenoiseweightsofbothneurons,one
canintroduceaskewintotheweightdistributionwhichcom-
pensatesforthedifferingdistancetothereversalpotentials:
(61)
A simple procedure to accomplish this is the follow-
ing:First,wedrawtheabsoluteweights w1 and w2 from
an arbitrary distribution and assign synapse types ran-
domlywithprobabilities pe/iafterwards.If w2isexcitatory,
we multiply w1 by
, otherwise by
.Thisway, (cid:104)w1(cid:105)remainsunchanged
andtheresultingweightssufficeEq.(61).
6.6 Statespaceswitchfrom{0,1}to{-1,1}
Toswitchfromthestatespace z ∈ {0, 1}to z(cid:48) ∈ {−1, 1}
whileconservingthestateprobabilities(i.e., p(z) = p(z(cid:48)))
rev − µ1)(Ee,2
(Ee,1
= −(Ee,1
e w2
rev − µ2)(cid:104)w1
rev − µ1)(Ei,2
rev − µ2)(cid:104)w1
e(cid:105)inputs
e w2
rev−µ2+piEe,2
rev−µ2+piEe,2
i (cid:105)inputs
peEi,2
peEi,2
rev−µ2
rev−µ2
rev−µ2
rev−µ2
Ee,2
Ei,2
14
i
i
j
i
i
2
4
i,j
i,j
i,j
i,j
i,j
i,j
=
+
=
zi
1
2
1
2
1
2
−
− 2
E(z(cid:48)) =
(cid:88)
(cid:88)
onehastoadequatelytransformthedistributionparameters
W and b. Since the distributions are of the form p(z) =
exp(cid:0)zTWz + zTb(cid:1),thisisequivalenttorequiringthatthe
energy E(z) = zTWz + zTbofeachstateremains,upto
aconstant,unchanged.
First,wecanwritetheenergyofastate z(cid:48)andusethe
transformation z(cid:48) = 2z − 1toget
(cid:88)
(cid:88)
(62a)
z(cid:48)iW(cid:48)ijz(cid:48)j +
z(cid:48)ib(cid:48)i
(cid:18)
(cid:88)
(cid:88)
ziW(cid:48)ijzj − 2
ziW(cid:48)ij − 2
W(cid:48)ijzj
(cid:19)
(cid:88)
(cid:88)
(cid:88)
(62b)
W(cid:48)ij
b(cid:48)i + 2
zib(cid:48)i
(cid:88)
(cid:88)
(cid:0)2b(cid:48)i
zi4W(cid:48)ijzj +
(cid:88)
(cid:1) + C ,
(62c)
W(cid:48)ij
(cid:80)
(cid:80)
ib(cid:48)iandweused
where Cisaconstant C = 1
i,jW(cid:48)ij −
thefactthatW(cid:48)ij issymmetric.Sinceconstanttermsinthe
energyleavetheprobabilitydistributioninvariant,wecan
simplycompare E(z(cid:48))and E(z)
(cid:88)
(63)
ibi ,
iWijzj +
andextractthecorrectparametertransformation:
(64)
Wij = 4W(cid:48)ij ,
(65)
bi = 2b(cid:48)i − 2
Fromthis,wecanalsocalculatetheinversetransformation
rulefor z = 1
2 (z(cid:48) + 1):
(66)
W(cid:48)ij =
(67)
b(cid:48)i =
6.7 TranslationfromBoltzmanntoneurosynap-
ticparameters
Asdiscussedinthemethodssection,following[44],theacti-
vationfunctionofLIFneuronsintheHCSisapproximately
logisticandcanbewrittenas
(68)
where z/kisthestatevectorofallotherneuronsexceptthe
k'th one and µ the mean membrane potential (Eq. (17b)).
u0and αaretheinflectionpointandtheinverseslope,re-
= (1 + exp (−(µ − u0)/α))−1 ,
p(zk = 1 z/k) = ϕ(µ)
Wij ,
bi +
W(cid:48)ij .
Wij .
(cid:88)
(cid:88)
E(z) =
1
4
1
2
zT
zT
1
4
1
2
i,j
j
j
i
=
j
,
0
El =
τm
τeff
Wkjzj − bk)
−1
p(zk = 1 z/k)
= (1 + exp (−bk))−1 ,
(cid:88)
spectively.Furthermore,theconditionalprobability p(zk =
1 z/k) of a Boltzmann distribution over binary random
variables zk,i.e., p(z) ∝ exp(cid:0) 1
2 zT W z + zT b(cid:1),isgivenby
1 + exp (−
(cid:88)
(69)
withsymmetricweightmatrix W, Wii = 0 ∀i,andbiases
b.Theseequationsallowatranslationfromtheparameters
ofaBoltzmanndistribution (bi, Wij)toparametersofLIF
neuronsandtheirsynapses (El, wij),suchthatthestate
dynamicofthenetworkapproximatessamplingfromthetar-
getBoltzmanndistribution.
First,thebiases bcanbemappedtoleakpotentials El
(or external currents) by requiring that, for W = 0 (that
is, no synaptic input from other neurons), the activity of
eachneuronequalstheconditionalprobabilityofthetarget
Boltzmanndistribution
(70)
leadingtothetranslationrule
(1 + exp (−(µ − u0)/α))−1 !
(71)
TomapBoltzmannweights Wij tosynapticweights wij,
we first have to rescale the Wij, as done for the biases
in Eq. (71). However, leaky integrator neurons have non-
rectangularPSPs,sotheirinteractionstrengthismodulated
overtime.ThisisdifferentfromtheinteractioninBoltzmann
machines,wherethePSPshapeisrectangular(Glauberdy-
namics).Nevertheless,wecanderiveaheuristictranslation
rulebyrequiringthatthemeaninteractionduringthere-
fractory period of the presynaptic neuron is the same in
bothcases,i.e.,(cid:90) τref
(72a)
(72b)
where P SP (t)isgivenbyEq.(19).Fromthis,wegetthe
translationruleforsynapticweights:
(cid:16)
(cid:17)(cid:16)
(73)
(cid:17)
(cid:16)
1 − τ syn
7 Acknowledgments
− τeff
We thank Luziwei Leng, Nico Gurtler and Johannes Bill
forvaluablediscussions.WefurtherthankEricMullerand
Christian Mauch for maintenance of the computing clus-
ter we used for simulations and Luziwei Leng for provid-
(cid:17)−1
(cid:17)(cid:105) .
Erev
kj − µ
e−
τeff − 1
τref
(αb + u0) −
(cid:90) τref
αWkjCm
τref
τ syn
(cid:104)gsyn
x (cid:105)
gl
Erev
x
.
τ syn
e−
τref
τ syn − 1
dt P SP (t)
!
=
dt αWij
0
= αWijτref ,
x∈{e,i}
(cid:104)
(cid:16)
wkj =
τeff
ingcodeimplementingtheCASTalgorithm.Thisworkhas
receivedfundingfromtheEuropeanUnion7thFramework
Programmeundergrantagreement604102(HBP),theHori-
zon 2020 Framework Programme under grant agreement
720270,785907(HBP)andtheManfredStarkFoundation.
8 References
[1] GHHenry,POBishop,RMTupper,andBDreher.Ori-
entationspecificityandresponsevariabilityofcellsin
the striate cortex. Vision research, 13(9):1771 -- 1779,
1973.
[2] PeterHSchiller,BarbaraLFinlay,andSusanFVol-
man.Short-termresponsevariabilityofmonkeystriate
neurons. Brainresearch,105(2):347 -- 349,1976.
[3] Rufin Vogels, Werner Spileers, and Guy A Orban.
Theresponsevariabilityofstriatecorticalneuronsin
the behaving monkey. Experimental brain research,
77(2):432 -- 436,1989.
[4] Robert J Snowden, Stefan Treue, and Richard A An-
dersen. Theresponseofneuronsinareasv1andmtof
thealertrhesusmonkeytomovingrandomdotpatterns.
ExperimentalBrainResearch,88(2):389 -- 400,1992.
[5] AmosArieli,AlexanderSterkin,AmiramGrinvald,and
ADAertsen.Dynamicsofongoingactivity:explanation
of the large variability in evoked cortical responses.
Science,273(5283):1868,1996.
[6] Rony Azouz and Charles M Gray. Cellular mecha-
nisms contributing to response variability of cortical
neuronsinvivo. JournalofNeuroscience,19(6):2209 --
2223,1999.
[7] Rajesh PN Rao, Bruno A Olshausen, and Michael S
Lewicki. Probabilisticmodelsofthebrain:Perception
andneuralfunction. MITpress,2002.
[8] KonradPKordingandDanielMWolpert. Bayesian
integration in sensorimotor
Nature,
learning.
427(6971):244 -- 247,2004.
[9] Jan W Brascamp, Raymond Van Ee, Andre J Noest,
RichardHAHJacobs,andAlbertVvandenBerg. The
timecourseofbinocularrivalryrevealsafundamental
roleofnoise. Journalofvision,6(11):8 -- 8,2006.
[10] Gustavo Deco, Edmund T Rolls, and Ranulfo Romo.
Stochastic dynamics as a principle of brain function.
Progressinneurobiology,88(1):1 -- 16,2009.
[11] J´ozsef Fiser, Pietro Berkes, Gergo Orb´an, and M´at´e
Lengyel.Statisticallyoptimalperceptionandlearning:
frombehaviortoneuralrepresentations.Trendsincog-
nitivesciences,14(3):119 -- 130,2010.
15
[12] Wolfgang Maass. Searching for principles of brain
computation. CurrentOpinioninBehavioralSciences,
11:81 -- 92,2016.
[13] RichardBStein. Somemodelsofneuronalvariability.
Biophysicaljournal,7(1):37 -- 68,1967.
[14] NicolasBrunel. Dynamicsofsparselyconnectednet-
works of excitatory and inhibitory spiking neurons.
Journal of computational neuroscience, 8(3):183 -- 208,
2000.
[15] NicolasFourcaudandNicolasBrunel.Dynamicsofthe
firing probability of noisy integrate-and-fire neurons.
Neuralcomputation,14(9):2057 -- 2110,2002.
[16] Wulfram Gerstner, Werner M Kistler, Richard Naud,
and Liam Paninski. Neuronal dynamics: From single
neurons to networks and models of cognition. Cam-
bridgeUniversityPress,2014.
[17] DKSmettersandAnthonyZador. Synaptictransmis-
sion:noisysynapsesandnoisyneurons. CurrentBiol-
ogy,6(10):1217 -- 1218,1996.
[18] Wolfgang Maass and Anthony M Zador. Dynamic
stochastic synapses as computational units.
In Ad-
vancesinneuralinformationprocessingsystems,pages
194 -- 200,1998.
[19] Peter N Steinmetz, Amit Manwani, Christof Koch,
MichaelLondon,andIdanSegev.Subthresholdvoltage
noise due to channel fluctuations in active neuronal
membranes.
Journal of computational neuroscience,
9(2):133 -- 148,2000.
[20] YosefYaromandJornHounsgaard.Voltagefluctuations
in neurons: signal or noise? Physiological reviews,
91(3):917 -- 929,2011.
[21] Rub´en Moreno-Bote. Poisson-like spiking in circuits
withprobabilisticsynapses. PLoScomputationalbiol-
ogy,10(7):e1003522,2014.
[22] Emre O Neftci, Bruno U Pedroni, Siddharth
Joshi, Maruan Al-Shedivat, and Gert Cauwenberghs.
Stochastic synapses enable efficient brain-inspired
learningmachines. Frontiersinneuroscience,10:241,
2016.
[23] CharlesFStevensandAnthonyMZador. Whenisan
integrate-and-fireneuronlikeapoissonneuron?InAd-
vancesinneuralinformationprocessingsystems,pages
103 -- 109,1996.
[24] Hans E Plesser and Wulfram Gerstner. Noise in
integrate-and-fireneurons:fromstochasticinputtoes-
caperates. Neuralcomputation,12(2):367 -- 384,2000.
[25] EJChichilnisky. Asimplewhitenoiseanalysisofneu-
ronallightresponses.Network:ComputationinNeural
Systems,12(2):199 -- 213,2001.
[26] Wulfram Gerstner and Werner M Kistler. Spiking
neuronmodels:Singleneurons,populations,plasticity.
Cambridgeuniversitypress,2002.
[27] Peter Dayan, LF Abbott, et al. Theoretical neuro-
science:computationalandmathematicalmodelingof
neural systems.
Journal of Cognitive Neuroscience,
15(1):154 -- 155,2003.
[28] SrdjanOstojicandNicolasBrunel. Fromspikingneu-
ronmodelstolinear-nonlinearmodels.PLoScomputa-
tionalbiology,7(1):e1001056,2011.
[29] Gergo Orb´an, Pietro Berkes, J´ozsef Fiser, and M´at´e
Lengyel. Neuralvariabilityandsampling-basedprob-
abilisticrepresentationsinthevisualcortex. Neuron,
92(2):530 -- 543,2016.
[30] LaurenceAitchisonandM´at´eLengyel.Thehamiltonian
brain: efficient probabilistic inference with excitatory-
inhibitoryneuralcircuitdynamics.PLoScomputational
biology,12(12):e1005186,2016.
[31] Moritz Deger, Moritz Helias, Clemens Boucsein, and
Stefan Rotter. Statistical properties of superimposed
stationaryspiketrains.JournalofComputationalNeu-
roscience,32(3):443 -- 463,2012.
[32] JINelson,PASalin,MH-JMunk,MArzi,andJBullier.
Spatialandtemporalcoherenceincortico-corticalcon-
nections:across-correlationstudyinareas17and18
inthecat. Visualneuroscience,9(1):21 -- 37,1992.
[33] Bruno B Averbeck, Peter E Latham, and Alexandre
Pouget. Neural correlations, population coding and
computation. Nature reviews neuroscience, 7(5):358,
2006.
[34] EmilioSalinasandTerrenceJSejnowski. Correlated
neuronal activity and the flow of neural information.
Naturereviewsneuroscience,2(8):539,2001.
[35] RonenSegev,MorrisBenveniste,EyalHulata,Netta
Cohen,AlexanderPalevski,EliKapon,YoashShapira,
andEshelBen-Jacob.Longtermbehavioroflithograph-
ically prepared in vitro neuronal networks. Physical
reviewletters,88(11):118102,2002.
[36] J´ozsefFiser,ChiayuChiu,andMichaelWeliky. Small
modulationofongoingcorticaldynamicsbysensoryin-
putduringnaturalvision.Nature,431(7008):573,2004.
[37] RobertRosenbaum,TatjanaTchumatchenko,andRub´en
Moreno-Bote.Correlatedneuronalactivityanditsrela-
tionshiptocoding,dynamicsandnetworkarchitecture.
Frontiersincomputationalneuroscience,8:102,2014.
16
Theory of
[38] Rub´en Moreno-Bote, Alfonso Renart, and N´estor
Parga.
input spike auto-and cross-
correlationsandtheireffectontheresponseofspiking
neurons. Neuralcomputation,20(7):1651 -- 1705,2008.
[39] Zachary F Mainen and Terrence J Sejnowski. Relia-
bilityofspiketiminginneocorticalneurons. Science,
268(5216):1503,1995.
[40] AnthonyZador. Impactofsynapticunreliabilityonthe
informationtransmittedbyspikingneurons. Journalof
Neurophysiology,79(3):1219 -- 1229,1998.
[41] Alexander Rauch, Giancarlo La Camera, Hans-Rudolf
Luscher,WalterSenn,andStefanoFusi. Neocortical
pyramidalcellsrespondasintegrate-and-fireneurons
toinvivo -- likeinputcurrents. Journalofneurophysiol-
ogy,90(3):1598 -- 1612,2003.
[42] Henry Markram, Eilif Muller, Srikanth Ramaswamy,
MichaelWReimann,MarwanAbdellah,CarlosAguado
Sanchez,AnastasiaAilamaki,LidiaAlonso-Nanclares,
Nicolas Antille, Selim Arsever, et al. Reconstruc-
tionandsimulationofneocorticalmicrocircuitry. Cell,
163(2):456 -- 492,2015.
[43] Lars Buesing, Johannes Bill, Bernhard Nessler, and
Wolfgang Maass. Neural dynamics as sampling: a
modelforstochasticcomputationinrecurrentnetworks
ofspikingneurons.PLoSComputBiol,7(11):e1002211,
2011.
[44] Mihai A. Petrovici, Johannes Bill, Ilja Bytschok, Jo-
hannes Schemmel, and Karlheinz Meier. Stochastic
inferencewithspikingneuronsinthehigh-conductance
state. Phys.Rev.E,94:042312,Oct2016.
[45] Dejan Pecevski, Lars Buesing, and Wolfgang Maass.
Probabilistic inference in general graphical models
throughsamplinginstochasticnetworksofspikingneu-
rons. PLoS computational biology, 7(12):e1002294,
2011.
[46] Dimitri Probst, Mihai A Petrovici, Ilja Bytschok, Jo-
hannesBill,DejanPecevski,JohannesSchemmel,and
Karlheinz Meier. Probabilistic inference in discrete
spacescanbeimplementedintonetworksoflifneurons.
Frontiersincomputationalneuroscience,9,2015.
[47] Luziwei Leng, Roman Martel, Oliver Breitwieser, Ilja
Bytschok,WalterSenn,JohannesSchemmel,Karlheinz
Meier, and Mihai A Petrovici. Spiking neurons with
short-termsynapticplasticityformsuperiorgenerative
networks. ScientificReports,8(1):10651,2018.
[48] MichaelMonteforteandFredWolf.Dynamicfluxtubes
formreservoirsofstabilityinneuronalcircuits. Physi-
calReviewX,2(4):041007,2012.
Spatiotempo-
[49] Ryan Pyle and Robert Rosenbaum.
ral dynamics and reliable computations in recurrent
spiking neural networks.
Physical review letters,
118(1):018103,2017.
Idan Segev, Henry Markram, and
[50] Galit Fuhrmann,
Misha Tsodyks. Coding of temporal information by
activity-dependent synapses.
Journal of neurophysi-
ology,87(1):140 -- 148,2002.
[51] AndrewDavison,DanielBrÃČÂijderle,JochenEppler,
Jens Kremkow, Eilif Muller, Dejan Pecevski, Laurent
Perrinet,andPierreYger. Pynn:acommoninterface
forneuronalnetworksimulators. FrontiersinNeuroin-
formatics,2:11,2009.
[52] Marc-Oliver Gewaltig and Markus Diesmann. Nest
(neuralsimulationtool).Scholarpedia,2(4):1430,2007.
[53] AlainDestexhe,MichaelRudolph,andDenisPar´e.The
high-conductancestateofneocorticalneuronsinvivo.
Naturereviewsneuroscience,4(9):739,2003.
[54] ArvindKumar,SvenSchrader,AdAertsen,andStefan
Rotter.Thehigh-conductancestateofcorticalnetworks.
Neuralcomputation,20(1):1 -- 43,2008.
[55] Solomon Kullback and Richard A Leibler. On infor-
mation and sufficiency. The annals of mathematical
statistics,22(1):79 -- 86,1951.
[56] DavidHAckley,GeoffreyEHinton,andTerrenceJSe-
jnowski. Alearningalgorithmforboltzmannmachines.
Cognitivescience,9(1):147 -- 169,1985.
[57] Ruslan Salakhutdinov. Learning deep boltzmann ma-
chinesusingadaptivemcmc.InProceedingsofthe27th
InternationalConferenceonMachineLearning(ICML-
10),pages943 -- 950,2010.
[58] Johannes Schemmel, Johannes Fieres, and Karlheinz
Meier. Wafer-scale integration of analog neural net-
works. InNeuralNetworks,2008.IJCNN2008.(IEEE
WorldCongressonComputationalIntelligence).IEEE
International
Joint Conference on, pages 431 -- 438.
IEEE,2008.
[59] Mihai A Petrovici, Bernhard Vogginger, Paul Muller,
Oliver Breitwieser, Mikael Lundqvist, Lyle Muller,
Matthias Ehrlich, Alain Destexhe, Anders Lansner,
Ren´eSchuffny,etal. Characterizationandcompensa-
tionofnetwork-levelanomaliesinmixed-signalneuro-
morphicmodelingplatforms. PloSone,9(10):e108590,
2014.
[60] Sebastian Schmitt, Johann Klahn, Guillaume Bellec,
Andreas Grubl, Maurice Guettler, Andreas Hartel,
StephanHartmann,DanHusmann,KaiHusmann,Se-
bastian Jeltsch, et al. Neuromorphic hardware in
the loop: Training a deep spiking network on the
17
brainscales wafer-scale system.
In Neural Networks
(IJCNN),2017InternationalJointConferenceon,pages
2227 -- 2234.IEEE,2017.
[61] JakobJordan,MihaiAPetrovici,OliverBreitwieser,Jo-
hannesSchemmel,KarlheinzMeier,MarkusDiesmann,
andTomTetzlaff. Stochasticneuralcomputationwith-
outnoise. arXivpreprintarXiv:1710.04931,2017.
[62] Mihai A Petrovici, Sebastian Schmitt, Johann Klahn,
DavidStockel,AnnaSchroeder,GuillaumeBellec,Jo-
hannesBill,OliverBreitwieser,IljaBytschok,Andreas
Grubl, et al. Pattern representation and recognition
withacceleratedanalogneuromorphicsystems. InCir-
cuits and Systems (ISCAS), 2017 IEEE International
Symposiumon,pages1 -- 4.IEEE,2017.
[63] Jun Haeng Lee, Tobi Delbruck, and Michael Pfeiffer.
Trainingdeepspikingneuralnetworksusingbackprop-
agation. Frontiersinneuroscience,10:508,2016.
[64] FriedemannZenkeandSuryaGanguli.Superspike:Su-
pervisedlearninginmultilayerspikingneuralnetworks.
Neuralcomputation,30(6):1514 -- 1541,2018.
[65] SaeedRezaKheradpisheh,MohammadGanjtabesh,Si-
monJThorpe,andTimoth´eeMasquelier. Stdp-based
spikingdeepconvolutionalneuralnetworksforobject
recognition. NeuralNetworks,99:56 -- 67,2018.
[66] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and
Andr´e van Schaik. Emnist: an extension of mnist to
handwritten letters. arXiv preprint arXiv:1702.05373,
2017.
[67] YannLeCun,L´eonBottou,YoshuaBengio,andPatrick
Haffner. Gradient-basedlearningappliedtodocument
recognition. Proceedings of the IEEE, 86(11):2278 --
2324,1998.
[68] A Aldo Faisal, Luc PJ Selen, and Daniel M Wolpert.
Noise in the nervous system. Nature reviews neuro-
science,9(4):292,2008.
[69] Tiago Branco and Kevin Staras. The probability of
neurotransmitterrelease:variabilityandfeedbackcon-
trolatsinglesynapses.NatureReviewsNeuroscience,
10(5):373,2009.
[70] John A White, Jay T Rubinstein, and Alan R Kay.
Channel noise in neurons. Trends in neurosciences,
23(3):131 -- 137,2000.
[71] H Sebastian Seung. Learning in spiking neural net-
worksbyreinforcementofstochasticsynaptictransmis-
sion. Neuron,40(6):1063 -- 1073,2003.
[72] Xiaohui Xie and H Sebastian Seung.
Learning in
neuralnetworksbyreinforcementofirregularspiking.
PhysicalReviewE,69(4):041909,2004.
18
[73] ZhangJChen,YongHe,PedroRosa-Neto,JurgenGer-
mann,andAlanCEvans. Revealingmodulararchitec-
tureofhumanbrainstructuralnetworksbyusingcor-
ticalthicknessfrommri. Cerebralcortex,18(10):2374 --
2381,2008.
[74] EdBullmoreandOlafSporns.Complexbrainnetworks:
graphtheoreticalanalysisofstructuralandfunctional
systems. Nature Reviews Neuroscience, 10(3):186,
2009.
[75] DavidMeunier,RenaudLambiotte,andEdwardTBull-
more.Modularandhierarchicallymodularorganization
of brain networks. Frontiers in neuroscience, 4:200,
2010.
[76] Maxwell A Bertolero, BT Thomas Yeo, and Mark
DâĂŹEsposito.Themodularandintegrativefunctional
architecture of the human brain. Proceedings of the
NationalAcademyofSciences,112(49):E6798 -- E6807,
2015.
[77] SenSong,PerJesperSjostrom,MarkusReigl,Sacha
Nelson, and Dmitri B Chklovskii. Highly nonrandom
features of synaptic connectivity in local cortical cir-
cuits. PLoSbiology,3(3):e68,2005.
[78] WeiJiMa,JeffreyMBeck,PeterELatham,andAlexan-
drePouget.Bayesianinferencewithprobabilisticpop-
ulationcodes. Natureneuroscience,9(11):1432,2006.
[79] Mark M Churchland, M Yu Byron, Stephen I Ryu,
GopalSanthanam,andKrishnaVShenoy.Neuralvari-
abilityinpremotorcortexprovidesasignatureofmo-
torpreparation.JournalofNeuroscience,26(14):3697 --
3712,2006.
[80] MihaiAlexandruPetrovici.FormVersusFunction:The-
ory and Models for Neuronal Substrates. Springer,
2016.
[81] PeterHaunggiandPeterJung.Colorednoiseindynam-
ical systems. Advances in chemical physics, 89:239 --
326,1994.
[82] ManuelOC´aceres.Harmonicpotentialdrivenbylong-
rangecorrelatednoise.PhysicalReviewE,60(5):5208,
1999.
[83] NorbertWiener. Generalizedharmonicanalysis. Acta
mathematica,55(1):117 -- 258,1930.
[84] Alexander Khintchine. Korrelationstheorie der sta-
tionarenstochastischenprozesse. MathematischeAn-
nalen,109(1):604 -- 615,1934.
[85] IljaBytschok,DominikDold,JohannesSchemmel,Karl-
heinzMeier,andMihaiA.Petrovici.Spike-basedprob-
abilisticinferencewithcorrelatednoise. BMCNeuro-
science,2017.
wpre
[86] IljaBytschok,DominikDold,JohannesSchemmel,Karl-
heinzMeier,andMihaiAPetrovici.Spike-basedprob-
abilisticinferencewithcorrelatednoise.arXivpreprint
arXiv:1707.01746,2017.
[87] LaurensvanderMaatenandGeoffreyHinton. Visual-
izing data using t-sne.
Journal of Machine Learning
Research,9(Nov):2579 -- 2605,2008.
9 Supportinginformation
9.1 Experimentdetails
AllsimulationsweredonewithPyNN0.8andNEST2.4.2.
The LIF model was integrated with a time step of dt =
0.1 ms. Since the SSN is a time-continuous system, we
could, in principle, retrieve a sample at every integration
step.However,asindividualneuronsonlychangetheirstate
onthetimescaleofrefractoryperiods,andhencenewstates
emergeonasimilartimescale,wereadoutthestatesin
intervalsof τref
2 .Ifnotstatedotherwise,weused USE = 1.0
and τrec = 10 ms as short term plasticity parameters for
connectionswithineachSSNtoensurepostsynapticpoten-
tialswithequalheight,asdiscussedin[80].Forbackground
connections,i.e.,Poissoninputorbackgroundinputcoming
fromotherSSNsinanensemble,weusedstaticsynapses
(USE = 1.0and τrec → 0)insteadtofacilitatethemathe-
maticalanalysis.Fortheinterconnectionsofanensemble,
weexpectthatshort-termdepressionwillnotaltertheper-
formanceofindividualSSNsinadrasticway,astheeffect
willberathersmallonaverageifmostneuronsarefaraway
fromtonicbursting.Thus,toallowaclearcomparabilityto
SSNs receiving Poisson input, we chose to neglect short-
termdepressionforensembleinterconnections.Theneuron
parameters used throughout all simulations are listed in
TableS1.
9.1.1 DetailstoFigure2B,C,Dofmaintext
Theparametersforthetargetdistributionsofallnetworks
were randomly drawn from beta distributions, i.e., W ∼
2 · (beta(0.5, 0.5) − 0.5)and b ∼ 1.2 · (beta(0.5, 0.5) − 0.5).
Thebiasofeachbackground-providingneuronwasadjusted
toyieldthedesiredfiringrate p ∈ {0.1, 0.6, 0.9}incaseof
(cid:17).Forthediffer-
nosynapticinput(W = 0) b = log
ent activity cases, we used N0.1 = 260, N0.6 = 50 and
N0.9 = 34networksasbackgroundinputforeachneuron
toreachthedesiredbackgroundnoisefrequency.ThePois-
son frequency of the noise-providing networks was set to
3000 Hz.Activationfunctionswererecordedbyprovidingev-
eryneuronwithbackgroundnoisefor 5 × 105 msandvary-
ingitsleakpotential.Fortheautocorrelationfunctions,we
first merged all individual noise spike trains and binned
the resulting spike train with a bin size of 0.5 ms before
calculatingtheautocorrelationfunction.
(cid:16) p
1−p
19
ij ∼ 4 · (beta(0.5, 5.0) − 0.5)
9.1.2 DetailstoFigure2F,Gofmaintext
Background input was generated from a pool of pairwise
connected neurons (i.e., small subnetworks with two neu-
ronseach)withstrongpositiveornegativeweightstoyield
highlypositivelyornegativelycorrelatedspiketrains.Each
pair of neurons in the main network (i.e, the network re-
ceiving no Poisson input) received the spikes of 80 such
subnetworksasbackgroundinput.Theweightsofthenoise-
generatingsubnetworksweredrawnfrombetadistributions
ij ∼ 4·(beta(5.0, 0.5)−0.5)(distributionstronglyskewed
to positive weights) or wpre
(skewedtonegativeweights).Theparametersofthemain
network were randomly generated as in Fig. 2D. The ab-
solute values of the weights Wnoise projecting from the
noise-generating subnetworks to the main network were
randomly generated from a (Gaussian-like) beta distribu-
tion Wnoise ∼ (beta(4.0, 4.0) − 0.5) · 2 · 0.001 + 0.001 µS.
The synapse type of each weight Wnoise was determined
randomly with equal probability. Furthermore, inhibitory
synapseswerescaledbyafactorof 1.35suchthatinhibitory
andexcitatoryweightshavethesamemeanvalueasforthe
simulationswithPoissonnoise(seeTableS1).Forthethree
traces shown, the absolute value of each synaptic weight
wasdrawnindependently.Synapsetypeswereeitherdrawn
accordingtoapattern(Fig.2F,leftandmiddle)orindepen-
dently(Fig.2F,right).
9.1.3 DetailstoFigure3A,B,CofmaintextandFigure
Foreverynetwork,thetargetdistributionparameterswere
againdrawnfromabetadistributionasinFig.2.Thecon-
nectivityofthenoiseconnectionswassetto = 0.05,i.e.,
eachneuronreceivedthespikesof 5%oftheremainingsub-
networks' neurons as stochastic background input, for the
ensembleof4003-neuronnetworks(notraining,FigureS2
intheSupportinginformation)andto = 0.10fortheen-
sembleof1006-neuronnetworks(training,maintext).
The training was done for subnetworks with 6 neurons
each, where every subnetwork was initialized with differ-
ent weights and biases than the target parameters, also
generated randomly. As an initial guess for the neurons'
activation functions, we used the activation function of a
neuronreceiving 2000 HzexcitatoryandinhibitoryPoisson
input,leadingtoaslopeof α = 1.47 mVandamid-point
at −52.97 mV.Theseparametersweresubsequentlyused
totranslatetheweightandbiasupdatesgivenbytheHeb-
bianwake-sleepalgorithm(seemaintext,Eqn.8and9)to
updatesofsynapticweightsandleakpotentials.Thesub-
networks were all trained simultaneously with contrastive
divergence,wherethemodeltermwasapproximatedbysam-
plingfor 1 × 105 ms.Thetrainingwasdonefor2000steps
andwithalearningrateof
t+2000.Asareference,50sub-
networksreceivingonlyPoissonnoise(2000 Hz)werealso
trainedinthesamewayfor2000steps.
S2intheSupportinginformation
400
20
Self-activation of the network can be observed when a
largeenoughfractionofneuronshaveasuprathresholdrest
potential,inourcasearound30%.
9.1.4 DetailstoFigure3D,Eofmaintext
The emulated ensemble consists of 15 4-neuron networks
whichwererandomlyinitializedontwoHICANNchips(HI-
CANN367and376onWafer33oftheBrainScaleSsys-
tem).Theanaloghardwareparameterswhichdeterminethe
physical range of weights adjustable by the 4bit setting
weresetto gmax = 500and gmax div = 1.Giventhecur-
rent state of development of the BrainScaleS system and
itssurroundingsoftware,welimitedtheexperimenttosmall
ensemblesinordertoavoidpotentialcommunicationbottle-
necks.BiaseswereimplementedbyassigningeverySSNneu-
ronabiasneuron,withitsrestpotentialsetabovethreshold
toforcecontinuousspiking.Whileamoreresource-efficient
implementation of biases is possible, this implementation
allowed an easier mapping of neuron - bias pairs on the
neuromorphic hardware. Bias strengths can then be ad-
justed by modifying the synaptic weights between SSN
neurons and their allocated bias neurons. The networks
weretrainedwiththecontrastivedivergencelearningrule
tosamplefromtheirrespectivetargetdistributions.Biases
wererandomlydrawnfromanormaldistributionwith µ = 0
and σ = 0.25.Theweightmatriceswererandomlydrawn
from W ∝ 2 · (beta(0.5, 0.5) − 0.5)andsubsequentlysym-
metrized by averaging with their respecetive transposes
0.5 · (W + W T).
Sincetherefractoryperiodsofhardwareneuronsvary,we
further measured the refractory period of every neuron in
theensemble,whichwaslaterusedtocalculatethebinary
neuron states from spike raster plots. Refractory periods
were measured by setting biases to large enough values
todriveneuronstotheirmaximalfiringrate.Afterrunning
the experiment, the duration of the refractory period can
be approximated by dividing the experiment time by the
numberofmeasuredspikes.
During the whole experiment, the ensemble did not re-
ceive external Poisson noise.
individual SSNs
received spikes from 20% of the remaining ensemble as
backgroundinput,withnoiseconnectionshavinghardware
weights of ±4 with the sign chosen randomly with equal
probability.Translationbetweentheoreticaland4bithard-
wareparameterswasdonebyclippingthevaluesintothe
range [−15, 15]androundingtointegervalues.Calculation
ofweightandbiasupdateswasperformedonahostcom-
puter.Thelearningratewassetto η = 1.0forallnetworks
performingbetterthanthemedianandtwicethisvaluefor
networksperformingworse.Foreverytrainingstep,theen-
semblewasrecordedfor 1 × 105 msbiologicaltimebefore
applyingaparameterupdate.
FortheexperimentswithPoissonnoise,everyneuronre-
Instead,
20
intheSupportinginformation
ceived 300 HzexternalPoissonnoiseprovidedbythehost
computer.
Theneuronparametersusedforallhardwareexperiments
arelistedinTableS2.
9.1.5 DetailstoFigure4ofmaintextandFigureS4-6
To reduce the training time, we pretrained classical re-
stricted Boltzmann machines on their respective datasets,
followed by direct translation to spiking network parame-
ters. To obtain better generative models, we utilized the
CASTalgorithm[57]whichcombinescontrastivedivergence
withsimulatedtempering.Eachsubnetworkwastrainedfor
200000stepswithaminibatchsizeof100,alearningrate
of
t+2000, an inverse temperature range β ∈ [1., 0.6] with
20 equidistant intervals and an adaptive factor γt = 9
1+t.
Statesbetweenthefastandslowchainwereexchangedev-
ery50samples.Tocollectthebackgroundstatisticsofthese
subnetworks,wefirstsimulatedallnetworkswithstochastic
Poissoninput.ToimprovethePoisson-stimulatedreference
networks'mixingproperties,weutilizedshort-termdepres-
sion to allow an easier escape from local energy minima
faster (USE = 0.01, τrec = 280 ms, global weight rescale
δW = 0.014−1). For classification, the grayscale value of
imagepixelswastranslatedtospikingprobabilitiesofthe
visibleunits,whichcanbeadjustedbysettingthebiases
(cid:19). Spike probabilities of 0 and 1 were
(cid:18)
as ln
mappedtobiasesof-50and50.Furthermore,duringclas-
sification,theconnectionsprojectingbackfromthehidden
neurons to the visible neurons were silenced in order to
preventthehiddenlayerfrominfluencingtheclampedinput.
Forpatternrivalry,thenon-occludedpixelswerebinarized.
Intotal,eachSSNreceivedbackgroundinputfrom20%of
the other networks' hidden neurons. For the classification
results,theexperimentwasrepeated10timesfordifferent
randomseeds(leadingtodifferentconnectivitymatricesbe-
tween the SSNs). For training and testing, we used 400
and200imagesperclass.InFig.4D,consecutiveimages
are 400 msapart.InFig.4E,fortheclamped"B",consecu-
tiveimagesare2sapart,forthe"L"1.5s.
The experiments with MNIST used an ensemble of five
networkswith784visibleneuronsand500hiddenneurons
each(Fig.S5andVideoS2intheSupportinginformation),
trained on 6 · 103 images per digit class (where we took
the digits provided by the EMNIST set to have balanced
classes). Since the generative properties of larger SSNs
depend heavily on the synaptic interaction, we also used
short-term plasticity for the case without Poisson noise
(USE = 0.01, τrec = 280 ms, δW = 0.01−1)toallowfluent
mixing between digit classes [47]. For MNIST, each SSN
receivedbackgroundinputfrom30%oftheothernetworks'
hidden neurons. Furthermore, the excitatory noise weight
wassetto wnoise
The network-generated images were obtained by aver-
grey value
1−grey value
= 0.0009 µS.
e
agingthefiringactivityofthevisibleneurons(forpattern
rivalry ±90 ms,forclassificationanddreamingofEMNIST
±80 msandforMNIST ±140 ms).Thetimeintervalswere
chosentoreducetheblurcausedbymixingwhenplotting
averagedspikingactivity.
9.2 Neuronparameters
Table 1: Neuron parameters used in simulations through-
outthemaintext.Themembranetimeconstantwaschosen
smallsuchthatsmallernoisefrequenciessufficetoreacha
high-conductancestate,allowingustousesmallerensem-
blesandhencereducesimulationtime.
0.1nF
1.0ms
-65.0mV
0.0mV
-90.0mV
-52.0mV
-53.0mV
10.0ms
10.0ms
10.0ms
0.001 µS
-0.00135 µS
membranecapacitance
membranetimeconstant
leakpotential
exc.reversalpotential
inh.reversalpotential
thresholdpotential
resetpotential
exc.synaptictimeconstant
inh.synaptictimeconstant
refractorytime
exc.Poissonweights
inh.Poissonweights
Cm
τm
El
Erev
Erev
e
i
ϑ
τ syn
e
τ syn
i
τref
wnoise
wnoise
e
i
Cm
τm
El
Erev
Erev
e
Table2:Neuronparametersusedfortheimplementationsin
anartificialneuralsubstrate.Notethattheseareintended
parametersandtherealizedonescanvaryfromneuronto
neuron.
biasneurons
60.0mV
-30.0mV
5.0ms
5.0ms
1.5ms
ensembleneurons
0.2nF
7ms
-20.0mV
60.0mV
-100.0mV
-20.0mV
-35.0mV
8.0ms
8.0ms
4.0ms
9.3 Videocaptions
9.3.1 VideoS1
Sampled distribution over time from an autonomous en-
semble (no explicit noise) of 15 4-neuron networks on an
artificial neural substrate (the BrainScaleS system). (top)
Median DKL of the ensemble (red) and the individual
networks (red, transparent) as a function of time after
training. The median DKL pre-training is shown in black.
i
ϑ
τ syn
e
τ syn
i
τref
(Bottom) Comparison between target distribution (blue)
and sampled distribution (red)
for all networks. Most
networksareabletoapproximatetheirtargetdistribution
well(e.g.,thenetworksatposition(1,1),(4,1)and(4,2)with
(row,column))oratleastapproximatethegeneralshapeof
thetargetdistribution(e.g.,(0,1)and(0,2)).Thenetworks
at position (2,2) and (3,0) show strong deviations from
their respective target distributions due to single neuron
deficiencies. Because of the speed-up of the BrainScaleS
system,itonlytakes100mstoemulate 106msofbiological
time.
9.3.2 VideoS2
Video of the data shown in Fig. S5 of the Supporting in-
formation. An ensemble of five hierarchical networks with
784-500-10 (visible-hidden-label) neurons trained on the
MNIST handwritten dataset generating digits without ex-
plicitnoisesourcesisshown(insimulation).Everynetwork
in the ensemble receives the spiking activity from hidden
neuronsoftheothernetworksasstochasticinputonly.(top)
Activityofthehiddenlayerofeachnetwork.(middle)Class
labelcurrentlypredictedbythelabelneurons.(Bottom)Ac-
tivityofthevisiblelayer.Togenerategrayscaleimagesfrom
spikes,weaveragedthespikingactivityofeachneuronover
awindowofsize ±90ms.
9.3.3 VideoS3
Video of the data shown in Fig. S7 of the Supporting
information. A single hierarchical network with 784-200
(visible-hidden)neuronsgeneratingsamplesoftheMNIST
handwrittendigitsdatasetwithoutexplicitnoisesourcesis
shown(insimulation).Toinitializethenetwork,wetrained
aBoltzmannmachineandtranslatedtheweightsandbiases
toneurosynapticparameterstoreducesimulationtime.(top)
Illustrationoftheusednetworkarchitecture.Lateral(non-
plastic)connectionsineachlayerwereutilizedasanoise
source(red),withaninterconnectivityof = 0.2.(bottom)
Averaged activity (average window ±90ms) of the visible
layer(left)afterinitializingthenetwork,(middle)afterfur-
thertrainingthenetworkand(right)forthecaseofexplicit
Poissonnoiseinsteadoflateralinterconnections.Afterini-
tialization,thenetworkisabletogeneraterecognizableim-
agesbutdoesnotmixwellbetweendifferentdigitclasses.
FurthertrainingthenetworkonimagesoftheMNISTtrain-
ingsetimprovesbothimagequalityandmixing,rivalingthe
quality of the reference setup with explicit Poisson noise.
During the second training phase, neurosynaptic parame-
tersareadjustedsuchthateveryneuronisabletoperform
itstaskwiththeavailablebackgroundactivityitreceives.
21
9.4 Supportingfigures
Figure S1: Compensation of input correlations by adjustment of weights and biases in an SSN [85, 86]. For simplicity, this is illustrated here for the
caseofsharedinputcorrelations,buttheresultsholdforalltypesofstaticallycorrelatedinputs.(A)Exemplaryarchitectureofanetworkwith3neurons
thatsamplesfromaBoltzmanndistributionwithparameters Wand b.Inordertoachievetherequiredstochasticregime,eachneuronreceivesexternal
noiseintheformofPoissonspiketrains(notshown).(B)-(D)Exemplarysampleddistributionsforanetworkoftwoneurons.The"default"caseisthe
onewhereallweightsandbiasesaresettozero(uniformdistribution,bluebars).(B)Sharednoisesourceshaveacorrelatingeffect,shiftingprobability
massintothe(1,1)and(0,0)states(redbars).(C)Inthe {0, 1}2 space,increasedweightsintroducea(positive)shiftofprobabilitymassfromallother
states towards the (1,1) state (red bars), which is markedly different from the effect of correlated noise. (D) In the {−1, 1}2 space, increased weights
have the same effect as correlated noise (red bars). (E) Dependence of the correlation coefficient r between the states of two neurons on the change
in synaptic weight ∆W (cid:48) (red) and the shared noise ratio s (blue). These define bijective functions g and h that can be used to compute the weight
change (∆W (cid:48) = f (s), with f := g−1 ◦ h) needed to compensate the effect of correlated noise in the {−1, 1}N space. (F) Study of the optimal
compensationruleinanetworkwithtwoneurons.Forsimplicity,theordinaterepresentsweightchangesforanetworkwithstatesinthe {−1, 1}2space,
whicharethentranslatedtocorrespondingparameters(W, b)forthe {0, 1}2 statespace.Thecolormapshowsthedifferencebetweenthesampledand
(cid:0)pnet (cid:107) ptarget(cid:1).Themappingprovidedbythecompensationrule f (see(E))
thetargetdistributionmeasuredbytheKullback-Leiblerdivergence DKL
isdepictedbythegreencurve.Notethatthecompensationrule ∆W (cid:48) = f (s)providesanearlyoptimalparametertranslation.Remainingdeviationsare
duetodifferencesbetweenLIFandGlauberdynamics.(G)CompensationofnoisecorrelationsinanSSNwithtenneurons.Theresultsaredepictedfor
asetoftenrandomlydrawnBoltzmanndistributionsover z ∈ {0, 1}10 (errorbars).ForasetofrandomlychosenBoltzmanndistributions,aten-neuron
networkperformssamplinginthepresenceofpairwise-sharednoiseratios s(x-axis).Thebluelinemarksthesamplingperformancewithoutnoise-induced
correlations(s = 0).Foranincreasingsharednoiseratio,uncompensatednoise(green)inducesasignificantincreaseinsamplingerror.Aftercompensation,
thesamplingperformanceisnearlycompletelyrestored.Asbefore,remainingdeviationsareduetodifferencesbetweenLIFandGlauberdynamics.(H)An
LIF-basedten-neuronnetworkwithsharednoisesources(s = 0.3foreachneuronpair)istrainedwithdatasamplesgeneratedfromatargetBoltzmann
distribution(bluebars).Duringtraining,thesampleddistributionbecomesanincreasinglybetterapproximationofthetargetdistribution(redline).For
comparison,wealsoshowthedistributionsampledbyanSSNwithparameterstranslateddirectlyfromtheBoltzmannparameters(purple).Thetrained
networkisabletoimproveuponthisresultbecauselearningimplicitlycompensatesfortheabovementioneddifferencesbetweenLIFandGlauberdynamics.
22
z1z2z3W=WTb1b2b3H0.00.10.20.30.40.5shared noise ratio s00.20.40.60.8corr. coeff. rhg-1E00011011z10-1p(z)00011011z10-1p(z)-1-1-111-111z010-1p(z0)10-210-1p(z)0500100015002000training steps10-210-1100DKL(pnetpdata)0.00.10.20.30.40.5shared noise ratio s10-210-1GA0.00.10.20.30.4shared noise ratio s-0.36-0.24-0.120.00.12∆W0FDBC0.00.10.20.30.40.50.6∆W010-310-210-1DKL(pnetptarget)z1z2zk∈{0,1}W12=W21>0z1z2z0k∈{−1,1}W012=W021>0z01z02 FigureS2:Translatingtheparametersoftargetdistributionstoneurosynapticparameters.(A)Astraightforwardwaytosetuptheparametersofeach
network(wij and El)istousetheparametertranslationasdescribedinthemaintext,i.e.,usethecorrespondingactivationfunctionofeachneuronto
correctlyaccountforthebackgroundnoisestatistics.Thisisdemonstratedhereforthecaseof(left)399networks(onlytwoshown)receivingPoisson
noiseandonenetworkonlyreceivingensembleinputand(right)allnetworksonlyreceivingensembleinput.Inbothcases,theresultingactivationfunction
isthesameandwecanindeeduseittotranslatetheparametersofthetargetdistributiontoneurosynapticparameters.(B)Usingthecorresponding
activationfunctionstosetuptheensemble(butnotraining),eachnetworkintheensembleisindeedabletoaccuratelysamplefromitstargetdistribution
withoutexplicitnoise,asexpectedfromourconsiderationsin(A)andthemaintext.Thisisshownhere(insoftwaresimulations)foranensembleof400
3-neuronSSNswithaninterconnectionprobabilityof 0.05,reachingamedian DKL of 12.8+6.4−5.0 × 10−3 (blue),whichisclosetotheidealresultwith
Poissonnoiseof 6.2+2.0−2.0 × 10−3(black,errorsgivenasthefirstandthirdquartile).
FigureS3:TheBrainScalesneuromorphicsystem.(A)AsingleHICANNchip(HighInputCountAnalogNeuralNetwork),theelementalbuildingblock
of the BrainScaleS wafer. The HICANN consists of two symmetric halves and harbors analog implementations of adaptive exponential integrate-and-
fire(AdEx)neuronsandconductance-basedsynapsesin180nmCMOStechnology.Floatinggatesnexttotheneuroncircuitsareusedtostoreneuron
parameters.Spikesarerouteddigitallythroughhorizontalandverticalbuses(notshown)andtranslatedintopostsynapticconductancesinthesynapse
array. Unlike in simulations on general-purpose CPUs, here neurons and synapses are physically implemented, with no numeric computations being
performedtocalculatenetworkdynamics.Asinglewaferconsistsof384HICANNchips.(B)IndividualcomponentsoftheBrainScaleSsystem,including
bothwaferandsupportstructure.Forinstance,FPGAboardsprovideanI/OinterfaceforwaferconfigurationandspikedataandGiga-Ethernetslots
provideaconnectionbetweenFPGAsandthecontrolclusterfromwhichusersconducttheirexperimentsviaPythonscriptsusingthePyNNAPI.(C)
CompletelyassembledwaferoftheBrainScaleSneuromorphicsystem.
23
A101102103104105106sim. time [ms]103102101100DKL(pnetptarget)EnsemblePoisson referenceBBCA Figure S4: t-SNE representation [87] of consecutively generated images of two of the four SSNs trained on EMNIST. Both SSNs smoothly traverse
severalregionsofthestatespacerepresentingimageclasseswhiledreaming.Thereddiamondmarksthefirstimageinthesequence,graylinesconnect
consecutiveimages.Consecutiveimagesare 200 msapart.
24
FigureS5:DreamingofMNISTinanensembleofnetworks.(A)DreamingensembleoffivehierarchicalSSNswith784visible,500hiddenand10label
neurons(withoutexplicitnoise).Eachrowrepresentssamplesfromasinglenetworkoftheensembles,withsamplesbeing 375 msapart.Tosetupthe
ensemble,arestrictedBoltzmannmachinewastrainedontheMNISTdatasetandtheresultingparameterstranslatedtocorrespondingneurosynaptic
parametersoftheensemble.Here,tofacilitatemixing,weusedshort-termdepressiontomodulatesynapticinteractionsandweakenattractorstatesthat
wouldbeotherwisedifficulttoescape[47].(B)t-SNErepresentation[87]ofconsecutivelygeneratedimagesoftwoofthefiveSSNstrainedonMNIST
digits.BothSSNsareabletogenerateandmixbetweendiverseimagesofdifferentdigitclasseswhiledreaming.Thereddiamondmarksthefirstimage
inthesequence,graylinesconnectconsecutiveimages.Consecutiveimagesare 400 msapart.
25
BA FigureS6:PatterncompletioninanensembletrainedonEMNIST.(A)Relativeabundanceofthelabeloutputwhileclampingpartsofa"B".Mostofthe
time(79.85%),theimageiscorrectlyclassifiedasa"B".Theclosestalternativeexplanation,an"R",isgeneratedsecondmost(17.45%).Theremaining
classesareexploredsignificantlylessoftenbythenetwork(0.43%,0.70%,1.57%).(B)Examplesofthevisiblelayeractivitywhilethelabellayerclassifies
thepartiallyclampedimageseitherasa"B"(top)oran"R"(bottom).(C)Examplesofthevisiblelayeractivitywhileclassifyingtheimageasa"T","X"
or"V".Inthesecases,theimagesgeneratedbythevisibleneuronsshowprominentfeaturesoftheseletters.
26
V102101100relative abundanceBRTXVlabel predictionABBCRTX Figure S7: A single hierarchical network with 784-200 (visible-hidden) neurons generating samples of the MNIST handwritten digits dataset without
explicitnoisesources(insimulation),representedviat-SNE.(A)Illustrationoftheusednetworkarchitecture.Lateral(non-plastic)connectionsineach
layerwereutilizedasanoisesource(red),withaninterconnectivityof = 0.2.(B,C)Averagedactivity(averagewindow ±90ms)ofthevisiblelayer(B)
afterinitializingthenetworkand(C)afterfurthertrainingthenetwork.Afterinitialization,thenetworkisabletogeneraterecognizableimagesbutdoes
notmixwellbetweendifferentdigitclassessincethenetworkisnotabletocorrectlyutilizeitsownbackgroundactivityasnoiseyet(B).Furthertraining
thenetworkonimagesoftheMNISTtrainingset(usingstandardContrastiveDivergencewithbatchsize 100,learningrate
t+2000 with tthenumberof
updates,for 1000trainingupdatesandapresentationtimepertrainingsampleof 200ms)improvesbothimagequalityandmixing(C).Duringthesecond
trainingphase,neurosynapticparametersareadjustedsuchthateveryneuronisabletoperformitstaskwiththeavailablebackgroundactivityitreceives.
40
27
CBA |
1608.03616 | 1 | 1608 | 2016-08-11T20:48:18 | Benchmarking confound regression strategies for the control of motion artifact in studies of functional connectivity | [
"q-bio.NC"
] | Since initial reports regarding the impact of motion artifact on measures of functional connectivity, there has been a proliferation of confound regression methods to limit its impact. However, recent techniques have not been systematically evaluated using consistent outcome measures. Here, we provide a systematic evaluation of 12 commonly used confound regression methods in 193 young adults. Specifically, we compare methods according to three benchmarks, including the residual relationship between motion and connectivity, distance-dependent effects of motion on connectivity, and additional degrees of freedom lost in confound regression. Our results delineate two clear trade-offs among methods. First, methods that include global signal regression minimize the relationship between connectivity and motion, but unmask distance-dependent artifact. In contrast, censoring methods mitigate both motion artifact and distance-dependence, but use additional degrees of freedom. Taken together, these results emphasize the heterogeneous efficacy of proposed methods, and suggest that different confound regression strategies may be appropriate in the context of specific scientific goals. | q-bio.NC | q-bio |
Benchmarking confound regression strategies for the control of motion artifact in
studies of functional connectivity
Rastko Cirica, Daniel H. Wolfa, Jonathan D. Powerb, David R. Roalfa, Graham Bauma, Kosha Ruparela, Russell T.
Shinoharac, Mark A. Elliottd, Simon B. Eickhoffe,f, Christos Davatzikosd, Ruben C. Gura, Raquel E. Gura, Danielle S.
Bassettg,h, Theodore D. Satterthwaitea,∗
aDepartment of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia PA, USA
bSackler Institute for Developmental Psychobiology, Weill Medical College of Cornell University, New York, NY, USA
cDepartment of Biostatistics and Epidemiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA
dDepartment of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia PA, USA
eDepartment of Clinical Neuroscience and Medical Psychology, Heinrich-Heine University Dusseldorf
fInstitute of Neuroscience and Medicine (INM-1), Research Center Julich
gDepartment of Bioengineering, University of Pennsylvania, Philadelphia PA, USA
hDepartment of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia PA, USA
Abstract
Since initial reports regarding the impact of motion artifact on measures of functional connectivity, there has been a
proliferation of confound regression methods to limit its impact. However, recent techniques have not been systematically
evaluated using consistent outcome measures. Here, we provide a systematic evaluation of 12 commonly used confound
regression methods in 193 young adults. Specifically, we compare methods according to three benchmarks, including
the residual relationship between motion and connectivity, distance-dependent effects of motion on connectivity, and
additional degrees of freedom lost in confound regression. Our results delineate two clear trade-offs among methods. First,
methods that include global signal regression minimize the relationship between connectivity and motion, but unmask
distance-dependent artifact.
In contrast, censoring methods mitigate both motion artifact and distance-dependence,
but use additional degrees of freedom. Taken together, these results emphasize the heterogeneous efficacy of proposed
methods, and suggest that different confound regression strategies may be appropriate in the context of specific scientific
goals.
Keywords:
fMRI, functional connectivity, artifact, confound, motion, noise
Introduction
Resting-state (intrinsic) functional connectivity (rsfc-
MRI) has evolved to become one of the most common
brain imaging phenotypes (Craddock et al., 2013; Fox and
Raichle, 2007; Power et al., 2014b; Smith et al., 2013; Van
Dijk et al., 2010), and has been critical for understand-
ing fundamental properties of brain organization (Damoi-
seaux et al., 2006; Fox et al., 2005; Power et al., 2011;
Yeo et al., 2011), brain development over the lifespan (Di-
Martino et al., 2014; Dosenbach et al., 2011; Fair et al.,
2008), and abnormalities associated with diverse clinical
conditions (Baker et al., 2014; Buckner et al., 2008; Fair
et al., 2010). rsfc-MRI has numerous advantages, includ-
ing ease of acquisition and suitability for a wide and ex-
panding array of analysis techniques. However, despite
knowledge that in-scanner motion can influence measures
of activation from task-related fMRI (Friston et al., 1996),
∗Corresponding author
Email address: [email protected] (Theodore D.
Satterthwaite )
Preprint submitted to arXiv
the impact of in-scanner motion on measures of functional
connectivity was not explored for 16 years after its initial
discovery. However, since the near-simultaneous publica-
tion of three independent reports in early 2012 (Van Dijk
et al., 2012; Power et al., 2012; Satterthwaite et al., 2012),
it has been increasingly recognized that motion can have a
large impact on rsfc-MRI measurements, and can system-
atically bias inference. This bias is particularly problem-
atic in developmental or clinical populations where mo-
tion is correlated with the independent variable of inter-
est (age, diagnosis) (Satterthwaite et al., 2012; Fair et al.,
2012), and has resulted in the re-evaluation of numerous
published findings.
In response to this challenge, there has been a recent
proliferation of methods aimed at mitigating the impact
of motion on functional connectivity (Yan et al., 2013a;
Power et al., 2015). These methods can be broadly
grouped into several categories. First, high-parameter con-
found regression strategies use expansions of realignment
parameters or tissue-compartment signals, often including
derivative and quadratic regressors (Friston et al., 1996;
Satterthwaite et al., 2013; Yan et al., 2013a). Second, prin-
August 15, 2016
cipal component analysis (PCA) based methods (Comp-
Cor; Behzadi et al. (2007); Muschelli et al. (2014)) find the
primary directions of variation within high-noise areas de-
fined by anatomy (e.g., aCompCor) or temporal variance
(tCompCor). Third, whole-brain independent component
analysis (ICA; Beckmann et al. (2005)) of single-subject
time series has increasingly been used for de-noising, with
noise components selected either by a trained classifier
(ICA-FIX; Griffanti et al. (2014); Salimi-Khorshidi et al.
(2014)) or using a priori heuristics (ICA-AROMA; Pruim
et al. (2015b,a)). Fourth, temporal censoring techniques
identify and remove (or de-weight) specific volumes con-
taminated by motion artifact, often followed by interpo-
lation. These techniques include scrubbing (Power et al.,
2012, 2014a, 2015), spike regression (Satterthwaite et al.,
2013), and de-spiking (Jo et al., 2013; Patel et al., 2014).
Censoring techniques have been reported to attenuate mo-
tion artifact, but at the cost of a shorter time series and
variably reduced degrees of freedom. Fifth, one recent re-
port emphasized the relative merits of spatially-tailored
confound regression using local white matter signals (wm-
Local; Jo et al. (2013)). Finally, the inclusion of global
signal regression (GSR) (Macey et al., 2004) in confound
regression models remains a source of controversy (Fox
et al., 2009; Murphy et al., 2009; Chai et al., 2012; Saad
et al., 2012; Yan et al., 2013c). While several studies have
suggested its utility in de-noising (Fox et al., 2009; Power
et al., 2015; Satterthwaite et al., 2013; Yan et al., 2013a),
other studies have emphasized the risk of removing a valu-
able signal (Yang et al., 2014; Hahamy et al., 2014), po-
tentially biasing group differences (Gotts et al., 2013; Saad
et al., 2012), or exacerbating distance-dependent motion
artifact. Distance-dependent artifact (Power et al., 2012;
Satterthwaite et al., 2012) manifests as increased connec-
tivity in short-range connections, and reduced connectiv-
ity in long-range connections, which has the potential to
impact measures of network topology (Yan et al., 2013b).
This recent proliferation of de-noising techniques has
prompted excitement but also sowed confusion. Unsurpris-
ingly, new de-noising pipelines have often tended to em-
phasize outcome measures that suggest their relative supe-
riority. As a result, investigators often anecdotally report
substantial uncertainty regarding which pipeline should be
used. Such uncertainty has been exacerbated by the lack of
common outcome measures used across studies, which has
hampered direct comparison among pipelines. While one
review paper has summarized recent developments in this
rapidly-evolving sub-field (Power et al., 2015), systematic
evaluation of de-noising pipelines according to a range of
benchmarks remains lacking.
Accordingly, in this report we compare a dozen of the
most commonly used confound regression strategies in a
large (N = 193) dataset of young adults. Pipelines eval-
uated include standard techniques, high-parameter con-
found regression, PCA-based techniques such as aComp-
Cor and tCompcor, ICA-based approaches such as ICA-
AROMA, spatially-tailored local white matter regression,
and three different censoring techniques (spike regression,
de-spiking, and scrubbing); GSR is included in many
pipelines as well. Critically, we compare these pipelines ac-
cording to three intuitive benchmarks, including the resid-
ual relationship between functional connectivity and sub-
ject motion, the degree of distance-dependent artifact, and
the loss of temporal degrees of freedom. As described be-
low, results underscore the relative strengths and weak-
nesses among these methods, and outline clear trade-offs
among commonly used confound regression approaches.
Materials and methods
Participants and data acquisition
The task-free BOLD data used in this study (N = 193)
were selected from the Philadelphia Neurodevelopmental
Cohort (PNC) (Satterthwaite et al., 2014, 2016) on the ba-
sis of age, health, and data quality. In order to minimize
the potential impact of developmental effects, subjects se-
lected for inclusion were aged at least 18 years at the date
of scan. All participants were free of medical problems
that could impact brain function (Merikangas et al., 2010),
lacked gross structural brain abnormalities (Gur et al.,
2013), and were not taking psychotropic medication. Fur-
thermore, subjects were excluded from the study if they
failed to satisfy any of three criteria for functional image
quality:
if subject movement (relative root mean square
displacement) averaged over all frames exceeded 0.2mm;
if over 20 individual frames featured movement in excess
of 0.25mm (Satterthwaite et al., 2012); or if brain cov-
erage during acquisition was incomplete. As a result of
these criteria, participants with gross in-scanner motion
were excluded, allowing us to evaluate the utility of con-
found regression strategies for the mitigation of artifact
due to micro-movements.
Structural and functional subject data were acquired
on a 3T Siemens Tim Trio scanner with a 32-channel
head coil (Erlangen, Germany), as previously described
(Satterthwaite et al., 2014, 2016). High-resolution struc-
tural images were acquired in order to facilitate align-
ment of individual subject images into a common space.
Structural images were acquired using a magnetization-
prepared, rapid-acquisition gradient-echo (MPRAGE) T1-
weighted sequence (TR = 1810ms; TE = 3.51ms; FoV
= 180 × 240mm; resolution 1mm isotropic). Approxi-
mately 6min of task-free functional data were acquired
for each subject using a blood oxygen level-dependent
(BOLD-weighted) sequence (TR = 3000ms; TE = 32ms;
FoV = 192 × 192mm; resolution 3mm isotropic; 124 spa-
tial volumes). Prior to scanning, in order to acclimate
subjects to the MRI environment and to help subjects
learn to remain still during the actual scanning session,
a mock scanning session was conducted using a decom-
missioned MRI scanner and head coil. Mock scanning
was accompanied by acoustic recordings of the noise pro-
duced by gradient coils for each scanning pulse sequence.
2
During these sessions, feedback regarding head movement
was provided using the MoTrack motion tracking system
(Psychology Software Tools, Inc, Sharpsburg, PA). Motion
feedback was only given during the mock scanning session.
In order to further minimize motion, prior to data acquisi-
tion subjects' heads were stabilized in the head coil using
one foam pad over each ear and a third over the top of
the head. During the resting-state scan, a fixation cross
was displayed as images were acquired. Subjects were in-
structed to stay awake, keep their eyes open, fixate on the
displayed crosshair, and remain still.
Structural image processing
A study-specific template was generated from a sam-
ple of 120 PNC subjects balanced across sex, race, and
age bins using the buildTemplateParallel procedure in
ANTs (Avants et al., 2011a). Study-specific tissue priors
were created using a multi-atlas segmentation procedure
(Wang et al., 2014). Next, each subject's high-resolution
structural image was processed using the ANTs Cortical
Thickness Pipeline (Tustison et al., 2014). Following bias
field correction (Tustison et al., 2010), each structural im-
age was diffeomorphically registered to the study-specific
PNC template using the top-performing SyN deformation
(Klein et al., 2009). Study-specific tissue priors were used
to guide brain extraction and segmentation of the subject's
structural image (Avants et al., 2011b).
BOLD time series processing
Task-free functional images were processed using the
XCP Engine (Ciric et al., In Preparation), which was con-
figured to support the 12 pipelines evaluated in this study
(see Figure 1). Each pipeline was based on de-noising
strategies previously described in the neuroimaging lit-
erature. A number of preprocessing procedures were in-
cluded across all de-noising pipelines. Common elements
of preprocessing included (1) correction for distortions in-
duced by magnetic field inhomogeneities, (2) removal of
the 4 initial volumes of each acquisition, (3) realignment of
all volumes to a selected reference volume using mcflirt
(Jenkinson et al., 2002), (4) demeaning and removal of
any linear or quadratic trends, (5) co-registration of func-
tional data to the high-resolution structural image using
boundary-based registration (Greve and Fischl, 2009), and
(6) temporal filtering using a first-order Butterworth fil-
ter with a passband between 0.01 and 0.08 Hz. These
preliminary processing stages were then followed by the
confound regression procedures described below.
In or-
der to prevent frequency-dependent mismatch during con-
found regression (Hallquist et al., 2013), all regressors were
band-pass filtered to retain the same frequency range as
the data. As in our prior work (Satterthwaite et al., 2012,
2013), the primary summary metric of subject motion used
was the mean relative root-mean-square displacement cal-
culated during time series realignment using mcflirt.
Overview of confound regression strategies
The primary objective of the current study was to evalu-
ate the performance of common de-noising strategies. We
selected 12 de-noising models, labelled 1–12 below, for
evaluation (Figure 1). Models 1–5 used nuisance parame-
ters derived from 6 movement estimates and 3 physiolog-
ical time series, as well as their temporal derivatives and
quadratic expansions.
• Model 1. (2P) Used only the 2 physiological time se-
ries: mean signal in WM and mean signal in CSF, and
functioned as a base model for comparison to other
more complex confound regression models.
• Model 2. (6P) Used only the 6 motion estimates de-
rived from mcflirt realignment as explanatory vari-
ables.
• Model 3.
(9P) Combined the 6 motion estimates
and 2 physiological time series with global signal re-
gression. This model has been widely applied to func-
tional connectivity studies (Fox et al., 2005, 2009).
• Model 4.
(24P) Expansion of model 2 that in-
cludes 6 motion parameters, 6 temporal derivatives,
6 quadratic terms, and 6 quadratic expansions of the
derivatives of motion estimates for a total 24 regres-
sors (Friston et al., 1996).
• Model 5.
(36P) Similar expansion of model 3:
9 regressors, their derivatives, quadratic terms, and
squares of derivatives (Satterthwaite et al., 2013).
Models 6–8 further expanded upon this maximal 36P
strategy by incorporating censoring approaches.
• Model 6. (36P+despike) Included 36 regressors as
well as despiking (Cox, 1996).
• Model 7.
(36P+spkreg) Included 36 regressors as
well as spike regression, as in Satterthwaite et al.
(2013).
• Model 8. (36P+scrub) Included 36 regressors as well
as motion scrubbing, as in Power et al. (2014a).
Models 9 and 10 adapted variants of the PCA-based
CompCor approach.
• Model 9. (aCompCor) Used 5 principal components
each from the WM and CSF, in addition to motion
estimates and their temporal derivatives (Muschelli
et al., 2014).
• Model 10.
(tCompCor) Used 6 principal compo-
nents from high-variance voxels (Behzadi et al., 2007).
The final two models evaluated used a regressor de-
rived from local WM signals, and subject-specific ICA de-
noising, respectively.
3
• Model 11.
(wmLocal) Used a voxelwise, localised
WM regressor in addition to motion estimates and
their temporal derivatives and despiking (Jo et al.,
2013).
• Model 12. (ICA-AROMA) Used a recently devel-
oped ICA-based procedure for removal of motion-
related variance from BOLD data, together with mean
WM and CSF regressors (Pruim et al., 2015a,b).
We explicitly limited our scope to models that did not
require training a classifier, and did not evaluate confound
regression strategies that require extensive parameter op-
timization (Salimi-Khorshidi et al., 2014; Griffanti et al.,
2014; Patel et al., 2014). Furthermore, in order to con-
strain the parameter space, we did not examine unpub-
lished combinations of de-noising approaches.
Confound regression using realignment parameters
Time series of six realignment parameters (three transla-
tional and three rotational) for each subject were returned
by mcflirt as part of time series realignment (motion cor-
rection). Additionally, the temporal derivative, quadratic
terms, and quadratic of the temporal derivative of each of
the realignment parameters were calculated, yielding 24 re-
alignment regressors in total. The original six realignment
parameters were included in confound regression models 2
and 3. Models 9 and 11 included 12 realignment regressors
– the 6 realignment parameters and their temporal deriva-
tives – while the full set of 24 expanded realignment regres-
sors were included as part of confound regression models
4–8.
Global signal regression
The mean global signal was computed by averaging
across all voxelwise time series located within a subject-
specific mask covering the entire brain. The global signal
was included in model 3, while the expanded models 5–8
included 4 global signal regressors: the global signal, its
derivative, its square, and the derivative of its square.
Tissue class regressors
Mean white matter (WM) and cerebrospinal fluid (CSF)
signals were computed by averaging within masks derived
from the segmentation of each subject's structural image;
these masks were eroded using AFNI's 3dmask tool (Cox,
1996) to prevent inclusion of gray matter signal via partial-
volume effects. The WM mask was eroded at the 2-voxel
level, while the CSF mask was eroded at the 1-voxel level.
More liberal erosion often resulted in empty masks in our
data. Temporal derivatives, quadratic terms, and squares
of the derivative were computed as above. Two tissue class
regressors (WM and CSF) were included in models 3 and
12, whereas their expansions (8 regressors) were included
in models 4–8.
Local white matter regression
Model 11 used a local WM regressor (Jo et al., 2013).
This was computed in AFNI using 3dLocalstat (Cox,
1996). Unlike the regressors described above, which were
voxel-invariant, the value of the local WM regressor was
computed separately at each voxel. For each voxel, a
sphere of radius 45mm was first centered on that voxel;
this sphere defined that voxel's local neighborhood. Next,
this spherical neighborhood was intersected with an eroded
WM mask to produce a local WM mask, which included
only the fraction of the WM that was also in the voxel's
neighborhood. The mean signal within this new local WM
mask was then used to model the local WM signal at the
voxel (Jo et al., 2013). This process was repeated at every
voxel in order to generate the local WM regressor. This
local WM regressor was included in model 11 along with
realignment parameters and their derivatives (12 total);
this model also included voxelwise de-spiking.
CompCor
Principal component analysis (PCA) can be used to
model noise in BOLD time series (Behzadi et al., 2007;
Muschelli et al., 2014). Broadly, the use of PCA-derived
regressors to model noise is called component-based cor-
rection (CompCor). Numerous variants of CompCor have
been developed; here, our focus will be on anatomical
CompCor (aCompCor, model 9) and temporal CompCor
(tCompCor, model 10).
In aCompCor, a PCA is per-
formed within an anatomically defined tissue class of in-
terest. We extracted 5 components for WM and CSF
each, yielding 10 compcor components (Muschelli et al.,
2014). As part of model 9, as in Muschelli et al. (2014),
these 10 aCompCor components were combined with 12
re-alignment parameters (raw and temporal derivative).
In tCompCor, the temporal variance of the BOLD signal
is first computed at each voxel. Subsequently, a mask is
generated from high-variance voxels, and principal compo-
nents are extracted from the time series at these voxels.
In confound regression model 10, tCompcor was imple-
mented using ANTs, with 6 tCompCor components used
as confound regressors for each participant.
ICA-AROMA
ICA-AROMA (automatic removal of motion artifact) is
a recently-introduced, widely-used method for de-noising
using single-subject ICA (Pruim et al., 2015a,b); we eval-
uated ICA-AROMA in confound regression model 12. In
contrast to other ICA based methods (e.g., ICA-FIX:
Salimi-Khorshidi et al. (2014)), it does not require dataset-
specific training data. The input to ICA-AROMA is a vox-
elwise time series that has been smoothed at 6mm FWHM
using a Gaussian kernel. After decomposing this time se-
ries using FSL's melodic (with model order estimated us-
ing the Laplace approximation) (Beckmann et al., 2005),
ICA-AROMA uses four features to determine whether
each component corresponds to signal or noise. The first 2
features are spatial characteristics of the signal source: (1)
4
the fraction of the source that falls within a CSF compart-
ment and (2) the fraction of the source that falls along
the edge or periphery of the brain. The remaining fea-
tures are derived from the time series of the source: (3) its
maximal robust correlation with time series derived from
realignment parameters and (4) its high-frequency spec-
tral content. ICA-AROMA includes two de-noising steps.
The first de-noising step occurs immediately after classi-
fication. All component time series (signal and noise) are
included as predictors in the linear model, and the resid-
ual BOLD time series is obtained via partial regression of
only the noise time series. A second confound regression
step occurs after temporal filtering, wherein mean signals
from WM and CSF were regressed from the data.
Temporal censoring:
scrubbing
de-spiking, spike regression, and
In addition to regression of nuisance time series, a num-
ber of 'temporal censoring' approaches were used to iden-
tify motion-contaminated volumes in the BOLD time se-
ries and reduce their impact on further analysis. These ap-
proaches included despiking, spike regression, and scrub-
bing. Despiking is a procedure that identifies outliers in
the intensity of each voxel's detrended BOLD time series
and then interpolates over these outliers. Despiking was
implemented in AFNI using the 3dDespike utility (Cox,
1996) as part of confound regression model 6.
Unlike despiking, which identifies outliers on a voxel-
wise basis, spike regression and scrubbing censor complete
volumes based on metrics of subject movement defined
a priori. For spike regression, as in Satterthwaite et al.
(2013), volumes were flagged for spike regression if their
volume-to-volume RMS displacement exceeded 0.25mm.
Next, as part of confound regression model 7, k 'spike'
regressors were included as predictor variables in the de-
noising model, where k equalled the number of volumes
flagged (Satterthwaite et al., 2013). For each flagged time
point, a unit impulse function that had a value of 1 at
that time point and 0 elsewhere was included as a spike
regressor.
For scrubbing, the framewise displacement (FD) (Power
et al., 2012) was computed at each time point as the sum
of the absolute values of the derivatives of translational
and rotational motion estimates.
If framewise displace-
ment (FD) at any point in time exceeded 0.2mm, then
that time point was flagged for scrubbing.
It should be
noted that the conversion of FD to RMS displacement is
approximately 2:1, and thus the published criterion for
scrubbing has a lower threshold for flagging high-motion
volumes than does spike regression. Scrubbing of BOLD
data was performed iteratively (Power et al., 2014a) as
part of confound regression model 8. At any stage where
a linear model was applied to the data (for instance, dur-
ing detrending procedures), high-motion epochs were tem-
porally masked out of the model so as not to influence
fit. During temporal filtering, a frequency transform was
used to generate surrogate data with the same phase and
spectral characteristics as the unflagged data. This sur-
rogate data was used to interpolate over flagged epochs
prior to application of the filter. During confound regres-
sion, flagged timepoints were excised from the time series
so as not to contribute to the model fit. For scrubbing (but
not spike regression) if fewer than five contiguous volumes
had unscrubbed data, these volumes were scrubbed and
interpolated as well.
Overview of outcome measures
We evaluated each de-noising pipeline according to three
benchmarks. Residual QC-FC correlations and distance-
dependence provided a metric of each pipeline's efficacy,
while loss of temporal DOF provided an estimate of each
pipeline's efficiency.
QC-FC correlations
In order to estimate the residual relationship between
subject movement and connectivity after de-noising, we
computed QC-FC correlations (quality control / func-
tional connectivity) (Power et al., 2015; Satterthwaite
et al., 2012, 2013; Power et al., 2012). While other met-
rics have been used in prior reports, including FD-DVARS
correlations, we favor QC-FC as the most useful metric of
interest as it directly quantifies the relationship between
motion and the primary outcome of interest (rather than
two quality metrics, as in FD-DVARS). For an extended
discussion of the rationale for this measure, see Power et al.
(2015).
We
evaluated QC-FC relationships within two
commonly-used whole-brain networks, the first consisting
of spherical nodes distributed across the brain (Power
et al., 2011) and the second comprising an areal par-
cellation of the cerebral cortex (Gordon et al., 2016).
For each network, the mean time series for each node
was calculated from the denoised residual data, and the
pairwise Pearson correlation coefficient between node
time series was used as the network edge weight (Biswal
et al., 1995). For each edge, we then computed the
correlation between the weight of that edge and the
mean relative RMS motion. To eliminate the potential
influence of demographic factors, QC-FC relationships
were calculated as partial correlations that accounted
for participant age and sex. We thus obtained, for each
de-noising pipeline, a distribution of QC-FC correlations.
This distribution was used to obtain two measures of the
pipeline's ability to mitigate motion artifact, including:
1) the number of edges significantly related to motion,
which was computed after using the false discovery rate
(FDR) to account for multiple comparisons; and 2) the
median absolute value of all QC-FC correlations. All
graphs were generated using ggplot2 in R version 3.2.3;
brain renderings were prepared in BrainNet Viewer (Xia
et al., 2013).
5
Distance-dependent effects of motion
Early work on motion artifact demonstrated that in-
scanner motion can bias connectivity estimates between
two nodes in a manner that is related to the distance
between those nodes (Satterthwaite et al., 2012; Power
et al., 2012). Under certain processing conditions, subject
movement enhances short-distance connections while re-
ducing long-distance connections. To determine the resid-
ual distance-dependence of subject movement, we first
used the center of mass of each node to obtain a distance
matrix D where entry Dij indicates the Euclidean dis-
tance between the centers of mass of nodes i and j. We
then obtained the correlation between the distance sepa-
rating each pair of nodes and the QC-FC correlation of
the edge connecting those nodes; this correlation served as
an estimate of the distance-dependence of motion artifact.
Additional degrees of freedom lost in confound regression
Including additional regressors in a confound model re-
duces the impact of motion on future analyses, but it is
not without cost. Confound regressors and censoring both
reduce the temporal degrees of freedom (DOF) in data.
This loss in temporal DOF may diminish the ability of
functional data acquisitions to sample a subject's connec-
tome or may introduce bias if the loss is variable across
subjects. Thus, de-noising strategies ideally limit the loss
of temporal DOF, for instance by including fewer, more
efficacious regressors.
In the present study, we also as-
sessed the number of temporal DOF lost in each confound
regression approach.
As in previous work (Pruim et al., 2015a), we assumed
that each time series regressed out and each volume ex-
cised from the data constituted a single temporal DOF.
Consequently, the loss of temporal DOF was estimated
as the sum of the number of regressors in each confound
model and the number of volumes flagged for excision un-
der that model.
It should be emphasized that the val-
ues thus obtained are imperfect estimates. First, because
functional MR time series typically exhibit temporal au-
tocorrelation, the actual loss in DOF will be less than the
estimated loss in DOF. Accordingly, censoring adjacent
volumes does not remove the same number of DOF as
does censoring volumes separated in time. Furthermore,
a temporal bandpass filter was uniformly applied to all
data prior to confound regression; this filtering procedure
would itself have removed additional temporal DOF and
elevated the autocorrelation of the data. Because this fil-
ter was uniform across all de-noising strategies, it was not
considered when estimating the loss of additional temporal
DOF in each model.
Results
Heterogeneity in confound regression performance
Confound regression strategies typically remove some,
but not all, of the artifactual variance that head motion
6
introduces into the BOLD signal. The motion-related arti-
fact that survives de-noising can be quantified to provide a
metric of pipeline performance. Here, our primary bench-
mark of confound regression efficacy was the residual rela-
tionship between brain connectivity and subject motion, or
the QC-FC correlation. We measured QC-FC correlations
using two metrics: the percentage of network connections
where a significant relationship with motion was present
(Figure 2), and the absolute median correlation (AMC)
between connection strength and head movement across
all connections (Figure 3).
No preprocessing strategy was completely effective in
abolishing the relationship between head movement and
connectivity. However, different approaches exhibited
widely varying degrees of efficacy. The top four con-
found regression strategies included 36 parameters, com-
prising an expansion of GSR, tissue-specific regressors
(WM, CSF), and realignment parameters. Beyond this
base 36-parameter model, all censoring techniques pro-
vided provided some additional benefit, reducing the num-
ber of edges that were significantly related to motion to
less than 1%. Convergent results were present across both
QC-FC measures (% edges, AMC) and networks (Power,
Gordon) that were evaluated.
In contrast, many pipelines performed relatively poorly,
leaving a majority of network edges with a residual re-
lationship with motion. Specifically, 82% of edges were
impacted by motion when the least effective method was
used (6 realignment parameters). The commonly used 24-
parameter expansion of realignment parameters originally
suggested by Friston et al. (1996) did not provide much
of an improvement (79% edges). Similarly, the local WM
regressor model (69% edges) and tCompCor (48% edges)
also resulted in substantial residual QC-FC correlations.
In fact, these methods performed worse than a basic 2-
parameter model composed of mean WM and CSF signals
(33% edges). Finally, several methods demonstrated in-
termediate performance, with 1-20% of edges impacted by
motion. This middle group included methods as disparate
as aCompCor (5% edges), ICA-AROMA (12% edges), and
the classic 9-parameter confound regression model which
included GSR (4% edges).
Variability in distance-dependent motion artifact after
confound regression
Our
second benchmark quantified the distance-
dependent motion artifact that was present in data pro-
cessed by each pipeline (Figure 4). We observed that
distance-dependence was present even under conditions
where artifact magnitude was attenuated. For example,
though the 36-parameter model was among the most ef-
fective in attenuating QC-FC relationships, its application
revealed strongly distance-dependent artifact. Examina-
tion of graphs that plot QC-FC by Euclidean distance (see
Figure 4C) revealed that this is due to effective mitiga-
tion of motion artifact for long-range but not short-range
connections.
Distance-dependence was most prominent in models
that included GSR, but did not include censoring (e.g.,
9-parameter and 36-parameter models). However, despite
the lack of global signal in the aCompCor or tCompcor
models, data returned from both of these component-
based approaches revealed substantial distance-dependent
artifact. Notably, inclusion of censoring consistently re-
duced distance-dependence, although scrubbing was some-
what more effective than spike regression or voxelwise de-
spiking.
The top performing method according to this bench-
mark was ICA-AROMA, which completely abolished any
distance-dependence of residual motion artifact. In other
words, the motion artifact that was still present in the data
after ICA-AROMA impacted all connections in a manner
that was not dependent on the spatial separation between
nodes. There was similar lack of distance-dependence in
the wmLocal model, although as noted above this model
did not provide effective de-noising according to QC-FC
benchmarks. Despite the presence of the global signal,
the 36-parameter model that included scrubbing also per-
formed quite well.
Effective preprocessing strategies use many additional de-
grees of freedom
Perhaps unsurprisingly,
the preprocessing strategies
that consistently reduced both QC-FC correlations and
distance-dependence were also among the costliest in
terms of loss of temporal degrees of freedom (Figure 5).
By definition, the 36-parameter models included a high
fixed number of regressors. Furthermore, models that
additionally included censoring resulted in a substantial
additional loss of data that varied across subjects. ICA-
AROMA also had a variable loss of DOF, but of a lower
magnitude than censoring or high-parameter confound re-
gression.
Discussion
In response to rapid evolution of confound regression
strategies available for the mitigation of motion artifact,
in this report we evaluated 12 commonly-used pipelines.
Results indicate that there is substantial heterogeneity in
the performance of these confound regression techniques
across all measures evaluated. The context, implications,
and limitations of these findings are discussed below.
Confound regression techniques have substantial perfor-
mance variability
We evaluated confound regression strategies according
to three intuitive benchmarks that were selected to capture
different domains of effectiveness. These included QC-FC
associations, distance-dependence of motion artifact, and
additional degrees of freedom lost in confound regression.
Across each benchmark, there was a striking heterogeneity
in pipeline performance.
Notably, five of the top six confound regression ap-
proaches included GSR. This effect was consistent in both
networks we evaluated. The effectiveness of GSR is most
likely due to the nature of motion artifact itself: in-scanner
head motion tends to induce widespread reductions in sig-
nal intensity across the entire brain parenchyma (see Sat-
terthwaite et al. (2013), Figure 4). As discussed in de-
tail elsewhere (Power et al., Under Revision) this effect
is highly reproducible across datasets, and is effectively
captured by time series regression of the global signal.
Beyond GSR, a second strategy that clearly minimizes
QC-FC relationships is temporal censoring. We evaluated
three censoring variants, including scrubbing, spike regres-
sion, and de-spiking. Compared to spike regression and
de-spiking, scrubbing appears to be more effective in re-
moving distance-dependent artifact in this dataset. This is
most likely due to the explicit tension between data qual-
ity and data quantity: because of the lower threshold for
scrubbing than spike regression (due to differences in FD
vs. RMS measures of motion; see Figure 9C in Yan et al.
(2013a)), more low-quality data was excised during scrub-
bing. Furthermore, scrubbing includes a criterion to not
leave isolated epochs (< 5 volumes) of un-scrubbed data.
Consequently, this leads to clear differences in the addi-
tional degrees of freedom lost by each method. In contrast
to spike regression and scrubbing, which eliminate high
motion volumes completely, time series de-spiking identi-
fies and interpolates large changes in signal intensity on a
voxelwise basis (Cox, 1996). This allows for spatial adap-
tivity (see Patel et al. (2014)) but also renders quantifi-
cation of data loss and comparisons with volume-based
censoring techniques more difficult.
Critically, while both GSR and censoring appeared ef-
fective in minimizing QC-FC relationships, they exhibited
opposite effects on distance-dependence. While censor-
ing techniques appear to consistently reduce the presence
of distance-dependence, GSR is associated with increased
distance-dependence. However, it should be emphasized
that the distance-dependence associated with GSR is not
the result of worsening associations with motion in certain
connections. Rather, the distance-dependence seen with
GSR stems from differential de-noising efficacy, whereby
motion artifact is more effectively minimized for long-
distance connections than for short-range connections.
Certain models such as the local WM regression approach
(Jo et al., 2013) thus have minimal distance-dependence,
but this is a consequence of lack of efficacy across all dis-
tances. In contrast, ICA-AROMA (Pruim et al., 2015b,a)
reduced motion to a moderate degree over all connec-
tion distances, resulting in almost no distance-dependence.
However, while clearly an improvement over some other
methods, data processed using ICA-AROMA was nois-
ier than other methods which included GSR or censor-
ing, and resulting networks contained a substantial num-
ber of edges impacted by motion. Somewhat to our sur-
prise, benchmark results for aCompCor (Behzadi et al.,
2007; Muschelli et al., 2014) were most similar to models
7
that included GSR. Alone among models where GSR was
not included, aCompCor was both relatively effective in
the mitigation of residual motion (5% of edges impacted)
and also exhibited substantial distance-dependence (e.g.,
r = −0.26). This suggests that while aCompCor does not
explicitly include GSR, the practical results of its applica-
tion are in fact quite similar.
Trade-offs of confound regression approaches: implications
for investigators
The current results emphasize two clear trade-offs in
the choice of confound regression strategy. First, pipelines
that include global signal regression tend to be more effec-
tive at minimizing QC-FC relationships, but at the cost
of some increase in distance-dependence. As noted above,
for minimizing QC-FC relationships, nearly all of the top
strategies (except aCompCor) included GSR. Conversely,
the two techniques that had the most substantial distance-
dependence (the 9-regressor and 36-regressor methods)
both included GSR. Second, censoring techniques provide
a clear benefit in reducing QC-FC relationships and addi-
tionally tend to attenuate distance-dependence. However,
by definition, removing contaminated volumes results in
less data and loss of degrees of freedom.
These two trade-offs suggest that a single confound re-
gression strategy is unlikely to be optimal for every study.
For example, in studies of network organization, the pres-
ence of both anti-correlations and distance-dependent ar-
tifact resulting from GSR may result in an undesirable
impact on nodal degree distribution (Yan et al., 2013b).
In contrast,
for studies of group or individual differ-
ences, minimizing QC-FC relationships is likely to be of
paramount importance so as to limit the possibility that
inference is driven by artifactual signals. This concern is
particularly relevant for studies of development or clinical
sub-groups where motion is systematically related to the
subject-level variable of interest (e.g., age, disease status).
Co-varying for motion at the group level is unlikely to be
a panacea for such studies when inadequate subject-level
time series de-noising is employed, as prior work (Power
et al., 2014a) has suggested that motion effects at the
group level may potentially both be nonlinear and vary
across sub-samples in a manner that is difficult to predict.
However, aggressive volume censoring may be problematic
in datasets with relatively brief acquisitions. In datasets
where long time series are acquired, such as multi-band ac-
quisitions (Feinberg et al., 2010) and intensive acquisitions
of single subjects (Laumann et al., 2015), loss of tempo-
ral degrees of freedom is less likely to be a major concern.
The 36-parameter models without volume censoring offer
uniformity, as does randomly or systematically censoring
additional volumes until all subjects retain approximately
the same degrees of freedom.
Limitations
Several limitations of the current approach should be
noted. One of the principal challenges in evaluating the
performance of de-noising approaches is the lack of a noise-
free ground truth. Our primary benchmark of confound
regression performance assumes that mitigation of the re-
lationship between QC (participant motion) and FC (i.e.,
the imaging measurement) is desirable. To the degree that
in-scanner motion itself represents a biologically informa-
tive phenotype, this approach will mistake signal for noise.
Indeed, prior data suggests that this may sometimes be
the case. For example, Zeng et al. (2014) found specific
changes in connectivity for participants who had generally
high levels of motion, even on scans where motion was low.
However, without multiple scans which allow such careful
dissociation, most studies are incapable of disambiguating
the large confounding effects of motion on connectivity.
Second, in place of QC-FC relationships, one could focus
on alternative benchmarks such as test-retest reliability
(Zuo et al., 2014). Reliability is certainly of interest, but
to the degree that motion tends to be highly correlated
within individuals across scanning sessions, there is a sub-
stantial potential for the presence of consistent motion ar-
tifact across sessions to artificially inflate reliability, and
diminish the biological relevance of observed results. A
third and related concern is that certain de-noising meth-
ods could conceivably both minimize QC-FC relationships
and even enhance reliability by aggressively removing both
signal and noise, but in the process diminish sensitivity to
meaningful individual differences. Indeed, one prior study
demonstrated the association between canonical resting
state networks and randomly generated confound parame-
ters (Bright and Murphy, 2015). This concern is somewhat
mitigated by other studies in independent datasets, which
suggest that higher-order confound regressors improve the
confound regression model fit (Yan et al., 2013a; Satterth-
waite et al., 2013), while random regressors do not (see Fig-
ure 8 in Satterthwaite et al. (2013)). Nonetheless, the ten-
sion between the goals of noise reduction and sensitivity to
individual differences remains an outstanding issue for the
field. Fourth, while our evaluation included many of the
most commonly used techniques, other approaches which
require substantial training or parameter selection (i.e.,
ICA-FIX (Salimi-Khorshidi et al., 2014; Griffanti et al.,
2014), wavelet de-spiking (Patel et al., 2014)) may be valu-
able and merit further consideration. Fifth and finally, it
should be noted that while improvements in image acqui-
sition (including multi-echo techniques) may not salvage
existing motion-contaminated data, it is likely that they
will change the methodological landscape of connectivity
research moving forward (Kundu et al., 2012, 2013; Bright
and Murphy, 2013).
Conclusions
Taken together, the present results underline the per-
formance heterogeneity of recently-introduced, commonly-
used confound regression methods.
In selecting among
these methods, investigators should be aware of the rel-
ative strengths and weaknesses of each approach, and un-
derstand how processing strategy may impact inference.
8
Clearly, the relative merit of each approach will vary by
research question and study design. Perhaps most im-
portantly, as has been emphasized in nearly every other
study of motion artifact, the choice of confound regres-
sion strategy is often dwarfed in importance by the need
to transparently report and evaluate the impact of mo-
tion in each dataset. At a minimum, this includes report-
ing the relationship between motion artifact and not only
subject phenotypes (e.g., group, age, symptom or cogni-
tive score) but also the functional connectivity phenotypes
being considered. In the context of such data, the distinc-
tion between observed results and the impact of motion
artifact can be understood. Such transparency bolsters
confidence in reported findings, but also will likely tend
to emphasize the remaining challenges for de-noising go-
ing forward. Especially when considered in the context of
the the rapid evolution of available techniques since 2012,
there is no doubt that innovations in post-processing con-
found regression strategies will continue.
Acknowledgements
Thanks to the acquisition and recruitment team,
in-
cluding Karthik Prabhakaran and Jeff Valdez. Thanks to
Chad Jackson for data management and systems support.
Thanks to Monica Calkins for phenotyping expertise. Sup-
ported by grants from the National Institute of Men-
tal Health: R01MH107703 (TDS), R01MH107235 (RCG),
and R01NS089630 (CD). The PNC was funded through
NIMH RC2 grants MH089983, and MH089924 (REG).
Additional support was provided by R21MH106799 (DSB
& TDS), R01MH101111 (DHW), K01MH102609 (DRR),
P50MH096891 (REG), R01NS085211 (RTS), and the
Dowshen Program for Neuroscience. DSB acknowledges
support from the John D. and Catherine T. MacArthur
Foundation, the Alfred P. Sloan Foundation, the Army Re-
search Laboratory and the Army Research Office through
contract numbers W911NF-10-2-0022 and W911NF-14-
1-0679, the National Institute of Mental Health (R01-
DC-009209-11), the National Institute of Child Health
and Human Development (R01HD086888-01), the Of-
fice of Naval Research, and the National Science Foun-
dation (BCS-1441502 and PHY-1554488). Support for
developing statistical analyses (RTS & TDS) was pro-
vided by a seed grant by the Center for Biomedi-
cal Computing and Image Analysis (CBICA) at Penn.
Data deposition: The data reported in this paper have
been deposited in database of Genotypes and Pheno-
types (dbGaP), www.ncbi.nlm.nih.gov/gap (accession no.
phs000607.v1.p1).
References
References
Avants, B.B., Tustison, N.J., Song, G., Cook, P.A., Klein, A., Gee,
J.C., 2011a. A reproducible evaluation of ANTs similarity metric
9
performance in brain image registration. NeuroImage 54, 2033–
2044. doi:10.1016/j.neuroimage.2010.09.025.
Avants, B.B., Tustison, N.J., Wu, J., Cook, P.A., Gee, J.C., 2011b.
An open source multivariate framework for N-tissue segmenta-
tion with evaluation on public data. Neuroinformatics 9, 381–400.
doi:10.1007/s12021-011-9109-y.
Baker, J.T., Holmes, A.J., Masters, G.a., Yeo, B.T.T., Krienen, F.,
Buckner, R.L., Ongur, D., 2014. Disruption of cortical association
networks in schizophrenia and psychotic bipolar disorder. JAMA
psychiatry 71, 109–18. doi:10.1001/jamapsychiatry.2013.3469,
arXiv:15334406.
Beckmann, C.F., Deluca, M., Devlin, J.T., Smith, S.M., 2005. Inves-
tigations into resting-state connectivity using independent compo-
nent analysis. Philos Trans R Soc Lond B Biol Sci 360, 1001–13.
doi:10.1098/rstb.2005.1634.
Behzadi, Y., Restom, K., Liau, J., Liu, T.T., 2007. A compo-
nent based noise correction method (CompCor) for BOLD and
perfusion based fMRI. NeuroImage 37, 90–101. doi:10.1016/j.
neuroimage.2007.04.042, arXiv:NIHMS150003.
Biswal, B., Yetkin, F.Z., Haughton, V.M., Hyde, J.S., 1995. Func-
tional connectivity in the motor cortex of resting human brain
using echo-planar MRI. Magnetic Resonance in Medicine 34, 537–
541. doi:10.1002/mrm.1910340409.
Bright, M.G., Murphy, K., 2013. Removing motion and physiolog-
ical artifacts from intrinsic BOLD fluctuations using short echo
data. NeuroImage 64, 526–537. doi:10.1016/j.neuroimage.2012.
09.043.
Bright, M.G., Murphy, K., 2015.
Is fMRI "noise" really noise?
Resting state nuisance regressors remove variance with network
structure. NeuroImage 114, 158–169. doi:10.1016/j.neuroimage.
2015.03.070.
Buckner, R.L., Andrews-Hanna, J.R., Schacter, D.L., 2008. The
brain's default network: Anatomy, function, and relevance to dis-
ease. Annals of the New York Academy of Sciences 1124, 1–38.
doi:10.1196/annals.1440.011.
Chai, X.J., Castan´an, A.N., Ongur, D., Whitfield-Gabrieli, S., 2012.
Anticorrelations in resting state networks without global signal re-
gression. NeuroImage 59, 1420–1428. doi:10.1016/j.neuroimage.
2011.08.048.
Cox, R., 1996. AFNI: Software for analysis and visualization of
functional magnetic resonance neuroimages. Comput Biomed Res
29, 162–173.
Craddock, R.C., Jbabdi, S., Yan, C.G., Vogelstein, J.T., Castel-
lanos, F.X., Di Martino, A., Kelly, C., Heberlein, K., Colcombe,
S., Milham, M.P., 2013.
Imaging human connectomes at the
macroscale. Nature Methods 10, 524–539. doi:10.1038/nmETh.
2482, arXiv:NIHMS150003.
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 of the United States of America
103, 13848–53. doi:10.1073/pnas.0601417103.
DiMartino, A., Fair, D.A., Kelly, C., Satterthwaite, T.D., Castel-
lanos, F.X., Thomason, M.E., Craddock, R.C., Luna, B., Lev-
enthal, B.L., Zuo, X.N., Milham, M.P., 2014. Unraveling the
miswired connectome: A developmental perspective. Neuron 83,
1335–1353. doi:10.1016/j.neuron.2014.08.050.
Dosenbach, N.U.F., Nardos, B., Cohen, A.L., Fair, D.A., Power,
D., Church, J.a., Nelson, S.M., Wig, G.S., Vogel, A.C., Lessov-
schlaggar, C.N., Barnes, K.A., Dubis, J.W., Feczko, E., Coalson,
R.S., Jr, J.R.P., Barch, D.M., Petersen, S.E., Schlaggar, B.L.,
2011. Prediction of individual brain maturity using fMRI. Science
329, 1358–1361. doi:10.1126/science.1194144.Prediction.
Fair, D.A., Cohen, A.L., Dosenbach, N.U.F., Church, J.A., Miezin,
F.M., Barch, D.M., Raichle, M.E., Petersen, S.E., Schlaggar, B.L.,
2008. The maturing architecture of the brain's default network.
Proceedings of the National Academy of Sciences of the United
States of America 105, 4028–4032. doi:10.1073/pnas.0800376105.
Fair, D.A., Nigg, J.T., Iyer, S., Bathula, D., Mills, K.L., Dosenbach,
N.U.F., Schlaggar, B.L., Mennes, M., Gutman, D., Bangaru, S.,
Buitelaar, J.K., Dickstein, D.P., Di Martino, A., Kennedy, D.N.,
Kelly, C., Luna, B., Schweitzer, J.B., Velanova, K., Wang, Y.F.,
Mostofsky, S., Castellanos, F.X., Milham, M.P., 2012. Distinct
neural signatures detected for ADHD subtypes after controlling
for micro-movements in resting state functional connectivity MRI
data. Frontiers in systems neuroscience 6, 80. doi:10.3389/fnsys.
2012.00080.
Fair, D.A., Posner, J., Nagel, B.J., Bathula, D., Dias, T.G.C., Mills,
K.L., Blythe, M.S., Giwa, A., Schmitt, C.F., Nigg, J.T., 2010.
Atypical default network connectivity in youth with attention-
deficit/hyperactivity disorder. Biological psychiatry 68, 1084–91.
doi:10.1016/j.biopsych.2010.07.003.
Feinberg, D.A., Moeller, S., Smith, S.M., Auerbach, E.J., Ramanna,
S., Glasser, M.F., Miller, K.L., Ugurbil, K., Yacoub, E.S., 2010.
Multiplexed echo planar imaging for sub-second whole brain fMRI
and fast diffusion imaging. PLoS ONE 5, e15710. doi:10.1371/
journal.pone.0015710.
Fox, M.D., Raichle, M.E., 2007. Spontaneous fluctuations in brain
activity observed with functional magnetic resonance imaging.
Nat Rev Neurosci 8, 700–711. doi:10.1038/nrn2201.
Fox, M.D., Snyder, A.Z., Vincent, J.L., Corbetta, M., Van Es-
sen, D.C., Raichle, M.E., 2005. The human brain is intrinsi-
cally organized into dynamic, anticorrelated functional networks.
Proceedings of the National Academy of Sciences of the United
States of America 102, 9673–8. doi:10.1073/pnas.0504136102,
arXiv:NIHMS150003.
Fox, M.D., Zhang, D., Snyder, A.Z., Raichle, M.E., 2009. The
global signal and observed anticorrelated resting state brain net-
works. Journal of neurophysiology 101, 3270–3283. doi:10.1152/
jn.90777.2008.
Friston, K., Williams, S., Howard, R., Frackowiak, R.S.J., Turner,
R., 1996. Movement-related effects in {fMRI} time-series. Mag-
netic Resonance in Medicine 35, 346–355.
Gordon, E.M., Laumann, T.O., Adeyemo, B., Huckins, J.F., Kel-
ley, W.M., Petersen, S.E., 2016. Generation and evaluation of a
cortical area parcellation from resting-state correlations. Cerebral
Cortex 26, 288–303. doi:10.1093/cercor/bhu239.
Gotts, S.J., Saad, Z.S., Jo, H.J., Wallace, G.L., Cox, R.W., Martin,
A., 2013. The perils of global signal regression for group compar-
isons: a case study of Autism Spectrum Disorders. Frontiers in
Human Neuroscience 7, 356. doi:10.3389/fnhum.2013.00356.
Greve, D.N., Fischl, B., 2009. Accurate and robust brain image
alignment using boundary-based registration. NeuroImage 48, 63–
72. doi:10.1016/j.neuroimage.2009.06.060.
Griffanti, L., Salimi-Khorshidi, G., Beckmann, C.F., Auerbach, E.J.,
Douaud, G., Sexton, C.E., Zsoldos, E., Ebmeier, K.P., Filippini,
N., Mackay, C.E., Moeller, S., Xu, J., Yacoub, E., Baselli, G.,
Ugurbil, K., Miller, K.L., Smith, S.M., 2014. ICA-based artefact
removal and accelerated fMRI acquisition for improved resting
state network imaging. NeuroImage 95, 232–247. doi:10.1016/j.
neuroimage.2014.03.034.
Gur, R.E., Kaltman, D., Melhem, E.R., Ruparel, K., Prabhakaran,
K., Riley, M., Yodh, E., Hakonarson, H., Satterthwaite, T., Gur,
R.C., 2013. Incidental findings in youths volunteering for brain
MRI research. American Journal of Neuroradiology 34, 2021–2025.
doi:10.3174/ajnr.A3525.
Hahamy, A., Calhoun, V., Pearlson, G., Harel, M., Stern, N., Attar,
F., Malach, R., Salomon, R., 2014. Save the global: global signal
connectivity as a tool for studying clinical populations with func-
tional magnetic resonance imaging. Brain connectivity 4, 395–403.
doi:10.1089/brain.2014.0244.
Hallquist, M.N., Hwang, K., Luna, B., 2013. The nuisance of
nuisance regression: Spectral misspecification in a common ap-
proach to resting-state fMRI preprocessing reintroduces noise
and obscures functional connectivity. NeuroImage 82, 208–225.
doi:10.1016/j.neuroimage.2013.05.116, arXiv:NIHMS150003.
Jenkinson, M., Bannister, P., Brady, M., Smith, S., 2002.
the robust and accurate linear
Neu-
doi:10.1016/S1053-8119(02)91132-8,
Improved optimization for
registration and motion correction of brain images.
roImage 17, 825–841.
arXiv:arXiv:1011.1669v3.
Cox, R.W., Saad, Z.S., 2013. Effective preprocessing procedures
virtually eliminate distance-dependent motion artifacts in resting
state FMRI. Journal of Applied Mathematics 2013. doi:10.1155/
2013/935154.
Klein, A., Andersson, J., Ardekani, B.A., Ashburner, J., Avants,
B., Chiang, M.C., Christensen, G.E., Collins, D.L., Gee, J., Hel-
lier, P., Song, J.H., Jenkinson, M., Lepage, C., Rueckert, D.,
Thompson, P., Vercauteren, T., Woods, R.P., Mann, J.J., Parsey,
R.V., 2009. Evaluation of 14 nonlinear deformation algorithms ap-
plied to human brain MRI registration. NeuroImage 46, 786–802.
doi:10.1016/j.neuroimage.2008.12.037.
Kundu, P., Brenowitz, N.D., Voon, V., Worbe, Y., V´ertes, P.E.,
Inati, S.J., Saad, Z.S., Bandettini, P.A., Bullmore, E.T., 2013.
Integrated strategy for improving functional connectivity mapping
using multiecho fMRI. Proceedings of the National Academy of
Sciences of the Unites States of America 110, 16187–16192. doi:10.
1073/pnas.1301725110, arXiv:arXiv:1408.1149.
Kundu, P., Inati, S.J., Evans, J.W., Luh, W.M., Bandettini, P.A.,
2012. Differentiating BOLD and non-BOLD signals in fMRI time
series using multi-echo EPI. NeuroImage 60, 1759–1770. doi:10.
1016/j.neuroimage.2011.12.028.
Laumann, T.O., Gordon, E.M., Adeyemo, B., Snyder, A.Z., Joo,
S.J., Chen, M.Y., Gilmore, A.W., McDermott, K.B., Nelson,
S.M., Dosenbach, N.U.F., Schlaggar, B.L., Mumford, J.A., Pol-
drack, R.A., Petersen, S.E., 2015. Functional system and areal
organization of a highly sampled individual human brain. Neuron
87, 658–671. doi:10.1016/j.neuron.2015.06.037.
Macey, P.M., Macey, K.E., Kumar, R., Harper, R.M., 2004. A
method for removal of global effects from fMRI time series. Neu-
roImage 22, 360–366. doi:10.1016/j.neuroimage.2003.12.042.
Merikangas, K.R., He, J.P., Brody, D., Fisher, P., Bourdon, K., Ko-
retz, D.S., 2010. Prevalence and treatment of mental disorders
among US children in the 20012004 NHANES. Pediatrics 125,
75–81. doi:10.1542/peds.2008-2598.Prevalence.
Murphy, K., Birn, R.M., Handwerker, D.A., Jones, T.B., Bandet-
tini, P.A., 2009. The impact of global signal regression on resting
state correlations: Are anti-correlated networks introduced? Neu-
roImage 44, 893–905. doi:10.1016/j.neuroimage.2008.09.036,
arXiv:NIHMS150003.
Muschelli, J., Nebel, M.B., Caffo, B.S., Barber, A.D., Pekar, J.J.,
Mostofsky, S.H., 2014. Reduction of motion-related artifacts in
resting state fMRI using aCompCor. NeuroImage 96, 22–35.
doi:10.1016/j.neuroimage.2014.03.028.
Patel, A.X., Kundu, P., Rubinov, M., Jones, P.S., V´ertes, P.E.,
Ersche, K.D., Suckling, J., Bullmore, E.T., 2014. A wavelet
method for modeling and despiking motion artifacts from resting-
state fMRI time series. NeuroImage 95, 287–304. doi:10.1016/j.
neuroimage.2014.03.012.
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. Neu-
roImage 59, 2142–2154. doi:10.1016/j.neuroimage.2011.10.018,
arXiv:NIHMS150003.
Power, J.D., Cohen, A.L., Nelson, S.M., Wig, G.S., Barnes, K.A.,
Church, J.A., Vogel, A.C., Laumann, T.O., Miezin, F.M., Schlag-
gar, B.L., Petersen, S.E., 2011. Functional network organization
of the human brain. Neuron 72, 665–678. doi:10.1016/j.neuron.
2011.09.006.
Power, J.D., Mitra, A., Laumann, T.O., Snyder, A.Z., Schlag-
gar, B.L., Petersen, S.E., 2014a. Methods to detect, charac-
terize, and remove motion artifact in resting state fMRI. Neu-
roImage 84, 320–341. doi:10.1016/j.neuroimage.2013.08.048,
arXiv:NIHMS150003.
Power, J.D., Schlaggar, B.L., Petersen, S.E., 2014b. Studying brain
organization via spontaneous fMRI signal. Neuron 84, 681–696.
doi:10.1016/j.neuron.2014.09.007, arXiv:NIHMS150003.
Power, J.D., Schlaggar, B.L., Petersen, S.E., 2015. Recent progress
and outstanding issues in motion correction in resting state fMRI.
NeuroImage 105, 536–551. doi:10.1016/j.neuroimage.2014.10.
044, arXiv:NIHMS150003.
Jo, H.J., Gotts, S.J., Reynolds, R.C., Bandettini, P.A., Martin, A.,
Pruim, R.H.R., Mennes, M., Buitelaar, J.K., Beckmann, C.F., 2015a.
10
visualization tool for human brain connectomics. PLoS ONE 8.
doi:10.1371/journal.pone.0068910.
Yan, C.G., Cheung, B., Kelly, C., Colcombe, S., Craddock, R.C., Di
Martino, A., Li, Q., Zuo, X.N., Castellanos, F.X., Milham, M.P.,
2013a. A comprehensive assessment of regional variation in the
impact of head micromovements on functional connectomics. Neu-
roImage 76, 183–201. doi:10.1016/j.neuroimage.2013.03.004,
arXiv:NIHMS150003.
Yan, C.G., Craddock, R.C., He, Y., Milham, M.P., 2013b. Ad-
dressing head motion dependencies for small-world topologies in
functional connectomics. Frontiers in human neuroscience 7, 910.
doi:10.3389/fnhum.2013.00910.
Yan, C.G., Craddock, R.C., Zuo, X.N., Zang, Y.F., Milham, M.P.,
2013c. Standardizing the intrinsic brain: Towards robust measure-
ment of inter-individual variation in 1000 functional connectomes.
NeuroImage 80, 246–262. doi:10.1016/j.neuroimage.2013.04.
081, arXiv:NIHMS150003.
Yang, G.J., Murray, J.D., Repovs, G., Cole, M.W., Savic, A.,
Glasser, M.F., Pittenger, C., Krystal, J.H., Wang, X.J., Pearl-
son, G.D., Glahn, D.C., Anticevic, A., 2014. Altered global brain
signal in schizophrenia. Proceedings of the National Academy of
Sciences of the United States of America 111, 7438–43. doi:10.
1073/pnas.1405289111.
Yeo, B.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D.,
Hollinshead, M., Roffman, J.L., Smoller, J.W., Zollei, L., Poli-
meni, J.R., Fischl, B., Liu, H., Buckner, R.L., 2011. The organi-
zation of the human cerebral cortex estimated by intrinsic func-
tional connectivity. Journal of neurophysiology 106, 1125–1165.
doi:10.1152/jn.00338.2011.
Zeng, L.L., Wang, D., Fox, M.D., Sabuncu, M., Hu, D., Ge, M.,
Buckner, R.L., Liu, H., 2014. Neurobiological basis of head motion
in brain imaging. Proceedings of the National Academy of Sciences
of the United States of America 111, 6058–62. doi:10.1073/pnas.
1317424111.
Zuo, X.N., Anderson, J.S., Bellec, P., Birn, R.M., Biswal, B.B.,
Blautzik, J., Breitner, J.C.S., Buckner, R.L., Calhoun, V.D.,
Castellanos, F.X., Chen, A., Chen, B., Chen, J., Chen, X., Col-
combe, S.J., Courtney, W., Craddock, R.C., Di Martino, A.,
Dong, H.M., Fu, X., Gong, Q., Gorgolewski, K.J., Han, Y., He,
Y., He, Y., Ho, E., Holmes, A., Hou, X.H., Huckins, J., Jiang, T.,
Jiang, Y., Kelley, W., Kelly, C., King, M., LaConte, S.M., Lain-
hart, J.E., Lei, X., Li, H.J., Li, K., Li, K., Lin, Q., Liu, D., Liu, J.,
Liu, X., Liu, Y., Lu, G., Lu, J., Luna, B., Luo, J., Lurie, D., Mao,
Y., Margulies, D.S., Mayer, A.R., Meindl, T., Meyerand, M.E.,
Nan, W., Nielsen, J.A., O'Connor, D., Paulsen, D., Prabhakaran,
V., Qi, Z., Qiu, J., Shao, C., Shehzad, Z., Tang, W., Villringer,
A., Wang, H., Wang, K., Wei, D., Wei, G.X., Weng, X.C., Wu,
X., Xu, T., Yang, N., Yang, Z., Zang, Y.F., Zhang, L., Zhang, Q.,
Zhang, Z., Zhang, Z., Zhao, K., Zhen, Z., Zhou, Y., Zhu, X.T.,
Milham, M.P., 2014. An open science resource for establishing re-
liability and reproducibility in functional connectomics. Scientific
data 1, 140049. doi:10.1038/sdata.2014.49.
Evaluation of ICA-AROMA and alternative strategies for motion
artifact removal in resting state fMRI. NeuroImage 112, 278–287.
doi:10.1016/j.neuroimage.2015.02.063.
Pruim, R.H.R., Mennes, M., van Rooij, D., Llera, A., Buitelaar, J.K.,
Beckmann, C.F., 2015b. ICA-AROMA: A robust ICA-based strat-
egy for removing motion artifacts from fMRI data. NeuroImage
112, 267–277. doi:10.1016/j.neuroimage.2015.02.064.
Saad, Z.S., Gotts, S.J., Murphy, K., Chen, G., Jo, H.J., Martin, A.,
Cox, R.W., 2012. Trouble at rest: how correlation patterns and
group differences become distorted after global signal regression.
Brain connectivity 2, 25–32. doi:10.1089/brain.2012.0080.
Salimi-Khorshidi, G., Douaud, G., Beckmann, C.F., Glasser, M.F.,
Griffanti, L., Smith, S.M., 2014. Automatic denoising of func-
tional MRI data: Combining independent component analysis
and hierarchical fusion of classifiers. NeuroImage 90, 449–468.
doi:10.1016/j.neuroimage.2013.11.046, arXiv:NIHMS150003.
Satterthwaite, T.D., Connolly, J.J., Ruparel, K., Calkins, M.E.,
Jackson, C., Elliott, M.A., Roalf, D.R., Hopsona, R., Prab-
hakaran, K., Behr, M., Qiu, H., Mentch, F.D., Chiavacci, R.,
Sleiman, P.M.A., Gur, R.C., Hakonarson, H., Gur, R.E., 2016.
The Philadelphia Neurodevelopmental Cohort: A publicly avail-
able resource for the study of normal and abnormal brain devel-
opment in youth. NeuroImage 124, 1115–1119. doi:10.1016/j.
neuroimage.2015.03.056.
Satterthwaite, T.D., Elliott, M.A., Gerraty, R.T., Ruparel, K., Loug-
head, J., Calkins, M.E., Eickhoff, S.B., Hakonarson, H., Gur,
R.C., Gur, R.E., Wolf, D.H., 2013. An improved framework
for confound regression and filtering for control of motion arti-
fact in the preprocessing of resting-state functional connectivity
data. NeuroImage 64, 240–256. doi:10.1016/j.neuroimage.2012.
08.052, arXiv:NIHMS150003.
Satterthwaite, T.D., Elliott, M.A., Ruparel, K., Loughead, J., Prab-
hakaran, K., Calkins, M.E., Hopson, R., Jackson, C., Keefe, J.,
Riley, M., Mentch, F.D., Sleiman, P., Verma, R., Davatzikos, C.,
Hakonarson, H., Gur, R.C., Gur, R.E., 2014. Neuroimaging of
the Philadelphia Neurodevelopmental Cohort. NeuroImage 86,
544–553. doi:10.1016/j.neuroimage.2013.07.064.
Satterthwaite, T.D., Wolf, D.H., Loughead, J., Ruparel, K., Elliott,
M.A., Hakonarson, H., Gur, R.C., Gur, R.E., 2012. Impact of in-
scanner head motion on multiple measures of functional connec-
tivity: Relevance for studies of neurodevelopment in youth. Neu-
roImage 60, 623–632. doi:10.1016/j.neuroimage.2011.12.063,
arXiv:NIHMS150003.
Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenk-
inson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-
Khorshidi, G., Woolrich, M.W., Barch, D.M., U??urbil, K., Van
Essen, D.C., 2013. Functional connectomics from resting-state
fMRI. Trends in Cognitive Sciences 17, 666–682. doi:10.1016/j.
tics.2013.09.016.
Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A.,
Yushkevich, P.A., Gee, J.C., 2010. N4ITK: Improved N3 bias cor-
rection. IEEE Transactions on Medical Imaging 29, 1310–1320.
doi:10.1109/TMI.2010.2046908.
Tustison, N.J., Cook, P.A., Klein, A., Song, G., Das, S.R., Duda,
J.T., Kandel, B.M., van Strien, N., Stone, J.R., Gee, J.C., Avants,
B.B., 2014. Large-scale evaluation of ANTs and FreeSurfer cortical
thickness measurements. NeuroImage 99, 166–179. doi:10.1016/
j.neuroimage.2014.05.044.
Van Dijk, K.R.A., Hedden, T., Venkataraman, A., Evans, K.C.,
Intrinsic functional con-
Lazar, S.W., Buckner, R.L., 2010.
nectivity as a tool for human connectomics:
theory, proper-
ties, and optimization. Journal of neurophysiology 103, 297–321.
doi:10.1152/jn.00783.2009.
Van Dijk, K.R.A., Sabuncu, M.R., Buckner, R.L., 2012. The in-
fluence of head motion on intrinsic functional connectivity MRI.
NeuroImage 59, 431–438. doi:10.1016/j.neuroimage.2011.07.
044, arXiv:NIHMS150003.
Wang, H., Cao, Y., Syeda-mahmood, T., 2014. Multi-atlas segmen-
tation with learning-based label fusion. Machine Learning in Med-
ical Imaging 35, 256–263. doi:10.1007/978-3-319-10581-9_32.
Xia, M., Wang, J., He, Y., 2013. BrainNet Viewer: A network
11
Figure 1: Schematic of the 12 de-noising models evaluated in the present study. For each of the 12 models indexed at left,
the table details what processing procedures and confound regressors were included in the model. De-noising models were selected
from the functional connectivity literature and represented a range of strategies.
12
Figure 2: Number of edges significantly related to motion after de-noising. Successful de-noising strategies reduced the
relationship between connectivity and motion. The number of edges (network connections) for which this relationship persists provides
evidence of a pipeline's efficacy. A, The percentage of edges significantly related to motion in a 264-node network defined by Power
et al. (2011). Fewer significant edges is indicative of better performance. B, The percentage of edges significantly related to motion in
a second, 333-node network defined by Gordon et al. (2016). C, Renderings of significant edges for each de-noising strategy, ranked
according to efficacy. Strategies that include regression of the mean global signal are framed in blue and consistently ranked as the
best performers.
13
Figure 3: Residual QC-FC correlations after de-noising. The absolute median QC-FC correlation is another measure of the
relationship between connectivity and motion. A, The absolute median correlation between functional connectivity and motion in a 264-
node network defined by Power et al. (2011). A lower absolute median correlation indicates better performance. B, The absolute median
correlation between functional connectivity and motion in a second, 333-node network defined by Gordon et al. (2016). C, Distributions
of all edgewise QC-FC correlations after each de-noising strategy, ranked according to efficacy. Results largely recapitulated those
reported in Figure 2, with GSR-based approaches (blue frame) collectively exhibiting the best performance. Whereas approaches that
included more regressors generally yielded a narrower distribution, those approaches that included GSR tended to shift the distribution's
center toward 0.
14
Figure 4: Distance-dependence of motion artifact after de-noising. The magnitude of motion artifact varies with the Euclidean
distance separating a pair of nodes, with closer nodes generally exhibiting greater impact of motion on connectivity. A, The residual
distance-dependence of motion artifact in a 264-node network defined by Power et al. (2011) following confound regression. B, The
residual distance-dependence of motion artifact in a second, 333-node network defined by Gordon et al. (2016). C, Density plots
indicating the relationship between the Euclidean distance separating each pair of nodes (x-axis) and the QC-FC correlation of
the edge connecting those nodes (y-axis). The overall trend lines for each de-noising strategy, from which distance-dependence is
computed, are indicated in red. The best performing models either excised high-motion volumes (36-parameter + scrubbing) or used
more localized regressors (ICA-AROMA and wmLocal). In general, approaches that made use of GSR without censoring resulted in
substantial distance-dependence. This effect was driven by differential efficacy of de-noising, with effective de-noising for long range
connections but not short-range connections.
15
Figure 5: Estimated loss of temporal degrees of freedom for each pipeline evaluated. Bars indicate mean number of
additional regressors per confound model; error bars indicate standard deviation for models where the number of confound regressors
varies by subject. High-parameter models and framewise censoring performed well overall on other benchmarks, but were also costliest
in terms of temporal degrees of freedom.
16
|
1810.11362 | 1 | 1810 | 2018-10-18T21:03:48 | On The Ignition, Propagation and Termination Of The Neuronal Bursting Activity During Ictogenesis In Epileptic Patients | [
"q-bio.NC"
] | Epilepsy creates a persistent increase in the probability of spontaneous seizures. An ictal episode evolves due to acute disturbance of the fine-tuned balance between excitatory vs. inhibitory inputs within a neural network in favor of excitation. The current literature that proposes the activity-dependent disinhibition as a valid mechanism of chronic epilepsy, does not provide clues on why this mechanism emerges only in epileptic patients and how the vicious circle resulting of an activity-dependent disinhibition in over-active ictogenic network would end. A new model, which presents chronic epilepsy as a disease of faulty architecture of the neural circuit, is discussed. Wherein; variable genetic or acquired predisposing factors drive abnormalities in the construction of multiple neural circuits resulting in an activity-dependent positive feedback excitatory loops which transform normal neural circuits into ictal foci. Such new mechanism, for igniting an activity-dependent unstable excitation with subsequent relatively stable disinhibition, leads to an ictal escape rhythm. The propagation of such bursting activity occurs either electrochemically via synaptic communication to remote susceptible circuits, or chemically via a trigger wave which recruits the non-connected proximal neurons. Termination occurs abruptly when the inhibitory interneurons functionally recover and reimpose their inhibitory effect on the ictogenic circuit to transform the escape rhythm into a normal, under-control output. The proposed model elucidates various enigmatic features of the disease; and illustrates both the end-result ictogenic mechanism arising from the wide variety of etiologies of human spontaneous and acquired epilepsy, and the timing of episodic transitions from normal activity to seizures. | q-bio.NC | q-bio |
On The Ignition, Propagation and Termination Of The
Neuronal Bursting Activity During Ictogenesis In Epileptic
Patients
Eslam Abbas
Harmel Hospital, 987 Nile Corniche, Al Kafor, Cairo, 11559, Egypt.
Abstract
Epilepsy creates a persistent increase in the probability of spontaneous seizures. An ictal episode evolves due to acute
disturbance of the fine-tuned balance between excitatory vs. inhibitory inputs within a neural network in favor of excitation.
The current literature that proposes the activity-dependent disinhibition as a valid mechanism of chronic epilepsy, does not
provide clues on why this mechanism emerges only in epileptic patients and how the vicious circle resulting of an activity-
dependent disinhibition in over-active ictogenic network would end. A new model, which presents chronic epilepsy as a disease
of faulty architecture of the neural circuit, is discussed. Wherein; variable genetic or acquired predisposing factors drive
abnormalities in the construction of multiple neural circuits resulting in an activity-dependent positive feedback excitatory
loops which transform normal neural circuits into ictal foci. Such new mechanism, for igniting an activity-dependent unstable
excitation with subsequent relatively stable disinhibition, leads to an ictal escape rhythm. The propagation of such bursting
activity occurs either electrochemically via synaptic communication to remote susceptible circuits, or chemically via a trigger
wave which recruits the non-connected proximal neurons. Termination occurs abruptly when the inhibitory interneurons
functionally recover and reimpose their inhibitory effect on the ictogenic circuit to transform the escape rhythm into a normal,
under-control output. The proposed model elucidates various enigmatic features of the disease; and illustrates both the end-
result ictogenic mechanism arising from the wide variety of etiologies of human spontaneous and acquired epilepsy, and the
timing of episodic transitions from normal activity to seizures.
Key words: Epilepsy; Ictogenesis, Excitatory Feedbacks; Competition Zone; and Escape Rhythm
Introduction
Brain electrical activity is non-synchronous (McPhee
and Hammer, 1995) and regulated by factors within the
neuron and the extraneuronal environment. Neuronal
factors comprise the type, number and distribution of
ion channels; changes to receptors of neurotransmitters
and modulation of gene expression (Bromfield et al.,
2006). The extraneuronal factors include ion concentra-
tion, synaptic plasticity and regulation of transmitters'
breakdown and reuptake (Blumenfeld, 2005).
During seizures; a discrete group of neurons begins
abnormal excessive firing in a synchronized manner
(Da Silva et al., 2003)(Margineanu, 2010) which then
can propagate to neighboring regions (Gotz-Trabert
et al., 2008). The characteristics of said firing are high
frequency bursts of action potentials and hypersyn-
chronization (Fisher et al., 2005). At the level of sin-
gle neurons, the ictal discharge shows a characteristic
paroxysmal depolarization shift, which consists of a se-
quence of sustained neuronal depolarization resulting in
a burst of action potentials, a plateau-like phase associ-
ated with completion of the action potential bursts, and
then rapid repolarization followed by hyperpolarization
(Misulis and Head, 2003).
The exact mechanism by which such smooth normal
brain activity is suddenly shifted to the bursting ictal
firing is still unclear (Noebels et al., 2012). The ax-
iomatic mechanism, which involves an acute imbalance
between the excitatory discharge and the inhibitory
control, needs further elaboration on the cellular and
molecular levels. Mutagenic neurochemical changes in
the neurotransmitters, receptors and ion channels are
a valid mechanism to address severe forms of epilep-
tic encephalopathies wherein the brain activity is dis-
torted with no or little normal epochs. Nevertheless;
neurochemical changes on such a molecular level
is
time-invariant and don't provide an explanation for
the unpredictable episodic, and relatively rare, seizures
(Schulze-Bonhage and Kuhn, 2008) that occurs in
chronic epilepsy. Besides; non-mutagenic changes in the
number and distribution of ion channels and neuro-
transmitter receptors can not be asserted as a primary
mechanism of epilepsy's pathogenesis, or a reactive
remodeling as a result of repeated ictal behavior of
epileptogenic foci.
Models of activity-dependent disinhibition for ictogen-
esis solve the timing problem; wherein disinhibition oc-
curred only at the extremes of network activity (Bracci
et al., 2001)(Wester and McBain, 2014). Therefore; the
probability of ictal changeover would depend on the
probability of that exceptional level of activity. Likely;
the same range of probabilities characterizing sponta-
neous seizures can agree with the odds of these levels of
activity (Staley, 2015). Once a seizure is induced by an
activity-dependent disinhibition, the seizure itself can
continue to produce activity levels that are sufficient
to suppress inhibition, providing the necessary positive
feedback to sustain the seizure. Yet; Said models do not
provide the fundamental mechanism by which a robust
input can transform a network with apparently normal
activity into an ictal focus. Moreover; they do not give
clues on how the vicious positive feedback circle will end,
adding more ambiguity on the poorly understood mech-
anism of termination of an ictal activity.
A model for the pathogenesis of chronic epilepsy shall
provide insights on the most enigmatic aspects of the dis-
ease like the basic fundamental mechanism that emerges
from the wide varieties of epileptogenic etiologies, prop-
agation, and termination of ictal behavior; besides the
unpredictable episodic timing of ictal episodes.
MODEL:
The ictogenesis process involves two main checkpoint
steps, the initiation of the burst, and the propagation of
the bursting activity. The initiation step involves abrupt
imbalance between excitation and inhibition within the
neural environment leading to the acute transition of
normal brain activity to ictal rhythm. The proposed
model functionally categorizes the origin of such abrupt
imbalance into three different levels of neuronal hyper-
excitability with ictal discharge, with special emphasis
on the second intermediate level; the chronic epilepsy.
While also proposes that the propagation checkpoint de-
pends on the initial conditions; wherein if a large-enough
number of neurons burst synchronously, the propagation
of a depolarization wave becomes mandatory.
The first level is seizures in otherwise healthy brain tis-
sue, which result from the simultaneous mass-excitation
of large numbers of neurons by external insult or stim-
uli. Electric stimulation during ECT and chemical
stimulation, either during acute overdosing or sudden
withdrawal of a drug, are valid examples. During ECT;
electrodes deliver an electrical impulse which traverses
through intermediary brain tissue to simultaneously
stimulate neurons by altering their internal electrical
milieu and concentration of ions (Swartz, 2014). Chem-
ical stimuli may involve toxic exposure to domoic acid,
which activates excitatory GluK1 glutamate receptors
(Jett, 2012), or by overdoses of theophylline, which
blocks the inhibitory adenosine A1 receptor (Boison,
2011). Also; abrupt withdrawal of GABAergic-acting
sedative -- hypnotic drugs can cause seizure due to chronic
GABA receptor downregulation as well as glutamate
overactivity, which lead to neurotransmitter sensitiza-
tion and neuronal hyperexcitability (Allison and Pratt,
2003). Accordingly and as a result of both electrical and
chemical stimulation; a large focus of hyperexcitable
neurons bursts synchronously bypassing the initiation
checkpoint to induce a depolarization wave which effec-
tively propagates to induce an ictal episode, or series of
episodes in case of chemical insults.
The third level comprises severe epileptic encephalopathies,
which result from severe genetic abnormalities that
compromise important inhibitory pathways. Such mu-
tations are most frequently associated with continuous
altering of brain functions; wherein there are no nor-
mal epochs of brain activity, with subsequent frequent
seizures (Allen et al., 2013)(Veeramah et al., 2013).
Said seizures are usually multiform and intractable and
usually accompanied by relentless cognitive, behavioral
and neurological deficits (Khan and Al Baradie, 2012).
The second intermediate level is chronic epilepsy, which
is emphasized in this model as it is responsible for the
vast majority of seizures in humans. Chronic epilepsy
can be subdivided into spontaneous, and acquired due
to an acute injury of the normal brain tissue as trauma,
strokes and infections. Despite different pathological ori-
gin; both subdivisions leads to extemporaneous activity-
dependent shifts in the balance between inhibition and
excitation in one or more neural circuits to increase the
probability of seizures break out. This model tracks the
poorly-understood initiation, propagation and termina-
tion of the ictal episode in chronic epilepsy.
Within a normal neural circuit; the excitatory output
of principal cells is usually put under control by the in-
terneurons which mainly exert an inhibitory effect on the
incoming inputs; wherein the triggering of an action po-
tential is determined by summation of excitatory and in-
hibitory signals. A neural circuit can be analyzed accord-
ing to its neuronal content, to determine the feedback,
feedforward balance between the inhibitory and the ex-
citatory neurons, to yield a block diagram as shown in
Fig. 1; wherein (I) represents the input excitatory sig-
nal, (H) represents the input inhibitory signal from the
interneurons, (e) is the summation effect and represents
the actual input, (P ) is the number of excitatory prin-
cipal neurons receiving the input (I) within the neural
circuit, (B) is the average number of inhibitory interneu-
rons for each excitatory principal neuron within the neu-
2
Fig. 1. A block diagram describes the architecture of the normal neural circuit vs. the ictogenic circuit.
ral circuit, and (O) represents the output of the neural
circuit. The output of the circuit is:
within the neural circuit in response to the input (I).
O = eP
and the actual input after summation is:
e = I − H
then, the output of the circuit can be described as:
O(1 + BP ) = IP
so, the ratio between the output and the input is:
O
I
=
P
1 + BP
This equation represents a negative feedback system,
wherein the emergent output of the circuit is lesser than
the input signal by the effect of (B, the average number
of inhibitory interneurons for each excitatory principal
neuron within the neural circuit). Herein and as a conse-
quence of all or none law; the O
I ratio actually represents
the firing probability of each excitatory principal neuron
3
O
I
< 1
Negative feedback is a widespread inhibitory tuning
mechanism within neural circuits, but negative feedfor-
ward usually augments a robust control as it's charac-
terized by inhibition certainty and minimal time lag.
The inhibitory interneurons within a neural circuit not
only put the excitatory output of the principal neu-
rons under check, but also assure desynchronous firing
of said neurons by decreasing the firing probability of
each excitatory principal neuron within the neural cir-
cuit in response to a certain input. The presence of
feed inhibition guarantees that the firing probability, of
an excitatory neuron in response to an input signal, is
usually less than one.
An ictogenic focus is characterized by defective neuronal
architecture which leads to priming of positive feedback
excitatory pathways consequential to interlaced excita-
tory interneurons or sprouting of axonal collaterals. Said
defective neuronal architecture can result from the faulty
neuronal arrangement due to dysgenesis guided by ge-
netic predispositions, or as an end result of the compen-
satory reconstruction that occurs after brain injury. The
positive feedback excitatory pathways are triggered off in
an activity-dependent manner. When an episodic surge
of the circuit activity reaches a critical value, the positive
feedback excitatory pathways are fully activated. Then;
the system enters a positive feedback loop wherein the
excitatory impulses reverberate to self-reinforce further
excitation.
The activity-dependent positive feedback loop is ex-
tremely unstable, as the value (Oj) is usually equal or
more than unity (Oj ≥ 1), and collapses spontaneously;
wherein (j) represents the average number of excitatory
feedback interconnections for each excitatory principal
neuron within the neural circuit. Yet, as a consequence
of impulses reverberation and self-reinforced excitation;
the interneuronal inhibitory network enters a state of
stupor. When the inhibitory control is defunct (H = 0),
the output of the ictogenic circuit is released. So; for
each excitatory principal cell within the ictogenic cir-
cuit (P = 1), each input will result in a subsequent
output (I = O) forming a synchronous escape rhythm;
wherein all the excitatory principal neurons receiving
the input (I) within the neural circuit will burst simul-
taneously upon receiving the input (I). However; such
self-reinforced signal magnification and the subsequent
escape rhythm (ictal discharge) occurs only at the ex-
tremes of the circuit activity; wherein the probability
of output release would depend on the probability of an
exceptional level of a vigorous activity that reaches the
critical value.
The bursting activity resulting from the escape rhythm
can be confined due to the integral surrounding zone
of inhibition, which cannot be overcome by the current
level of ictal discharge, to produce a focal abnormality of
the brain electrical activity; or can propagate to distort
the electrical activity of the whole or a large portion of
the brain. The bursting ictal discharge will not overcome
the zone of inhibition as long as the number of excitatory
principal neurons, receiving an input (I) within an icto-
genic neural circuit, is small (P is small) leading to fail-
ure of electrochemical recruitment via synaptic commu-
nication; and/or failure of chemical recruitment of the
non-connected proximal neurons as (r (cid:28) φr); wherein
(r) is the radius of the recruited focus, and (φr) is the
minimum critical focal radius from which the chemical
recruitment can begin and the trigger wave can get off.
The propagation of the paroxysmal depolarization shift,
as a characteristic bursting activity, can occur via synap-
tic transmission to the susceptible connected neurons of
other remote networks; wherein the repetitive discharge
from the presynaptic ictogenic neuron leads to Ca++ ac-
cumulation in the presynaptic terminals enhancing neu-
rotransmitters release. Besides; it can propagate to the
non-connected adjacent neurons by altering the extra-
cellular ionic concentrations to produce a trigger wave
which gradually recruits the neurons in the proximity
to initiate a bursting activity by augmenting their ex-
citability.
The neuronal resting membrane potential (Vresting) is
mainly maintained by the role of K + and N a+ leak chan-
nels (Purves et al., 2001). Due to the high concentration
level of N a+ ions in the extracellular fluid and dimin-
ished membrane permeability; the effect of changes of
the ion concentration in the extraneuronal environment
on the resting membrane potential, differs greatly be-
tween K + and N a+ ions. Such effect can be represented
by the following two equations:
Vresting ∝ − K +
Vresting ∝ +
N a+
out
in − K +
K +
out
out − N a+
N a+
out
in
Both equations show the two different effects of fluctua-
tions in the extraneuronal concentration of K + and N a+
ions on the resting membrane potential. Accordingly and
regardless of the osmotic effect; while the electrochemi-
cal effect of hypo/hypernatremia is negligible, neuronal
resting membrane potential is very sensitive to trivial
fluctuation of extraneuronal potassium concentration.
The bursting activity resulting from the escape rhythm
leads to a transient increase in the extraneuronal K +
concentrations within the proximity of the neural net-
work. Said K + fluctuations increase the excitability of
adjacent neurons due to depolarizing the resting mem-
brane potential. An increase in the neuronal excitability
sets the affected neurons into bursting activity, leading
to the propagation of the PDS through the brain tissue.
As a trigger wave, each affected neuron augments further
propagation, so that the propagation of PDS doesn't
slow down or lose amplitude as it travels through the
brain tissue. The rate of recruitment of neurons into the
bursting activity, propagation of the trigger wave, can
be represented by the following equation:
d(N )
dt
= ρn
(cid:19)
(cid:18) d(Vol)
(cid:18)
dt
(cid:19)
r2.
d(r)
dt
or
wherein
d(N )
dt
= 4π.ρn
r ≥ φr
wherein (N ) is the number of recruited neurons in unit
of time (t), (ρn) is the neuronal density per unit of vol-
ume (Vol), and (r) is the radius of the recruited focus,
or the radius of the propagating trigger wave. (φr) is
the minimum critical focal radius from which the chemi-
cal recruitment can launch. The bursting activity of the
ictogenic circuit ends abruptly when the inhibitory in-
terneurons within the ictogenic focus become function-
ally effective to transform the stable escape rhythm into
a normal, under-control output.
4
RESULTS:
Many results can be inferred from the proposed model,
but hereinafter are some that elucidate various enig-
matic features of chronic epilepsy. By presenting
epilepsy as a disease caused by mal-architecture of the
neural circuits within the brain tissue; the probabil-
ity of formation of spontaneous ictogenic foci
increases within the brain regions characterized
by high-rate of continuous active modulation
of tissue micro-structure. Additionally; the model
probes into the genesis of the ictal activity in epilep-
tic patients; wherein the initiation involves igniting an
activity-dependent unstable positive feedback excita-
tory loop which spontaneously collapses, yet leads to a
synchronous escape rhythm forming a bursting activity.
The action potential generated by uninhibited ig-
nition tends to have echoic reverberations; so that
the resulting escape rhythm shows the characteristic
multi-spiking paroxysmal depolarization shift as a hall-
mark of epilepsy on the cellular level. The ictal bursting
activity can be either confined by an intact zone of in-
hibition, which can't be overcome by the current level
of ictal discharge (P is small and/or r (cid:28) φr), to pro-
duce a focal abnormality of the brain electrical activ-
ity; or propagating to distort the electrical activity of
the whole or a large portion of the brain. The propa-
gation of the bursting discharge occurs either electro-
chemically via synaptic communication to remote sus-
ceptible circuits, or chemically via a trigger wave which
recruit the non-connected proximal neurons. Termina-
tion occurs abruptly when the inhibitory interneurons
functionally recover and reimpose their inhibitory effect
on the ictogenic circuit to transform the escape rhythm
into a normal, under-control output.
Two architectural factors contribute to the aggressive-
ness of an ictogenic focus; a major factor which is the
average density, or average number, of the excitatory
feedback pathways for each excitatory principal neuron
within the neural circuit (j), and a minor factor which
is the ratio between the neuronal density to the
(cid:1) within the ictal focus. Addition-
ally; the velocity of the chemical neuronal recruitment
via a trigger wave is proportional to the neuronal den-
sity of the affected brain region.
synaptic density(cid:0) ρn
ρs
Moreover; the model shows that spontaneous epilepsy
is a dynamic non-static chronic disease; wherein
untreated recurrent attacks can lead to formation of
neo-foci or exaggeration of the principle one via mal-
architecting resulting of ectogenic excitotoxicity and
subsequent faulty rewiring driven by genetic predispo-
sitions. On the other hand; a rare possibility of long-
term decomposition of an ictogenic focus, due to neural
plasticity, can occur if consolidation with further ictal
attacks is prevented. Besides; paradoxical therapeutic
5
effect of combined anti-epileptic drugs, which enhance
the GABAergic hyperpolarization, can occur, due to the
higher probability of main input magnification, if the
synaptic plasticity between the main presynap-
tic input and the postsynaptic membrane within
the competition zone is anti-Hebbian.
DISCUSSION:
Several decades of experiments, which depend on
pharmacologically-induced seizures, have established
the idea that an imbalance between inhibitory and ex-
citatory activity leads to ictal episodes (Scharfman,
2007). An acute imbalance between excitation and in-
hibition is thus a valid ictogenic mechanism. Problems
arise when such mechanism is extended to cover the
chronic process of epileptogenesis, which creates a per-
sistent rise in the probability of spontaneous seizures
(Staley, 2015). The timing of seizures in chronic epilepsy
is unpredictable and ictal episodes are relatively rare,
representing much less than 1% of the total brain ac-
tivity (Moran et al., 2004). Thus; in chronic epilepsy,
an ictogenic mechanism shall also explain the timing of
episodic transitions from normal to ictal brain activity.
Additionally; another difficulty arises when applying the
model of imbalanced inhibition and excitation to the eti-
ology of chronic epilepsy, which does not usually suggest
a chronic imbalance. In genetic predisposed epilepsy,
analyses of the genetic etiology have occasionally found
causal loss-of-functions mutations in inhibitory path-
ways (Macdonald and Kang, 2012), but loss-of-function
mutations are also found in several excitatory pathways
(Frank et al., 2006)(Carvill et al., 2014), and the ma-
jority of causal mutations involve genes that do not di-
rectly alter the balance of inhibition and excitation (Ran
et al., 2014). In the acquired epilepsies, seizures start af-
ter an insult to a normal brain tissue as a result of stroke,
trauma, or infection. Steady-state imbalances in excita-
tion against inhibition, in established animal models of
acquired epilepsy, are difficult to demonstrate. In the Pi-
locarpine model; a damage to the inhibitory neurons is
compensated by an increase in GABAergic synaptoge-
nesis before the onset of seizures (Zhang et al., 2009).
Compensatory glutamatergic synaptogenesis also hap-
pens (Buckmaster, 2014), but steady-state network im-
balances, between excitation and inhibition, are not ev-
ident in experimental and human epilepsy (Sutula and
Dudek, 2007).
The current literature that proposes the activity-
dependent disinhibition as a valid mechanism in chronic
epilepsy, does not provide clues on why this mechanism
emerges only in epileptic patients and how the vicious
circle resulting of an activity-dependent disinhibition in
over-active ictogenic network would end. The proposed
model presents chronic epilepsy as a disease of faulty
architecture of the micro-structure of the brain tissue;
wherein variable genetic or acquired predisposing fac-
tors drive abnormalities in the construction of multiple
neural circuits resulting in an activity-dependent posi-
tive feedback excitatory loops which transform normal
neural circuits into ictal foci. Such new mechanism for
igniting an activity-dependent unstable excitation with
subsequent relatively stable disinhibition leads to an
ictal escape rhythm. Additionally; Said model provides
insights about the mechanism of propagation and ter-
mination of such bursting activity. Besides; it illustrates
both the end-result ictogenic mechanism arising from
the wide variety of etiologies of human spontaneous and
acquired epilepsy, and the timing of episodic transitions
from normal activity to seizures.
MOLECULAR MECHANISM OF IGNITION:
Either genetic predispositions drives defective wiring in
spontaneous epilepsy or after-insult dysgenesis in the
acquired one; both lead to architectural flaws in one or
more neural circuits transforming them into ictal foci;
wherein positive feedback excitatory loops are formed
via sprouting axonal collaterals and/or interlaced exci-
tatory interneurons. Consequently; a competition zone
is formed, which comprises presynaptic long-term po-
tentiated main excitatory input synapse(s), presynaptic
inhibitory synapse(s) (formed by feedback or feedfor-
ward inhibitory interneurons), presynaptic long-term
depressed feedback excitatory synapse(s) (formed by
positive feedback interneurons or axonal collaterals),
and a postsynaptic membrane of a principle neuron. The
competition zone is not an anatomical but a functional
entity; wherein a dynamic contest among synaptic plas-
ticities determine the polarization of the postsynaptic
membrane and further generation of an action potential.
Normally; when a main excitatory input comes into the
competition zone, the activity of the inhibitory feed-
forward or feedback control dictates the probability of
action potential generation and firing of the postsynap-
tic neuron, whereas the long-term depressed excitatory
feedback plays a negligible role.
Spike-timing-dependent plasticity causes a difference in
the homosynaptic plasticity between the presynaptic
main excitatory input, which evolves a long-term po-
tentiation, and the presynaptic excitatory feedback(s),
which evolve long-term depression. Due to time lag; the
presynaptic feedback excitatory synapse(s) are long-
term depressed because they always fires immediately
after the firing of the supplied postsynaptic principal
neuron, hence this particular feedback excitatory stim-
ulus is made weaker (Song et al., 2000)(Debanne et al.,
1994). Additionally; the rise in the synaptic weight
of the presynaptic main excitatory pathway dictates
a reduction in the synaptic weight of the presynap-
tic feedback excitatory synapse(s) to keep the average
synaptic weight approximately conserved (Lynch et al.,
1977)(Chistiakova et al., 2014).
A strong input form the presynaptic main excitatory
pathway can be cumulated via spatial or temporal
summation to surpass a threshold that ignites an ictal
activity. Upon receiving a robust input, the activity-
dependent depressed positive feedbacks are fully acti-
vated, due to Ca++ rush, and the accumulated vesicles
fuse into the presynaptic membrane to release an ex-
cessive amount of excitatory neurotransmitters into
the competition zone. The positive feedback excitatory
loops are unstable and collapse spontaneously; yet the
excitatory neurotransmitters released into the competi-
tion zone render the presynaptic feedback or feedforward
inhibitory synapse(s) functionally obtunded. Additional
molecular mechanism can also contribute to sudden in-
hibitory stupor within the competition zone; excessive
unstable positive feedbacks that ignite the ictal behav-
ior cause the inhibitory feedbacks, not the inhibitory
feedforwards, to oscillate rapidly. Intensely activated,
dendritic GABAA receptors excite rather than inhibit
the postsynaptic membrane (Staley et al., 1995) of the
principal neuron(s) with the competition zone.
The molecular mechanism of GABA-mediated mem-
brane depolarization involves differential anionic con-
centration shift during intense GABAA receptor activa-
tion. The membrane potential is positive to the Cl− re-
versal potential (Bormann et al., 1987), so that the Cl−
ions flux is inward and the intracellular Cl− concentra-
tion increases, but the membrane potential is negative
to the HCO−
3 reversal potential (Huguenard and Alger,
1986), so that the intracellular HCO−
3 concentration
decreases due to the efflux of HCO−
3 ions. The decrease
in the intracellular HCO−
3 concentration, drives the dif-
fusion of CO2 across the dendritic membrane, allowing
continual regeneration of intracellular HCO−
3 ions by
carbonic anhydrase at a rate that exceeds the removal
of intracellular Cl− ions (Staley, 1994). Accordingly;
The electrochemical gradient for Cl− collapses more
significantly than does the HCO−
3 gradient, producing
a shift in the GABAA reversal potential toward the
HCO−
3 reversal potential that is sufficient to explain
the depolarizing response.
Due to the functional stupor of the inhibitory pathways;
the presynaptic main excitatory input is released to si-
multaneously activate the postsynaptic principle neu-
rons forming an escape rhythm. The action potential
generated by uninhibited ignition tends to have echoic
reverberations along the neuronal membrane. NMDA re-
ceptors are very sensitive to changes in the membrane
potential; wherein extracellular M g++ and Zn++ ions
bind to specific sites on the receptor, blocking the pas-
sage of other cations through the ion channel. Reverber-
ating depolarization of the neuronal membrane, mainly
the somatic and dendritic membrane, dislodges and re-
pels the M g++ and Zn++ ions from the pore, thus al-
lowing a voltage-dependent flow of Ca++ and N a+ ions
into the cell and K + out of the cell (Cull-Candy et al.,
6
2001)(Liu and Zhang, 2000). The influx of cations cre-
ates a local signal that can; in the absence of anion-
dependent inhibitory control which, besides type-A K +
channels, prevent reverberating depolarization; retrigger
N a+ voltage-gated ion channels. Such behavior enables
multiple waves of N a+ influx, creating a train of repet-
itive spikes which is followed by a subsequent hyperpo-
larization generated by Ca++-dependent K + channels.
This response is the paroxysmal depolarization shift,
which is a hallmark of epilepsy on the cellular level.
MOLECULAR MECHANISM OF PROPAGATION
AND TERMINATION:
If the number of principal neurons of an ictogenic focus
is not large enough to overcome the zone of inhibition
exerted onto their axon terminals, the bursting activity
would not be electrochemically transmitted via synaptic
communication to the connected remote brain loci. Ad-
ditionally; the synchronous activity of said small num-
ber of neurons would not significantly disturb the extra-
neuronal ionic environment, thus a trigger wave would
not be created leading to failure of recruitment of ad-
jacent non-connected neurons of nearby loci. However;
when the number of principal neurons of an ictogenic fo-
cus is large enough, the bursting activity makes its way
through the confinement and tends to propagate. Said
large number of neurons will likely overcome the zone
of inhibition and transmit the ictal activity via synaptic
communication to the connected remote loci.
Besides; synchronous activity of a large number of princi-
pal neurons within an ictal focus can potentially disturb
the extraneuronal ionic environment, generating a trig-
ger wave which recruits adjacent non-connected neurons
of the nearby neural networks. Ictal activity, of a large
number of neurons within an ictogenic focus, induces re-
gional elevation of extracellular K + ions due to persis-
tent efflux. As mentioned before; neuronal resting mem-
brane potential is very sensitive to changes in extracel-
lular K + concentration. Regional rise in the levels of ex-
tracellular K + ions reduces the magnitude of potassium
gradient across the cell membrane of the adjacent neu-
rons; and therefore, shifts the absolute value of the rest-
ing membrane potential to depolarization. The adjacent
neurons become hyperexcitable and respond by produc-
ing an ictal activity. The process of recruitment is self-
reinforced generating a trigger wave that promulgates
spatially with a defined speed. Active synaptic clear-
ance brings about rebalance among the opposed neuro-
transmitters within the competition zone, so that the
inhibitory control reimposes its effect and decrease the
probability of firing of the principle neurons within the
ictal circuit demolishing the synchronous escape rhythm;
consequently the ictal episode is terminated. Reactive
inhibitory reimposition contributes to the altered state
of consciousness after an epileptic seizure.
7
EFFECT OF ANTIEPILEPTICS ON THE COMPE-
TITION ZONE:
The competition zone is a functional, not an anatomical,
entity; wherein the mechanism of action of antiepilep-
tics can be applied to. This subsection discusses the ef-
fects of antiepileptics on the competition zone because
of the paradox that results from a molecular mechanism
that is believed to contribute to the functional stupor of
the inhibitory synapse(s), which results from the exces-
sive unstable positive feedbacks that ignite an ictal be-
havior. The mechanism implicates that an intensely ac-
tivated, dendritic GABAA receptors excite rather than
inhibit the postsynaptic membrane (Staley et al., 1995).
Accordingly; drugs that enhance the effect of the neuro-
transmitter GABA at the GABAA receptors, as benzo-
diazepines, can effectively prevent seizures as they pre-
vent the main input summation, so they prevent the ig-
nition of the activity-dependent positive feedback exci-
tatory loops in the first place. But; if the main input
bypasses such inhibitory effect and ignites the unstable
positive feedbacks, the overstimulated GABAA recep-
tor, by the drug, would enhance the excitation of the
postsynaptic membrane of the principal neuron(s) and
functionally contribute, and further prolong, the presy-
naptic inhibitory stupor.
Consequently; drugs that enhance the effect of the neu-
rotransmitter GABA at the GABAA receptors, as ben-
zodiazepines and the like, exacerbate epileptic seizures
that develop during using them as a preventive anti-
convulsant. Whilst paradoxically; such medications are
more effective if they are used to terminate an active
attack as they boost the effect of inhibitory synapse(s)
on the stable escape rhythm; which results after sponta-
neous collapse of the unstable excitatory positive feed-
backs. Besides, they inhibit the electrochemical propa-
gation of the ictal activity by consolidating the integrity
of the inhibitory zone on the synaptic terminals of the
ictogenic principal neurons of the primary ictal focus.
CONCLUSION:
In short; chronic epilepsy is a disease of mal-architecture
of the neural circuits that results from genetically-
predisposed, defective wiring in the spontaneous type;
or neural dysgenesis in the acquired type. Said mal-
architecture leads to formation of activity-dependent
excitatory positive feedbacks that can ignite an ictal
activity in response to robust stimulation. Such ictal ac-
tivity tends to propagate electrochemically via synaptic
communication to the connected distant loci, and chem-
ically via a trigger wave to the adjacent non-connected
loci. The ictal activity is terminated upon reimposition
of the functionally obtunded inhibitory control over the
competition zone within the primary ictal focus.
REFERENCES:
Allen, A. S., S. F. Berkovic, P. Cossette, N. Delanty,
D. Dlugos, E. E. Eichler, M. P. Epstein, T. Glauser,
D. B. Goldstein, Y. Han, et al.
2013.
cephalopathies. Nature, 501(7466):217.
De novo mutations
in epileptic
en-
Allison, C. and J. Pratt
2003. Neuroadaptive processes in gabaergic and glu-
tamatergic systems in benzodiazepine dependence.
Pharmacology & therapeutics, 98(2):171 -- 195.
Blumenfeld, H.
2005. Cellular and network mechanisms of spike-wave
seizures. Epilepsia, 46:21 -- 33.
Boison, D.
2011. Methylxanthines, seizures, and excitotoxicity.
In Methylxanthines, Pp. 251 -- 266. Springer.
Bormann, J., O. P. Hamill, and B. Sakmann
1987. Mechanism of anion permeation through chan-
nels gated by glycine and gamma-aminobutyric acid
in mouse cultured spinal neurones. The Journal of
physiology, 385(1):243 -- 286.
Bracci, E., M. Vreugdenhil, S. P. Hack, and J. G. Jefferys
2001. Dynamic modulation of excitation and inhibi-
tion during stimulation at gamma and beta frequen-
cies in the ca1 hippocampal region. Journal of Neu-
rophysiology, 85(6):2412 -- 2422.
Bromfield, E. B., J. E. Cavazos, and J. I. Sirven
2006. Basic mechanisms underlying seizures and
epilepsy.
Buckmaster, P. S.
2014. Does mossy fiber sprouting give rise to the
epileptic state? In Issues in Clinical Epileptology: A
View from the Bench, Pp. 161 -- 168. Springer.
Carvill, G. L., S. Weckhuysen, J. M. McMahon, C. Hart-
mann, R. S. Møller, H. Hjalgrim, J. Cook, E. Ger-
aghty, B. J. Oroak, S. Petrou, et al.
2014. Gabra1 and stxbp1: novel genetic causes of
dravet syndrome. Neurology, 82(14):1245 -- 1253.
Chistiakova, M., N. M. Bannon, M. Bazhenov, and
M. Volgushev
2014. Heterosynaptic plasticity: multiple mechanisms
and multiple roles. The Neuroscientist, 20(5):483 -- 498.
Da Silva, F. L., W. Blanes, S. N. Kalitzin, J. Parra,
P. Suffczynski, and D. N. Velis
2003. Epilepsies as dynamical diseases of brain sys-
tems: basic models of the transition between normal
and epileptic activity. Epilepsia, 44:72 -- 83.
Debanne, D., B. Gahwiler, and S. M. Thompson
1994. Asynchronous pre-and postsynaptic activity in-
duces associative long-term depression in area ca1 of
the rat hippocampus in vitro. Proceedings of the Na-
tional Academy of Sciences, 91(3):1148 -- 1152.
Fisher, R. S., W. V. E. Boas, W. Blume, C. Elger, P. Gen-
ton, P. Lee, and J. Engel Jr
2005. Epileptic seizures and epilepsy: definitions pro-
posed by the international league against epilepsy
(ilae) and the international bureau for epilepsy (ibe).
Epilepsia, 46(4):470 -- 472.
Frank, H. Y., M. Mantegazza, R. E. Westenbroek, C. A.
Robbins, F. Kalume, K. A. Burton, W. J. Spain, G. S.
McKnight, T. Scheuer, and W. A. Catterall
2006. Reduced sodium current in gabaergic interneu-
rons in a mouse model of severe myoclonic epilepsy in
infancy. Nature neuroscience, 9(9):1142.
Gotz-Trabert, K., C. Hauck, K. Wagner, S. Fauser, and
A. Schulze-Bonhage
2008. Spread of ictal activity in focal epilepsy. Epilep-
sia, 49(9):1594 -- 1601.
Huguenard, J. and B. Alger
1986. Whole-cell voltage-clamp study of the fad-
ing of gaba-activated currents in acutely dissociated
hippocampal neurons. Journal of Neurophysiology,
56(1):1 -- 18.
Jett, D. A.
2012. Chemical toxins that cause seizures. Neurotox-
icology, 33(6):1473 -- 1475.
Khan, S. and R. Al Baradie
2012.
Epilepsy research and treatment, 2012.
Epileptic encephalopathies: an overview.
Liu, Y. and J. Zhang
2000. Recent development in nmda receptors. Chinese
medical journal, 113(10):948 -- 956.
Lynch, G. S., T. Dunwiddie, and V. Gribkoff
1977.
tic correlate of
266(5604):737.
Heterosynaptic depression: a postsynap-
Nature,
long-term potentiation.
Cull-Candy, S., S. Brickley, and M. Farrant
2001. Nmda receptor subunits: diversity, develop-
ment and disease. Current opinion in neurobiology,
11(3):327 -- 335.
Macdonald, R. L. and J.-Q. Kang
2012. mrna surveillance and endoplasmic reticu-
lum quality control processes alter biogenesis of mu-
tant gabaa receptor subunits associated with genetic
epilepsies. Epilepsia, 53:59 -- 70.
Staley, K.
2015. Molecular mechanisms of epilepsy. Nature neu-
roscience, 18(3):367.
Staley, K. J., B. L. Soldo, and W. R. Proctor
1995. Ionic mechanisms of neuronal excitation by in-
hibitory gabaa receptors. Science, 269(5226):977 -- 981.
Sutula, T. P. and F. E. Dudek
2007. Unmasking recurrent excitation generated by
mossy fiber sprouting in the epileptic dentate gyrus:
an emergent property of a complex system. Progress
in brain research, 163:541 -- 563.
Swartz, C. M.
2014. A mechanism of seizure induction by electricity
and its clinical implications. The journal of ECT,
30(2):94 -- 97.
Veeramah, K. R., L. Johnstone, T. M. Karafet, D. Wolf,
R. Sprissler, J. Salogiannis, A. Barth-Maron, M. E.
Greenberg, T. Stuhlmann, S. Weinert, et al.
2013. Exome sequencing reveals new causal mutations
in children with epileptic encephalopathies. Epilepsia,
54(7):1270 -- 1281.
Wester, J. C. and C. J. McBain
2014. Behavioral state-dependent modulation of
distinct interneuron subtypes and consequences for
circuit function. Current opinion in neurobiology,
29:118 -- 125.
Zhang, W., R. Yamawaki, X. Wen, J. Uhl, J. Diaz, D. A.
Prince, and P. S. Buckmaster
2009. Surviving hilar somatostatin interneurons en-
large, sprout axons, and form new synapses with gran-
ule cells in a mouse model of temporal lobe epilepsy.
Journal of Neuroscience, 29(45):14247 -- 14256.
Margineanu, D. G.
2010. Epileptic hypersynchrony revisited. Neurore-
port, 21(15):963 -- 967.
McPhee, S. J. and G. D. Hammer
1995. Pathophysiology of Disease: An Introduction to
Clinical Medicine, (Lange Medical Books). USA.
Misulis, K. E. and T. C. Head
2003. Essentials of clinical neurophysiology, volume 1.
Garland Science.
Moran, N., K. Poole, G. Bell, J. Solomon, S. Kendall,
M. McCarthy, D. McCormick, L. Nashef, J. Sander,
and S. Shorvon
2004. Epilepsy in the united kingdom: seizure fre-
quency and severity, anti-epileptic drug utilization
and impact on life in 1652 people with epilepsy.
Seizure, 13(6):425 -- 433.
Noebels, J., M. Avoli, M. Rogawski, R. Olsen, and
A. Delgado-Escueta
2012. Jasper's Basic Mechanisms of the Epilepsies.
Oxford University Press.
Purves, D., G. Augustine, D. Fitzpatrick, L. Katz,
A. LaMantia, J. McNamara, and S. Williams
2001. Neuroscience 2nd edition. sunderland (ma) sin-
auer associates.
Ran, X., J. Li, Q. Shao, H. Chen, Z. Lin, Z. S. Sun, and
J. Wu
2014. Epilepsygene: a genetic resource for genes and
mutations related to epilepsy. Nucleic acids research,
43(D1):D893 -- D899.
Scharfman, H. E.
2007. The neurobiology of epilepsy. Current neurology
and neuroscience reports, 7(4):348 -- 354.
Schulze-Bonhage, A. and A. Kuhn
2008. Unpredictability of seizures and the burden of
epilepsy. Seizure prediction in epilepsy: from basic
mechanisms to clinical applications.
Song, S., K. D. Miller, and L. F. Abbott
2000. Competitive hebbian learning through spike-
timing-dependent synaptic plasticity. Nature neuro-
science, 3(9):919.
Staley, K.
1994. The role of an inwardly rectifying chloride con-
ductance in postsynaptic inhibition. Journal of neu-
rophysiology, 72(1):273 -- 284.
9
|
1707.08240 | 2 | 1707 | 2018-03-23T19:05:12 | The impact of epilepsy surgery on the structural connectome and its relation to outcome | [
"q-bio.NC"
] | Temporal lobe surgical resection brings seizure remission in up to 80% of patients, with long-term complete seizure freedom in 41%. However, it is unclear how surgery impacts on the structural white matter network, and how the network changes relate to seizure outcome. We used white matter fibre tractography on preoperative diffusion MRI to generate a structural white matter network, and postoperative T1-weighted MRI to retrospectively infer the impact of surgical resection on this network. We then applied graph theory and machine learning to investigate the properties of change between the preoperative and predicted postoperative networks. Temporal lobe surgery had a modest impact on global network efficiency, despite the disruption caused. This was due to alternative shortest paths in the network leading to widespread increases in betweenness centrality post-surgery. Measurements of network change could retrospectively predict seizure outcomes with 79% accuracy and 65% specificity, which is twice as high as the empirical distribution. Fifteen connections which changed due to surgery were identified as useful for prediction of outcome, eight of which connected to the ipsilateral temporal pole. Our results suggest that the use of network change metrics may have clinical value for predicting seizure outcome. This approach could be used to prospectively predict outcomes given a suggested resection mask using preoperative data only. | q-bio.NC | q-bio | NeuroImage: Clinical 18 (2018) 202–214
Contents lists available at ScienceDirect
NeuroImage: Clinical
journal homepage: www.elsevier.com/locate/ynicl
The impact of epilepsy surgery on the structural connectome and its relation
to outcome
Peter N. Taylora,b,c,⁎, Nishant Sinhaa,b, Yujiang Wanga,b,c, Sjoerd B. Vosd,e, Jane de Tisic,
Anna Miserocchic, Andrew W. McEvoyc, Gavin P. Winstonc,e,1, John S. Duncanc,e,1
a Interdisciplinary Computing and Complex BioSystems Group, School of Computing Science, Newcastle University, UK
b Institute of Neuroscience, Faculty of Medical Science, Newcastle University, UK
c NIHR University College London Hospitals Biomedical Research Centre, UCL Institute of Neurology, Queen Square, London WC1N 3BG, UK
d Translational Imaging Group, Centre for Medical Image Computing, University College London, UK
e Chalfont Centre for Epilepsy, Chalfont St Peter SL9 0LR, UK
T
A R T I C L E I N F O
A B S T R A C T
Keywords:
Connectome
Network
Temporal lobe epilepsy
Surgery
Machine learning
Support vector machine (SVM)
Background: Temporal lobe surgical resection brings seizure remission in up to 80% of patients, with long-term
complete seizure freedom in 41%. However, it is unclear how surgery impacts on the structural white matter
network, and how the network changes relate to seizure outcome.
Methods: We used white matter fibre tractography on preoperative diffusion MRI to generate a structural white
matter network, and postoperative T1-weighted MRI to retrospectively infer the impact of surgical resection on
this network. We then applied graph theory and machine learning to investigate the properties of change be-
tween the preoperative and predicted postoperative networks.
Results: Temporal lobe surgery had a modest impact on global network efficiency, despite the disruption caused.
This was due to alternative shortest paths in the network leading to widespread increases in betweenness
centrality post-surgery. Measurements of network change could retrospectively predict seizure outcomes with
79% accuracy and 65% specificity, which is twice as high as the empirical distribution. Fifteen connections
which changed due to surgery were identified as useful for prediction of outcome, eight of which connected to
the ipsilateral temporal pole.
Conclusion: Our results suggest that the use of network change metrics may have clinical value for predicting
seizure outcome. This approach could be used to prospectively predict outcomes given a suggested resection
mask using preoperative data only.
1. Introduction
Epilepsy is a serious neurological disorder characterised by re-
current unprovoked seizures affecting 1% of
the population.
Neurosurgical resection can bring remission in up to 80% of those with
refractory focal epilepsy, with 41% remaining entirely seizure free for
years (De Tisi et al., 2011). The most common type of epilepsy surgery
is anterior temporal lobe resection, in which the amygdala, anterior
hippocampus, and anterior temporal neocortex are removed. The
commonest neurological sequelae of temporal lobe surgery are memory
impairment, visual field deficits and word-finding difficulties (Jutila
et al., 2002; Gooneratne et al., 2017).
Recent studies have investigated surgical outcome by considering
the brain as a network of connected regions. Such networks can then be
subjected to quantitative analysis techniques, which measure local and
global properties in networks (see Bernhardt et al. (2015) for review).
Network measures that have been found to be altered in temporal lobe
epilepsy (TLE) include the clustering coefficient of a region, which
captures the connectedness of neighbours of a region (Bernhardt et al.,
2011). Furthermore, the strength of a connection (e.g. the number of
streamlines connecting two areas), or the strength of a region's con-
nectivity (e.g. the number of streamlines connecting a region to all
other regions) may also be altered in TLE (Besson et al., 2014a, 2014b;
Taylor et al., 2015). Another measure of a network is its efficiency,
which is a measure of network integration - i.e. how easy it is to travel
between one region to another via direct and indirect paths, and has
been shown to be altered in patients with TLE (Liu et al., 2014). Finally,
regression analysis and machine learning approaches have also been
⁎ Corresponding author at: Interdisciplinary Computing and Complex BioSystems Group, School of Computing Science, Newcastle University, UK.
E-mail address: [email protected] (P.N. Taylor).
1 Equal contribution.
https://doi.org/10.1016/j.nicl.2018.01.028
Received 26 July 2017; Received in revised form 5 December 2017; Accepted 21 January 2018
2213-1582/ © 2018 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/).
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
applied to brain networks of TLE to relate them to surgical outcome
(Bonilha et al., 2013; Munsell et al., 2015; Bonilha et al., 2015; Ji et al.,
2015).
A challenge in comparing networks across subjects is the choice of
an appropriate baseline or benchmark. There are two common ap-
proaches to this. One is to threshold the connectivity so that all subjects
have the same number of connections. A range of thresholds are then
checked, and the most significant results reported across thresholds
(Zhang et al., 2011). This has the drawback of removing 'weak' but
potentially important connections. A second approach is to compare the
network to a random network with the same number of regions and
connections. Typically this is done by either rewiring the existing net-
work (Maslov and Sneppen, 2002), or by generating a new network
according to predefined rules (Betzel et al., 2016; Bauer and Kaiser,
2017). Many different types of baseline networks can be used and this
will therefore influence results.
Recently Kuceyeski et al. (2013) introduced the network modifica-
tion (NeMo) tool in the context of stroke (Kuceyeski et al., 2015a,
2015b) and multiple sclerosis (Kuceyeski et al., 2015a, 2015b). The
NeMo tool is a method to enable a direct comparison between networks
that undergo change. For example, in their study of stroke, the authors
drew masks over stroke affected areas and overlaid this mask with data
from healthy subjects. Normal connectivity from the healthy subjects,
and altered connectivity (i.e. tracts which pass through the stroke
mask) were calculated. This approach allowed the authors to calculate a
change in connectivity metric (ChaCo), which was shown to correlate
with outcomes. Since the authors use the pre-stroke network as a
baseline to investigate the implied post-stroke differences, the analysis
is possible without the need to generate random networks or threshold
the connectivities. This obviates the need for arbitrarily chosen surro-
gate networks by effectively using the patient's own network as the
surrogate instead – a distinct advantage of the technique. A drawback of
that study is that the tractography was derived from a cohort of healthy
controls, rather than the stroke patients. Nonetheless, this framework is
ideally suited to investigate changes in networks, given a well-defined
alteration such as a stroke or surgery.
In this study we used a ChaCo-like approach in the context of epi-
lepsy surgery and addressed the following questions: What is the impact
of surgery on the patient's network? How does this impact graph the-
oretic properties such as region strength, network efficiency? Do these
changes to patient networks correlate with surgical outcome?
Although the resection masks we use in this study are derived ret-
rospectively from postoperative data, our methods could in future be
applied preoperatively using a mask of the intended resection.
2. Materials and methods
2.1. Patients & MRI acquisition
We retrospectively studied 53 patients who underwent temporal
lobe epilepsy surgery at the National Hospital for Neurology and
Neurosurgery, London, United Kingdom. Full patient details can be
found in Table S11, a summary is given in Table 1. Patient outcomes
were defined at 12 months postoperatively, according to the ILAE
classification of surgical outcomes (Wieser et al., 2001) and separated
into two groups. Group 1 includes patients who were completely sei-
zure free (ILAE 1), and group 2 incorporates all other possibilities (ILAE
2–6). No patient had any prior history of neurosurgery. We used a χ2
test to check for differences between outcome groups in gender, side of
surgery, and evidence of hippocampal sclerosis. We applied Kruskal-
Wallis test to check for differences in age between outcome groups.
All patients underwent preoperative anatomical T1-weighted MRI
and preoperative diffusion MRI. Postoperative T1-weighted MRI was
obtained within 12 months of surgery with the exception of one patient,
who was rescanned later.
MRI studies were performed on a 3T GE Signa HDx scanner (General
Table 1
Patient demographics and relation to outcome group.
ILAE 1
ILAE 2–6
Significance
N
36 (68%)
17 (32%)
Male/female
Left/right TLE
16/20
22/14
4/13
8/9
Age (mean, S.D./median,
I.Q.R.)
Hippocampal sclerosis
37, 11.6/
39.6, 19.25
25 (69%)
41.5, 10.6/
42.3, 10.8
10 (59%)
p = 0.3597,
χ2=0.839
p = 0.3353,
χ2=0.923
p = 0.2374
p = 0.4460,
χ2=0.5808
Electric, Waukesha, Milwaukee, WI). Standard imaging gradients with a
maximum strength of 40mT m−1 and slew rate 150 T m−1 s−1 were
used. All data were acquired using a body coil for transmission, and 8-
channel phased array coil for reception. Standard clinical sequences
were performed including a coronal T1-weighted volumetric acquisi-
tion with 170 contiguous 1.1 mm-thick slices (matrix, 256 × 256; in-
plane resolution, 0.9375 × 0.9375 mm).
Diffusion MRI data were acquired using a cardiac-triggered single-
shot spin-echo planar imaging sequence (Wheeler-Kingshott et al.,
2002) with echo time = 73 ms. Sets of 60 contiguous 2.4 mm-thick
axial slices were obtained covering the whole brain, with diffusion
sensitizing gradients applied in each of 52 noncollinear directions (b
value of 1,200mm2 s−1 [δ = 21 ms, Δ = 29 ms, using full gradient
strength of 40 mT m−1]) along with 6 non-diffusion weighted scans.
The gradient directions were calculated and ordered as described
elsewhere (Cook et al., 2007). The field of view was 24 cm, and the
acquisition matrix size was 96 × 96, zero filled to 128 × 128 during
reconstruction,
of
1.875 × 1.875 × 2.4 mm. The DTI acquisition time for a total of 3480
image slices was approximately 25 min (depending on subject heart
rate).
reconstructed
voxel
giving
a
size
2.2. Image processing
2.2.1. T1 processing
Preoperative anatomical MRI was used to generate parcellated re-
gions of interest (network nodes: ROIs). We used two different ap-
proaches to do this, generating two different parcellation schemes.
First, we used the FreeSurfer recon-all pipeline (https://surfer.nmr.
mgh.harvard.edu/), which performs intensity normalization, skull
stripping, subcortical volume generation, gray/white segmentation,
and parcellation (Fischl, 2012). The default parcellation scheme from
FreeSurfer (the Desikan-Killiany atlas (Fischl et al., 2002; Desikan et al.,
2006)) contains 82 cortical ROIs and subcortical ROIs and is widely
used in the literature (e.g. Munsell et al., 2015; Taylor et al., 2015). The
method FreeSurfer uses to generate its ROIs uses anatomical priors
based on a manually annotated dataset from healthy controls. However,
this may be suboptimal in the case of disease and therefore, we use a
second approach based on geodesic information flow (GIF) to generate
ROIs which has the advantage of performing well even in the presence
of neuropathology (Cardoso et al., 2015). Using the GIF approach, we
generate 114 cortical and subcortical ROIs (Table 2). A drawback of
using the GIF approach is comparison to previous studies is less
straightforward since most previous work use alternative atlases. The
results presented in the main manuscript use the GIF derived ROIs,
while we include results using FreeSurfer derived ROIs in supplemen-
tary materials to aid comparison to previous studies.
2.2.2. DWI processing
Preoperative diffusion MRI data were first corrected for signal drift
(Vos et al., 2016), then eddy current and movement artefacts were
203
Occipital fusiform gyrus
Opercular part of the
inferior frontal gyrus
Orbital part of the inferior
frontal gyrus
Posterior cingulate gyrus
Precuneus
Parahippocampal gyrus
Posterior insula
Parietal operculum
Postcentral gyrus
Posterior orbital gyrus
Planum polare
Precentral gyrus
Planum temporale
Subcallosal area
Superior frontal gyrus
Supplementary motor
cortex
Supramarginal gyrus
Superior occipital gyrus
Superior parietal lobule
Superior temporal gyrus
Temporal pole
OCP
OFuG
OpIFG
OrIFG
PCgG
PCu
PHG
PIns
PO
PoG
POrG
PP
PrG
PT
SCA
SFG
SMC
SMG
SOG
SPL
STG
TMP
gyrus
Anterior insula
Anterior orbital
gyrus
Angular gyrus
Calcarine cortex
Central operculum
Cuneus
Entorhinal area
Frontal operculum
Frontal pole
Fusiform gyrus
Gyrus rectus
Inferior occipital
gyrus
Inferior temporal
gyrus
AIns
AOrG
AnG
Calc
CO
Cun
Ent
FO
FRP
FuG
GRe
IOG
ITG
Lingual gyrus
Lateral orbital gyrus
Middle cingulate
LiG
LOrG
MCgG
gyrus
Medial frontal cortex MFC
MFG
Middle frontal gyrus
Middle occipital
MOG
gyrus
Medial orbital gyrus
Postcentral gyrus
medial segment
Precentral gyrus
medial segment
Superior frontal
gyrus medial
segment
Middle temporal
gyrus
MOrG
MPoG
MPrG
MSFG
MTG
P.N. Taylor et al.
Table 2
Full names of the abbreviated regions of interest (ROIs).
Full name
Abbreviation
Full name
Abbreviation
Anterior cingulate
ACgG
Occipital pole
Triangular part of the
inferior frontal gyrus
Transverse temporal gyrus
TrIFG
TTG
corrected using the FSL eddy_correct tool (Andersson and Sotiropoulos,
2016). The b vectors were then rotated appropriately using the 'fdt-
rotate-bvecs' tool as part of FSL (Jenkinson et al., 2012; Leemans and
Jones, 2009). The diffusion data were reconstructed using generalized
q-sampling imaging (Yeh et al., 2010) with a diffusion sampling length
ratio of 1.2. A deterministic fibre tracking algorithm (Yeh et al., 2013)
was then used, allowing for crossing fibres within voxels, with seeds
placed at the whole brain. Probabilistic approaches to fibre tracking
(e.g. Tournier et al., 2012) have been shown to perform less well with
respect to accuracy and number of false positive connections inferred.2
The choice of tractography algorithm is therefore related to whether
sensitivity or specificity is more important. It was recently shown
(Zalesky et al., 2016) that the introduction of false positive connections
is substantially more detrimental to the calculation of graph theoretic
measures such as efficiency and clustering than the introduction of false
negatives. We therefore use deterministic tractography since this has
been shown to have fewer false positive connections1. Default tracto-
graphy parameters from the 14 February 2017 build of DSI studio
software were used as follows. The angular threshold used was 60 de-
grees and the step size was set to 0.9375 mm. The anisotropy threshold
was determined automatically by DSI Studio. Tracks with length < 10
mm and > 300 mm were discarded. A total of 1,000,000 tracts were
calculated per subject and saved in diffusion space.
2.2.3. Resection mask
To draw resection masks we linearly registered the postoperative T1
2 http://www.tractometer.org/downloads/downloads/ismrm_presentation/Challenge_
ResultTractometer_updated_for_pdf_generation.pdf.
204
NeuroImage: Clinical 18 (2018) 202–214
MRI to the preoperative MRI using the FSL FLIRT tool (Jenkinson and
Smith, 2001; Jenkinson et al., 2002). Resection masks were then
manually drawn using the FslView software by overlaying the post-
operative MRI with the preoperative MRI starting at the most anterior
coronal slice, then proceeding posteriorly every three slices. Attention
was given to ensure masks did not extend beyond the Sylvian fissure
into inferior frontal lobe since this is unrealistic for anterior temporal
resection and there was no evidence for this on post-operative MRI.
Once complete, coronal slices were then joined by drawing in every
sagittal slice. Masks were saved in the preoperative T1 space. Linear
registration is justified in this instance because nonlinear deformation
of the tissue around the resection site may lead to misleading bound-
aries due to local deformations induced as part of the processing. All
registrations were visually inspected in detail during the drawing of the
resection mask.
2.3. Network generation
To align the tracts with the ROIs we linearly registered the pre-
operative T1 image to the first non-diffusion-weighted image and saved
the transformation matrix using FSL FLIRT. We then multiplied every
coordinate in every tract by the inverse of this transformation matrix to
get the tracts in T1 space. The quality of the registration between tracts,
ROIs, and the resection mask was confirmed through visual inspection
for all subjects. Since networks are constructed in native space, this
removes any mismatching of track types due to potential nonlinear
registration issues which is advantageous compared to previous studies
of network change (Kuceyeski et al., 2016). To generate preoperative
connectivity matrices, we looped through all tracts and deemed two
regions as connected if the two endpoints of the tract terminate in those
regions. This generated a weighted connectivity (adjacency) matrix in
which each entry in the matrix represents the number of streamlines
connecting two regions. To generate predicted postoperative con-
nectivity matrices we performed the same process as above with one
exception. Any tract that had any point within the resection mask was
excluded from building the matrix. The inferred postoperative network
therefore always had fewer streamlines than the preoperative network
and makes the assumption that remaining portions of removed tracts
subserve no functionality. This reduction in streamlines alone did not
explain outcome (Fig. S1). For analysis we applied a log10 transfor-
mation to connection weights. Following this, right hemisphere regions
(rows and columns in the matrix) were flipped in patients with a right
sided resection as we investigated ipsi- and contralateral differences. All
subjects had 72 regions per hemisphere.
A summary of the processing pipeline to generate the GIF con-
nectivity matrices is shown in Fig. 1.
2.4. Visualisation
To visualise the resection mask we first linearly registered the pre-
operative T1 to the MNI brain template using FSL FLIRT to generate a
transformation matrix. Next, we nonlinearly registered the preoperative
T1-weighted image to the MNI template using FSL FNIRT initialised
with the aforementioned transformation to generate a nonlinear warp.
Finally, we applied this warp to both the preoperative T1-weighted
image, and mask. All images were visually inspected to check for re-
gistration quality. We repeated this for all patients to generate a mask in
MNI space. These masks were then loaded into MatLab, binarised
(thresholded at 0.5), and summed across all subjects. This was then
saved, and overlaid with the MNI brain using FSLView. Masks in MNI
space were generated for visualisation purposes only and not used in
the analysis.
Three-dimensional projections of brain regions were produced using the
centre of mass for each region and visualised using BrainNet Viewer (Xia
et al., 2013). We created the scatter plots using the UniVarScatter function
in Matlab (https://github.com/GRousselet/matlab_visualisation).
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 1. Processing pipeline summary for GIF matrix generation.
The pipeline is applied to each subject.
2.5. Network analysis
We investigated properties of the networks using well established
graph theory measures that assessed the properties of a region, the
properties of a connection, and the properties of the network as a whole
(i.e. global). We excluded self-connections from all measures while
applying the weighted and undirected versions of the functions in the
Brain Connectivity Toolbox (Rubinov and Sporns, 2010). We in-
vestigated measures of strength, regional communicability (sum of
communicability matrix Estrada and Hatano, 2008), betweenness cen-
trality, clustering, and global efficiency, since several of these have
been suggested to be altered in previous studies (Bernhardt et al., 2011;
Besson et al., 2014a, 2014b; Taylor et al., 2015; Liu et al., 2014).
Region volume, region strength, connection strength changes were
computed by dividing the post-operative network property by the pre-
operative network property. For example, if node strength reduces from
10 pre-operatively to 2 in the inferred post-operative network the
change measurement will be 0.2. Other measures (betweenness, clus-
tering) can increase or decrease after surgery and so the difference was
computed by subtracting the pre-operative value from the post-opera-
tive value.
2.6. Elastic net regularisation and support vector machine classification
framework
We implemented a machine learning framework to investigate the
association of metrics indicating changes in the region strength, region
volumes, and connection strength with seizure free (ILAE 1) and non-
seizure free (ILAE 2–6) outcomes. We computed the aforementioned
metrics and treated them as feature vectors. Specifically, the feature
space included 380, 85, and 14 features computed from proportional
reduction in the connection strength, region strength, and region vo-
lumes respectively for each subject. This resulted in a high-dimensional
feature matrix, A ∈ ℝn×m, where m = 479 indicates the total number of
features and n = 53 denotes the total number of subjects in our study.
Since the number of features, m, is much larger than the number of
subjects, n, we implemented a two-step classification paradigm
(Munsell et al., 2015; Casanova et al., 2011). In the first step, we per-
formed feature selection by implementing logistic regression with
elastic net regularisation (Zou and Hastie, 2005). This gives a sparse m-
dimensional weight vector, x ∈ ℝm, which minimises the following
regularised logistic regression problem:
min
x
n
∑
i
1
=
w
i
log(1
e
+
i T i
x a
y
(
(
−
c
))
+
)
+
ρ
2
x
2
2
λ
+
x
1
(1)
where, y = (y1,y2,⋯,yn) is the n-dimensional vector representing the
surgical outcomes (+1 for seizure-free outcome and −1 for nonseizure-
free outcome); a Ti denotes the i − th row of feature vector A ∈ ℝn×m; wi
is the weight for the i − th sample (all samples were equally weighted
in our implementation); c is the scalar intercept; λ ∥ x∥1 is the l1 reg-
2 is the l2 regularisation (smooth-
ularisation (sparsity) term; and
2
ness) term. We optimised the cost function in Eq. 1 by applying the
LogisticR method in the sparse learning with efficient projections
(SLEP) software package (Liu et al., 2009a, 2009b; Liu and Ye, 2009;
Liu and Ye, 2009; Duchi et al., 2008).
xρ
2
The weight vector x computed from Eq. 1 is a sparse representation
of the training data set, i.e., the zero weights indicate the features
which are not associated with surgical outcome, whereas the non-zero
weights indicate the features associated with the surgical outcome. We
derived a binary mask by setting all the non-zero weights in x to 1.
Finally, we applied this mask to the high-dimensional feature space
while selecting only the features associated with surgical outcome, re-
sulting in a low-dimensional feature space.
In the second step of our classification framework, we incorporated
the aforementioned reduced feature representation in a support vector
machine (SVM) classifier with a linear kernel. We designed the SVM
classifier in MATLAB using the 'fitcsvm' class. Specifically, we divided
our data samples into test and training sets, incorporated a leave-one-
out cross-validation scheme, and computed the average accuracy, sen-
sitivity, and specificity. To prevent the classifier from evaluating
skewed classification performance due to the between-class im-
balance-36 in the seizure-free class as opposed to 17 in the non-sei-
zure-free class-we implemented a class-weighted SVM by setting all
class prior probabilities as uniform similar to previous studies (e.g.
Wang et al., 2012). This sets a higher misclassification penalty for the
minority class as compared to the majority class, thereby preventing the
classifier from learning simply from the underlying empirical data
distribution (Osuna et al., 1997; Veropoulos et al., 1999; Wu and
Rohini, 2004).
Since our objective was to determine a minimum set of features that
would most accurately separate the seizure-free outcome and the
nonseizure-free outcome classes, we incorporated a leave-one-out cross-
validated grid search on the l1 and l2 regularisation parameters (λ,ρ)
(Munsell et al., 2015; Casanova et al., 2011). We varied the λ parameter
205
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 2. Resection mask and region volume changes after surgery.
(a) Region mask in MNI space for visualisation only (all other analyses were performed in preoperative T1 space) shows similar spatial profile between outcome groups. (b) Absolute
amount of tissue resected does not relate with outcome. (c) Percentage of whole brain removed, does not relate to outcome. (d) Proportion of tissue remaining after surgery varies
between subjects and regions. Here, 1 represents that the region remained intact, 0 represents that region was removed entirely. Full region names can be found in Table 2 Each dot per
region is a single subject (53 per region in total).
sequentially from 0.01 to 0.96 in steps of 0.05 and the ρ parameter from
0.01 to 1.96 in steps of 0.05. At each grid point, we incorporated the
machine learning framework detailed above, obtained a reduced fea-
ture space, and computed the performance metrics indicating the ac-
curacy, sensitivity, and specificity.
3. Results
Firstly, we investigate the inferred impact of surgery on network
regions and secondly, on network connections. Finally, using machine
learning we investigate how network change metrics relate to outcomes
from surgery.
3.1. Impact of surgery on network regions
3.1.1. Region volumes
All patients had similar surgical resections with the most variation
between subjects on the boundary of the resection site (Fig. 2a). The
absolute (Fig. 2b), and relative (Fig. 2c) volume of the resected tissue
did vary between subjects, but did not explain outcome (p = 0.56 and
p = 0.62 respectively – permutation test for difference between mean,
10,000 permutations used). Fig. 2d shows the proportion of tissue re-
maining after resection in different brain regions. Unsurprisingly the
regions which were most reduced in volume were amygdala, hippo-
campus, entorhinal cortex, parahippocampal gyrus, and the temporal
pole. This suggests partial disruption to multiple regions, rather than
complete disruption to specific regions. Although there is variation
between subjects in the proportion of each region removed, this did not
predict outcome (Fig. S2). These results are reproduced for the Free-
Surfer atlas in Fig. S3.
3.1.2. Region strength
Each brain region is connected to other regions with streamlines.
The sum of all streamlines directly connected to a region is termed the
region strength - i.e. how strongly a region is directly connected to all
other regions. Fig. 3 shows the impact of removing all streamlines that
pass into the resected tissue by plotting the resultant change in strength.
As expected, those regions that are partially removed (i.e. those
showing a reduction in Fig. 2d) also have reduced strengths. The effect
of surgery on other brain regions can be demonstrated with this ap-
proach. For example, the ipsilateral basal forebrain, although it is ex-
tratemporal and not resected, has a substantial reduction in strength for
many patients. This suggests that preoperatively the basal forebrain has
many of its connections with the resected tissue, rather than other
areas. Several other areas beyond the resected regions also undergo
alterations in strength in many subjects including the ipsilateral lateral,
medial and posterior, orbital gyri among others. These results are
shown in detail in Table S12(a, b), and repeated for the FreeSurfer
regions in Fig. S4. It is worth noting that region strength changes shown
here may not be limited to only regions partly, or wholly resected, but
may also change if a portion of a tract travels through the resection site,
effectively disconnecting two non-resected areas.
Broadly speaking, the strengths of contralateral regions were not as
substantially affected by the surgery which suggests most connections
are intrahemispheric. None of the changes in node strength met sig-
nificance for association with seizure-free outcome or not.
3.1.3. Region network measures
The region volume change (Fig. 2) captures information solely
about individual regions, while the region strength change (Fig. 3)
captures information about a region and its directly connected neigh-
bours. Other measures of network regions capture information re-
garding aspects of their role in the wider network.
The betweenness centrality of a region represents how many of the
shortest paths between other regions traverse the region. Hub regions
tend to occur in many of the shortest paths and have been implicated in
206
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 3. Changes in region strength are predominantly ipsilateral.
(a) Proportion of strength remaining after surgery. A value of 1 (0) indicates all streamlines connecting to that region are kept (removed). (b) An example region shows ipsilateral
hippocampal connection change in the ILAE > 1 group, and the ILAE 1 group. (c) Substantially affected regions visualised in 3D MNI space. Size and colour correspond to median
reduction across all patients. Colour bars in panel (c) and (a) are consistent and correspond to the proportion of strength remaining between 0 and 1.
a wide range of neurological disorders (Crossley et al., 2014). If a region
is surgically removed, the shortest path between regions may be via
other areas. The betweenness centrality of a region can therefore either
increase or decrease following surgery.
Fig. 4 shows the change in betweenness centrality after surgery.
Regions involved in the resection have reductions in betweenness
centrality suggesting reduced integration of those areas with the wider
network. However, if those regions are no longer occurring in shortest
paths, the following question remains: which ones are? Fig. 4a shows
that many regions have substantial increase in betweenness centrality.
The redirection of shortest paths in the network following surgery ap-
pears to be widely distributed among many other regions, rather than
just the hubs - suggesting there are many alternative pathways (i.e.
redundancy). Panel (b) in Fig. 4 shows the spatial distribution of the
median change across all patients. This confirms that the increases
(shown in blue) are widespread, even in contralateral regions.
Because of the widespread redistribution of the shortest paths, we
also evaluated global efficiency, which is the inverse of the average
shortest path length of the network. Fig. 4c demonstrates the reduction
in global efficiency following surgery. Given the redirection of shortest
paths shown in panel (a) of the same figure, it is now easy to see why
the reduction on global efficiency is only < 10% in most cases. This
change does not relate to seizure outcome (p = 0.49 – permutation test
for difference between mean, 10,000 permutations used). We observed
similar results when using the FreeSurfer atlas (Fig. S5).
We additionally investigated the change in regional clustering
coefficient, a measure of the interconnectedness of neighbours of a
region, following surgery. We found very subtle differences in this
measure beyond the resection site following surgery (Fig. S6), which
did not relate to outcome. Reductions of regional network communic-
ability were most pronounced in ipsilateral amygdala, entorhinal, and
temporal pole regions but also did not relate to outcome (p > 0.05,
Kruskal-Wallis test, supplementary results S7).
3.2. Impact of surgery on network connections
3.2.1. Connection strength
Fig. 5a shows the connections which are affected by surgery in all
patients. These connections typically involve those areas which are
resected. Panel (b) shows the median reduction in connectivity fol-
lowing surgery. Connections in panel (a) are a subset of those in panel
(b). There are many connections which are only partly (as opposed to
entirely) reduced in strength. Partial reduction in strength can be
considered as removing some, but not all, of the connectivity between
two areas. The greatest impact was seen in connections into the ipsi-
lateral posterior temporal lobe and the inferior frontal lobe, with in
some patients, connections to the ipsilateral parietal lobe and con-
tralateral temporal lobe and medial frontal and parietal lobes as would
be expected given Fig. 3.
3.2.2. Connection network measures
Similar to the regional betweenness centrality, connection (edge)
betweenness centrality captures information regarding how often a
connection occurs in the shortest path between other regions – i.e. how
frequently that connection is traversed in the shortest paths. This value
can increase or decrease following surgery, as alternative paths is taken.
207
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 4. Widespread changes in region betweenness centrality following surgery.
(a) Enhancements (positive values) and reductions (negative values) in betweenness centrality following surgery. A value of 0 indicates no change. (b) Median change of betweenness
centrality shows that decreases are generally constrained to the resected regions (red) whereas increases are widespread across many other regions (blue). Colour bars in panel (b) and (a)
are consistent and correspond to the change in betweenness centrality (y axis in panel a). Red (blue) indicates a decrease (increase) in betweenness centrality after surgery. (c) Global
network efficiency is typically reduced by < 10% following surgery, and does not predict outcome.
Fig. 6 shows the median change in betweenness centrality across all
subjects. Decreases in connection betweenness centrality were broadly
limited to ipsilateral temporal areas whereas the increases were more
widely distributed.
3.3. Network change association with seizure outcome
As an outlook to prospective applications, we used a machine
learning strategy to retrospectively investigate the association between
Fig. 5. Connectivity is disrupted by surgery.
(a) Connections which are reduced in strength following surgery in all patients. (b) Median reduction in connectivity strength across all patients. A value of 1 (0) indicates that the median
change to a connection following surgery is a 0% (100%) reduction in the number of streamlines. Only values < 1 are shown and therefore shows only connections which are reduced in
strength in the majority of patients (as opposed to connections which are reduced in all patients - shown in panel a). Connections in panel (b) are a superset of those in panel (a).
208
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 6. Widespread changes in connection betweenness centrality following surgery.
Median decreases (increases) in connection betweenness centrality across all subjects shown in red (blue).
network changes – region volume, region strength, and connection
strength – with seizure outcome.
Our machine learning approach minimises the number of features
that give a high accuracy of outcome classification. As input features we
incorporated all changes in volume, region strength, and connection
strength and applied an elastic net algorithm to select those which were
most useful. Fig. 7 shows those features selected by the algorithm.
Fifteen features were selected, all of which were connections. This
Fig. 7. Features derived with the elastic net algorithm.
(a) Fifteen features which were found to be most informative using our machine learning pipeline. All are connections which are reduced in strength following surgery. Colour coding
shows the difference in mean reduction between groups. Connections in red are reduced by surgery more in seizure-free patients than in not-seizure-free patients. (b) Correspondence
with resected tissue.
209
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
suggests that incorporating knowledge of regional properties (strength
and volume change) does not improve the prediction substantially. All
fifteen connections in Fig. 7 are coloured according to the difference in
mean value between groups. In other words, a negative value shows
that the mean ILAE > 1 patient has a greater reduction in connectivity
than the mean ILAE 1 patient (and vice-versa). The greatest reduction
any group can have is 1 (i.e. a 100% reduction). Values are therefore
bound between ± 1, and consequently not normally distributed. We
therefore show the difference in mean, rather than the effect size. Eight
of the features are connections involving the temporal pole. The list of
selected features is given in Table S13 and are visualised in movie S14.
Replication of these results using normalised data, preoperative data, or
the alternative FreeSurfer parcellation did not lead to substantial im-
provements in accuracy (Supplementary results S10). Those who be-
came seizure free had greater reductions of connectivity into the ipsi-
lateral superior temporal lobe and frontal lobe.
Table 3 shows our findings including accuracy, sensitivity and
specificity when using these features. Our specificity of 64.7% reflects
the finding that of all patients that the model predicted would have a
not-seizure-free outcome, 64.7% actually had a not-seizure-free out-
come. This is substantially higher than the empirical distribution
(32%).
4. Discussion
In this study we have used detailed neuroanatomical information
from both pre- and postoperative MRI to infer the impact of surgery on
patient-specific brain networks. Our three main contributions are as
follows.
First, we find that the impact of surgery leads to a reduction
of < 10% in efficiency in the majority of patients, due to a redirection
of shortest paths. Second, there is no single feature which fully accounts
for outcome with 100% accuracy. This means that there is likely not a
single 'target' to aim for with surgery in this patient group that is
captured by our data. Third, we have demonstrated that machine
learning, in conjunction with connectivity change metrics, can produce
predictions with high accuracy, and specificity which is twice as high as
the empirical distribution. This means that around two thirds of the
patients we predict are not seizure free, are actually not seizure free.
This could help preoperatively in identifying which patients might have
lower chances of becoming seizure free and put extra consideration in
offering surgery or when consenting those patients.
Most previous network-based studies in epilepsy typically perform
group comparisons of patients relative to controls using either struc-
tural networks (Kamiya et al., 2016; Taylor et al., 2015; Lemkaddem
et al., 2014; de Salvo et al., 2014; Bonilha et al., 2012), functional
networks (Chavez et al., 2010; Liao et al., 2010), or both (Zhang et al.,
2011; Douw et al., 2015; Wirsich et al., 2016). Few have investigated
alterations in global network properties with respect to surgical out-
come (although see He et al., 2017 and Morgan et al., 2017 for recent
examples). Bonilha and colleagues utilised preoperative structural
networks inferred from diffusion MRI to investigate network properties
with respect to surgical outcomes (Bonilha et al., 2013). This includes a
multi-centre study using machine learning which is perhaps most si-
milar to our work (Munsell et al., 2015). That study analysed pre-
operative structural networks and compared several machine learning
algorithms for feature selection and classification for prediction of
surgical outcomes. They found that the best performing algorithm uti-
lised an elastic net/support vector machine combination, and achieved
a cross-site accuracy of 66%, with a same-site accuracy of 70%. We
therefore chose to use broadly the same technique for classification
here. Instead of investigating preoperative networks, however, we in-
vestigate the difference networks - i.e. the inferred changes brought
about by patient-specific surgery. These difference networks have led us
to achieve a same-site accuracy of 79% in our data.
We have performed a retrospective analysis of connectivity change
using a mask drawn from postoperative data – i.e. the tissue that was
resected. However, we envisage that in future a prospective prediction
could be performed. This should be possible using the preoperative data
and a mask of what the surgeon intends to resect (e.g. through the
manual adaptation of an 'average mask' in the case of a standardised
surgery such as anterior temporal lobe resection). Indeed, in theory a
prediction is possible for any given presurgical resection plan. Fig. 8
suggests how this could be incorporated into such an evaluation, with
the aim to address the question of, 'will this proposed resection render
the patient seizure free?' We do not address the separate question of,
'what is the optimal resection for this patient?', which has to integrate
the need to avoid damage to critical structures (e.g. Winston et al.,
2012). This study has developed a pipeline for answering the first of
those two questions, and retrospectively validated it with 79% accu-
racy. Approaches to investigate the second of those questions have been
proposed recently (Sinha et al., 2017; Goodfellow et al., 2016).
4.1. Regional change following surgery
A potentially surprising result (Fig. 2), is that we found that the
amount of resected tissue did not appear to predict outcome. This is in
contrast to the work of (Wyler et al., 1995) and (Shamim et al., 2009)
who have shown that larger resections do correlate with seizure out-
comes. However, other studies have found conflicting results which
accord with our own findings (Hardy et al., 2003; Keller et al., 2015;
Liao et al., 2016). The variations in the literature may be due to dif-
ferent populations of left/right TLE, evidence of hippocampal sclerosis,
and follow-up duration. Here we have controlled for these aspects
where possible with no significant differences between outcome groups
(Table 1).
Our finding that there is variation between subjects, in terms of
proportion of regions removed, can be explained by two potential fac-
tors. First, there are differences in tissue removed between subjects, and
second, there are differences in regions between subjects. The latter was
one of our motivations for using two different parcellations to aid
comparison (the second of which is shown in supplementary materials).
Table 3
Confusion matrix indicating performance of machine learning framework for predicting surgical outcomes.
Actual surgical outcome
Predicted
outcome
Seizure free = 37
Not seizure free = 16
Accuracy = 0.792
Seizure free = 36
True positive = 31
False negative = 5
(Type II Error)
True Positive Rate,
or Sensitivity = 0.861
False Negative Rate,
or Miss rate = 0.139
210
Not seizure free = 17
False positive = 6
(Type I Error)
True negative = 11
False Positive Rate
or Fall out = 0.353
True Negative Rate,
or Specificity = 0.647
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Fig. 8. Potential incorporation of pipeline for prospective evaluation.
Suggestion of how our approach could be used as part of the pre-surgical evaluation. From left to right: Multimodal data is acquired and evaluated at the multi-disciplinary team meeting.
If surgery is recommended a proposed surgical resection mask is drawn, then fed into the connection change pipeline. The pipeline then gives a predicted post-operative network given the
patient-specific mask and patient-specific presurgical DTI & T1w MRI. The pipeline uses these predicted connectivity changes, along with the pre-existing feature set to predict seizure
outcome. In this example the predicted outcome would be seizure free.
Furthermore, the resolution of the parcellation (i.e. the number of re-
gions that the brain network is divided into) will also affect results.
Here we use fairly low resolution parcellations of ≈ 100 regions in
contrast to recently proposed higher resolutions (Besson et al., 2014a,
2014b; Taylor et al., 2017; Besson et al., 2017). Our reasoning for this is
that in this study we compare regional properties between subjects;
thus it is important that we are confident that one region in one subject
is the same region in another, and can be reliably reproduced. One of
the reasons that the FreeSurfer software is so popular for this is that the
authors have demonstrated confidence maps at region boundaries
(Fischl et al., 2004). One of the drawbacks of using such low resolution
connectivity matrices however is the inability to investigate detailed
within-area architecture as in (Taylor et al., 2017). Indeed, this may be
an important and productive area of future research.
4.2. Network change following surgery
We suggest that our finding of only a < 10% reduction in network
efficiency following surgery (Fig. 4c) may explain why surgery usually
leads to relatively few severe side effects. The reason for only a subtle
change in efficiency is the redirection of shortest paths via other con-
nections as shown in Fig. 6. We speculate that the decreases in edge
betweenness correspond to those connections which undergo Wallerian
degeneration. Furthermore, we speculate that connections with in-
creased betweenness centrality will undergo 'strengthening', potentially
reflected by increases in FA on structural imaging postoperatively.
Qualitative evidence to support this has been shown by (Winston et al.,
2014) who found decreases in FA in tracts involving the resected tissue
and increased FA in more distant ipsilateral superior tracts within fairly
short timescales of three months. Further support for this hypothesis
comes from a study by (Jeong et al., 2016) who report increases in the
FA of connections following frontal lobe epilepsy surgery far from the
surgical site. Other studies have also made similar findings (Yogarajah
et al., 2010; Pustina et al., 2014; Faber et al., 2013). Taken together,
this suggests a potential rerouting of connectivity, via new shortest
paths, leading to increased FA and thus avoiding major cognitive def-
icits following surgery.
The relationship between network efficiency, shortest path lengths,
and function is not entirely clear. However, path transitivity, which is
strongly related to path length has been shown to strongly correlate to
functional connectivity (Goñi et al., 2014). Network efficiency itself has
also been shown to relate to intelligence in functional networks in
healthy controls (van den Heuvel et al., 2009). It remains to be seen
how changes to network efficiency metrics following surgery relate to
patient changes in brain function beyond seizure outcome.
In addition to widespread network changes such as those described
above, the network change pipeline also allows the inference of more
localised differences. As shown in Fig. 3, we found decreases in node
strength in extratemporal areas, particularly the basal forebrain in the
frontal lobe. This leads us to suggest that frontal lobe function may be
altered following surgery. Several studies are in agreement with this.
For example, Stretton et al. (2014) found improvements in frontal lobe
working memory function after surgery, while Martin et al. (2000)
showed significant
increases in verbal fluency following surgery.
However, other studies have also found conflicting results (see Stretton
and Thompson, 2012 for review). We hypothesise these improvements
are related to post-surgical alterations in connectivity, and the varia-
bility in the results may be down to patient variation (i.e. the large
spread in Fig. 3). This effect is typically referred to as a process of
functional improvement once nociferous cortex has been removed.
Future studies correlating these changes in connectivity with changes in
frontal lobe function will help to elucidate these relationships, poten-
tially predicting not only seizure outcome, but also functional out-
comes. Indeed, this has already been demonstrated in stroke where
connectivity change metrics have been found to relate to functional
outcomes (Kuceyeski et al., 2015a, 2015b).
4.3. Machine learning for outcome prediction
Although very few have used brain network data, several studies in
211
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
recent years have begun to apply machine learning strategies for sur-
gical outcome prediction. The benefits of the approach appear pro-
mising. For example, (Bernhardt et al., 2015), using hippocampal
measurements as features, made outcome predictions with 81% accu-
racy. In another study of surgical outcomes in temporal lobe epilepsy
(Memarian et al., 2015) demonstrated accuracy of 95% using various
features derived from both imaging and demographic data. In addition,
(Feis et al., 2013) achieved accuracy of outcome prediction of up to
96% using preoperative T1 measurements of white matter volumes.
Other studies have also used machine learning algorithms, along with
imaging data, to classify lateralisation of seizure onset (Focke et al.,
2012; Bennett et al., 2017). Our prediction accuracy is in a broadly
similar range to many of these and may be improved by including other
features such as those used for generating predictive nomograms (Jehi
et al., 2015).
Multiple approaches exist in machine learning to attempt to im-
prove the accuracy of the outcome prediction by selecting the most
predictive features from a high-dimensional feature space (Tibshirani,
1996, Hoerl and Kennard, 2004, Breiman et al., 1984, Breiman et al.,
1984). Here we have adopted a fairly conservative approach in which
we implement elastic net regularisation (Zou and Hastie, 2005) while
incorporating regularisation parameters which identify a minimum
number of features important for the prediction. Even using this con-
servative approach we obtain a prediction accuracy of 79.2%. It is
worth noting that we impose a prior probability expectation of 50%. If
we did not impose this, and instead used the empirical distribution of
outcomes (68%: 32% - row 1 in Table 1), it would be easy to get ac-
curacies of at least 68% by simply suggesting all patients will be seizure-
free. The prediction accuracy of 79.2%, even using our highly con-
servative approach, is therefore encouraging.
4.4. Strengths & weaknesses
There are several strengths to this work. For example, here we have
a fairly large cohort of 53 patients which were all scanned on the same
scanner using the same scanning protocol. We also have good con-
sistency in terms of operation type (all anterior temporal lobe resection
performed by the same 2 surgeons using the same technique), and
consistency in follow-up duration for outcome - 12 months. The con-
sistency in follow-up duration is particularly important because out-
comes can change over time (De Tisi et al., 2011). A further strength of
our study, in comparison to some other network connectivity change
studies, is that our analysis is performed in subject (native) space. This
means nonlinear deformations of the scans to a common space, which
may decrease the accuracy of the region mapping, are not necessary.
Our investigation of changes in structural connectivity following surgery
using a ChaCo-like approach is, to our knowledge, novel in epilepsy.
Moreover, the ChaCo approach has the advantage of not requiring ar-
bitrary thresholding of the network, or comparison to arbitrarily chosen
random networks as a baseline – instead the patient's own pre-operative
network is used.
A possible weakness of this study is that we have flipped the con-
nectivity matrices in right TLE patients, simplifying our analysis into
ipsi- and contralateral hemispheres. Although there is precedence for
this approach (Bonilha et al., 2013; Munsell et al., 2015; Ji et al., 2015)
it has been suggested that left and right TLE are not simply the mirror
image of each other (Besson et al., 2014a, 2014b), especially with re-
spect to post-surgical cognitive deficits and resection size which is ty-
pically larger in the non-dominant hemisphere. We did try to mitigate
this however by ensuring the proportion of left:right patients in each
outcome group are similar (p = 0.3353; Table 1).
Another potential bias is the resection mask, which was manually
drawn for all subjects by the same rater, with a majority subset drawn
by a second rater (results S9). Variations between raters exist, but are
small. The correlation of network features between raters is > 0.98,
giving confidence in the reproducibility of the masks drawn here. Our
choice of tractography algorithm was influenced by the intention to
limit false positive connections as suggested by Zalesky et al. (2016).
Despite this, there are known limitations with all approaches and this
study should also be considered with the known limitations of the
imaging technologies in mind. For example, the inability to resolve
connection direction and potential bias for tractography to favour
shorter, straighter connections (Jones, 2010). Our use of number of
streamlines as a connectivity metric is based on the extensive prior
literature of ChaCo analysis by Kuceyeski et al. (2013, 2015a, 2015b).
However, other connection strength metrics, such as density (Taylor
et al., 2015), or volume (Irimia and Van Horn, 2016) may also yield
informative results.
Our network change pipeline allows to infer the immediate network
change brought about by surgery. It is worth highlighting that this does
not necessarily represent the long-term brain network changes in the
months/years that occur as a result of e.g. plasticity or degeneration.
This is both a strength and a weakness of our approach and is one
reason why we assess outcomes at only 12 months. The mechanisms
underlying why some patients relapse to seizure recurrence only at
several years after surgery (De Tisi et al., 2011) are unclear and would
require longitudinal diffusion MRI data gathered over several years to
investigate in this framework. The potential to use our approach pro-
spectively (Fig. 8) is a distinct advantage however.
4.5. Outlook and wider implications
Although we have focussed exclusively on temporal lobe epilepsy
surgery here, this method could be generalized to other types of neu-
rosurgery such as those for other epilepsies such as neocortical epilepsy,
tumours or movement disorders. Additionally, other outcome criteria
could also be predicted. For example, instead of predicting seizure
outcome, functional deficits arising due to damage to known functional
subnetworks (e.g. somatomotor, attention, visual) could be predicted
using existing atlases (Yeo et al., 2011). Our approach would be pos-
sible if preoperative imaging data is available and a proposed resection
mask drawn before the surgery. Deficits for the given mask (surgery)
could then be predicted for individual subjects. Additionally, multiple
proposed masks (surgeries) could be tested against each other in-silico,
the results for each compared, and the optimal surgery predicted. Al-
gorithms for a similar approach are already being developed for in-
vasive electrode recordings (Rodionov et al., 2013; Sinha et al., 2017).
4.6. Conclusion
It is to be expected that outcome from surgery will depend not only
on the network connectivity of the brain before surgery, but the dis-
ruption to this network caused by the surgery itself. This can be mea-
sured as the change in connectivity, which can be predicted for a given
resection preoperatively. Our analysis shows that such connectivity
change metrics can be used to retrospectively predict seizure outcome
with high accuracy, and may relate to functional outcomes.
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.nicl.2018.01.028.
Acknowledgments
We are grateful to the Wolfson Foundation and the Epilepsy Society
for supporting the Epilepsy Society MRI scanner.
PNT was funded by Wellcome Trust (105617/Z/14/Z). PNT thanks
Rob Forsyth and Helen Marley for discussion. GPW was supported by
the MRC (G0802012, MR/M00841X/1). SBV is funded by the National
Institute for Health Research University College London Hospitals
Biomedical Research Centre (NIHR BRC UCLH/UCL High Impact
Initiative).
212
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
This study was carried out in part at UCLH and the support of the
NIHR-funded UCLH/UCL BRC is gratefully acknowledged. JD is sup-
ported by the NIHR Senior investigator scheme.
References
Andersson, Jesper L.R., Sotiropoulos, Stamatios N., 2016. An integrated approach to
correction for off-resonance effects and subject movement in diffusion mr imaging.
NeuroImage 125, 1063–1078 (Elsevier).
Bauer, Roman, Kaiser, Marcus, 2017. Nonlinear growth: an origin of hub organization in
complex networks. Royal Soc. Open Sc. 4 (3), 160691 (The Royal Society).
Bennett, O., Cardoso, J., Duncan, J.S., Winston, G., Ourselin, S., 2017. Learning how to
See the Invisible - Using Machine Learning to Find Underlying Abnormality Patterns
in Reportedly Normal MR Brain Images from Patients with Epilepsy. Proceedings.
ISMRM.
Bernhardt, Boris C., Chen, Zhang, He, Yong, Evans, Alan C., Bernasconi, Neda, 2011.
Graph-theoretical analysis reveals disrupted small-world organization of cortical
thickness correlation networks in temporal lobe epilepsy. Cereb. Cortex 21 (9),
2147–2157 (Oxford Univ Press).
Bernhardt, Boris C., Hong, Seok-Jun, Bernasconi, Andrea, Bernasconi, Neda, 2015.
Magnetic resonance imaging pattern learning in temporal lobe epilepsy: classification
and prognostics. Ann. Neurol. 77 (3), 436–446 (Wiley Online Library).
Besson, Pierre, Dinkelacker, Vera, Valabregue, Romain, Thivard, Lionel, Leclerc, Xavier,
Baulac, Michel, Sammler, Daniela, et al., 2014a. Structural connectivity differences in
left and right temporal lobe epilepsy. NeuroImage 100, 135–144 (Elsevier).
Besson, Pierre, Lopes, Renaud, Leclerc, Xavier, Derambure, Philippe, Tyvaert, Louise,
2014b. Intra-subject reliability of the high-resolution whole-brain structural con-
nectome. NeuroImage 102, 283–293 (Elsevier).
Besson, Pierre, Kathleen Bandt, S., Proix, Timothée, Lagarde, Stanislas, Jirsa, Viktor K.,
Ranjeva, Jean-Philippe, Bartolomei, Fabrice, Guye, Maxime, 2017. Anatomic con-
sistencies across epilepsies: a stereotactic-EEG informed high-resolution structural
connectivity study. Brain 140 (10), 2639–2652.
Betzel, Richard F., Avena-Koenigsberger, Andrea, Goñi, Joaquín, He, Ye, De Reus, Marcel
A., Griffa, Alessandra, Vértes, Petra E., et al., 2016. Generative models of the human
connectome. NeuroImage 124, 1054–1064 (Elsevier).
Bonilha, Leonardo, Nesland, Travis, Martz, Gabriel U., Joseph, Jane E., Spampinato,
Maria V., Edwards, Jonathan C., Tabesh, Ali, 2012. Medial temporal lobe epilepsy is
associated with neuronal fibre loss and paradoxical increase in structural connectivity
of limbic structures. J. Neurol. Neurosurg. Psychiatry 83 (9), 903–909 (BMJ
Publishing Group Ltd).
Bonilha, Leonardo, Helpern, Joseph A., Sainju, Rup, Nesland, Travis, Edwards, Jonathan
C., Glazier, Steven S., Tabesh, Ali, 2013. Presurgical connectome and postsurgical
seizure control in temporal lobe epilepsy. Neurology 81 (19), 1704–1710 (AAN
Enterprises).
Bonilha, Leonardo, Jensen, Jens H., Baker, Nathaniel, Breedlove, Jesse, Nesland, Travis,
Lin, Jack J., Drane, Daniel L., Saindane, Amit M., Binder, Jeffrey R., Kuzniecky,
Ruben I., 2015. The brain connectome as a personalized biomarker of seizure out-
comes after temporal lobectomy. Neurology 84 (18), 1846–1853 (AAN Enterprises).
Breiman, L., Friedman, J.H., Alshen, R.A., Stone, C.J., 1984. CART: Classification and
Regression Trees. Wadsworth, Belmont, CA.
Cardoso, M. Jorge, Modat, Marc, Wolz, Robin, Melbourne, Andrew, Cash, David,
Rueckert, Daniel, Ourselin, Sebastien, 2015. Geodesic information flows: spatially-
variant graphs and their application to segmentation and fusion. IEEE Trans. Med.
Imaging 34 (9), 1976–1988 (IEEE).
Casanova, Ramon, Wagner, Benjamin, Whitlow, Christopher, Williamson, Jeff, Shumaker,
Sally, Maldjian, Joseph, Espeland, Mark, 2011. High dimensional classification of
structural Mri Alzheimer's disease data based on large scale regularization. Front.
Neuroinform. 5, 22. http://dx.doi.org/10.3389/fninf.2011.00022.
Chavez, Mario, Valencia, Miguel, Navarro, Vincent, Latora, Vito, Martinerie, Jacques,
2010. Functional modularity of background activities in normal and epileptic brain
networks. Phys. Rev. Lett. 104 (11), 118701 (APS).
Cook, Philip A., Symms, Mark, Boulby, Philip A., Alexander, Daniel C., 2007. Optimal
acquisition orders of diffusion-weighted mri measurements. J. Magn. Reson. Imaging
25 (5), 1051–1058 (Wiley Online Library).
Crossley, Nicolas A., Mechelli, Andrea, Scott, Jessica, Carletti, Francesco, Fox, Peter T.,
McGuire, Philip, Bullmore, Edward T., 2014. The hubs of the human connectome are
generally implicated in the anatomy of brain disorders. Brain 137 (8), 2382–2395
(Oxford Univ Press).
De Tisi, Jane, Bell, Gail S., Peacock, Janet L., McEvoy, Andrew W., Harkness, William F.J.,
Sander, Josemir W., Duncan, John S., 2011. The long-term outcome of adult epilepsy
surgery, patterns of seizure remission, and relapse: a cohort study. Lancet 378 (9800),
1388–1395 (Elsevier).
Desikan, Rahul S., Ségonne, Florent, Fischl, Bruce, Quinn, Brian T., Dickerson, Bradford
C., Blacker, Deborah, Buckner, Randy L., et al., 2006. An automated labeling system
for subdividing the human cerebral cortex on Mri scans into gyral based regions of
interest. NeuroImage 31 (3), 968–980 (Elsevier).
Douw, Linda, DeSalvo, Matthew N., Tanaka, Naoaki, Cole, Andrew J., Liu, Hesheng,
Reinsberger, Claus, Stufflebeam, Steven M., 2015. Dissociated multimodal hubs and
seizures in temporal lobe epilepsy. Ann. Clin. Transl. Neurol. 2 (4), 338–352 (Wiley
Online Library).
Duchi, John, Shalev-Shwartz, Shai, Singer, Yoram, Chandra, Tushar, 2008. Efficient
projections onto the L1-ball for learning in high dimensions. In: Proceedings of the
25th International Conference on Machine Learning. ACM, New York, NY, USA, pp.
272–279. (ICML '08). https://doi.org/10.1145/1390156.1390191.
Estrada, E., Hatano, N., 2008. Communicability in complex networks. Phys. Rev. E 77 (3),
036111.
Faber, Jennifer, Schoene-Bake, Jan-Christoph, Trautner, Peter, Lehe, Marec, Elger,
Christian E., Weber, Bernd, 2013. Progressive fiber tract affections after temporal
lobe surgery. Epilepsia 54 (4), e53-e57.
Feis, Delia-Lisa, Schoene-Bake, Jan-Christoph, Elger, Christian, Wagner, Jan,
Tittgemeyer, Marc, Weber, Bernd, 2013. Prediction of post-surgical seizure outcome
in left mesial temporal lobe epilepsy. Neurol. Clin. 2, 903–911 (Elsevier).
Fischl, Bruce, 2012. FreeSurfer. NeuroImage 62 (2), 774–781 (Elsevier).
Fischl, Bruce, Salat, David H., Busa, Evelina, Albert, Marilyn, Dieterich, Megan,
Haselgrove, Christian, van der Kouwe, Andre, et al., 2002. Whole brain segmentation:
automated labeling of neuroanatomical structures in the human brain. Neuron 33 (3),
341–355 (Elsevier).
Fischl, Bruce, van der Kouwe, André, Destrieux, Christophe, Halgren, Eric, Ségonne,
Florent, Salat, David H., Busa, Evelina, et al., 2004. Automatically parcellating the
human cerebral cortex. Cereb. Cortex 14 (1), 11–22 (Oxford Univ Press).
Focke, Niels K., Yogarajah, Mahinda, Symms, Mark R., Gruber, Oliver, Paulus, Walter,
Duncan, John S., 2012. Automated Mr image classification in temporal lobe epilepsy.
NeuroImage 59 (1), 356–362 (Elsevier).
Goñi, J., van den Heuvel, M.P., Avena-Koenigsberger, A., Velez de Mendizabal, N., Betzel,
R.F., Griffa, A., ... Sporns, O., 2014. Resting-brain functional connectivity predicted
by analytic measures of network communication. Proc. Natl. Acad. Sci. U. S. A. 111
(2), 833–838. http://dx.doi.org/10.1073/pnas.1315529111.
Goodfellow, M., Rummel, C., Abela, E., Richardson, M.P., Schindler, K., Terry, J.R., 2016.
Estimation of brain network ictogenicity predicts outcome from epilepsy surgery. Sci.
Rep. 6.
Gooneratne, I.K., Mannan, S., de Tisi, J., Gonzalez, J.C., McEvoy, A.W., Miserocchi, A.,
Diehl, B., Wehner, T., Bell, G.S., Sander, J.W., Duncan, J.S., 2017. Somatic compli-
cations of epilepsy surgery over 25 years at a single center. Epilepsy Res. 132, 70–77.
Hardy, Steven G., Miller, John W., Holmes, Mark D., Born, Donald E., Ojemann, George
A., Dodrill, Carl B., Hallam, Danial K., 2003. Factors predicting outcome of surgery
for intractable epilepsy with pathologically verified mesial temporal sclerosis.
Epilepsia 44 (4), 565–568 (Wiley Online Library).
He, X., Doucet, G.E., Pustina, D., Sperling, M.R., Sharan, A.D., Tracy, J.I., 2017.
Presurgical thalamic "hubness" predicts surgical outcome in temporal lobe epilepsy.
Neurology 10–1212.
van den Heuvel, M.P., Stam, C.J., Kahn, R.S., Pol, H.E.H., 2009. Efficiency of functional
brain networks and intellectual performance. J. Neurosci. 29 (23), 7619–7624.
Hoerl, A.E., Kennard, R.W., 2004. Ridge Regression. 9780471667193 John Wiley and
Sons, Inc.http://dx.doi.org/10.1002/0471667196.ess2280.
Irimia, Andrei, Van Horn, John Darrell, 2016. Scale-dependent variability and quantita-
tive regimes in graph-theoretic representations of human cortical networks. Brain
Connect. 6 (2), 152–163.
Jehi, L., Yardi, R., Chagin, K., Tassi, L., Russo, G.L., Worrell, G., ... Chauvel, P., 2015.
Development and validation of nomograms to provide individualised predictions of
seizure outcomes after epilepsy surgery: a retrospective analysis. Lancet Neurol. 14
(3), 283–290.
Jenkinson, Mark, Smith, Stephen, 2001. A global optimisation method for robust affine
registration of brain images. Med. Image Anal. 5 (2), 143–156 (Elsevier).
Jenkinson, Mark, Bannister, Peter, Brady, Michael, Smith, Stephen, 2002. Improved op-
timization for the robust and accurate linear registration and motion correction of
brain images. NeuroImage 17 (2), 825–841 (Elsevier).
Jenkinson, Mark, Beckmann, Christian F., Behrens, Timothy E.J., Woolrich, Mark W.,
Smith, Stephen M., 2012. Fsl. NeuroImage 62 (2), 782–790 (Elsevier).
Jeong, Jeong-Won, Asano, Eishi, Juhász, Csaba, Behen, Michael E., Chugani, Harry T.,
2016. Postoperative axonal changes in the contralateral hemisphere in children with
medically refractory epilepsy: a longitudinal diffusion tensor imaging connectome
analysis. Hum. Brain Mapp. 37 (11), 3946–3956 (Wiley Online Library).
Ji, Gong-Jun, Zhang, Zhiqiang, Xu, Qiang, Wei, Wei, Wang, Jue, Wang, Zhengge, Yang,
Fang, et al., 2015. Connectome reorganization associated with surgical outcome in
temporal lobe epilepsy. Medicine 94 (40), e1737 (LWW).
Jones, D.K., 2010. Challenges and limitations of quantifying brain connectivity in vivo
with diffusion MRI. Quant. Imaging Med. Surg. 2, 341–355.
Jutila, L., Immonen, A., Mervaala, E., Partanen, J., Partanen, K., Puranen, M., Kälviäinen,
R., et al., 2002. Long term outcome of temporal lobe epilepsy surgery: analyses of 140
consecutive patients. J. Neurol. Neurosurg. Psychiatry 73 (5), 486–494 (BMJ
Publishing Group Ltd).
Kamiya, Kouhei, Amemiya, Shiori, Suzuki, Yuichi, Kunii, Naoto, Kawai, Kensuke, Akira,
Kunimatsu, Nobuhito, Saito, Shigeki, Aoki, Kuni, Ohtomo, 2016. Machine learning of
Dti structural brain connectomes for lateralization of temporal lobe epilepsy. Magn.
Reson. Med. Sci. 15 (1), 121–129 (Japanese Society for Magnetic Resonance in
Medicine).
Keller, Simon S., Richardson, Mark P., Schoene-Bake, Jan-Christoph, O'muircheartaigh,
Jonathan, Elkommos, Samia, Kreilkamp, Barbara, Goh, Yee Yen, Marson, Anthony G.,
Elger, Christian, Weber, Bernd, 2015. Thalamotemporal alteration and postoperative
seizures in temporal lobe epilepsy. Ann. Neurol. 77 (5), 760–774 (Wiley Online
Library).
Kuceyeski, Amy, Maruta, Jun, Relkin, Norman, Raj, Ashish, 2013. The network mod-
ification (nemo) tool: elucidating the effect of white matter integrity changes on
cortical and subcortical structural connectivity. Brain Connect. 3 (5), 451–463 (Mary
Ann Liebert, Inc. 140 Huguenot Street, 3rd Floor New Rochelle, NY 10801 USA).
Kuceyeski, A.F., Vargas, W., Dayan, M., Monohan, E., Blackwell, C., Raj, A., Fujimoto, K.,
Gauthier, S.A., 2015a. Modeling the relationship among gray matter atrophy, ab-
normalities in connecting white matter, and cognitive performance in early multiple
sclerosis. Am. J. Neuroradiol. 36 (4), 702–709 (Am Soc Neuroradiology).
Kuceyeski, Amy, Navi, Babak B., Kamel, Hooman, Relkin, Norman, Villanueva, Mark, Raj,
213
P.N. Taylor et al.
NeuroImage: Clinical 18 (2018) 202–214
Ashish, Toglia, Joan, O'dell, Michael, Iadecola, Costantino, 2015b. Exploring the
brain's structural connectome: a quantitative stroke lesion-dysfunction mapping
study. Hum. Brain Mapp. 36 (6), 2147–2160 (Wiley Online Library).
Kuceyeski, Amy, Navi, Babak B., Kamel, Hooman, Raj, Ashish, Relkin, Norman, Toglia,
Joan, Iadecola, Costantino, O'dell, Michael, 2016. Structural connectome disruption
at baseline predicts 6‐months post‐stroke outcome. Hum. Brain Mapp. 37 (7),
2587–2601.
Leemans, Alexander, Jones, Derek K., 2009. The B-matrix must be rotated when cor-
recting for subject motion in DTI data. Magn. Reson. Med. 61 (6), 1336–1349.
Lemkaddem, Alia, Daducci, Alessandro, Kunz, Nicolas, Lazeyras, François, Seeck,
Margitta, Thiran, Jean-Philippe, Vulliémoz, Serge, 2014. Connectivity and tissue
microstructural alterations in right and left temporal lobe epilepsy revealed by dif-
fusion spectrum imaging. Neuroimaging Clin. N. Am. 5, 349–358 (Elsevier).
Liao, Wei, Zhang, Zhiqiang, Pan, Zhengyong, Mantini, Dante, Ding, Jurong, Duan, Xujun,
Cheng, Luo, Lu, Guangming, Chen, Huafu, 2010. Altered functional connectivity and
small-world in mesial temporal lobe epilepsy. PLoS One 5 (1), e8525 (Public Library
of Science).
Liao, Wei, Ji, Gong-Jun, Xu, Qiang, Wei, Wei, Wang, Jue, Wang, Zhengge, Yang, Fang,
et al., 2016. Functional connectome before and following temporal lobectomy in
mesial temporal lobe epilepsy. Sci. Rep. 6, 23153 (Nature Publishing Group).
Liu, Jun, Ye, Jieping, 2009. Efficient Euclidean projections in linear time. In: Proceedings
of the 26th Annual International Conference on Machine Learning. ACM, New York,
NY, USA, pp. 657–664. ICML '09. https://doi.org/10.1145/1553374.1553459.
Liu, J., Ji, S., Ye, J., 2009a. SLEP: Sparse Learning with Efficient Projections. Arizona
State University. http://www.public.asu.edu/~jye02/Software/SLEP.
Liu, Jun, Chen, Jianhui, Ye, Jieping, 2009b. Large-scale sparse logistic regression. In:
Proceedings of the 15th Acm Sigkdd International Conference on Knowledge
Discovery and Data Mining. ACM, New York, NY, USA, pp. 547–556. KDD '09.
https://doi.org/10.1145/1557019.1557082.
Liu, Min, Chen, Zhang, Beaulieu, Christian, Gross, Donald W., 2014. Disrupted anatomic
white matter network in left mesial temporal lobe epilepsy. Epilepsia 55 (5), 674–682
(Wiley Online Library).
Martin, R.C., Sawrie, S.M., Edwards, R., Roth, D.L., Faught, E., Kuzniecky, R.I., Morawetz,
R.B., Gilliam, F.G., 2000. Investigation of executive function change following
anterior temporal lobectomy: selective normalization of verbal fluency.
Neuropsychology 14 (4), 501.
Maslov, Sergei, Sneppen, Kim, 2002. Specificity and stability in topology of protein
networks. Science 296 (5569), 910–913 (American Association for the Advancement
of Science).
Memarian, Negar, Kim, Sally, Dewar, Sandra, Engel, Jerome, Staba, Richard J., 2015.
Multimodal data and machine learning for surgery outcome prediction in compli-
cated cases of mesial temporal lobe epilepsy. Comput. Biol. Med. 64, 67–78
(Elsevier).
Morgan, Victoria L., Englot, Dario J., Rogers, Baxter P., Landman, Bennett A., Cakir,
Ahmet, Abou-Khalil, Bassel W., Anderson, Adam W., 2017. Magnetic resonance
imaging connectivity for the prediction of seizure outcome in temporal lobe epilepsy.
Epilepsia 58 (7), 1251–1260.
Munsell, Brent C., Wee, Chong-Yaw, Keller, Simon S., Weber, Bernd, Elger, Christian, da
Silva, Laura Angelica Tomaz, Nesland, Travis, Styner, Martin, Shen, Dinggang,
Bonilha, Leonardo, 2015. Evaluation of machine learning algorithms for treatment
outcome prediction in patients with epilepsy based on structural connectome data.
NeuroImage 118, 219–230. http://dx.doi.org/10.1016/j.neuroimage.2015.06.008.
Osuna, Edgar, Freund, Robert, Girosi, Federico, 1997. Support vector machines: training
and applications. Libr. Technol. Rep. AIM-1602.
Pustina, Dorian, Doucet, Gaelle, Evans, James, Sharan, Ashwini, Sperling, Michael,
Skidmore, Christopher, Tracy, Joseph, 2014. Distinct types of white matter changes
are observed after anterior temporal lobectomy in epilepsy. PLoS One 9 (8), e104211.
Rodionov, Roman, Vollmar, Christian, Nowell, Mark, Miserocchi, Anna, Wehner, Tim,
Micallef, Caroline, Zombori, Gergely, et al., 2013. Feasibility of multimodal 3d
neuroimaging to guide implantation of intracranial eeg electrodes. Epilepsy Res. 107
(1), 91–100 (Elsevier).
Rubinov, Mikail, Sporns, Olaf, 2010. Complex network measures of brain connectivity:
uses and interpretations. NeuroImage 52 (3), 1059–1069 (Elsevier).
de Salvo, Matthew N., Douw, Linda, Tanaka, Naoaki, Reinsberger, Claus, Stufflebeam,
Steven M., 2014. Altered structural connectome in temporal lobe epilepsy. Radiology
270 (3), 842–848. http://dx.doi.org/10.1148/radiol.13131044.
Shamim, Sadat, Wiggs, Edythe, Heiss, John, Sato, Susumu, Liew, Clarissa, Solomon,
Jeffrey, Theodore, William H., 2009. Temporal lobectomy: resection volume, neu-
ropsychological effects, and seizure outcome. Epilepsy Behav. 16 (2), 311–314
(Elsevier).
Sinha, N., Dauwels, J., Kaiser, M., Cash, S.S., Brandon Westover, M., Wang, Y., Taylor,
P.N., 2017. Predicting neurosurgical outcomes in focal epilepsy patients using com-
putational modelling. Brain 140 (2), 319–332.
Stretton, J., Thompson, P.J., 2012. Frontal lobe function in temporal lobe epilepsy.
Epilepsy Res. 98 (1), 1–13.
Stretton, J., Sidhu, M.K., Winston, G.P., Bartlett, P., McEvoy, A.W., Symms, M.R., Koepp,
M.J., Thompson, P.J., Duncan, J.S., 2014. Working memory network plasticity after
anterior temporal lobe resection: a longitudinal functional magnetic resonance
imaging study. Brain 137 (5), 1439–1453.
Taylor, Peter N., Han, Cheol E., Schoene-Bake, Jan-Christoph, Weber, Bernd, Kaiser,
Marcus, 2015. Structural connectivity changes in temporal lobe epilepsy: spatial
features contribute more than topological measures. Neuroimaging Clin. N. Am. 8,
322–328 (Elsevier).
Taylor, Peter N., Wang, Yujiang, Kaiser, Marcus, 2017. Within brain area tractography
suggests local modularity using high resolution connectomics. Sci. Rep. 7 (Nature
Publishing Group).
Tibshirani, Robert, 1996. Regression shrinkage and selection via the lasso. J. R. Stat. Soc.
Ser. B Methodol. 267–288.
Tournier, J., Calamante, F., Connelly, A., 2012. MRtrix: diffusion tractography in crossing
fiber regions. Int. J. Imaging Syst. Technol. 22 (1), 53–66.
Veropoulos, Konstantinos, Campbell, Colin, Cristianini, Nello, 1999. Controlling the
sensitivity of support vector machines. In: Proceedings of the International Joint
Conference on AI, pp. 55–60.
Vos, Sjoerd B., Tax, Chantal M.W., Luijten, Peter R., Ourselin, Sebastien, Leemans,
Alexander, Froeling, Martijn, 2016. The importance of correcting for signal drift in
diffusion Mri. In: Magnetic Resonance in Medicine. Wiley Online Library.
Wang, Zuoguan, Gunduz, Aysegul, Brunner, Peter, Ritaccio, Anthony L., Ji, Qiang, Schalk,
Gerwin, 2012. Decoding onset and direction of movements using electrocortico-
graphic (ECoG) signals in humans. Front. Neuroeng. 5.
Wheeler-Kingshott, Claudia A.M., Hickman, Simon J., Parker, Geoffrey J.M., Ciccarelli,
Olga, Symms, Mark R., Miller, David H., Barker, Gareth J., 2002. Investigating cer-
vical spinal cord structure using axial diffusion tensor imaging. NeuroImage 16 (1),
93–102 (Elsevier).
Wieser, H.G., Blume, W.T., Fish, D., Goldensohn, E., Hufnagel, A., King, D., Sperling,
M.R., Lüders, H., Pedley, Timothy A., 2001. Proposal for a new classification of
outcome with respect to epileptic seizures following epilepsy surgery. Epilepsia 42
(2), 282–286 (Wiley Online Library).
Winston, G.P., Daga, P., Stretton, J., Modat, M., Symms, M.R., McEvoy, A.W., Ourselin, S.,
Duncan, J.S., 2012. Optic radiation tractography and vision in anterior temporal lobe
resection. Ann. Neurol. 71 (3), 334–341.
Winston, Gavin P., Stretton, Jason, Sidhu, Meneka K., Symms, Mark R., Duncan, John S.,
2014. Progressive white matter changes following anterior temporal lobe resection
for epilepsy. Neuroimaging Clin. N. Am. 4, 190–200 (Elsevier).
Wirsich, Jonathan, Perry, Alistair, Ridley, Ben, Proix, Timothée, Golos, Mathieu, Bénar,
Christian, Ranjeva, Jean-Philippe, et al., 2016. Whole-brain analytic measures of
network communication reveal increased structure-function correlation in right
temporal lobe epilepsy. Neuroimaging Clin. N. Am. 11, 707–718 (Elsevier).
Wu, Xiaoyun, Rohini, Srihari, 2004. Incorporating prior knowledge with weighted margin
support vector machines. In: Proceedings of the Tenth ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM.
Wyler, Allen R., Hermann, Bruce P., Somes, Grant, 1995. Extent of medial temporal re-
section on outcome from anterior temporal lobectomy: a randomized prospective
study. Neurosurgery 37 (5), 982–991 (LWW).
Xia, Mingrui, Wang, Jinhui, He, Yong, 2013. BrainNet viewer: a network visualization
tool for human brain connectomics. PLoS One 8 (7), e68910 (Public Library of
Science).
Yeh, Fang-Cheng, Wedeen, Van Jay, Tseng, W.-Y.I., 2010. Generalized-sampling imaging.
Med. Imaging IEEE Trans. 29 (9), 1626–1635 (IEEE).
Yeh, Fang-Cheng, Verstynen, Timothy D., Wang, Yibao, Fernández-Miranda, Juan C.,
Tseng, Wen-Yih Isaac, 2013. Deterministic diffusion fiber tracking improved by
quantitative anisotropy. PLoS One 8 (11), e80713 (Public Library of Science).
Yeo, B.T. Thomas, Krienen, Fenna M., Sepulcre, Jorge, Sabuncu, Mert R., Lashkari,
Danial, Hollinshead, Marisa, Roffman, Joshua L., et al., 2011. The organization of the
human cerebral cortex estimated by intrinsic functional connectivity. J.
Neurophysiol. 106 (3), 1125–1165.
Yogarajah, M., Focke, N.K., Bonelli, S.B., Thompson, P., Vollmar, C., McEvoy, A.W.,
Alexander, D.C., Symms, M.R., Koepp, M.J., Duncan, J.S., 2010. The structural
plasticity of white matter networks following anterior temporal lobe resection. Brain
133 (8), 2348–2364.
Zalesky, Andrew, Fornito, Alex, Cocchi, Luca, Gollo, Leonardo L., van den Heuvel, Martijn
P., Breakspear, Michael, 2016. Connectome sensitivity or specificity: which is more
important? NeuroImage 142, 407–420.
Zhang, Z., Liao, W., Chen, H., Mantini, D., Ding, J.R., Xu, Q., Wang, Z., et al., 2011.
Altered functional–structural coupling of large-scale brain networks in idiopathic
generalized epilepsy. Brain 134 (10), 2912–2928 (Oxford Univ Press).
Zou, Hui, Hastie, Trevor, 2005. Regularization and variable selection via the elastic net. J.
R. Stat. Soc. Ser. B (Stat Methodol.) 67 (2), 301–320. (Royal Statistical Society,
Wiley). http://www.jstor.org/stable/3647580.
214
|
1310.5968 | 1 | 1310 | 2013-10-22T15:53:16 | A Cell-Level Mechanism of Contrast Gain Control | [
"q-bio.NC",
"physics.bio-ph"
] | The gain of neurons' responses in the auditory cortex is sensitive to contrast changes in the stimulus within a spectrotemporal range similar to their receptive fields, which can be interpreted to represent the tuning of the input to a neuron. This indicates a local mechanism of contrast gain control, which we explore with a minimal mechanistic model here. Gain control through noisy input has been observed in vitro and in a range of computational models. We investigate the behaviour of the simplest of such models to showcase gain control, a stochastic leaky integrate-and-fire (sLIF) neuron, which exhibits gain control through divisive normalisation of the input both with and without accompanying subtractive shift of the input-response curve, depending on whether input noise is proportional to or independent of its mean. To get a more direct understanding of how the input statistics change the response, we construct an analytic approximation to the firing rate of a sLIF neuron constituted of the expression for the deterministic case and a weighted average over the derived approximate steady-state distribution of conductance due to poissonian synaptic inputs. This analytic approximation qualitatively produces the same behaviour as simulations and could be extended by spectrotemporally tuned inputs to give a simple, physiological and local mechanism of contrast gain control in auditory sensing, building on recent experimental work that has hitherto only been described by phenomenological models. By comparing our weighted average firing rate curve with the commonly used sigmoidal input-response function, we demonstrate a nearly linear relationship between both the horizontal shift (or stimulus inflection point) and the inverse gain of the sigmoid and statistics derived from the sLIF model parameters, thus providing a structural constraint on the sigmoid parameter choice. | q-bio.NC | q-bio | A Cell-Level Mechanism of Contrast Gain Control
Linus J. Schumacher
Wolfson Centre for Mathematical Biology,
Mathematical Institute, University of Oxford,
Radcliffe Observatory Quarter, Woodstock Road,
Oxford, OX2 6GG, United Kingdom and
Computational Biology Group, Department of Computer Science,
University of Oxford, Wolfson Building, Oxford, OX1 3QD, United Kingdom
Geoff K. Nicholls
Department of Statistics, University of Oxford,
1 South Parks Road, Oxford, OX1 3TG, United Kingdom
(Dated: October 17, 2018)
3
1
0
2
t
c
O
2
2
]
.
C
N
o
i
b
-
q
[
1
v
8
6
9
5
.
0
1
3
1
:
v
i
X
r
a
1
Abstract
The gain of neurons' responses in the auditory cortex is sensitive to contrast changes in the stimulus
within a spectrotemporal range similar to their receptive fields (Rabinowitz et al., 2012), which can be
interpreted to represent the tuning of the input to a neuron. This indicates a local mechanism of contrast
gain control, which we explore with a minimal mechanistic model here. Gain control through noisy input
has been observed in vitro (Chance et al., 2002) and in a range of computational models (Ayaz and Chance,
2009; Longtin et al., 2002). We investigate the behaviour of the simplest of such models to showcase gain
control, a stochastic leaky integrate-and-fire (sLIF) neuron, which exhibits gain control through divisive
normalisation of the input both with and without accompanying subtractive shift of the input-response
curve, depending on whether input noise is proportional to or independent of its mean. To get a more direct
understanding of how the input statistics change the response, we construct an analytic approximation to
the firing rate of a sLIF neuron constituted of the expression for the deterministic case and a weighted
average over the derived approximate steady-state distribution of conductance due to poissonian synaptic
inputs. This analytic approximation qualitatively produces the same behaviour as simulations and could be
extended by spectrotemporally tuned inputs to give a simple, physiological and local mechanism of contrast
gain control in auditory sensing, building on recent experimental work that has hitherto only been described
by phenomenological models (Rabinowitz et al., 2012). By comparing our weighted average firing rate curve
with the commonly used sigmoidal input-response function, we demonstrate a nearly linear relationship
between both the horizontal shift (or stimulus inflection point) and the inverse gain of the sigmoid and
statistics derived from the sLIF model parameters, thus providing a structural constraint on the sigmoid
parameter choice.
I.
INTRODUCTION -- CONTRAST GAIN CONTROL
Animals can process sensory information over a wide range of stimulus intensity contrasts. To
do so, the gain of their neural response adjusts itself to variations in spatiotemporal (in vision)
or spectrotemporal contrast (in hearing). One such mechanism of gain control, divisive normal-
isation, has been successfully applied in models of the visual system for a long time, but been
studied much less in the auditory system. Recently it has been shown (Rabinowitz et al., 2012,
2011) that a large part of gain control of neurons in the auditory cortex can be accounted for by
horizontally shifting and divisively normalising the neural response, i.e. firing rate, in proportion
to the stimulus contrast (Fig. 1). The exact relationship can be described by "gain contrast ker-
nels" (Rabinowitz et al., 2011). Furthermore, it was found that the gain control is most sensitive
to contrast in spectrotemporal regions where the neuron's response is also high in magnitude (Ra-
binowitz et al., 2012, 2011). In other words, the contrast gain control seems to operate mainly
2
within the spectrotemporal receptive field (STRF) and be insensitive to variations in stimulus con-
trast outside the STRF. This phenomenological description of gain control is indicative of local
computation and raises the possibility that the underlying physiological or biophysical mechanism
operates on a single cell rather than on a population level.
Gain control has also been observed in mathematical and computational neuron models, the
simplest relevant one being the leaky integrate-and-fire (LIF) neuron with noisy synaptic input.
This model springs to mind as a good candidate for a local mechanism of contrast gain control.
Starting from a LIF model (as implemented in, for example, (Ayaz and Chance, 2009), but without
constructing any pools of normalising or modulatory neurons) we examine gain control in response
to change in input statistics, aiming to find analytical approximations to the simulation results
wherever possible in order to get a better handle on the model. By identifying qualitative and
quantitative similarities between the stochastic LIF model and the gain contrast kernel model, we
suggest their equivalence. The stochastic LIF model is presented as a hypothesis for the mechanism
that could underlie the gain contrast kernels
FIG. 1. Spectrotemporal contrast kernel model by Rabinowitz et al. (2012): The stimulus spec-
trogram Ltf is convolved with a linear spectrotemporal kernel kf h (equivalent to the STRF) and passed
through sigmoidal response function to give the predicted firing rate yt. The horizontal shift, or stimulus
inflection point, c and the inverse gain d of the sigmoid are functions of stimulus contrast σtf , and their
relationship is described by a single contrast kernel κ(cd)
f g ≈ kf h. Reproduced with permission.
3
II. METHODS
A. Leaky integrate and fire (LIF) neurons
We start with a basic neuron model that is analytically tractable: The leaky integrate-and-fire
neuron model is generally regarded as the simplest description of a nerve cell that produces any
realistic output. The membrane potential V follows the differential equation
C
dV (t)
dt
= I(t) − gV (t)
where C is the membrane capacitance, g the membrane conductance and I(t) the input current,
which could be current flowing across the membrane through ion channels or current injected
in a patch-clamp experiment. When the input current is large enough, the membrane potential
rises until it reaches a threshold Vth, upon which a spike is recorded and the potential is reset to
Vreset = 0, representing the firing of an action potential. For a constant input current, the firing
rate (number of spikes per time) is given by the expression (Koch and Segev, 1998)
0,
f (I) =
[tref − C
g log(1 − gVth
I )]
I ≤ gVth
I > gVth
(1)
−1,
where tref is the refractory period, i.e., the length of time spiking is suppressed after an action
potential has been fired. Note that the model as such does not exhibit a refractory period, it has
to be added to achieve saturation of the firing rate. One could set tref to a physiologically realistic
value on the order of 1 ms, but for the purpose of comparing to simulations we simply set it to the
simulation time step ∆t.
B.
stochastic LIF (sLIF) neurons
Having a firing rate expression for the deterministic LIF neuron, we now want to add noise to
the input. An extension to the LIF model that makes it more physiological is to let the membrane
conductance g vary with stochastic synaptic input, which can be excitatory or inhibitory. To
compare with simulations, we adapt the implementation from Ayaz and Chance (2009) in this
section, and stay close to their notation throughout this paper. Whenever a spike arrives at an
excitatory (inhibitory) synapse, e.g. through a Poisson process with rate λe (λi), the conductance
increases instantaneously by a fixed absolute amount ∆ge (∆gi), representing the assumption that
a fixed amount of neurotransmitter is released and opens a fixed amount of postsynaptic ion
4
channels. Between spikes, the excitatory (inhibitory) part of the conductance decays to zero with
time constant τge (τgi). The synaptic conductance thus evolves independently of the membrane
potential and any reversal or threshold values thereof. However, the conductance terms are now
no longer constant, so the differential equation describing the evolution of the membrane potential
becomes
C
dV (t)
dt
= IF F (t) + gL[VL − V (t)] + ge(t)[Ee − V (t)] + gi(t)[Ei − V (t)]
where Ee (Ei) is the reversal potential of the excitatory (inhibitory) ion channels, IF F is the feed-
forward current and VL (= 0 earlier) is the resting membrane potential, which is also assumed to
be the value of Vreset.
If we shift the potential by VL and separate the voltage dependent and independent terms, we
get back our original equation
C
dV (t)
dt
= I(t) − g(t)V (t)
only that now we have 'effective' current and conductance:
g(t) = gL + ge(t) + gi(t)
I(t) = IF F (t) + ge(t)[Ee − VL] − gi(t)[VL − Ei]
(2)
(3)
Ideally we would now directly derive a stochastic version of the deterministic firing rate equation
(1), but this is intractable. Instead, we will try and approximate the distribution of ge(t) and gi(t).
From this, we might be able to work out the distributions of g(t) and I(t), and so on, hopefully
arriving at a distribution of f (I) -- though a simpler approach will prove far more tractable, as we
will see in the rest of this section.
The model complexity is deliberately kept low for the sake of tractability and suffices for the
purpose of our argument. However, more complex neuron models, such as the Morris-Lecar have
been shown to be equivalent to a stochastic LIF neuron (Ditlevsen and Greenwood, 2012), giving
additional justification to our model choice.
1. Mean-field failure
Naıvely one might try a mean-field approach and substitute the steady-state values (cid:104)ge,i(cid:105) =
∆ge,iλe,iτge,i (assuming synaptic input is Poisson distributed with rate λ) into (1), but in practice
this will often give zero firing rate even when simulations show spiking. This is because even when
5
the condition (cid:104)I(cid:105) > (cid:104)g(cid:105)Vth is not satisfied, fluctations, which are not captured by the mean field
approach, can drive the membrane potential above threshold.
2. Analytical approaches in the literature
Decades of work have been dedicated to obtain analytical expressions for the firing rate of
an sLIF neuron (Koch and Segev, 1998). Approaches include use of the Fokker-Planck equation
((Ermentrout and Terman, 2010), section 10.2, and references therein), first order corrections to
the first passage time (Brunel and Sergi, 1998), the Volterra integral equation (Buonocore et al.,
2010) and coupling of excitatory and inhibitory noise (Longtin et al., 2002; L´ansk´a et al., 1994;
Lansk´y et al., 1995). And while these are all valuable achievements in one way or another, they
typically involve integrals that can only be evaluated numerically (such as error functions (Yu
and Lee, 2003; Ostojic and Brunel, 2011)), infinite sums, gamma functions or other mathematical
unpleasantries. These may allow the neuroscientist to do away with running exhaustive simulations
(at least for those special cases that these approaches have been tailored to), but they generally
don't provide a good insight into how a change in statistics of the noisy input affect the output.
For an Ornstein-Uhlenbeck (OU) neuron a linear approximation for the firing rate in terms of
input mean and variance can be found (Yu and Lee, 2003; L´ansk´y and Sacerdote, 2001; Sacerdote
and Giraudo, 2013). However, in the implementation of a stochastic LIF neuron here and else-
where (Ayaz and Chance, 2009), the membrane Voltage is not described by an OU process. Rather,
it is the membrane conductance that can be described by an OU process, as we will show in Sec-
tion II C. The membrane voltage differential equation is thus coupled to the solution of this OU
process. If one assumes the conductance becomes a random variable itself, we suspect one could
arrive at a Chan-Karolyi-Longstaff-Sanders (CKLS), or Brennan-Schwartz (BS), process (Chan
et al., 1992; Brennan and Schwartz, 1980), in which voltage fluctuations are dependent on the
input statistics and the voltage itself. To our knowledge, the mathematical behaviour of such a
CKLS/BS neuron has never been studied.
6
3. Parameter values used
When simulating the sLIF model to test our analytical approximations, we used the following
parameter values, as in (Ayaz and Chance, 2009).
Parameter Description
Value Units
C
gL
VL
Vth
Ee
Ei
∆ge
∆gi
τge,i
∆t
T
membrane capacitance
resting (or leak) conductance
resting (and reset) membrane potential
threshold potential for firing an action potential
reversal potential of excitatory conductance
reversal potential of inhibitory conductance
instantaneous change in excitatory conductance
instantaneous change in inhibitory conductance
time constant of synaptic conductances
time step
740
20
pF
nS
-70 mV
-52 mV
0 mV
-80 mV
0.16gL
0.48gL
[gL]
[gL]
5 ms
0.05 ms
typical length of each simulation of neural activity
1
s
C. Approximations of the steady-state distribution of stochastic synaptic input
To understand the influence of fluctuations in synaptic input on the firing rate of our sLIF
neuron, consider the distribution of conductances, excitatory or inhibitory, here generically denoted
by g. In an actual implementation of the sLIF model as described in section II B, time will be
discretised and the evolution of each of the excitatory and inhibitory conductance terms can be
described by an AR(1)-process of the form
gt = c + φgt−∆t + σξξt
where c is a constant, φ is the autoregressive parameter and ξt is white noise with variance σξ.
This is based on the assumption that poissonian synaptic input can be approximated as
∆g · P(λ∆t) ≈ ∆g · N (λ∆t,
√
λ∆t) = c + σξξt
where P and N are Poisson and normally distributed random variables, respectively (Note we use
P and N to denote both the random variable and its distribution). So c = σ2
ξ = (∆g)2λ∆t and
φ = e−∆t/τg ≈ 1−∆t/τg for time steps ∆t << τg. Substituting these values into the expressions for
7
the mean µ = c/(1 − φ) and variance σ2 = σ2
ξ /(1 − φ2) of an AR(1) process and expanding to first
order in ∆t/τg we see that steady-state value of each of the conductance terms is approximately
normally distributed with mean and variance
µg = ∆gλτg
σ2
g = (∆g)2λτg/2
(4)
(5)
One can also write the evolution of g in the continuous time limit as an Ornstein-Uhlenbeck (OU)
process. Starting from our discrete time AR(1)-process
gt = (1 − ∆t/τg)gt−∆t + ∆gλ∆t + ∆g
√
λ∆tξt
and taking the limit ∆t → 0 we get the expression
dg = Φ(µg − g)dt + σW dW
where Φ = 1/τg, the mean µg is as in (4) and W is a Wiener process with variance σ2
The variance around the steady-state value in an OU-process is σ2
W = ∆g2λ .
W /2Φ, which gives us the
g = σ2
same value for σg as in (5).
We make the steady-state assumption here, but LIF neurons with dynamic inputs have been
shown to have similar gain control (Ly and Doiron, 2009).
D. Weighted average of deterministic LIF neuron firing rate
We derived in section II C that the excitatory and inhibitory conductances are approximately
normally distributed over time. By ergodicity, or by wide-sense stationarity, the distribution of
a conductance value over time is equal to distribution of conductance values at a particular time
point. Hence, we can estimate the firing rate of a stochastic LIF neuron by taking the average of
the firing rate expression for the deterministic case (1) over the joint probability distribution of
ge and gi (assumed independent), and recall that we shifted all potential terms by VL in the sLIF
model construction, i.e.
f (I) ≈ (cid:104)[tref − C
g
log(1 − g(Vth−VL)
I
−1(cid:105)N (µge ,σge )·N (µgi ,σgi ), I > g(Vth − VL)
)]
(6)
where g and I are given by equations (2) and (3), respectively, and µge,i, σge,i are given by (4,5).
8
III. RESULTS
A. Agreement of analytical approximation with simulations
The steady state distribution of synaptic conductances under poissonian spike input is well
approximated by the normal and gamma distributions derived in the previous section (Fig. 2(a)).
We can also approximate the distributions of g, I and g/I well (Fig. 7(b-d)), but as discussed in
section A 2 this is really only of use if we could then go on to derive the distribution of log(1 −
g(Vth − VL)/I).
Approximating ge and gi as either normal or gamma distributed, one can then numerically inte-
grate (1) over the (separable) joint probability distribution P (ge, gi) and thus in turn approximate
FIG. 2. Steady-state distributions of the excitatory (ge) and inhibitory (gi) conductances in
a sLIF neuron model. Excitatory (Ie) and inhibitory (Ii) synaptic currents are approximated by a
C dV
dt = d− n
Vth−VL
normal and a gamma distribution. The model differential equation, in terms of the parameters plotted, is
−1.
Distribution parameters are not fitted, but derived as described in the main text (4,5). Data shown for
V and the firing rate for the deterministic case is f (I) = [tref− C(Vth−VL)
log(1−r(I, g))]
n(g)
synaptic input rates λe = λi = 1 kHz, but approximations were qualitatively similar for rates as low as
0.5 kHz, and better the higher the rates.
9
00.511.522.500.511.522.53I/nAP Ie = ge(Ee - VL)Ii = gi(VL - Ei)NormalGamma FIG. 3. Illustration of how the weighted average of LIF firing rates is calculated to approximate
the sLIF firing rate. (a) Deterministic LIF firing rates (1) for a range of conductance values (ge, gi) at
synaptic input rates λe = λi = 1 kHz, colour and line thickness according to the joint probability distribution
of the conductance values. The sampling density has been kept low for illustrative purposes. Black curve
is the weighted average (6) of the colored curves, with the relative weight of each curve given by its colour
(as in (b)) and (approximately) its line thickness. Only curves with relative weight > e−2 are shown. Other
parameters as in main text, section II B 3. (b) Joint probability distribution (2D Gaussian) of (ge, gi). (c,d)
As (a,b), but with λe = λi = 3 kHz.
the firing rate of a sLIF neuron to a given input, as illustrated in Fig. 3. The expression for the
weighted average of the deterministic LIF neuron firing rate (6) agrees well with the firing rates
estimated from direct simulation of the stochastic LIF equations and captures the decrease in gain
when noisiness of the synaptic input is increased (Fig. 4).
B. Relating parameters of the sigmoid non-linearity to those of the sLIF model
Having an analytical approximation for the firing rate of a sLIF neuron (6) that agrees well
with the simulations (Fig. 4) we would like to relate changes in the synaptic input statistics to the
10
11.522.5050100firing rate/HzI/nAage/nSgi/nSb 01020300204060801000.0050.010.0150.0222.533.5050100firing rate/HzI/nAcge/nSgi/nSd 0204060800501001502002500.010.020.03 FIG. 4. Firing rate of sLIF model simulations is approximated by weighted average of deter-
ministic LIF firing rate expression Firing rates of the sLIF model were estimated from simulations with
low (λe = λi = 1 kHz) and high (λe = λi = 3 kHz) noise balanced synaptic input (qualitatively similar
fits could be produced for rates as low as 0.5 kHz). Error bars are standard deviation over 100 simulations.
Simulation results are approximated by averaging the analytical firing rate expression for the deterministic
LIF model over a normal or gamma distribution for the conductances.
changes in the shape of the response curve. The weighted average curves resulting from (6) are
similar in shape to the sigmoid non-linearities used to describe gain control in recent experimental
work (Rabinowitz et al., 2012, 2011):
y(z) = a +
b
1 + e− z−c
d
(7)
where a, b, c and d are parameters. With this expression, gain control was studied by fitting the
parameters of the sigmoid function as functions of stimulus contrast (Rabinowitz et al., 2012, 2011)
(Fig. 1). We would like to relate the parameters of this phenomenological description of gain control
to parameters in our mechanistic model. Generally this is intractable, as (6) cannot be evaluated
in closed form, but we can make some progress by comparing the shape of the weighted average
firing rate with that of the sigmoid, and how their shapes change with respect to their parameters.
In the sLIF model described here, there is no background firing, so a = 0, although this would
11
00.511.522.533.5020406080100I/nAfiring rate/Hz λ = 1 kHzλ = 3 kHznormalgamma have to be fitted to some positive value when comparing against in vivo data.
Data from in vivo experiments mainly seem to lie on the left half of the fitted sigmoid function,
i.e. the abscissae are smaller than the abscissa of the inflection point (Rabinowitz et al., 2012,
2011). By inspecting Fig. 3 we see that the inflection abscissa of a sigmoid (were its left half fitted
to the numerical data) roughly coincides with the threshold input to get firing (the rheobase) in
the deterministic case, at mean conductance values. Intuitively, this can be explained by the fact
that the deterministic firing rate curve for the mean conductances has the highest weighting in the
average and hence strongly determines it shape. For abscissa greater than the mean threshold input,
this main curve will be linear, i.e. with no curvature, as will the resulting average, approximately.
Therefore we might be tempted to set c to the threshold vale of our input I, i.e.
µth = (µge + µgi + gL)(Vth − VL)
(8)
Similarly one can argue that the standard deviation in the input threshold will be inversely related
to the gain of the firing rate curve, because it controls the horizontal spread of deterministic firing
rate curves with high weighting (Fig. 3), so one might like to try and set d equal to
(cid:113)
σth =
σ2
ge + σ2
gi(Vth − VL)
(9)
Similar observations about the qualitative dependence of the firing rate curve shape on input
statistics have been made by Yu and Lee (2003), but they did not attempt to relate the f-I curve
parameters to the input statistics in a functional form.
To test these predictions, we fitted sigmoid functions to data from sLIF simulations using
Matlab's fit function. The fits were generally qualitatively indistinguishable when initialised
at random multiple times with c and d constrained to be of the same order of magnitude as µth
and σth, though fitting would occasionally get stuck in what appeared to be local minima. For
Figs 5&6 c and d were initialised at µth and σth, respectively.
A sigmoid with c given by (8) is in decent agreement with the sLIF neuron model data (Fig. 5).
When d is given by (9) agreement is less good but the qualitative characteristics as the curve shape
and its shifting and scaling are still captured. The quality of fits is similar when both c and d are
fixed and only parameter b is fitted.
12
FIG. 5. sLIF firing rates are approximated by sigmoid functions Data from simulations of the sLIF
model for five balanced synaptic input rates (λ = λe = λi) were used to fit a sigmoid (7) with a = 0 and
the parameters b, c, d fitted, or with either or both c and d fixed at µth (8) and σth (9), respectively. Error
bars are standard deviations of 100 simulations.
13
0.811.21.41.6010203040λ = 1f/HzI/nA1.21.41.61.820102030λ = 1.75f/HzI/nA1.822.22.42.6051015λ = 2.5f/HzI/nA2.22.42.62.830510λ = 3.25f/HzI/nA2.62.833.23.402468λ = 4f/HzI/nA sLIF datab,c,d fittedb,c fitted; d = σthb,d fitted; c = µthb fitted; c = µth, d = σth FIG. 6. Relationship between fitted sigmoid parameters and statistics of sLIF model parameters
Parameters b, c and d of a sigmoid (7) on data from sLIF model simulations for a range of balanced synaptic
input rates from λ = 1 kHz to 4 kHz. Error bars are 95% confidence intervals on fitted parameters, as
given by Matlab's fit function (curve fitting toolbox). Red diagonal crosses without error bars indicate
that the confidence interval was greater than the range of abscissae. µth = (µge + µgi + gL)(Vth − VL) and
σth =
(Vth − VL).
(cid:113)
σ2
ge
+ σ2
gi
14
020004000600010002000300040005000µth/pAcb c d fitted200300400500600100200300400500600σth/pAdb c d fitted02000400060000200040006000800010000µth/pAcb c fitted, d fixed200300400500600100200300400500600σth/pAdb d fitted, c fixed IV. DISCUSSION
We have presented here a stochastic LIF neuron as a mechanistic model for contrast gain
control in neural auditory processing. In the model, subtractive and divisive normalisation of the
firing rate curve arise simply from the dynamics of noisy synaptic input. This in itself is no new
observation, but it has, to our knowledge, not been made use of to understand auditory contrast
gain control. Only recently has the tuning of contrast sensitivity in vivo been found to be similar
to the spectrotemporal receptive field (STRF) (Rabinowitz et al., 2012, 2011). This is indicative
of local computation and leaves open the possibility of contrast gain control occurring intrinsically
by feed-forward mechanisms only, i.e. without recurrent feedback -- which is exactly the case in the
sLIF model discussed here. Thus a neuron in the auditory cortex may not compute the stimulus
contrast directly, but exhibit at least part of its gain control just through its biophysics, in response
to a change in contrast in the input it receives from lower levels of auditory processing.
Others have drawn similar conclusions in a more general context, and with a slightly different
interpretation: Yu and Lee (2003) suggest that "the change of effective stimulus threshold in various
statistical stimulus environments is the key factor underlying variance or contrast gain control".
Hong et al. (2008) attribute a neuron's intrinsic gain control due to the fact that "inputs with
different statistics interact with the nonlinearity of the system in different ways".
A. Predictions
This hypothesis makes testable predictions on the relationship between mean and variance of
the input to the auditory cortex -- they ought to be proportional, as in a Poisson process -- which
is something that might be measurable in the inferior colliculus, which sits upstream of the cortex
in the auditory pathway.
In addition, effective membrane conductance and current distributions could be measured in-
tracellularly and compared to our approximations (Fig. 2).
If mean conductances and internal
model parameter values (section II B 3) of a cell are known, one could examine their relationship to
the parameters of the sigmoid nonlinearity (fitted to in vivo data) and their dependence on stim-
ulus contrast, as in (8) or (9). This could in turn motivate (or criticise) the particular choice of
a sigmoid function for a neuron's response non-linearity, rather than a curve of similar shape but
different functional form.
The best fits of sigmoids to spike rate data seem to be achieved for models were the divisive and
15
subtractive parameters, c and d in (7), have the same stimulus contrast dependence (Rabinowitz
et al., 2011). This emerges naturally from a local mechanism of gain control, as the subtrac-
tive and divisive parameters are determined by the same synaptic input and hence have the same
spectrotemporal tuning. Based on the arguments presented in section III B, we proposed relation-
ships for the sigmoid parameters c ∝ µg and d ∝ σg, which is in qualitative agreement with other
work (Ayaz and Chance, 2009; Longtin et al., 2002; Ostojic and Brunel, 2011). If these relation-
ships hold, the relationship between spectrotemporal contrast kernels and receptive fields ought to
be κ(c)
tf ≈ ktf, as Rabinowitz et al. (2011) sug-
tf = κ(d)
tf ≈ ktf and κ(d)
tf ≈ ktf, rather than κ(c)
gested. Whether this can be confirmed upon re-inspection of previous results (Rabinowitz et al.,
2011) remains to be seen.
B. Model criticism
To achieve purely divisive normalisation without a subtractive shift of the firing rate curve, the
sLIF model needs balanced excitatory and inhibitory input (Longtin et al., 2002). This might seem
like physiologically unlikely fine-tuning, but balanced synaptic input has been observed experimen-
tally and simple mechanisms for maintaining the balance have been proposed (Vogels et al., 2011).
Furthermore, purely divisive normalisation is rarely seen in vivo (Rabinowitz et al., 2011).
Hong et al. (2008) have described gain control as a diffusion of the noiseless f-I curve, which
is similar in form to our weighted average (6) for normal distributions. However, Hong et al.
(2008) never see a decrease in LIF firing rate at higher noise levels, which we and others (Ayaz
and Chance, 2009) do observe in our LIF neurons. This disagreement might come down to the
particular relationship of mean vs. variance in the input statistics, and the resulting subtractive vs.
divisive scaling of the f-I curve. However, our aim was to understand and explain the functional
form of gain control seen in the auditory cortex (Rabinowitz et al., 2012, 2011), which does exhibit
decreased firing rate at increased noise levels.
Fitting a sigmoid non-linearity with some derived, rather than fitted, parameters seems to give
good agreement with model data (Fig. 5). However, the suggested correspondences c = µth (8)
and d = σth (9) appear not to be exact (Fig. 6), though (nearly) linear. Thus we might have
to (a) rethink the argument presented in section III B and consider the possibility that c and d
are determined by the statistics of a different set of model parameters (which might well have a
similar dependence on the synaptic input rates, but with different constants of proportionality) or
(b) make the fitting of the sigmoid parameters more robust.
16
C.
Impact beyond auditory neuroscience
Beyond the application to auditory contrast gain control, we have derived an analytical approx-
imation (1) to the firing rate of a stochastic LIF neuron that, if still not in closed form, is easy to
evaluate (involving numerical integration over only two dimensions) and hopefully lends itself more
easily to an intuitive understanding of how the input statistics affect the firing rate than other ex-
pressions (Longtin et al., 2002; Ermentrout and Terman, 2010; Brunel and Sergi, 1998; Buonocore
et al., 2010; L´ansk´a et al., 1994; Lansk´y et al., 1995).
D. Further work
Promising directions for further work exist, most notably to fit the sLIF model on stimulus
data that have been filtered through estimated STRFs (as in (Rabinowitz et al., 2012, 2011)) and
to refit the gain kernel model on data generated from the sLIF model. Mathematically, one could
investigate whether membrane voltage distributions can be found when its evolution is described
by a stochastic process with fluctuations of voltage-dependent magnitude.
E. Acknowledgements
LJS would like to thank Neil Rabinowitz, Ben Willmore and Jan Schnupp for their collaboration
and helpful discussions.
N. C. Rabinowitz, B. D. B. Willmore, J. W. H. Schnupp, and A. J. King, The Journal of neuroscience 32,
11271 (2012).
F. S. Chance, L. F. Abbott, and A. D. Reyes, Neuron 35, 773 (2002).
A. Ayaz and F. S. Chance, Journal of neurophysiology 101, 958 (2009).
A. Longtin, B. Doiron, and A. R. Bulsara, Bio Systems 67, 147 (2002).
N. C. Rabinowitz, B. D. B. Willmore, J. W. H. Schnupp, and A. J. King, Neuron 70, 1178 (2011).
C. Koch and I. Segev, Computational neuroscience (MIT Press, 1998).
S. Ditlevsen and P. E. Greenwood, Journal of mathematical biology (2012), 10.1007/s00285-012-0552-7.
B. Ermentrout and D. H. Terman, Mathematical Foundations of Neuroscience (Springer, 2010).
N. Brunel and S. Sergi, Journal of theoretical biology 195, 87 (1998).
A. Buonocore, L. Caputo, E. Pirozzi, and L. M. Ricciardi, Neural computation 22, 2558 (2010).
V. L´ansk´a, P. L´ansk´y, and C. E. Smith, Journal of theoretical biology 166, 393 (1994).
17
P. Lansk´y, L. Sacerdote, and F. Tomassetti, Biological cybernetics 465, 457 (1995).
Y. Yu and T. Lee, Physical Review E 68, 011901 (2003).
S. Ostojic and N. Brunel, PLoS computational biology 7, e1001056 (2011).
P. L´ansk´y and L. Sacerdote, Physics Letters A 285, 132 (2001).
L. Sacerdote and M. T. Giraudo, Stochastic Biomathematical Models Lecture Notes in Mathematics, 2058
(2013), 10.1007/978-3-642-32157-3.
K. C. Chan, G. A. Karolyi, F. A. Longstaff, and A. B. Sanders, The Journal of Finance XLVII (1992).
M. Brennan and E. Schwartz, Journal of Financial and Quantitative Analysis XV (1980).
C. Ly and B. Doiron, PLoS computational biology 5, e1000365 (2009).
S. Hong, B. N. Lundstrom, and A. L. Fairhall, PLoS computational biology 4, e1000119 (2008).
T. P. Vogels, H. Sprekeler, F. Zenke, C. Clopath, and W. Gerstner, Science 334, 1569 (2011).
C. A. Coelho, Journal of Multivariate Analysis 64, 86 (1998).
D. V. Hinkley, Biometrika 56, 635 (1969).
R. C. Geary, Journal of the Royal Statistical Society 93, 442 (1930).
Appendix A: Other approximations for the conductance distribution and beyond
1. Gamma distribution
As seen in the results section (Fig. 2), the steady state distribution of the conductances seems
to be captured even better by a gamma distribution with shape k and scale θ, especially with
respect to its skew. Matching the mean (kθ) and variance (kθ2) of the gamma distribution to that
of the normal approximation (4,5), we require the shape parameter to be k = 2λτg and the scale
θ = ∆g/2. It is tempting to try to arrive at these parameters by deriving the gamma distribution
as the probability distribution of the sum of k exponentially distributed random variables of mean
θ, or as the distribution of waiting times until 'death' of random variables with poissonian arrival.
However, the analogy does not seem to work, the parameters would be out by factors of 2 or
1/2. For this reason approximating the steady-state conductances as gamma distributed it less
appealing. To arrive at an expression for the distribution of firing rates we would also like to know,
as a second step, the distribution of sums and differences of the conductances (as these appear
as g and I in the firing rate expression (1)), but there seems to be no established result for the
difference of two gamma distributed random variables, and the sum for different shape parameters
is an unwieldy expression (Coelho, 1998).
18
2. Further distributions
Approximating the distribution of ge and gi as normal, we can work out the distribution of g and
I, which will also be normal and (nearly) independent. The next term in the firing rate expression
(1) whose distribution we would like to know is g(t)(Vth − VL)/I(t). This is approximated by a
Gaussian ratio distribution (Hinkley, 1969), which can be made me made more Gaussian by a
Geary-Hinkley-transformation (Hinkley, 1969; Geary, 1930) of the random variable as long as I(t)
is unlikely to be negative. As seen in Fig. 7, this all works well, but going further is difficult.
Working out the distribution of log(1− g(t)(Vth− VL)/I(t)) while only considering values for which
g(t)(Vth − VL)/I(t) < 1 is tricky and goes beyond the scope of this paper. Instead, we turn to a
more pragmatic approach in the next section.
19
FIG. 7. Steady-state distributions of terms involving the excitatory (ge) and inhibitory (gi)
plotted, is C dV
r(I, g))]
n(g)
dt = d− n
Vth−VL
V and the firing rate for the deterministic case is f (I) = [tref− C(Vth−VL)
conductances in a sLIF neuron model. The model differential equation, in terms of the parameters
log(1−
−1. Distribution parameters are not fitted, but derived as described in the main text (4,5). Data
shown for synaptic input rates λe = λi = 1 kHz, but approximations were qualitatively similar for rates
as low as 0.5 kHz, and better the higher the rates. (a) Effective current d = IF F + Ie − Ii and effective
conductance (multiplied by a voltage term) n = g(Vth − VL) are approximately normally distributed, with
µd ∝ µIe − µIi, σ2
(b) The distribution of the ratio r of
the terms in (b) approximately follows the ratio distribution (Hinkley, 1969). Range shown is limited to
and µn ∝ µge + µgi , σ2
n = σ2
ge
+ σ2
gi
d = σ2
Ie
+ σ2
Ii
f (I) > 0. (c) The Geary-Hinkley-transform (Hinkley, 1969; Geary, 1930) makes r approximately normally
distributed with zero mean and unit variance. Maximum abscissa as in (c).
20
0.511.522.5300.511.5I/nAPa ndN. diff.N. sum00.5100.050.1ratioPb r = n/dRatio distrib.-2-1000.010.020.03Pxc GH(r)N(0,1) |
1611.04794 | 3 | 1611 | 2017-09-01T13:57:18 | Exploring the reproducibility of functional connectivity alterations in Parkinson's Disease | [
"q-bio.NC"
] | Since anatomic MRI is presently not able to directly discern neuronal loss in Parkinson's Disease (PD), studying the associated functional connectivity (FC) changes seems a promising approach toward developing non-invasive and non-radioactive neuroimaging markers for this disease. While several groups have reported such FC changes in PD, there are also significant discrepancies between studies. Investigating the reproducibility of PD-related FC changes on independent datasets is therefore of crucial importance. We acquired resting-state fMRI scans for 43 subjects (27 patients , 16 controls) and compared the observed FC changes with those obtained in 2 independent datasets, one made available by the PPMI consortium and a second one by the group of Tao Wu. Unfortunately, PD-related functional connectivity changes turned out to be non-reproducible across datasets. This could be due to disease heterogeneity, but also to technical differences. To distinguish between the two, we devised a method to directly check for disease heterogeneity using random splits of a single dataset. Since we still observe non-reproducibility in a large fraction of random splits of the same dataset, we conclude that functional heterogeneity may be a dominating factor behind the lack of reproducibility of FC alterations in different rs-fMRI studies of PD. While global PD-related functional connectivity changes were non-reproducible across datasets, we identified a few individual brain region pairs with marginally consistent FC changes across all three datasets. However, training classifiers on each one of the 3 datasets to discriminate PD scans from controls produced only low accuracies on the remaining two test datasets. Moreover, classifiers trained and tested on random splits of the same dataset (which are technically homogeneous) also had low test accuracies, directly substantiating disease heterogeneity. | q-bio.NC | q-bio | Exploring the reproducibility of functional connectivity
alterations in Parkinson's Disease
Liviu Badea1*, Mihaela Onu2,3, Tao Wu4,5, Adina Roceanu6, Ovidiu Bajenaru6,7
1 Artificial Intelligence and Bioinformatics Group, National Institute for Research and Development in Informatics,
Bucharest, Romania
2 Medical Imaging Department, Clinical Hospital Prof. Dr. Th. Burghele, 20 Panduri Street, Bucharest, Romania
3 University of Medicine and Pharmacy "Carol Davila", Biophysics Department, Bucharest, Romania
4 Department of Neurobiology, Key Laboratory on Neurodegenerative Disorders of Ministry of Education, Beijing
Institute of Geriatrics, Xuanwu Hospital, Capital Medical University, Beijing, China
5 Beijing Key Laboratory on Parkinson's Disease, Parkinson Disease Centre of Beijing Institute for Brain Disorders,
Beijing, China
6 University Emergency Hospital Bucharest, Neurology Department, Bucharest, Romania
7 University of Medicine and Pharmacy "Carol Davila", Department of Clinical Neurosciences, Bucharest, Romania
* Corresponding author
Abstract
Since anatomic MRI is presently not able to directly discern neuronal loss in Parkinson's Disease
(PD), studying the associated functional connectivity (FC) changes seems a promising approach
toward developing non-invasive and non-radioactive neuroimaging markers for this disease. While
several groups have reported such FC changes in PD, there are also significant discrepancies
between studies. Investigating the reproducibility of PD-related FC changes on independent
datasets is therefore of crucial importance.
We acquired resting-state fMRI scans for 43 subjects (27 patients and 16 normal controls, with 2
replicate scans per subject) and compared the observed FC changes with those obtained in two
independent datasets, one made available by the PPMI consortium (91 patients, 18 controls) and a
second one by the group of Tao Wu (20 patients, 20 controls).
Unfortunately, PD-related functional connectivity changes turned out to be non-reproducible across
datasets. This could be due to disease heterogeneity, but also to technical differences. To
distinguish between the two, we devised a method to directly check for disease heterogeneity using
random splits of a single dataset. Since we still observe non-reproducibility in a large fraction of
random splits of the same dataset, we conclude that functional heterogeneity may be a dominating
factor behind the lack of reproducibility of FC alterations in different rs-fMRI studies of PD.
While global PD-related functional connectivity changes were non-reproducible across datasets, we
identified a few individual brain region pairs with marginally consistent FC changes across all
three datasets. However, training classifiers on each one of the three datasets to discriminate PD
scans from controls produced only low accuracies on the remaining two test datasets. Moreover,
1
classifiers trained and tested on random splits of the same dataset (which are technically
homogeneous) also had low test accuracies, directly substantiating disease heterogeneity.
Keywords: Reproducibility; Resting state functional connectivity markers; Parkinson's Disease;
Data sharing
Abbreviations
PD Parkinson's Disease
NC Normal Controls
rs-fMRI resting state functional Magnetic Resonance Imaging
FC Functional Connectivity
ROI Region of Interest
H&Y Hoehn and Yahr score
PPMI Parkinson's Progression Markers Initiative
SVM Support Vector Machine classifier
GNB Gaussian Naïve Bayes classifier
2
Introduction
Although Parkinson's disease (PD) is the second most common neurodegenerative disease after
Alzheimer's disease, its diagnosis is still difficult, especially in the early premotor stages, as it is
mainly based on clinical evidence. To date, there is still no unique standard diagnostic test for PD,
despite the intense research efforts to develop accurate biomarkers based on blood tests or imaging
scans. The best current objective tests for PD evaluate dopaminergic function in the basal ganglia
by using various PET or SPECT radiotracers (e.g. DaTSCAN). But these tests make use of
radioactive substances, are performed only in specialized imaging centers and can also be very
expensive. Moreover, the loss of dopaminergic nigro-striatal neurons is a delayed pathological
event in the evolution of the disease, corresponding to Braak stages III-IV.
On the other hand, conventional (CT or MRI) brain scans of PD patients usually appear normal or
with minor non-specific changes, so that conventional imaging techniques are only useful for ruling
out other diseases that can be secondary causes of parkinsonism.
Therefore, since anatomic MRI is presently not able to directly discern (dopaminergic) neuronal
loss in PD [Tuite et al., 2013], studying the associated functional connectivity (FC) changes seems
to be a promising approach toward developing non-invasive and non-radioactive neuroimaging
markers for this disease.
While many groups have reported such FC changes in PD (see Supplementary Table 1 for a list of
such studies), an in-depth analysis of existing literature revealed significant discrepancies between
studies. Investigating the reproducibility of PD FC changes on independent datasets is therefore of
crucial importance.
A comprehensive review and analysis of the literature related to resting-state fMRI studies of
Parkinson's disease is out of the scope of the present paper [Gottlich et al., 2013; Long et al., 2012;
Skidmore et al., 2013; Baudrexel et al., 2011; Wu et al., 2011; Kwak et al., 2010; Kwak et al.,
2012; Helmich et al., 2010; Helmich et al., 2011; Luo et al., 2014; Yu et al., 2013; Hacker et al.,
2012; Kurani et al., 2015; Baggio et al., 2015; Esposito et al., 2013; Szewczyk-Krolikowski et al.,
2014; Tessitore et al., 2012; Sharman et al., 2013; Wu et al., 2012; Liu et al., 2013; Chen et al.,
2015; Wen et al., 2013; Zhang et al., 2015] (Supplementary Table 1; see also the review by
[Tahmasian et al., 2015]). We only mention some important inconsistencies of reported functional
connectivity changes in PD. Due to the crucial importance of the striatum in PD, we first discuss
some inconsistencies involving striatal seeds [Tahmasian et al., 2015]:
Contrary to [Hacker et al. 2012], [Helmich et al. 2010] observed no significant difference in
caudate functional connectivity in PD.
On the other hand, contrary to the study [Helmich et al., 2010], [Luo, Song, et al. 2014] did not
observe increased FC of the anterior putamen.
In contrast to [Hacker et al., 2012], [Luo, Song, et al. 2014] did not find a FC decrease between
the striatal seeds and the brainstem.
There are also inconsistencies involving non-striatal seeds. For example, [Wu et al., 2011] found
disrupted FC between the pre-SMA and the left putamen, as opposed to [Helmich et al., 2010], who
did not find a decreased FC between the putamen and pre-SMA in PD.
3
Since the motor symptoms are the most striking clinical manifestations in PD, many rs-fMRI
studies of PD concentrate on the sensorimotor system, including the basal ganglia, while
disregarding any other FC changes. On the other hand, other more unbiased studies tried to
determine a more global picture of the FC changes in PD. Some even tried to develop classifiers for
the disease based on rs-fMRI data [Long et al., 2012; Skidmore et al., 2013; Szewczyk-
Krolikowski et al., 2014; Chen et al., 2015], but most studies were not validated on independent
datasets.
There are also some gross discrepancies involving even the sign of the main FC changes in PD. For
example, [Luo et al. 2014] found only decreased FC in early stage PD, whereas most studies also
find FC increases.
The general picture one gets from the literature is complex and at times somewhat confusing due to
the numerous inconsistencies. Of course, these inconsistencies could be due to the different disease
stages analyzed, to the inherent functional heterogeneity of the disease, but also to technical
differences, or to the differences in the complex data (pre)processing workflows. Therefore, it is
crucial to use the same data processing workflow to check the reproducibility of PD-related FC
changes on as many independent datasets as possible.
In this work, we report a comparison between three different datasets obtained by completely
independent research groups (see Table 1). More precisely, we acquired resting-state scans for 43
Romanian subjects (27 patients and 16 normal controls, with 2 replicate scans per subject) and
compared the observed functional connectivity changes with those obtained in two independent
datasets, one made available by the PPMI consortium in the US (91 patients, 18 controls) and a
second one by the group of Tao Wu in China (20 patients and 20 normal controls).
Table 1. The 3 PD datasets compared in the present study (PD=Parkinson's Disease,
NC=normal controls).
PD
subjects
NC
Dataset
subjects
NEUROCON 27 (16 M) 16 (5 M)
Tao Wu
PPMI
PD
scans
54
20 (11 M) 20 (12 M) 20
91 (63 M) 18 (14 M) 134
NC
scans
31
20
19
age (NC)
mean (SD)
age PD
mean(SD)
68.7 (10.6) 67.6 (11.9) 0.76 1.92 (0.33) 4.6 (6.5)
65.2 (4.4) 64.8 (5.6) 0.78 1.88 (0.63) 5.4 (3.9)
61.3 (10.2) 64.7 (9.7) 0.17 1.72 (0.48) 1.9 (1.0)
p
(age
NC-
PD)
Hoehn &
Yahr
mean(SD)
disease
duration
mean(SD)
We briefly describe our approach in Fig 1A to better guide the reader through the remainder of the
paper. We started by comparing the PD-related global functional connectivity changes in the 3
datasets and found them to be non-reproducible. Of course, this could be due to disease
heterogeneity, but also to technical differences. To better distinguish between these two
possibilities, we devised a method to directly check for disease heterogeneity using random splits
of a single dataset. On the other hand, we searched for individual brain region pairs with consistent
connectivity changes across all three datasets. Finally, to more directly discriminate PD scans from
controls, we trained multivariate machine learning classifiers on one dataset and tested them on the
remaining two. Additionally, we trained and tested classifiers on technically homogeneous random
splits of the same dataset, to more directly check for disease heterogeneity.
4
Fig 1. Overview. (A) Main steps of the analysis. (B) Using random splits of a dataset with replicate
scans to check for disease (group) heterogeneity: (a, right) by placing different subjects (with all
their replicate scans) in the two splits ("split subjects") and respectively (b, left) by splitting the
replicates of the same subjects in the two splits ("split replicates"). The "split replicates" datasets
must show reproducible changes anyway, while non-reproducible changes across the "split
subjects" datasets are an indication of disease heterogeneity. (a,b,c,... correspond to subjects, while,
for instance, a' and a" are replicate scans for subject a.)
5
This is the first study investigating the reproducibility of functional connectivity changes in
Parkinson's disease on more than 2 datasets. Given the paucity of publicly available rs-fMRI PD
datasets, we advocate the critical importance of data sharing for enabling the discovery of
reproducible rs-fMRI biomarkers of PD.
Materials and Methods
Datasets
Three resting-state fMRI datasets of Parkinson's disease were compared in this study (see also
Table 1 for the main patient characteristics and the numbers of scans in each study):
-
-
-
the NEUROCON rs-fMRI study of 27 PD patients and 16 normal controls of the Neurology
Department of the University Emergency Hospital Bucharest (Romania),
a dataset of 20 PD patients and 20 normal controls provided by the group of Tao Wu (China),
a publicly available dataset of 91 PD patients and 18 controls of the Parkinson's Progression
Markers Initiative (PPMI) study in the US.
The datasets are somewhat similar, except for PPMI, which involved patients with a diagnosis of
PD for two years or less and who are not taking PD medications, while most patients from the other
two studies have been under treatment (most under levodopa). Also, PPMI patients were scanned in
the 'eyes open' condition. Still, we argue that our findings were not affected by these differences.
Since the datasets were compared in a pairwise manner, any putative discrepancies due to the
shorter disease durations in the PPMI dataset would only show up in the NEUROCON-PPMI and
Tao Wu-PPMI comparisons, but not in the NEUROCON-Tao Wu comparison. This was not
observed in reality.
NEUROCON
The NEUROCON study enrolled 27 patients with Parkinson's disease (mean age±SD 68.7±10.6
years) and 16 age-matched normal controls (67.6±11.9 years) with no history of neurological or
psychiatric disease. The patients were clinically assessed at the Neurology Department of the
University Emergency Hospital Bucharest (Romania) to be in the early or moderate stage of the
disease according to the Queen Square Brain Bank (QSBB) clinical criteria and met the
EFNS/MDS-ES (European Federation of Neurological Societies/Movement Disorder Society–
European Section) recommendations for diagnosis of Parkinson's disease. The mean disease
duration was 4.6 (±6.5) years for the entire patient cohort and respectively 2.75 (±2.15) years after
excluding three patients with particularly long disease durations (over 10 years: 11, 16 and 32 years,
respectively). Despite the longer disease durations, the above-mentioned 3 patients met the criteria
for moderately advanced disease (H&Y stage 2) and thus were included in the study. The mean
Hoehn and Yahr (H&Y) score [Hoehn and Yahr, 2001] was 1.93 (±0.33) and respectively 1.92
(±0.35) after excluding the 3 patients with long disease durations. All patients were in an early to
moderate stage of disease (stages 1 to 2.5). The mean score on the motor subset of the Unified
Parkinson's Disease Rating Scale (UPDRS) [Fahn, 1986] in the off medication condition was 28.3
(±9.3) for the entire patient cohort and 26.9 (±8.8) after excluding the 3 patients with long disease
6
durations. The study has been approved by the University Emergency Hospital Bucharest ethics
committee in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later
amendments. All patients gave their written informed consent to participate in the study.
Scanning. All subjects underwent two consecutive 8 min fMRI scans in a 1.5-Tesla Siemens
Avanto MRI scanner, in an awake resting state with their eyes closed. Two consecutive replicate
scans were acquired for each subject to enable the study of the reproducibility and respectively
homogeneity of FC changes in PD. (A single (control) subject could only be scanned once.) The
patients were scanned in the "off medication" state, at least 10 hours after the last intake of their
medication. The scanning protocol involved an Echo Planar sequence with repetition time (TR)
3480 ms, echo time (TE) 50 ms, axial orientation, voxel size 3.8×3.8×5 mm (without slice gaps),
flip angle 90 and number of averages=1. Each resting state session lasted 8.05 min, comprising
137 volumes. To enable better co-registration to the standard MNI template, high-resolution T1-
weighted images were also obtained for all subjects using an MPRAGE sequence (IR method,
TR=1940ms, TE=3.08ms, inversion time (IT)=1100ms, voxel size 0.97×0.97×1 mm, number of
averages=1).
Tao Wu
The dataset comprised 20 PD patients (11 males, mean age±SD 65.2±4.4 years) and 20 age-
matched normal controls (12 males, 64.8±5.6 years). All patients were in the early to moderate
stage of the disease (Hoehn and Yahr stages 1 to 2.5, except for a single patient with H&Y stage 3)
and had normal Mini-Mental State Examination (MMSE) scores. Moreover, there was no
statistically significant MMSE difference (p=0.43) between PD patients (28.8±1.1) and normal
controls (29.1±1.3).
Scanning. Both resting state fMRI and anatomic T1 scans were acquired for the 40 subjects in a
Siemens Magnetom Trio 3T equipment, in an awake resting state with their eyes closed. Each
resting state session lasted about 8 min (239 volumes, TR=2s, TE=40 ms, flip angle=90) with a
voxel size of 4×4×5 mm (6464 matrix, 28 slices, field of view=256mm×256mm). MPRAGE
scans were also obtained (voxel size 1×1×1 mm) for registration to the MNI template.
PPMI
Imaging data for 91 PD patients (63 males) and 18 normal controls (14 males) were downloaded
from the Parkinson's Progression Markers Initiative (PPMI) study data portal (http://www.ppmi-
info.org/access-data-specimens/download-data/,
https://ida.loni.usc.edu/home/projectPage.jsp?project=PPMI) with the kind permission of the PPMI
consortium. The study includes patients with a diagnosis of PD for two years or less and who are
not taking PD medications. The patients and controls are age-matched (p=0.17, mean age±SD
61.3±10.2 years for the PD patients and respectively 64.7±9.7 years for the normal controls). The
patients had a mean Hoehn & Yahr score of 1.72 (SD 0.48), with a mean disease duration (at the
time of the scan) of 1.9 years (SD 1.0). All patients had H&Y scores 1 to 2, except for only two,
who were classified as H&Y stage 3.
Scanning. The subjects were scanned in 8 different centers, but with a similar protocol on Siemens
Tim Trio 3Tesla scanners. Each resting state session lasted about 8.4 min (210 volumes, TR=2.4s,
TE=25ms, flip angle 80) with a voxel size of 3.3×3.3×3.3 mm (6866 matrix, 40 slices). Subjects
were instructed to rest quietly, keeping their eyes open and not to fall asleep. MPRAGE scans were
7
also obtained (voxel size 1×1×1 mm, TR=2.3s, TE=2.98ms, flip angle=9) for registration to the
MNI template.
Although our functional connectivity computations did not require any particular type of data
normalization (as only inter-region correlations are computed, rather than amplitudes), we also
considered a subset of scans acquired in a single center (center number 32, with the largest number
of PD patient and normal control scans), referred to in the following by the suffix 'center32'.
Preprocessing
to NIfTI using dcm2nii
(https://www.nitrc.org/projects/dcm2nii/) and
All datasets were preprocessed in a uniform manner. The raw scanner data in DICOM format was
converted
further
preprocessed using FSL (FMRIB Software Library v5 http://fsl.fmrib.ox.ac.uk/fsl/) as follows:
motion correction using MCFLIRT, brain extraction with BET, spatial smoothing (Gaussian kernel
FWHM 5mm) and denoising using nonlinear filtering (SUSAN), temporal high-pass filtering (with
a cutoff frequency of 0.01 Hz), registration to the standard Montreal Neurological Institute
MNI152 template via the anatomical T1 image (more precisely, BBR registration of the BOLD
image to the T1 image, followed by 12 DOF linear+nonlinear registration of the latter to the 2mm
MNI template). Nonlinear registration was performed at a resampling resolution of 4mm.
Besides the above 'standard' preprocessing workflow, we also considered alternative workflows
involving global signal regression (GS) and respectively a temporal bandpass filter (0.01-0.1Hz –
an ideal low-pass 0.1Hz filter was used in addition to the default FSL 0.01Hz highpass filter).
Since subject motion in the scanner has been observed to have significant influence on the
functional connectivities computed from rs-fMRI data, despite motion correction (e.g. [Power et al.,
2015]), we also considered subsets of scans with low in-scanner motion (marked by the suffix '0',
e.g. 'NC0' and 'PD0' – see also Supplementary Table 2).
PD-related functional connectivity changes
Functional connectivity [Friston and Buchel, 2003] is a rather loosely defined term, which
encompasses many different methods used to reveal temporal correlations of BOLD activity across
the brain. The simplest method consists in computing the correlations between all pairs of regions
of a given brain parcellation, but more sophisticated data decomposition methods, such as
Independent Component Analysis (ICA) are also widely used. Such data decomposition methods
do not assume a given brain parcellation, but instead construct spatial maps grouping voxels with
highly correlated timecourses. (Still, instead of being given a parcellation, such methods need to be
provided with a target number of components.)
In our study of the reproducibility of functional connectivity changes in PD, we used brain
parcellations constructed independently of the datasets under comparison, rather than applying
data decomposition methods, such as group-ICA, since the latter would be inherently biased
towards the "training dataset". Group-ICA may obtain a better functional parcellation for the
training dataset, but that parcellation would be less appropriate for any other independent dataset
("overfitting"), thereby introducing a bias in the analysis. To avoid these problems, we have chosen
to use brain parcellations constructed independently of the datasets under comparison, including
functional brain parcellations obtained by group-ICA on completely independent sets of subjects
(such as the 'Stanford' functional parcellation [Shirer et al., 2011]).
8
Moreover, to compensate for potential biases of any specific parcellation, we extended our analyses
to a number of 13 different parcellations employed in other rs-fMRI studies, two anatomical (AAL,
Talairach) and 11 functional (see Table 2 for more details).
Table 2. The brain parcellations used in the functional connectivity comparisons.
Number of
regions
Comments
Parcellation
Reference
AAL
[Tzourio-Mazoyer et
al., 2002]
[Craddock et al., 2012]
Craddock 130
Craddock 260
Craddock 500
Craddock 950
Shen 100
Shen 200
Shen 300
OASIS
Power
Gordon_surface [Gordon et al., 2014]
[Marcus et al., 2007]
[Power et al., 2011]
[Shen et al., 2013]
Talairach
[Talairach and
Tournoux, 1988]
Stanford
[Shirer et al., 2011]
90
116
130
260
500
950
93
183
278
97
264
333
695
anatomic atlas
spherical regions with a 10mm radius
anatomic atlas
functional parcellation obtained by
group-ICA
For each parcellation, we computed average timecourses for each region of interest (ROI) and the
resting state functional connectivity between each pair of ROIs (ROI1,ROI2) as the Fisher z-
transform of the temporal correlation between the corresponding ROI timecourses:
FC(ROI1,ROI2) = z(corr(ROI1,ROI2)).
For each dataset, we determined significant PD-related FC changes by applying two-sample t-tests
(with unequal sample sizes and unequal variances) to the functional connectivities of all ROI pairs.
ROI-pairs with significant group differences (NC versus PD) represent regions whose functional
connectivity was found to be significantly different in PD patients in that particular dataset. The
main aim of this study is to determine whether these changes are reproducible across datasets, to
enable the development of functional imaging biomarkers for PD.
Reproducibility of global functional connectivity changes in PD
Comparison of 3 different PD datasets
We first compared the global PD-related functional connectivity changes across the three
independent datasets NEUROCON, Tao Wu and PPMI to check to what extent these changes are
reproducible. More precisely, we performed pairwise comparisons for all dataset pairs as follows.
For each pair of datasets (i,j) and a fixed parcellation, we checked the extent to which the PD-
related FC changes in one dataset are correlated to the changes in the second dataset.
9
PD-related FC changes were quantified using t-values t(ROIk,ROIl) from group comparisons
(unpaired two-sample t-tests between NC and PD) of the functional connectivities between pairs of
regions of interest FC(ROIk,ROIl).
Then the reproducibility Rij across the two datasets i and j was determined as the correlation
between the corresponding t-values (viewed as a vector over all ROI pairs) for the two datasets:
where
Rij = corr(Ti,Tj),
(1)
Ti=(ti(ROI1,ROI2), ti(ROI1,ROI3), ti(ROI1,ROI4), )
with ti(ROIk,ROIl) the t-value corresponding to PD-related FC changes between ROIk and ROIl
with respect to dataset i (and similarly Tj for dataset j).
For a more intuitive graphical depiction of reproducibility across two datasets, we also constructed
the scatter-plot of ROI-pair t-values corresponding to group comparisons in the two datasets (see
Fig 2 for an example of such a scatter-plot).
Comparing PD-related FC changes (t-values) in the two datasets amounts to plotting for each ROI-
pair the t-value in dataset 1 against the t-value in dataset 2. We thereby obtain a scatter-plot with a
point for each ROI pair. The comparison of the FC changes in the two datasets thus involves
analyzing the distribution of points in the scatter-plot: ideally, perfect reproducibility would entail a
diagonal distribution of points in the scatter-plot, corresponding to perfectly correlated t-values in
the two datasets. Fig 4 depicts examples of good reproducibility, while Fig 2 shows cases of non-
reproducibility across datasets.
The correlation of t-values for the two datasets Rij = corr(Ti,Tj), as introduced above in (1), can be
viewed as an aggregate measure of the reproducibility across the two datasets i and j.
To obtain a more quantitative measure of the statistical significance of such a correlation Rij
between datasets, we performed permutations of the group labels (NC and PD) independently for
the two datasets and computed the p-value of the Rij value as the fraction of permutations for
which the dataset correlation w.r.t. the permuted data Rij
pij = { permutation Rij
() exceeds the real one (Rij):
() Rij } / N,
(2)
where N is the total number of permutations. All our permutation tests involved N=1000
permutations.
Various factors have been mentioned in the literature to affect functional connectivity measures:
-
subject motion in the scanner [Power et al., 2015],
- global signal regression (with or without) [Murphy et al., 2009; Hayasaka 2013],
-
the choice of the parcellation.
To study the influence of these factors on our reproducibility results, we also considered subsets of
scans with low in-scanner motion (marked by the suffix '0', e.g. 'NC0' and 'PD0'), repeated our
analyses with global signal regression, bandpass filtering and performed the comparisons using all
13 brain parcellations previously mentioned. Since the PPMI data has been acquired in several
different imaging centers, we also considered a potentially more homogeneous subset of scans
10
acquired in a single center (center number 32, with the largest number of PD patient and normal
control scans), referred in the following by the suffix 'center32'.
Comparison of random splits of the same PD dataset
As already mentioned in the Introduction, the observed lack of reproducibility of global FC
changes across datasets could be due to disease heterogeneity, but also to technical differences. To
distinguish between these two possibilities, we devised a method to directly check for disease
heterogeneity using random splits of a single dataset with replicate scans. Technical differences can
then be excluded since all the scans have been acquired under identical technical conditions. More
precisely, since in the NEUROCON study we acquired two replicate scans for each subject, we
constructed two homogeneous dataset splits simply by using (different) scans of the same subjects.
Additionally, two heterogeneous dataset splits can be obtained by placing different subjects (with
all their scans) in each split. In other words, instead of comparing two distinct datasets, we
compared two random splits of the same dataset, either:
(a) by placing different subjects in the two splits, with all the replicate scans of a subject in the
same split ("split subjects", heterogeneous split), or
(b) by placing each replicate scan of the same subject in a different split, so that the two splits
contain (different) scans of the same subjects ("split replicates", homogeneous split).
Dataset splits (b) are homogeneous since they contain scans of the same subjects, while splits (a)
they contain scans of different subjects. Therefore, consistent
are heterogeneous since
reproducibility across all random heterogeneous splits would indicate disease homogeneity, while
non-reproducibility in a large fraction of random heterogeneous splits would imply disease
heterogeneity. (In both cases, we expect to observe consistent reproducibility across the
homogeneous splits, at least as long as the technical noise is not dominating the biological signal.)
A diagram of our method is shown in Fig 1B.
As in the pairwise comparisons between different datasets, we used permutation tests and formula
(2) to compute p-values of the reproducibility across split datasets, for both the heterogeneous
("split subjects") and the homogeneous ("split replicates") datasets. Due to the random nature of
the splits, we repeated the analysis for Ns > 1000 different random splits of the original data. To
assess the fraction of (non-)reproducible splits, we determined the empirical cumulative
distribution function (CDF) of the reproducibility p-values for the Ns random splits.
The analysis was also repeated for the data with global signal regression.
Influence of technical factors, preprocessing and parcellation
We also studied the influence on reproducibility of certain key technical factors and preprocessing
steps, such as:
-
-
the repetition time (TR),
linear vs. nonlinear registration,
- global signal regression,
-
the specific brain parcellation used for evaluating functional connectivity.
The AAL parcellation (which is typical) was used whenever not specified otherwise.
11
Doubling the TR
The repetition time might, in principle, influence the measured low-frequency rs-fMRI fluctuations
and indirectly the functional connectivities (which are temporal correlations). To study the
influence of the repetition time on reproducibility, we constructed a synthetic dataset with a double
TR by leaving out every second time-point from the NEUROCON timeseries data (for each scan
and each voxel). We then analyzed with our method the reproducibility of group changes in
functional connectivity between the original NEUROCON dataset and the synthetic one with a
double TR.
Linear vs. nonlinear registration
To study the impact of registration on reproducibility, we preprocessed the NEUROCON data both
with linear and nonlinear registration to the MNI 152 template and determined the reproducibility
of group changes in functional connectivity between the two resulting datasets.
Global signal regression
Since global signal regression (GSR) has been observed to be very effective at removing scanning
artifacts [Hayasaka, 2013], including motion artifacts, we also studied the reproducibility of FC
changes between the NEUROCON dataset processed with GSR and the same dataset processed
without GSR.
Influence of parcellation
To avoid potential biases of any specific parcellation, we repeated our pairwise comparisons
between the 3 PD datasets using all 13 different parcellations mentioned above, including
functional and anatomic parcellations, with a wide range of numbers of regions of interest (90 to
950).
Reproducibility of the individual differentiating FC changes
The reproducibility analysis performed above involves global functional connectivity changes, i.e
changes in the FC of all ROI pairs, not just the ones that differentiate PD from normal controls.
Even with non-reproducible global FC changes, it might be in principle possible that only a very
few brain region pairs might still reproducibly differentiate PD patients from controls. To this end,
we also studied individual brain region pairs with FC changes that are significant w.r.t. all datasets.
More precisely, for each ROI-pair, we compute max(p), the largest (least significant) of the three p-
values obtained in the three datasets (separately for the FC increases and respectively decreases)
and sort the ROI-pairs in increasing order of this max(p). The most significant min(max(p)) of these
max(p) corresponds to the ROI-pair with the best overall significance with respect to the 3 datasets,
as all other ROI-pairs have larger (less significant) p-values with respect to at least one dataset.
Finally, to assess the statistical significance of such a best ROI-pair, we use a permutation test (of
the disease labels in each dataset) to check the fraction of random permutations with a more
significant (smaller) min(max(p)) than the real data (we performed N=1000 random permutations).
p(min(max(p+))) = { permutation min(max(p+))() min(max(p+)) } / N,
where min(max(p+)) corresponds to FC increases in NC versus PD. A similar relation holds for the
FC decreases min(max(p)).
The analysis was repeated for all 13 parcellations considered in this study (Table 2).
12
Learning classifiers for discriminating PD-related FC changes
We used machine learning techniques to learn classifiers that discriminate PD from controls using
functional connectivities between ROI pairs as features.
First, we trained classifiers on each one of the 3 datasets (NEUROCON, Tao Wu, PPMI) and tested
them on the other two datasets. Both Linear Support Vector Machines (SVM) and Gaussian Naïve
Bayes (GNB) classifiers were tested, with progressively increasing numbers of features:
N=10,50,100,500,5000 (functional connectivities between ROI pairs). The N features selected were
the best discriminating ROI pair functional connectivities, based on unpaired t-tests between
normal and PD scans. As the two classes (NC-Normal Controls and PD-Parkinson's Disease) are
not balanced in all 3 datasets, we employ the average accuracy Aacc = (acc(NC)+acc(PD))/2 for
assessing the performance of the classifiers (a random classifier is expected to have an average
accuracy of 0.5).
Since the different datasets are not technically homogeneous, we also trained and tested classifiers
on random splits of the same dataset, to check to what extent the low accuracies are due to
technical differences, or to disease heterogeneity. More precisely, we performed 10,000 random
splits in half of each dataset, trained a classifier on one half and tested it on the other.
13
Results
ROI-pairs with significant group differences (NC versus PD) in functional connectivity were found
in all three PD datasets: NEUROCON, Tao Wu and PPMI (Supplementary Tables 3 and 4).
However, these changes seemed at first sight to be distinct in each dataset. Our main aim in this
paper has been to systematically investigate the reproducibility of the PD-related FC changes
across independent validation datasets.
PD-related FC changes are non-reproducible across 3 datasets
The reproducibility of global PD-related functional connectivity changes was determined by
pairwise comparisons between three independent datasets: NEUROCON, Tao Wu and respectively
PPMI. Fig 2 shows the scatter-plots of ROI-pair t-values (corresponding to the group comparison
NC-PD) for the three dataset pairs, indicating a lack of reproducibility of global FC changes in PD.
(Perfect reproducibility would correspond to a diagonal distribution of points corresponding to ROI
pairs with perfectly correlated t-values with respect to both datasets.) Moreover, discriminating
ROI-pairs situated in the upper right and respectively lower left corners of one plot are not
discriminating in the other plots.
Fig 2. Scatter-plots of ROI-pair t-values for the three dataset pairs indicate non-
reproducibility of global PD-related FC changes.
For a more quantitative measure of the reproducibility of FC changes between two datasets, we
computed the Pearson correlation between t-values (viewed as vectors of over all ROI pairs) with
respect to each dataset (R values shown in Fig 2.) We also estimated the statistical significance (p-
values) of these reproducibility measures by permutation tests of the group labels independently for
the two datasets – Table 3 shows the reproducibility measure and associated p-value for various
pairwise comparisons between the three datasets, with standard highpass preprocessing (>0.01Hz),
global signal regression (GS) and respectively bandpass filter (0.01-0.1Hz). Since in-scanner
motion may influence FC measures, we present not only a comparison between the full patient and
normal control cohort, but also that corresponding to a subset of scans with low in-scanner motion
14
(denoted by the suffix '0'). Moreover, since PPMI data were acquired at many different centers, we
also considered the restriction of the PPMI data to the scans from a single center (suffix 'center32').
The AAL parcellation was used in this case, but we also study the influence of the parcellation later
on.
Table 3. Reproducibility measure R and associated p-value p for various pairwise
comparisons between datasets with standard highpass preprocessing ('standard'), global signal
regression (GS) and respectively bandpass filter (BP) ('0' indicates 'low in scanner motion',
'center32' – scans performed at a single PPMI center).
Dataset 1
NEUROCON
Group
contrast 1
NC-PD
NC0-PD0
NC-PD
NC0-PD0
Dataset
2
Tao Wu
NEUROCON
NC-PD
PPMI
NC0-PD0
NC-PD
NC0-PD0
Tao Wu
NC-PD
PPMI
NC0-PD0
Group contrast
2
NC-PD
NC0-PD0
NC-PD
NC0-PD0
NC_center32-
PD_center32
NC0_center32-
PD0_center32
NC-PD
NC0-PD0
NC_center32-
PD_center32
NC0_center32-
PD0_center32
standard
R
p
GS
BP
R
p
R
p
-0.145955 0.949 -0.0679698 0.804 0.000845579 0.476
-0.0832455 0.847 0.035725 0.300 0.0132475 0.437
0.0692377 0.257
-0.163251 0.937
0.0401831 0.352 0.0308243 0.373 -0.0824776 0.800
0.10039
0.138
-0.0122797 0.526 0.0188561 0.430
-0.142598 0.914
-0.0873994 0.813 -0.0894912 0.832
-0.152545 0.945
0.00673472 0.466 0.0228885 0.403 0.0948706 0.153
-0.0182896 0.560 0.044843 0.298 0.0545817 0.293
-0.0446155 0.687 -0.032841 0.666 0.0243423 0.400
-0.0712155 0.812 -0.0176842 0.597 -0.0436096 0.681
A clear lack of reproducibility of global PD-related FC changes is observed in all the three dataset
pairs. This is the first study comparing three independent rs-fMRI datasets of PD. The fact that we
compare 3 datasets is very important, as it lowers the probability that the lack of reproducibility is
due to a dataset that may be "faulty" in some sense – in that case, with 3 datasets we might still
observe reproducibility with respect to the remaining dataset pair (which we do not see in reality).
Inconsistent reproducibility of FC changes in heterogeneous
dataset splits indicate disease heterogeneity
The non-reproducibility across 3 datasets mentioned above seems to be due to disease
heterogeneity, but it could also be due to technical differences. To exclude the latter possibility, we
checked for disease heterogeneity using random splits of a single dataset with replicate scans
(NEUROCON), all of which have been acquired under identical technical conditions.
(a) We first constructed random splits by placing different subjects in the two splits, with all the
replicate scans of a subject in the same split ("split subjects", heterogeneous splits). As can be seen
in Fig 3A (blue curve), a large fraction of these random splits display non-reproducible functional
connectivity changes. More precisely, 88% of the random heterogeneous splits show non-
reproducibility at the p>0.01 level and 42% at the p>0.05 level. Fig 3B shows the complementary
cumulative distribution function (1-CDF) for the corresponding reproducibility measure R (blue
curve), while Fig 3C presents a typical scatter-plot of ROI-pair t-values for a random
heterogeneous split (one with R equal to the median).
15
Fig 3. Inconsistent reproducibility of PD-related FC changes in random heterogeneous dataset
splits and consistent reproducibility in random homogeneous dataset splits. (A)
Complementary cumulative distribution function (CCDF=1-CDF) of the reproducibility p-values
for Ns=2510 random heterogeneous splits and Ns=325 random homogeneous splits. (B) CCDF of
the reproducibility measure R. (C) A scatter-plot of ROI-pair t-values for a random heterogeneous
split. (D) A scatter-plot of ROI-pair t-values for a random homogeneous split.
(b) Next, we constructed random splits by placing each replicate scan of the same subject in a
different split, so that the two splits contain (different) scans of the same subjects ("split replicates",
homogeneous splits). In contrast to (a), all homogeneous splits showed reproducibility at the p<10-3
level (Fig 3A, red curve). Fig 3B shows the complementary CDF for the reproducibility measure
(red curve) – note the significantly higher reproducibility (R) values for the homogeneous splits
(red curve, median R=0.71) as compared to the heterogeneous splits (blue curve, median R=0.21).
Fig 3D displays a typical scatter-plot of ROI-pair t-values for a random homogeneous split.
The observed non-reproducibility in a large fraction of heterogeneous dataset splits indicates
disease heterogeneity, in line with the comparison between the 3 independent PD datasets. As a
control, we did indeed observe consistent reproducibility with respect to all homogeneous dataset
splits, demonstrating that the technical noise could not have been the dominating factor behind the
erratic non-reproducibility in heterogeneous splits.
16
The fact that the well-known clinical heterogeneity of Parkinson's disease is also accompanied by
heterogeneity in resting state functional connectivity may not retrospectively be a big surprise to an
experienced neurologist, although its exact extent could not have been estimated a priori, before
analyzing the data. However, does this FC heterogeneity in PD also imply the lack of practical
usefulness of rs-fMRI functional connectivity? Are there any other conditions that can be reliably
differentiated using resting state functional connectivity? To answer these questions, we applied
our approach to a different, potentially more homogeneous contrast, namely that between eyes
open and eyes closed resting state conditions in healthy volunteers. Repeating our analysis of
reproducibility of FC group changes on random splits of the Beijing eyes open-eyes closed dataset
[Liu et al., 2013] (see Supporting Information) revealed reproducibility (p<0.05) not just in the
homogeneous dataset splits, but also in the heterogeneous ones (Supplementary Fig 1 – only 6% of
the heterogeneous and just 0.8% of the homogeneous random splits were non-reproducible at the
p>0.05 level).
Summing up our findings, from the point of view of global FC changes, Parkinson's disease is
heterogeneous, as opposed to the eyes open-eyes closed contrast, which is much more
homogeneous (Table 4).
Table 4. Summary of reproducibility of global functional connectivity changes in Parkinson's
Disease and respectively the Eyes Open-Eyes Closed contrast.
Reproducibility of
global FC changes
different datasets
PD
no
split subjects
(heterogeneous splits)
inconsistent
(no in a large
fraction of splits)
split replicates
(homogeneous splits)
yes
EO-EC
yes
yes
Influence of technical factors and preprocessing on reproducibility
We found good reproducibility when changing various technical factors or processing options of
the NEUROCON data, such as (see Fig 4):
- doubling the repetition time (TR),
-
registration (linear versus nonlinear),
- global signal regression (with versus without).
17
Fig 4. Reproducibility when changing various technical factors or preprocessing options.
This is in line with our conclusion that functional heterogeneity, rather than these technical factors,
is the dominating factor behind the lack of reproducibility of FC changes in different rs-fMRI
studies of Parkinson's disease.
We also tested the influence of various rs-fMRI denoising methods on the reproducibility of PD-
related FC changes, such as ICA-FIX [Griffanti et al., 2014; Salimi-Khorshidi et al., 2014], or
regression of the mean white matter and/or cerebrospinal fluid signal – none of these denoising
methods changed the observed non-reproducibility (data not shown).
We also observed no improvement in reproducibility across random splits of the NEUROCON
dataset after regressing out potential confounders, such as age, gender, or disease duration (data not
shown).
Influence of parcellation on reproducibility
We have argued that functional connectivity must be computed with respect to an unbiased
parcellation (i.e. one that hasn't been constructed from any of the analyzed datasets). However, any
given parcellation has also specific biases that may in principle affect the capacity to discriminate
between PD and normal controls – an especially relevant factor is the average size and number of
the ROIs. Testing the reproducibility of the PD-related global FC changes using 13 different
parcellations, with varying numbers of ROIs (between 90 and 950, see Table 2), revealed a lack of
reproducibility regardless of parcellation, or dataset pair (Table 5). (The NEUROCON-PPMI
comparison was marginally significant (p=0.05) for the NC-PD contrast, but this significance
didn't survive perturbations such as selecting just the 'center32' scans from PPMI (p=0.259), or
restriction to the low motion scans NC0-PD0 (p=0.26), or NC0-PD0_center32 (p=0.555).)
18
Table 5. Reproducibility measure and associated p-value for 13 parcellations and all three
dataset pairs (NC-PD contrast).
NEUROCON-TaoWu NEUROCON-PPMI
TaoWu-PPMI
R
p
R
p
R
-0.011116
-0.0490741
-0.032232
-0.0247647
0.172 -0.0546894
0.554 0.0927652 0.191 -0.0613781
-0.00871389 0.560 0.0775174 0.205 -0.0770479
0.00488102 0.469 0.079955
0.00952484 0.409 0.0694818 0.174 -0.0347004
0.708 0.0958863 0.187 -0.0567748
0.641 0.0936098 0.172 -0.0618746
0.632 0.0829943 0.159 -0.0729234
0.00698018 0.482 0.0205063 0.411 -0.0155159
0.166 0.0812484 0.141 -0.0153593
0.262 0.025185
0.686 0.0263891 0.357 0.0598013
0.730 0.185859
0.949 0.0692377 0.257 0.00673472
0.391 -0.0435482
0.050
-0.081526
-0.0273596
-0.0600487
0.0533032
-0.145955
p
0.747
0.864
0.839
0.748
0.734
0.799
0.871
0.598
0.600
0.748
0.129
0.819
0.466
Parcellation
Craddock130
Craddock260
Craddock500
Craddock950
Shen100
Shen200
Shen300
OASIS
Power264
Talairach
Stanford
AAL
Gordon_surface 0.0381087
Marginally significant individual differentiating FC changes in PD
Despite non-reproducibility of PD-related global FC changes across different datasets, a small
number of ROI-pairs that distinguish PD from controls may nevertheless, in principle, show
reproducible changes across datasets. To check for this possibility, we concentrated on individual
brain region pairs with FC changes that are significant w.r.t. all datasets, by sorting the ROI-pairs
according to their least significance max(p) with respect to all datasets.
For example, Table 6 shows the ROI-pairs with FC decreases in PD (i.e. positive t-values,
corresponding to NC>PD) and max(p) < 0.05 for the Power264 parcellation, without global signal
regression. The best ROI-pair has max(p+)=0.0125, so min(max(p+))=0.0125.
To check whether this min(max(p+)) is statistically significant, we performed permutation tests as
described. Table 7 lists these min(max(p)) values as well as their associated significance
p(min(max(p))) for all 13 parcellations. Only two out of the 13 parcellations yielded significant
ROI-pairs at the p<0.05 significance level ('Power264' and 'Talairach'), while a third parcellation
produced only marginally significant ROI-pairs ('Shen100', p=0.055) - see Table 8 and Fig 5.
These (marginally) significant ROI-pairs involve visual-sensorimotor, respectively visual-parietal
association areas. Whether these ROI-pair changes are more widely reproducible or not will have to
await the release of more publicly-available PD rs-fMRI datasets.
19
Table 6. Best ROI-pairs with FC decreases in PD (t>0, corresponding to NC>PD) and max(p)
< 0.05 (for the Power264 parcellation, without global signal regression).
max(p)
ROI1-ROI2
NEUROCON
(NC-PD)
p
p TaoWu (NC-
p PPMI (NC-
PD)
PD)
t
NEUROCON
(NC-PD)
t TaoWu
(NC-PD)
t PPMI (NC-
PD)
0.012526
0.016354
0.025311
0.038357
0.038529
0.038634
0.044406
0.046861
sphere5(-21,-31,61)-
sphere5(15,-77,31)
sphere5(-42,45,-2)-
sphere5(43,-78,-12)
sphere5(-21,-31,61)-
sphere5(-24,-91,19)
sphere5(-38,-33,17)-
sphere5(20,-86,-2)
sphere5(37,-81,1)-
sphere5(38,-17,45)
sphere5(-21,-31,61)-
sphere5(-26,-90,3)
sphere5(-21,-31,61)-
sphere5(29,-77,25)
sphere5(-21,-31,61)-
sphere5(-40,-88,-6)
0.012526
0.0018915
0.011638
2.5541
3.3398
0.0026347
0.016354
0.0079673
3.1172
2.5153
0.0019939
0.025311
0.025232
3.2025
2.3298
0.038357
0.00041882
0.014314
2.1104
3.8665
0.038529
0.02727
0.038269
2.1123
2.297
0.012064
0.038634
0.033842
2.5863
2.1425
0.0085123
0.027058
0.044406
2.7193
2.3019
0.0077414
0.046861
0.017283
2.7342
2.0557
2.7402
2.905
2.3949
2.653
2.2003
2.2603
2.1243
2.55
Table 7. Significance of min(max(p)) values for all 13 parcellations (no global signal
regression).
min(max(p+)) min(max(p)) p(min(max(p+))) p(min(max(p)))
0.0607935
0.0503968
0.0294203
0.0202304
0.0172931
0.0371887
0.0425589
0.0324906
0.0697703
0.012526
0.0275321
0.00701726
0.0599084
0.112189
0.242712
0.148728
0.0909226
0.0542509
0.376232
0.151199
0.115015
0.3402
0.118716
0.11095
0.0339528
0.281706
0.208
0.152
0.170
0.206
0.316
0.055*
0.195
0.199
0.208
0.033*
0.185
0.032*
0.145
0.426
0.831
0.843
0.859
0.815
0.905
0.747
0.804
0.938
0.846
0.846
0.522
0.820
Parcellation
AAL
Craddock130
Craddock260
Craddock500
Craddock950
Shen100
Shen200
Shen300
OASIS
Power264
Gordon_surface
Talairach
Stanford
20
Table 8. Marginally significant FC changes w.r.t. all 3 datasets (decreased in PD).
Parcellation p(min(max(p+))) min(max(p+)) ROI1
Talairach 0.032
0.00701726
Power264 0.033
0.012526
Shen100
0.055
0.0371887
(-24,-58,4)
left visual association
area, lingual gyrus,
BA18
sphere5(-21,-31,61)
left postcentral /
precentral gyrus
L.BA19.3
left cuneus, precuneus
ROI2
(–38, –32,16)
left superior temporal gyrus,
BA41, planum temporale /
parietal operculum
sphere5(15, –77,31)
right cuneus
R.BA6.1
right SMA, middle cingulate
Fig 5. Marginally significant FC changes w.r.t. all 3 datasets. The ROIs were mapped onto the
brain surface using BrainNet Viewer [Xia et al., 2013] (http://www.nitrc.org/projects/bnv/).
Testing classifiers for discriminating PD-related FC changes
Training classifiers on functional connectivity data for each one of the 3 datasets (NEUROCON,
Tao Wu, PPMI) and testing them on the other two datasets produced average accuracies on test
data in the range 0.225 - 0.7 (mean 0.497, standard deviation 0.073), while a random classifier is
expected to have an average accuracy of 0.5. Fig 6 shows the corresponding average accuracies
Aacc = (acc(NC)+acc(PD))/2 for standard preprocessing (with the default FSL highpass filter at
0.01Hz), global signal regression and respectively bandpass filtering (0.01-0.1Hz) for both linear
SVM and Gaussian Naïve Bayes (GNB) classifiers with N=5000 features (out of the total of 6670
ROI pairs of the AAL parcellation). See Supplementary Fig 2 and Supplementary Table 5 for the
accuracies of classifiers with N=10,50,100,500,5000 features.
21
Since for each training dataset (for example NEUROCON), we have two different test datasets
(PPMI and TaoWu in our example), we also computed an aggregated average accuracy for each
dataset by taking the mean of the two average accuracies corresponding to the two remaining test
datasets (Aacc(dataset-dataset1)+Aacc(dataset-dataset2))/2. The resulting aggregated average
accuracies were low, in the range 0.336 - 0.591 (mean 0.497, standard deviation 0.0522, compared
to 0.5 for a random classifier; see also Fig 7).
Since the three datasets are not technically homogeneous, we also trained and tested classifiers on
random splits of the same dataset, to check to what extent the low accuracies are due to technical
differences, or to disease heterogeneity. Fig 8 shows the average accuracies for 10,000 random
splits in half of each dataset and various preprocessing options. Again, the means of the average
accuracies over the 10,000 tests were low, in the range 0.51 – 0.66, reinforcing the evidence for
disease heterogeneity.
Fig 6. Average accuracies for classifiers trained on dataset 1 and tested on dataset 2 for all
dataset pairs using standard preprocessing ('standard'), global signal regression (GS) and
respectively bandpass filtering (0.01-0.1Hz). Here, SVM (linear Support Vector Machine) and
GNB (Gaussian Naïve Bayes) classifiers used N=5000 features – see Supplementary Fig 2 for
classifier accuracies for varying N. As an example, NEUROCON-PPMI denotes classifiers trained
on NEUROCON and tested on PPMI data.
22
Fig 7. Aggregated average accuracies for classifiers trained on each of the 3 datasets using
standard preprocessing ('standard'), global signal regression (GS) and respectively bandpass
filtering (0.01-0.1Hz). Classifiers were trained with N=10,50,100,500,5000 features. As an example,
NEUROCON(10) refers to the aggregated accuracy (Aacc(NEUROCON-PPMI) +
Aacc(NEUROCON-TaoWu))/2 for classifiers trained on NEUROCON and tested on PPMI and
respectively TaoWu data using N=10 features. SVM – linear SVM classifier, GNB – Gaussian
Naïve Bayes classifier.
Fig 8. Average accuracies for classifiers trained and tested on split data from the same dataset
using standard preprocessing ('standard'), global signal regression (GS) and respectively bandpass
filtering (0.01-0.1Hz). An SVM classifier with N=5000 features was used.
23
Discussion
The accelerated increase in the number of functional connectivity studies of Parkinson's Disease
requires a consolidation of the knowledge in this field for enabling the development of clinically
relevant rs-fMRI markers for this disease. Unfortunately however, there are many inconsistencies
between published works and virtually no high confidence reproducibility studies.
This is the first study investigating the reproducibility of functional connectivity changes in
Parkinson's disease on more than two datasets. The fact that we use a uniform data processing
workflow for all datasets excludes a large number of technical factors as potential culprits for the
observed differences between datasets. Also, the fact that our comparison involves three datasets is
essential, as it lowers the probability that the observed lack of reproducibility is due to a
problematic dataset – in such a case, with 3 datasets we might still observe reproducibility with
respect to the remaining dataset pair, something which we do not see in reality.
To better clarify the issue, we devised a method to directly check for disease heterogeneity using
random splits of a single dataset with replicate scans. Technical differences can then be excluded
since all the scans have been acquired under identical technical conditions. The fact that we still
observe non-reproducibility in a significant fraction of random subsamples of each individual
dataset (these subsamples being technically homogeneous as they come from the same dataset),
suggests that functional heterogeneity may be a dominating factor behind the lack of
reproducibility of functional connectivity alterations in different resting state fMRI studies of
Parkinson's disease.
This could be due to the heterogeneous multi-lesional topography and progression of the
neurodegenerative process, possibly accompanied by variable compensatory functional circuit
changes, as well as by changes due to dopaminergic medication [Tahmasian et al., 2015].
For a a more direct graphical depiction of the heterogeneity of the functional connectomes of the
PD patients, we have applied consensus NMF clustering [Brunet et al., 2004] for a progressively
increasing number of clusters (k=2,…,18, Supplementary Fig 3). Note that besides the consistent
grouping of the replicate scan pairs for each patient, it is difficult to single out an optimal number
of clusters k.
While global PD-related functional connectivity differences were non-reproducible across datasets,
we identified a few individual ROI pairs with marginally consistent FC differences across all three
datasets. However, finding out whether these differences are more widely reproducible or not will
have to await the release of more public PD datasets.
Additionally, we applied more sophisticated multivariate machine learning techniques to learn
classifiers that discriminate PD from controls using functional connectivities between ROI pairs as
features. However, training classifiers on each one of the three datasets (NEUROCON, Tao Wu,
PPMI) produced only low accuracies on the remaining two (test) datasets, in line with the
preceding results. Furthermore, since the three datasets are not technically homogeneous, we also
trained and tested classifiers on random splits of the same dataset, to more directly check to what
extent the low accuracies are due to technical differences, or to disease heterogeneity. Again, we
obtained low average accuracies (with means in the range 0.51 – 0.66), reinforcing the evidence for
disease heterogeneity. Interestingly, these results are consistent with a recent study [Orban et al.,
2017] on multisite generalizability of schizophrenia diagnosis based on functional brain
connectivity, which reported multisite classification accuracies below 70%, in contrast to over 30
24
previously published,
classification accuracy exceeds 80%.
largely single-site schizophrenia studies, whose average reported
Therefore, given the paucity of publicly available rs-fMRI PD datasets, we advocate the critical
importance of data sharing for enabling the discovery of reproducible and clinically useful
functional imaging biomarkers of PD. In this regard, we view our study as an important first step
towards more refined reproducibility studies that would be possible only with more publicly
available datasets. In view of the many inconsistencies found in the published literature on PD-
related functional connectivity changes, we strongly argue for a direct computational comparison of
PD rs-fMRI datasets using a uniform data processing workflow, to avoid publication bias as well as
processing workflow differences in the separate studies.
Limitations. The present study has concentrated on PD-related changes in functional connectivity
(loosely viewed as correlations between different regions of interest), rather than changes in
fluctuations of the amplitude of the rs-fMRI signal. In a complementary study, [Wu et al., 2015]
observed PD-related changes in ALFF, but with rather limited reproducibility. An in-depth analysis
of the reproducibility of PD-related differences in the amplitude of fluctuations is out of the scope
of the present paper.
Acknowledgments
We are very grateful to the PPMI consortium for granting us access to the Parkinson's Progression
Markers Initiative (PPMI) imaging data. Data used in the preparation of this article were obtained
from the PPMI database (www.ppmi-info.org/data). For up-to-date information on the study, visit
www.ppmi-info.org. PPMI – a public-private partnership – is funded by the Michael J. Fox
Foundation
including AbbVie, Avid
Radiopharmaceuticals, Biogen, Bristol-Myers Squibb, Covance, GE Healthcare, Genentech,
GlaxoSmithKline, Eli Lilly and Company, Lundbeck, Merck, Meso Scale Discovery, Pfizer,
Piramal Imaging, Roche, Servier, UCB (Union Chimique Belge) and GOLUB Capital.
for Parkinson's Research and
funding partners,
We also thank Madalina Tivarus and the reviewers for some very useful suggestions for improving
the manuscript.
Compliance with Ethical Standards
Ethical approval: The study was carried out in accordance with The Code of Ethics of the World
Medical Association (Declaration of Helsinki) for experiments involving humans and was
approved by the ethics committee of the University Emergency Hospital Bucharest.
Informed consent: Informed consent was obtained in written form from all individual participants
included in the study.
25
References
Baggio HC, Segura B, Sala‐Llonch R, Marti MJ, Valldeoriola F, Compta Y, Tolosa E, Junque C.
Cognitive impairment and resting‐state network connectivity in Parkinson's disease. Human
brain mapping. 2015 Jan 1;36(1):199-212.
Baudrexel S, Witte T, Seifried C, von Wegner F, Beissner F, Klein JC, Steinmetz H, Deichmann R,
Roeper J, Hilker R. Resting state fMRI reveals increased subthalamic nucleus–motor cortex
connectivity in Parkinson's disease. Neuroimage. 2011 Apr 15;55(4):1728-38.
Brunet JP, Tamayo P, Golub TR, Mesirov JP. Metagenes and molecular pattern discovery using
the national academy of sciences. 2004 Mar
matrix factorization. Proceedings of
23;101(12):4164-9.
Chen Y, Yang W, Long J, Zhang Y, Feng J, Li Y, Huang B. Discriminative Analysis of
Parkinson's Disease Based on Whole-Brain Functional Connectivity. PloS one. 2015 Apr
17;10(4):e0124153.
Craddock RC, James GA, Holtzheimer PE, Hu XP, Mayberg HS. A whole brain fMRI atlas
generated via spatially constrained spectral clustering. Human brain mapping. 2012 Aug
1;33(8):1914-28.
Esposito F, Tessitore A, Giordano A, De Micco R, Paccone A, Conforti R, Pignataro G,
Annunziato L, Tedeschi G. Rhythm-specific modulation of the sensorimotor network in drug-
naive patients with Parkinson's disease by levodopa. Brain. 2013 Mar 1;136(3):710-25.
Fahn S. Recent developments in Parkinson's disease. Edited by S. Fahn, C. D. Mardsen, P. Jenner,
and P. Teychenne. Raven Press; 1986.
Friston KJ, Buchel C. Functional connectivity. Human Brain Function. 2003;2:999-1018.
Gordon EM, Laumann TO, Adeyemo B, Huckins JF, Kelley WM, Petersen SE. Generation and
evaluation of a cortical area parcellation from resting-state correlations. Cerebral cortex. 2014
Oct 14:bhu239.
Göttlich M, Münte TF, Heldmann M, Kasten M, Hagenah J, Krämer UM. Altered resting state
brain networks in Parkinson's disease. PloS one. 2013 Oct 28;8(10):e77336.
Griffanti L, Salimi-Khorshidi G, Beckmann CF, Auerbach EJ, Douaud G, Sexton CE, Zsoldos E,
Ebmeier KP, Filippini N, Mackay CE, Moeller S. ICA-based artefact removal and accelerated
fMRI acquisition for improved resting state network imaging. Neuroimage. 2014 Jul 15;95:232-
47.
Hacker C, Perlmutter J, Criswell S, Ances B, Snyder A. Resting state functional connectivity of the
striatum in Parkinson's disease. Brain. 2012 Nov 28:aws281.
Hayasaka S. Functional connectivity networks with and without global signal correction. Frontiers
in human neuroscience 7. 2013.
Helmich RC, Derikx LC, Bakker M, Scheeringa R, Bloem BR, Toni I. Spatial remapping of
cortico-striatal connectivity in Parkinson's disease. Cerebral cortex. 2010 May 1;20(5):1175-86.
Helmich RC, Janssen MJ, Oyen WJ, Bloem BR, Toni I. Pallidal dysfunction drives a
cerebellothalamic circuit into Parkinson tremor. Annals of neurology. 2011 Feb 1;69(2):269-81.
26
Hoehn MM, Yahr MD. Parkinsonism: Onset, progression and mortality. Neurology 57:S11–S26.
2001.
Kurani AS, Seidler RD, Burciu RG, Comella CL, Corcos DM, Okun MS, MacKinnon CD,
Vaillancourt DE. Subthalamic nucleus-sensorimotor cortex functional connectivity in de novo
and moderate Parkinson's disease. Neurobiology of aging. 2015 Jan 31;36(1):462-9.
Kwak Y, Peltier S, Bohnen N, Müller M, Dayalu P, Seidler RD. Altered resting state cortico-
striatal connectivity in mild to moderate stage Parkinson's disease. Frontiers in systems
neuroscience. 2010 Sep 15;4:143.
Kwak Y, Peltier S, Bohnen N, Müller M, Dayalu P, Seidler RD. L-DOPA changes spontaneous
low-frequency BOLD signal oscillations in Parkinson's disease: a resting state fMRI study.
Frontiers in systems neuroscience. 2012 Jul 4;6:52.
Liu D, Dong Z, Zuo X, Wang J, Zang Y. Eyes-open/eyes-closed dataset sharing for reproducibility
evaluation of resting state fMRI data analysis methods. Neuroinformatics. 2013 Oct
1;11(4):469-76.
Liu H, Edmiston EK, Fan G, Xu K, Zhao B, Shang X, Wang F. Altered resting-state functional
connectivity of the dentate nucleus in Parkinson's disease. Psychiatry Research: Neuroimaging.
2013 Jan 30;211(1):64-71.
Long D, Wang J, Xuan M, Gu Q, Xu X, Kong D, Zhang M. Automatic classification of early
Parkinson's disease with multi-modal MR imaging. PloS one. 2012 Nov 9;7(11):e47714.
Luo C, Song W, Chen Q, Zheng Z, Chen K, Cao B, Yang J, Li J, Huang X, Gong Q, Shang HF.
Reduced functional connectivity in early-stage drug-naive Parkinson's disease: a resting-state
fMRI study. Neurobiology of aging. 2014 Feb 28;35(2):431-41.
Marcus DS, Wang TH, Parker J, Csernansky JG, Morris JC, Buckner RL. Open Access Series of
Imaging Studies (OASIS): cross-sectional MRI data in young, middle aged, nondemented, and
demented older adults. Journal of cognitive neuroscience. 2007 Sep;19(9):1498-507.
Murphy K, Birn RM, Handwerker DA, Jones TB, Bandettini PA. The impact of global signal
regression on resting state correlations: are anti-correlated networks introduced?. Neuroimage.
2009 Feb 1;44(3):893-905.
Orban P, Dansereau C, Desbois L, Mongeau-Pérusse V, Giguère CÉ, Nguyen H, Mendrek A, Stip
E, Bellec P. Multisite generalizability of schizophrenia diagnosis classification based on
functional brain connectivity. Schizophrenia research. 2017 Jun 7.
Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, Vogel AC, Laumann TO,
Miezin FM, Schlaggar BL, Petersen SE. Functional network organization of the human brain.
Neuron. 2011 Nov 17;72(4):665-78.
Power JD, Schlaggar BL, Petersen SE. Recent progress and outstanding issues in motion correction
in resting state fMRI. Neuroimage. 2015 Jan 15;105:536-51.
Salimi-Khorshidi G, Douaud G, Beckmann CF, Glasser MF, Griffanti L, Smith SM. Automatic
denoising of functional MRI data: combining independent component analysis and hierarchical
fusion of classifiers. Neuroimage. 2014 Apr 15;90:449-68.
27
Sharman M, Valabregue R, Perlbarg V, Marrakchi‐Kacem L, Vidailhet M, Benali H, Brice A,
Lehéricy S. Parkinson's disease patients show reduced cortical‐subcortical sensorimotor
connectivity. Movement Disorders. 2013 Apr 1;28(4):447-54.
Shen X, Tokoglu F, Papademetris X, Constable RT. Groupwise whole-brain parcellation from
resting-state fMRI data for network node identification. Neuroimage. 2013 Nov 15;82:403-15.
Shirer WR, Ryali S, Rykhlevskaia E, Menon V, Greicius MD. Decoding subject-driven cognitive
states with whole-brain connectivity patterns. Cerebral cortex. 2012 Jan 1;22(1):158-65.
Skidmore FM, Yang M, Baxter L, Von Deneen KM, Collingwood J, He G, White K, Korenkevych
D, Savenkov A, Heilman KM, Gold M. Reliability analysis of the resting state can sensitively
and specifically identify the presence of Parkinson disease. Neuroimage. 2013 Jul 15;75:249-61.
Szewczyk-Krolikowski K, Menke RA, Rolinski M, Duff E, Salimi-Khorshidi G, Filippini N,
Zamboni G, Hu MT, Mackay CE. Functional connectivity in the basal ganglia network
differentiates PD patients from controls. Neurology. 2014 Jul 15;83(3):208-14.
Talairach J, Tournoux P. Co-planar stereotaxic atlas of the human brain. 3-Dimensional
proportional system: an approach to cerebral imaging. 1988
Tessitore A, Amboni M, Esposito F, Russo A, Picillo M, Marcuccio L, Pellecchia MT, Vitale C,
Cirillo M, Tedeschi G, Barone P. Resting-state brain connectivity in patients with Parkinson's
disease and freezing of gait. Parkinsonism & related disorders. 2012 Jul 31;18(6):781-7.
Tuite PJ, Mangia S, Michaeli S. Magnetic resonance imaging (MRI) in Parkinson's disease.
Journal of Alzheimer's disease & Parkinsonism. 2013 Mar 25:001.
Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O, Delcroix N, Mazoyer B,
Joliot M. Automated anatomical labeling of activations in SPM using a macroscopic anatomical
parcellation of the MNI MRI single-subject brain. Neuroimage. 2002 Jan 31;15(1):273-89.
Wen X, Wu X, Liu J, Li K, Yao L. Abnormal baseline brain activity in non-depressed Parkinson's
disease and depressed Parkinson's disease: a resting-state functional magnetic resonance
imaging study. PloS one. 2013 May 22;8(5):e63691.
Wu T, Long X, Wang L, Hallett M, Zang Y, Li K, Chan P. Functional connectivity of cortical
motor areas in the resting state in Parkinson's disease. Human brain mapping. 2011 Sep
1;32(9):1443-57.
Wu T, Ma Y, Zheng Z, Peng S, Wu X, Eidelberg D, Chan P. Parkinson's Disease-Related Spatial
Covariance Pattern Identified with Resting-State Functional MRI. Journal of Cerebral Blood
Flow & Metabolism. 2015 Nov 1;35(11):1764-70.
Wu T, Wang J, Wang C, Hallett M, Zang Y, Wu X, Chan P. Basal ganglia circuits changes in
Parkinson's disease patients. Neuroscience letters. 2012 Aug 22;524(1):55-9.
Xia M, Wang J, He Y. BrainNet Viewer: a network visualization tool for human brain
connectomics. PloS one. 2013 Jul 4;8(7):e68910.
Yu R, Liu B, Wang L, Chen J, Liu X. Enhanced functional connectivity between putamen and
supplementary motor area in Parkinson's disease patients. PLoS One. 2013 Mar 26;8(3):e59717.
28
Zhang D, Liu X, Chen J, Liu B, Wang J. Widespread increase of functional connectivity in
Parkinson's disease with tremor: a resting-state FMRI study. Frontiers in aging neuroscience.
2015;7.
29
Supporting Information
Resting-state fMRI studies of Parkinson's disease
Supplementary Table 1. Resting-state fMRI studies of Parkinson's disease
Study
Main study characteristics
eyes open /
closed
NC
PD Patient characteristics
Gottlich 2013
graphs, AAL, AAL subdivision in
343 cortical & subcortical ROIs
eyes closed
20
37
advanced disease, on medication
(L-dopa, agonists)
classifier: SVM (ALFF, ReHo,
RFCS, GM, WM, CSF) accuracy
86.96%, sensitivity 78.95%,
specificity 92.59%, precision 88%
classifier (ALFF): accuracy 88%,
sensitivity 92%, specificity 87%,
precision 87%
eyes closed
27
19
early stage (H&Y 1-2), OFF (12
h)
eyes closed
15
14 OFF(12-18 h)
Long 2012
Skidmore
2013
Baudrexel
2011
FC STN, M1 hand area
eyes closed
44
31
Wu 2011
FC pre-SMA - M1
seeds: pre-SMA, M1, PCC
eyes closed
18
18
early stage PD patients (n=31)
during the medication-off state
with healthy controls (n=44); 16
tremor, 15 non-tremor
akinesia right (at most mild
tremor), OFF (12h) H&Y 1.78+-
0.5
Kwak 2010
6 striatal seeds: 3 caudate (inferior
ventral striatum, superior ventral
striatum, dorsal caudate), 3
putamen (dorsal caudal putamen,
dorsal rostral putamen, ventral
rostral putamen)
eyes fixed
on cross
24
25
mild to moderate stage (H&Y 1-
2.5), ON and OFF(12–18 h)
Kwak 2012
ALFF (fALLF)
24
24
Helmich 2010
FC anterior, posterior putamen,
caudate
posterior cingulate (control)
eyes closed
36
41
Helmich 2011
tremor
eyes closed
36
41
mild to moderate stage (H&Y 1-
2.5) ON and OFF (12-18h)
13 PD without any tremor, 18
moderate to severe.
10 - never anti-PD medication
(median H&Y 2.1, max 5)
perfectly age-matched (57 years)
19 tremor, 23 nontremor, right-
handed, 12 no medication, OFF
(12 h)
Supporting information
30
Luo 2014
FC anterior, posterior putamen,
anterior caudate, amygdala
eyes closed
52
52
Yu 2013
FC putamen, caudate, and SMA
20
19
52 PD right handed, early stage
drug-naïve, H&Y 1.85 (max 3),
31 right onset, 21 left onset
OFF
obvious at least a mild tremor
Hacker 2012
FC striatum
6 ROIs: (i) caudate nucleus; (ii)
anterior putamen; and (iii) posterior
putamen
Kurani 2015
FC (restricted to motor areas)
seeds: STN (most affected part),
posterior cingulate (control seed)
eyes open
(fixate
cross)
eyes open
(focused
on the
word
"RELAX")
19
13
advanced PD ON
19
39
20 de novo
19 moderate OFF
Baggio 2015
cognitive impairement
ICA+dual regression (25 networks),
group comparison of: DMN, dorsal
attention network (DAN), bilateral
frontoparietal networks (FPN),
43 seeds (10 DAN, 18 DMN, 15
FPN)
36
65
ON state
65 nondemented: 34% mild
cognitive impairement (MCI)
Esposito 2013
ICA (fastICA BrainVoyager 40
components, restriction to
sensorimotor network-best fit with
previous template)
eyes closed
18
20
20 drug naïve PD, 10 before &
after (1H) levodopa, 10 before &
after placebo
Szewczyk-
Krolikowski
2014
ICA+dual regression – basal
ganglia (BG) network from 80
separate controls
- classifier based on average BG
component values in voxels
discriminating PD(OFF)-controls
eyes open
19
19
discovery cohort: 19 PD
ON/OFF, 19 controls
validation cohort: 13 PD (5 drug-
naïve), no controls (this
disallows evaluation of
specificity!)
80 elderly controls for BG
template (MELODIC groupICA
50 components)
all subjects right-handed
Tessitore
2012
fastICA, sogICA
40 components
select DMN, the frontoparietal
(right and left FPN), sensorimotor
network (SMN), visual, auditory
eyes closed
15
29
PD ON
16 FOG+
Supporting information
31
ROIs: caudate, putamen, globus
pallidus, thalamus
sensorimotor (M1, postcentral
gyrus)
associative (ventrolateral,
dorsolateral prefrontal)
limbic (orbitofrontal, rectus gyrus,
cingulate, insula, medial temporal
ctx, perirhinal & entorhinal cortex,
hippocampus, amygdala)
ROIs: SNc bilaterally
effective connectivity Granger
causality analysis (GCA)
with long TR=2000 & short
TR=400
Sharman 2013
Wu 2012
eyes closed
45
36
eyes closed
16
16
16 de novo
OFF/ON
Liu 2013
FC dentate nucleus (cerebellum) ->
cerebellar output
eyes closed
18
18
Chen 2015
SVM classifier LOOCV based on
FC between 116 ROIs of AAL
parcellation (150 features selected
by Kendall tau correlation) -
93.62% accuracy, 90.47%
sensitivity, 96.15 specificity
eyes closed
26
21
mild to moderate (1.34 H&Y
average)
in OFF state
8 rigidity & bradykinesia-
dominant (PD_AR)
10 tremor-dominant (PD_T)
21 PD (10 males, 11 females,
58.3 years), 26 HC (10 males, 16
females, 61.3 years) OFF state
(12 hours) UPDRS 29.8 (sd 9.1),
disease duration 3.2 years (sd
3.2)
Wen 2013
depression
ALFF
eyes closed
21
33
OFF
17 depression
Supporting information
32
Subject motion in scanner
Since subject motion in the scanner has been observed to have significant influence on the
functional connectivities computed from rs-fMRI data, despite motion correction (e.g. [Power et al.,
2015]), we also considered subsets of scans with low in-scanner motion (marked by the suffix '0',
e.g. 'NC0' and 'PD0' – see also Supplementary Table 2). The 3 datasets have different in-scanner
motion characteristics.
Tao Wu
Low motion condition
Supplementary Table 2. The subsets of scans with low in-scanner motion
Dataset
NEUROCON max abs motion 1/4 voxel = 0.958mm
max rel motion 1/6 voxel = 0.638mm
max abs motion 1/4 voxel = 1mm
max rel motion 1/6 voxel = 0.67mm
max abs motion 1/3 voxel = 1.098mm
max rel motion 1/3 voxel = 1.098mm
mean abs motion 1/6 voxel = 0.549mm
max rel motion 1/6 voxel = 0.549mm
PPMI
134
20
NC scans PD scans NC0 scans PD0 scans
31
39
54
27
20
19
13
13
14
89
The following Tables show the p-values corresponding to potential group differences in motion in
the 3 datasets analyzed:
NEUROCON
NC-PD (31-54) p-value (t-value)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.00243021
(-3.14036)
0.0545481
(-1.96039)
0.0945346
(-1.6961)
0.135713
(-1.50937)
NC0-PD0 (27-39)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.166962
(-1.39829)
0.043621
(-2.06138)
0.319776
(-1.00337)
0.100148
(-1.67026)
Note that in the NEUROCON cohort, patients move more than normal controls.
Supporting information
33
Tao Wu
NC-PD (20-20)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.0270361
(2.33514)
0.727106
(0.352521)
0.0074933
(2.96994)
0.543508
(0.613522)
NC0-PD0 (13-14)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.256895
(1.16397)
0.497445
(0.688602)
0.0859714
(1.80121)
0.455533
(0.758043)
In the Tao Wu cohort, patients move less than normal controls.
PPMI
NC-PD (19-134)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.0619124
(1.9849)
0.485967
(0.710965)
0.107866
(1.69169)
0.134632
(1.56558)
NC0-PD0 (13-89)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.0656489
(1.94269)
0.314195
(-1.03452)
0.224242
(1.26499)
0.53616
(0.63367)
NC_center32-PD_center32 (9-30)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.242348
(1.23728)
0.77672
(0.289486)
0.344483
(1.00257)
0.318376
(1.05993)
NC0_center32-PD0_center32 (7-23)
p(mean rel)
p(mean abs)
p(max rel)
p(max abs)
0.173686
(1.40637)
0.649951
(0.46271)
0.252684
(1.21065)
0.13755
(1.64485)
Supporting information
34
ROI-pairs with significant group differences in the separate datasets
Supplementary Table 3 below shows the numbers of significant ROI pairs for several significance
thresholds for the unpaired t-test between patient and control functional connectivities (for the AAL
parcellation and without correction for multiple comparisons, due to the limited sample sizes). A
single scan for each subject was considered in this comparison.
Supplementary Table 3. Number of significant ROI-pairs for the AAL parcellation
significance
NEUROCON(16-27) TaoWu(20-20) PPMI(18-91)
level
0.001
0.005
0.01
0.05
4
23
48
283
15
68
121
489
19
81
156
737
As expected, the power of the tests was greater for the datasets with more samples, but the effect
sizes (t-values) were similar, as can be seen in Supplementary Table 4 below, which shows the top
positive and respectively negative effect sizes observed.
Supplementary Table 4. Top effect sizes (t-values) and the corresponding p-values for the top
positive and respectively negative FC alterations
top t+
top p+
top t−
top p−
NEUROCON(16-27)
TaoWu(20-20)
PPMI(18-91)
3.05647
0.00429499
-4.495779
6.09E-05
4.632511
4.23E-05
-2.91725
0.0062734
4.70764
7.99E-05
-3.870344
0.0003765
Supporting information
35
Comparison of random splits of an 'eyes open-eyes closed' dataset
Besides the potentially heterogeneous contrast between PD and normal controls, we also tested our
method on a different, potentially more homogeneous contrast, namely 'eyes open' versus 'eyes
closed' resting state in healthy volunteers. We used the Beijing eyes-open-eyes closed (EO-EC)
dataset [Liu et al., 2013] (http://fcon_1000.projects.nitrc.org/indi/IndiPro.html), which involved 48
college students aged 19–31 years, 24 female with no history of neurological and psychiatric
disorders. Each participant underwent three 8 min resting state scanning sessions: an EC session
followed by two sessions counter-balanced across subjects: one EO resting state and one EC resting
state session.
The functional images were obtained on a Siemens Trio 3 Tesla scanner using an echo-planar
imaging sequence with the following parameters: 33 axial slices, thickness/gap=3.5/0.7 mm, in-
plane resolution=64×64, repetition time=2000 ms, echo time=30 ms, flip angle=90°, field of view
(FOV)=200×200mm2, 240 volumes per scan. In addition, a 3D T1-weighted MPRAGE image was
acquired with the following parameters: 128 sagittal slices, slice thickness/gap=1.33/0 mm, in-
plane resolution=256×192, TR=2530 ms, TE=3.39 ms, inversion time (TI)=1100 ms, flip angle=7°,
FOV=256×256 mm2. Note that the parameters used in this study are quite similar to the ones from
the PD datasets, including the scanning time (~8 min), with the exception of the 1.5 Tesla field
strength used in the NEUROCON study (all the other studies used 3 Tesla machines).
We repeated our analyses of reproducibility of group changes in functional connectivity on random
splits of the Beijing EO-EC dataset on both "split subjects" (heterogeneous) and "split replicates"
(homogeneous) datasets using the AAL parcellation. As in the case of PD, permutation tests were
employed to compute p-values of the reproducibility across split datasets. Additionally, we
repeated the analysis for the data with global signal regression.
'Eyes Open-Eyes Closed' FC changes are reproducible
The fact that the well-known clinical heterogeneity of Parkinson's disease is also accompanied by
heterogeneity in resting state functional connectivity may not retrospectively be a big surprise to an
experienced neurologist, although its exact extent could not have been estimated a priori, before
analyzing the data.
Supporting information
36
However, does this FC heterogeneity in PD also imply the lack of practical usefulness of rs-fMRI
functional connectivity? Are there any other conditions that can be reliably differentiated using
resting state functional connectivity? To answer these questions, we applied our approach to a
different, potentially more homogeneous contrast, namely that between eyes open and eyes closed
resting state conditions in healthy volunteers. Repeating our analysis of reproducibility of FC group
changes on random splits of the Beijing eyes open-eyes closed dataset [Liu et al., 2013] revealed
reproducibility (p<0.05) not just in the homogeneous dataset splits, but also in the heterogeneous
ones (Supplementary Fig 1 – only 6% of the heterogeneous and just 0.8% of the homogeneous
random splits were non-reproducible at the p>0.05 level). This implies that the EO-EC contrast
produces more homogeneous and reproducible global FC changes.
Supplementary Fig 1. Consistent reproducibility of 'eyes open'-'eyes closed' (EO-EC) FC
changes in random heterogeneous and homogeneous dataset splits. (A) Complementary
cumulative distribution function (CCDF=1-CDF) of the reproducibility p-values for Ns=1056
random heterogeneous splits and Ns=814 random homogeneous splits. (B) CCDF of the
reproducibility measure R. (C) A scatter-plot of ROI-pair t-values for a random heterogeneous split.
(D) A scatter-plot of ROI-pair t-values for a random homogeneous split.
Supporting information
37
Learning classifiers for discriminating PD-related FC changes
Supplementary Fig 2. Average accuracies Aacc = (acc(NC)+acc(PD))/2 for classifiers trained
on dataset 1 and tested on dataset 2 for all dataset pairs using standard preprocessing
('standard'), global signal regression (GS) and respectively bandpass filtering (0.01-0.1Hz).
Classifiers were trained with N=10,50,100,500,5000 features. For example, NEUROCON-
PPMI(10) shows average accuracies of classifiers trained on NEUROCON and tested on PPMI data
using N=10 features. SVM – linear SVM classifier, GNB – Gaussian Naïve Bayes classifier.
Supporting information
38
Supplementary Table 5. Detailed performance metrics for the various classifiers
TRAIN
DATASET
TEST
DATASET
classifier
N
pairs
Pre-
processing
TEST
AvgaccPer
Class
TEST
accPerCl
ass(NC)
TEST
accPerCla
ss(PD)
NEUROCON PPMI
gnb_pooled
10 standard
0.526709 0.1579
0.895522
NEUROCON PPMI
svm_linear
10 standard
0.452082 0.1579
0.746269
NEUROCON PPMI
gnb_pooled
50 standard
0.447761 0
NEUROCON PPMI
svm_linear
50 standard
0.462687 0
0.895522
0.925373
NEUROCON PPMI
gnb_pooled
100 standard
0.410644 0.05263 0.768657
NEUROCON PPMI
svm_linear
100 standard
0.421642 0
0.843284
NEUROCON PPMI
gnb_pooled
500 standard
0.433032 0.05263 0.813433
NEUROCON PPMI
svm_linear
500 standard
0.49293
0.10526 0.880597
NEUROCON PPMI
gnb_pooled 5000 standard
0.466614 0.05263 0.880597
NEUROCON PPMI
svm_linear
5000 standard
0.530833 0.26316 0.798507
NEUROCON TaoWu
gnb_pooled
10 standard
0.35
NEUROCON TaoWu
svm_linear
10 standard
0.4
NEUROCON TaoWu
gnb_pooled
50 standard
0.225
NEUROCON TaoWu
svm_linear
50 standard
0.5
NEUROCON TaoWu
gnb_pooled
100 standard
0.275
NEUROCON TaoWu
svm_linear
100 standard
0.45
NEUROCON TaoWu
gnb_pooled
500 standard
NEUROCON TaoWu
svm_linear
500 standard
NEUROCON TaoWu
gnb_pooled 5000 standard
NEUROCON TaoWu
svm_linear
5000 standard
0.325
0.425
0.35
0.45
0.2
0.15
0.15
0.15
0.15
0.05
0.2
0.05
0.25
0.1
0.5
0.65
0.3
0.85
0.4
0.85
0.45
0.8
0.45
0.8
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
NEUROCON
gnb_pooled
10 standard
0.491637 0.6129
0.37037
NEUROCON
svm_linear
10 standard
0.471326 0.3871
0.555556
NEUROCON
gnb_pooled
50 standard
0.486858 0.67742 0.296296
NEUROCON
svm_linear
50 standard
0.501493 0.35484 0.648148
NEUROCON
gnb_pooled
100 standard
0.519116 0.74194 0.296296
NEUROCON
svm_linear
100 standard
0.490143 0.25807 0.722222
NEUROCON
gnb_pooled
500 standard
0.576762 0.83871 0.314815
NEUROCON
svm_linear
500 standard
0.471924 0.12903 0.814815
NEUROCON
gnb_pooled 5000 standard
0.542413 0.67742 0.407407
NEUROCON
svm_linear
5000 standard
0.439665 0.06452 0.814815
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
gnb_pooled
10 standard
0.45
0.85
svm_linear
10 standard
0.7
gnb_pooled
50 standard
0.475
svm_linear
50 standard
0.55
gnb_pooled
100 standard
0.5
svm_linear
100 standard
0.625
gnb_pooled
500 standard
0.55
svm_linear
500 standard
0.575
gnb_pooled 5000 standard
0.6
svm_linear
5000 standard
0.575
0.6
0.8
0.3
0.85
0.35
0.9
0.15
0.85
0.15
0.05
0.8
0.15
0.8
0.15
0.9
0.2
1
0.35
1
TaoWu
TaoWu
TaoWu
TaoWu
NEUROCON
gnb_pooled
10 standard
0.397551 0.25807 0.537037
NEUROCON
svm_linear
10 standard
0.41368
0.29032 0.537037
NEUROCON
gnb_pooled
50 standard
0.462366 0.25807 0.666667
NEUROCON
svm_linear
50 standard
0.397252 0.3871
0.407407
Supporting information
39
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
NEUROCON
gnb_pooled
100 standard
0.480884 0.25807 0.703704
NEUROCON
svm_linear
100 standard
0.443548 0.3871
0.5
NEUROCON
gnb_pooled
500 standard
0.503883 0.32258 0.685185
NEUROCON
svm_linear
500 standard
0.41368
0.29032 0.537037
NEUROCON
gnb_pooled 5000 standard
0.476105 0.32258 0.62963
NEUROCON
svm_linear
5000 standard
0.448626 0.19355 0.703704
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
gnb_pooled
10 standard
0.493912 0.36842 0.619403
svm_linear
10 standard
0.539081 0.47368 0.604478
gnb_pooled
50 standard
0.629419 0.68421 0.574627
svm_linear
50 standard
0.499018 0.73684 0.261194
gnb_pooled
100 standard
0.614493 0.68421 0.544776
svm_linear
100 standard
0.518068 0.84211 0.19403
gnb_pooled
500 standard
0.599568 0.68421 0.514925
svm_linear
500 standard
0.483896 0.68421 0.283582
gnb_pooled 5000 standard
0.625687 0.68421 0.567164
svm_linear
5000 standard
0.547722 0.78947 0.30597
NEUROCON PPMI
gnb_pooled
10 GS
0.515318 0.10526 0.925373
NEUROCON PPMI
svm_linear
10 GS
0.458955 0
0.91791
NEUROCON PPMI
gnb_pooled
50 GS
0.575805 0.31579 0.835821
NEUROCON PPMI
svm_linear
50 GS
0.571485 0.1579
0.985075
NEUROCON PPMI
gnb_pooled
100 GS
0.526905 0.21053 0.843284
NEUROCON PPMI
svm_linear
100 GS
0.522584 0.05263 0.992537
NEUROCON PPMI
gnb_pooled
500 GS
0.49293
0.10526 0.880597
NEUROCON PPMI
svm_linear
500 GS
0.500393 0.10526 0.895522
NEUROCON PPMI
gnb_pooled 5000 GS
0.522977 0.1579
0.88806
NEUROCON PPMI
svm_linear
5000 GS
0.538885 0.42105 0.656716
NEUROCON TaoWu
gnb_pooled
10 GS
NEUROCON TaoWu
svm_linear
10 GS
NEUROCON TaoWu
gnb_pooled
50 GS
NEUROCON TaoWu
svm_linear
50 GS
NEUROCON TaoWu
gnb_pooled
100 GS
NEUROCON TaoWu
svm_linear
100 GS
NEUROCON TaoWu
gnb_pooled
500 GS
NEUROCON TaoWu
svm_linear
500 GS
NEUROCON TaoWu
gnb_pooled 5000 GS
NEUROCON TaoWu
svm_linear
5000 GS
0.5
0.425
0.5
0.4
0.475
0.425
0.45
0.6
0.375
0.575
0.2
0.15
0.4
0.05
0.45
0
0.6
0.25
0.55
0.15
0.8
0.7
0.6
0.75
0.5
0.85
0.3
0.95
0.2
1
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
NEUROCON
gnb_pooled
10 GS
0.623656 0.58065 0.666667
NEUROCON
svm_linear
10 GS
0.499403 0.25807 0.740741
NEUROCON
gnb_pooled
50 GS
0.549881 0.45161 0.648148
NEUROCON
svm_linear
50 GS
0.488053 0.16129 0.814815
NEUROCON
gnb_pooled
100 GS
0.462366 0.25807 0.666667
NEUROCON
svm_linear
100 GS
0.51135
0.09677 0.925926
NEUROCON
gnb_pooled
500 GS
0.595878 0.58065 0.611111
NEUROCON
svm_linear
500 GS
0.502091 0.09677 0.907407
NEUROCON
gnb_pooled 5000 GS
0.616487 0.67742 0.555556
NEUROCON
svm_linear
5000 GS
0.423536 0.03226 0.814815
TaoWu
gnb_pooled
10 GS
0.525
0.45
0.6
Supporting information
40
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
svm_linear
10 GS
gnb_pooled
50 GS
svm_linear
50 GS
gnb_pooled
100 GS
svm_linear
100 GS
gnb_pooled
500 GS
svm_linear
500 GS
gnb_pooled 5000 GS
svm_linear
5000 GS
0.575
0.45
0.575
0.575
0.65
0.475
0.6
0.55
0.575
0.4
0.45
0.3
0.45
0.35
0.35
0.2
0.55
0.2
0.75
0.45
0.85
0.7
0.95
0.6
1
0.55
0.95
NEUROCON
gnb_pooled
10 GS
0.459976 0.29032 0.62963
NEUROCON
svm_linear
10 GS
0.434588 0.25807 0.611111
NEUROCON
gnb_pooled
50 GS
0.413978 0.16129 0.666667
NEUROCON
svm_linear
50 GS
0.446237 0.22581 0.666667
NEUROCON
gnb_pooled
100 GS
0.434588 0.25807 0.611111
NEUROCON
svm_linear
100 GS
0.441458 0.29032 0.592593
NEUROCON
gnb_pooled
500 GS
0.436679 0.35484 0.518519
NEUROCON
svm_linear
500 GS
0.443548 0.3871
0.5
NEUROCON
gnb_pooled 5000 GS
0.408901 0.35484 0.462963
NEUROCON
svm_linear
5000 GS
0.452808 0.3871
0.518519
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
gnb_pooled
10 GS
0.536332 0.73684 0.335821
svm_linear
10 GS
0.513354 0.57895 0.447761
gnb_pooled
50 GS
0.520228 0.42105 0.619403
svm_linear
50 GS
0.388845 0.21053 0.567164
gnb_pooled
100 GS
0.512962 0.47368 0.552239
svm_linear
100 GS
0.535546 0.52632 0.544776
gnb_pooled
500 GS
0.630204 0.89474 0.365672
svm_linear
500 GS
0.532011 0.57895 0.485075
gnb_pooled 5000 GS
0.577965 0.89474 0.261194
svm_linear
5000 GS
0.562451 0.73684 0.38806
NEUROCON PPMI
gnb_pooled
10 bandpass
0.519049 0.10526 0.932836
NEUROCON PPMI
svm_linear
10 bandpass
0.437156 0.1579
0.716418
NEUROCON PPMI
gnb_pooled
50 bandpass
0.459348 0.10526 0.813433
NEUROCON PPMI
svm_linear
50 bandpass
0.462883 0.05263 0.873134
NEUROCON PPMI
gnb_pooled
100 bandpass
0.467203 0.21053 0.723881
NEUROCON PPMI
svm_linear
100 bandpass
0.440888 0.1579
0.723881
NEUROCON PPMI
gnb_pooled
500 bandpass
0.425766 0.10526 0.746269
NEUROCON PPMI
svm_linear
500 bandpass
0.481932 0.1579
0.80597
NEUROCON PPMI
gnb_pooled 5000 bandpass
0.414375 0.05263 0.776119
NEUROCON PPMI
svm_linear
5000 bandpass
0.497054 0.21053 0.783582
NEUROCON TaoWu
gnb_pooled
10 bandpass
0.475
NEUROCON TaoWu
svm_linear
10 bandpass
0.525
NEUROCON TaoWu
gnb_pooled
50 bandpass
0.475
NEUROCON TaoWu
svm_linear
50 bandpass
0.5
NEUROCON TaoWu
gnb_pooled
100 bandpass
0.525
NEUROCON TaoWu
svm_linear
100 bandpass
0.575
NEUROCON TaoWu
gnb_pooled
500 bandpass
0.475
NEUROCON TaoWu
svm_linear
500 bandpass
0.7
0.15
0.35
0.35
0.25
0.45
0.35
0.35
0.4
0.8
0.7
0.6
0.75
0.6
0.8
0.6
1
Supporting information
41
NEUROCON TaoWu
gnb_pooled 5000 bandpass
0.5
NEUROCON TaoWu
svm_linear
5000 bandpass
0.65
0.5
0.35
0.5
0.95
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
NEUROCON
gnb_pooled
10 bandpass
0.448626 0.19355 0.703704
NEUROCON
svm_linear
10 bandpass
0.464456 0.35484 0.574074
NEUROCON
gnb_pooled
50 bandpass
0.469235 0.29032 0.648148
NEUROCON
svm_linear
50 bandpass
0.432796 0.03226 0.833333
NEUROCON
gnb_pooled
100 bandpass
0.459976 0.29032 0.62963
NEUROCON
svm_linear
100 bandpass
0.483572 0.09677 0.87037
NEUROCON
gnb_pooled
500 bandpass
0.485364 0.32258 0.648148
NEUROCON
svm_linear
500 bandpass
0.488351 0.03226 0.944444
NEUROCON
gnb_pooled 5000 bandpass
0.476105 0.32258 0.62963
NEUROCON
svm_linear
5000 bandpass
0.481481 0
0.962963
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
TaoWu
gnb_pooled
10 bandpass
0.475
svm_linear
10 bandpass
0.55
gnb_pooled
50 bandpass
0.5
svm_linear
50 bandpass
0.475
gnb_pooled
100 bandpass
0.475
svm_linear
100 bandpass
0.45
gnb_pooled
500 bandpass
0.475
svm_linear
500 bandpass
0.525
gnb_pooled 5000 bandpass
svm_linear
5000 bandpass
0.45
0.55
0.2
0.5
0.4
0.15
0.45
0.2
0.4
0.1
0.45
0.1
0.75
0.6
0.6
0.8
0.5
0.7
0.55
0.95
0.45
1
NEUROCON
gnb_pooled
10 bandpass
0.543309 0.29032 0.796296
NEUROCON
svm_linear
10 bandpass
0.515532 0.29032 0.740741
NEUROCON
gnb_pooled
50 bandpass
0.515532 0.29032 0.740741
NEUROCON
svm_linear
50 bandpass
0.364994 0.32258 0.407407
NEUROCON
gnb_pooled
100 bandpass
0.499403 0.25807 0.740741
NEUROCON
svm_linear
100 bandpass
0.41129
0.32258 0.5
NEUROCON
gnb_pooled
500 bandpass
0.506272 0.29032 0.722222
NEUROCON
svm_linear
500 bandpass
0.506272 0.29032 0.722222
NEUROCON
gnb_pooled 5000 bandpass
0.503883 0.32258 0.685185
NEUROCON
svm_linear
5000 bandpass
0.490143 0.25807 0.722222
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
PPMI
gnb_pooled
10 bandpass
0.568932 0.47368 0.664179
svm_linear
10 bandpass
0.531422 0.42105 0.641791
gnb_pooled
50 bandpass
0.607031 0.68421 0.529851
svm_linear
50 bandpass
0.480754 0.84211 0.119403
gnb_pooled
100 bandpass
0.621956 0.68421 0.559701
svm_linear
100 bandpass
0.454635 0.84211 0.067164
gnb_pooled
500 bandpass
0.652003 0.73684 0.567164
svm_linear
500 bandpass
0.473095 0.78947 0.156716
gnb_pooled 5000 bandpass
0.633346 0.73684 0.529851
svm_linear
5000 bandpass
0.547722 0.78947 0.30597
Supporting information
42
Consensus NMF clustering
For a more direct graphical depiction of the heterogeneity of the functional connectomes of the PD
patient scans, we have applied consensus NMF clustering [Brunet et al., 2004] for a progressively
increasing number of clusters k=2,…,18 (Supplementary Fig 3). The Figure depicts the symmetric
consensus co-clustering matrices for the PD scans from the NEUROCON dataset. Note that besides
the consistent grouping of the replicate scan pairs for each patient, it is difficult to single out an
optimal number of clusters k.
Supplementary Fig 3. Consensus NMF clustering of functional connectomes of PD patient
scans from the NEUROCON dataset.
References
Brunet JP, Tamayo P, Golub TR, Mesirov JP. Metagenes and molecular pattern discovery using
matrix factorization. Proceedings of the national academy of sciences. 2004 Mar 23;101(12):4164-
9.
Supporting information
43
|
1606.08313 | 1 | 1606 | 2016-06-27T15:15:35 | Integrated Information and Metastability in Systems of Coupled Oscillators | [
"q-bio.NC"
] | It has been shown that sets of oscillators in a modular network can exhibit a rich variety of metastable chimera states, in which synchronisation and desynchronisation coexist. Independently, under the guise of integrated information theory, researchers have attempted to quantify the extent to which a complex dynamical system presents a balance of integrated and segregated activity. In this paper we bring these two areas of research together by showing that the system of oscillators in question exhibits a critical peak of integrated information that coincides with peaks in other measures such as metastability and coalition entropy. | q-bio.NC | q-bio |
Integrated Information and Metastability in Systems of Coupled Oscillators
Pedro A.M. Mediano,∗ Juan Carlos Farah, and Murray Shanahan
Department of Computing, Imperial College London
(Dated: June 28, 2016)
It has been shown that sets of oscillators in a modular network can exhibit a rich variety of
metastable chimera states, in which synchronisation and desynchronisation coexist. Independently,
under the guise of integrated information theory, researchers have attempted to quantify the extent
to which a complex dynamical system presents a balance of integrated and segregated activity. In
this paper we bring these two areas of research together by showing that the system of oscillators
in question exhibits a critical peak of integrated information that coincides with peaks in other
measures such as metastability and coalition entropy.
Keywords: Synchronisation, chimera states, metastability, integrated information
I.
INTRODUCTION
Systems of coupled oscillators are ubiquitous both in
nature and in the human-engineered environment, making
them of considerable scientific interest [1]. A variety of
mathematical models of such systems have been devised
and their synchronisation properties have been the sub-
ject of much study. Typical studies of this sort, such as
the classic work of Kuramoto [2], examine the conditions
under which the system converges on a stable state of
either full synchronisation or desynchronisation, perhaps
identifying an order parameter that determines a critical
phase transition from one state to the other. The collec-
tion of known attractors of such systems was enlarged
with the discovery of so-called chimera states, in which
the system of oscillators partitions into two stable subsets,
one of which is fully synchronised while the other remains
permanently desynchronised [3].
Although systems of coupled oscillators that converge
on stable states are both mathematically interesting and
scientifically relevant, they are by no means representative
of all real-world synchronisation phenomena. For exam-
ple, the brain exhibits synchronous rhythmic activity on
multiple spatial and temporal scales, but never settles into
a stable state. Although it enters chimera-like states of
high partial synchronisation, these are only temporary. A
system of coupled oscillators that continually moves from
one highly synchronised state to another under its own
intrinsic dynamics is said to be metastable. In [4], it was
shown that a modular network of phase-lagged Kuramoto
oscillators will exhibit metastable chimera states under
certain conditions. Variants of this model have since been
used to replicate the statistics of the brain under a variety
of conditions, including the resting state [5], cognitive
control [6], and anaesthesia [7].
In a separate line of enquiry, a number of researchers
have attempted to pin down the notion of dynamical
complexity. A system is said to have high dynamical com-
plexity if it exhibits a balance of integrated and segregated
activity, where a system's activity is integrated to the
∗ [email protected]
extent that its parts influence each other and segregated
to the extent that its parts act independently [8]. Promi-
nent among these attempts is the Integrated Information
Theory (IIT), originally proposed by Balduzzi and Tononi
[9], but extended by multiple authors ever since [10–12].
In the present paper, we connect these two lines of en-
quiry by demonstrating that modular networks of coupled
oscillators of the sort described in [4] not only exhibit
metastable chimera states, but also have high dynamical
complexity. Moreover, we show that measures of both
phenomena peak in the narrow critical region of the pa-
rameter space wherein the system is poised between order
and disorder, and IIT offers a rich picture of dynamical
complexity in this critical regime. To our knowledge, this
is the first description of a dynamical system in which the
three major complexity indicators of criticality, metasta-
bility, and integrated information all appear.
II. METHODS
We examine a system of coupled Kuramoto oscillators,
extensively used to study non-linear dynamics and syn-
chronisation processes [13]. We build upon the work of [4]
with a community-structured network of oscillators. The
network is composed of 8 communities of 32 oscillators
each, with every oscillator being coupled to all other os-
cillators in its community with probability 1 and to each
oscillator in the rest of the network with probability 1/32.
The state of each oscillator i is captured by its phase θi,
the evolution of which is governed by the equation
Kij sin (θj − θi − α) ,
(cid:88)
= ω +
(1)
dθi
dt
1
κ + 1
j
where ω is the natural frequency of each oscillator, κ is the
average degree of the network, K is the connectivity matrix
and α is a global phase lag. We set ω = 1 and κ = 63.
To reflect the community structure, the coupling between
two oscillators i, j is Kij = 0.6 if they are in the same
community or Kij = 0.4 otherwise. We tune the system by
modifying the value of the phase lag, parametrised by β =
π/2 − α. We note that the system is fully deterministic,
i.e. there is no noise injected in the dynamical equations.
A. Metastability
In this section we review the basic concepts behind
metastability and how it is quantified, following a similar
description to that of [4].
The building block of the dynamical quantities we study
in this article is the instantaneous synchronisation R, that
quantifies the dispersion in θ-space of a given set of oscil-
lators. In general, we denote as Rc(t) the instantaneous
synchronisation of a community c of oscillators at time t,
given by
Rc(t) = (cid:104)eiθj (t)(cid:105)j∈c .
(2)
To quantify metastability, we use the metastability index
λ, which is defined as the average temporal variance of
the synchrony of each community c, i.e.
λc = vart Rc(t)
λ = (cid:104)λc(cid:105)c .
(3a)
(3b)
Last, we also define global synchrony ξ as the average
across time and space of instantaneous synchrony,
ξ =(cid:12)(cid:12)(cid:10)Rc(t)(cid:11)
(cid:12)(cid:12) .
t,c
(4)
2
as a whole compared to when it is considered as the
sum of two subsystems M{1,2}. We refer to τ as the
integration timescale. In other words, ϕ evaluates how
much information is generated by the system but not
by the two subsystems alone. For a specific bipartition
B = {M 1, M 2}, the effective information of the system
beyond B is
ϕ[X; τ,B] = I(Xt−τ , Xt) − 2(cid:88)
I(M k
t−τ , M k
t ) .
(5)
k=1
The main idea behind the computation of Φ is to ex-
haustively search all possible partitions of the system and
calculate the effective information of each of them. Among
those we select the partition with lowest ϕ (under some
considerations, see below), termed the Minimum Informa-
tion Bipartition (MIB). Then, the integrated information
of the system is the effective information beyond its MIB.
Given the above expression for ϕ, Φ is defined as
Φ[X, τ ] = ϕ[X; τ,BMIB]
ϕ[X; τ,B]
BMIB = argB min
K(B) = min(cid:8)H(M 1), H(M 2)(cid:9) ,
K(B)
(6a)
(6b)
(6c)
According to Eq. (2), Rc (and therefore ξ) is bounded
in the [0, 1] interval. Rc(t) will be 1 if all oscillators in c
have the same phase at time t, and will be 0 if they are
maximally spread across the unit circle. This [0, 1] bound
on R allows us to place an upper bound on λ – assuming a
unimodal synchrony distribution, the maximum possible
value of λ is λmax = 1/9.
As defined in Eq. (3a), λc represents the size of the
fluctuations in the internal synchrony of a community.
A system that is either hypersynchronised or completely
desynchronised will have a very small λc, whereas one
whose elements fluctuate in and out of synchrony will
have a high λc. In other words, a system of oscillators
exhibits metastability if its elements remain in the vicinity
of a synchronised state without falling into such a state
permanently.
B.
Integrated Information
Although other information-theoretic quantities have
also been linked to complexity in a neuroscience context
[7, 14], we take integrated information Φ as the main
informational measure of complexity in our study [9].
There are more modern accounts of the theory [15, 16],
but the latest versions have not been as thoroughly studied
and are not amenable to easy estimation from time series
data. For these reasons, we focus on the methods in [10]
to empirically estimate Φ from an observed time series.
The building block of integrated information is effective
information, ϕ. Effective information quantifies how much
better a system X is at predicting its own future (or
decoding its own past) after a time τ when it is considered
where K is a normalisation factor to avoid biasing Φ to
excessively unbalanced bipartitions. Defined this way,
Φ can be understood as the minimum information loss
incurred by splitting the system into two subsystems.
It quantifies the collective emergent behaviour that is
present in the whole system but not in any bipartition.
III. RESULTS
We ran 1500 simulations with values of β distributed
uniformly at random in the range [0, 2π) using RK4 with
a stepsize of 0.05 for numerical integration. Each simula-
tion was run for 5 × 106 timesteps, of which the first 104
are discarded to avoid effects from transient states. All
information-theoretic measures are reported in bits.
We first study the system from a purely dynamical per-
spective, following the analysis in [4]. Global synchrony
and metastability are shown in Fig. 1. The first character-
istic we observe is that there are two well differentiated
dynamical regimes – one of hypersynchronisation and one
of complete desynchronisation, with strong metastability
appearing in the narrow transition bands between one
and the other.
It is in this transition region where the oscillators oper-
ate in a critical regime poised between order and disorder
and complex phenomena appear. As the system moves
from desynchronisation to full synchronisation there is a
sharp increase in metastability, followed by a smoother
decrease as the system becomes hypersynchronised. In
the region 0 < β < π/8, the system remains in a complex
equilibrium between an ordered and a disordered phase.
3
FIG. 1. Global synchrony and metastability for different phase
lags β for the whole [0, 2π) range (bottom) and around the
critical transition region (top). Rapid increase of metastability
marks the onset of the phase transition. Note the different β
ranges in both plots.
FIG. 2. Integrated information Φ and coalition entropy Hc in
the phase transition. Within the broad region between order
and disorder in which Hc rises there is a narrower band in
which complex spatiotemporal patterns generate high Φ.
A.
Information-theoretic analysis
One feature of Φ (and of any other information-theoretic
measure), compared to λ, is that it does not need to be cal-
culated directly from the state of the system. According
to its definition, Φ is substrate-agnostic, meaning that the
relevant quantity for the calculation of Φ is not the physi-
cal state of the system, but some informational state – the
configuration of the system that we consider to contain
information. Therefore, we must define an informational
state mapping, that extracts the information-bearing sym-
bols from the physical state of the system.
Although calculating Φ on the real-valued phases is
possible, for simplicity we choose the coalition configura-
tion of the system as the informational state, defined as
the set of communities that are highly internally synchro-
nised. To calculate the coalition configuration at time t
we calculate Rc(t) of each community and threshold it,
such that
(cid:40)
X c
t =
if Rc(t) > γ
1
0 otherwise.
We refer to γ as the coalition threshold. After calculat-
ing the coalitions, the history of the system is reduced to a
time series with 8 binary variables. Having a multivariate
discrete time series, it is now tractable to compute Φ. By
default, we use γ = 0.8 in all our analyses shown here.
As depicted in Fig. 2, Φ shows a similar behaviour
to λ – it peaks in the transition regions and shrinks in
the fully ordered and the fully disordered regimes. We
also compare Φ with arguably the simplest information-
theoretic measure – entropy H. The entropy of the state
of the network calculated on the coalitions Xt forms the
coalition entropy Hc.
Both Φ and Hc peak precisely at the same point. Al-
though both measures depend on the chosen coalition
threshold γ, the results are qualitatively the same for a
wide range of thresholds.
Although it peaks in the same region as λ and Hc,
we note that Φ reveals new properties of the system
by virtue of incorporating temporal information in its
definition. That is, to have a high Φ a system must
exhibit complex spatial and temporal patterns. We can
verify this by performing a random time-shuffle on the
time series. This shuffling leaves λ and Hc unaltered, as
they don't explicitly depend on time correlations, but has
a high impact on Φ, which shrinks to zero. This indicates
that Φ is sensitive to properties of the system that are
not reflected by other measures.
Furthermore, Φ can be used to investigate the behaviour
of the system at multiple timescales. Figure 3 shows the
behaviour of Φ for several values of τ , and compares it
with standard time-delayed mutual information (TDMI)
I(Xt−τ , Xt).
The first thing we note is that Φ and TDMI have oppo-
site trends with τ . TDMI decreases for longer timescales
while Φ increases. At short timescales the system is highly
predictable – thus the high TDMI – but this short-term
evolution does not involve any system-wide interaction
– thus the low Φ. Furthermore, at these timescales Φ is
negative, which can be interpreted as an indication of
redundancy [17] in the evolution of the system: the parts
share some information, such that they separately con-
tain more information about their past than the whole
system about its past. For larger τ TDMI decreases, as
the evolution of the system becomes harder to track using
the coalition configuration. Simultaneously, Φ becomes
higher, indicating that the remaining TDMI has a stronger
integrated component that is not accounted for by the
TDMI of the partitions of the system. Overall, we see a
00.20.40.600.020.04Metastabilityλ00.20.40.60.81Synchronyξλξ0π2π3π200.020.04Phaselagβλ00.20.40.60.81ξ00.10.20.30.40.500.511.522.5PhaselagβIntegratedInformationΦ02468CoalitionEntropyHcΦHc 4
Φ. This means that Φ is sensitive to more complicated
dynamic patterns than the other measures considered,
and is in that sense more discriminating.
We note that λ is a community-local quantity – that is,
the calculation of λc for each community is independent of
the rest. Conversely, Φ relies exclusively on the irreducible
interaction between communities. These two quantities
are nevertheless intrinsically related, insofar as internal
variability enables the system to visit a larger repertoire
of states in which system-wide interaction can take place.
B. Robustness of Φ against measurement noise
We will now consider the impact of measurement noise
on Φ, wherein the system runs unchanged but our record-
ing of it is imperfect. For this experiment we run the
(deterministic) simulation as presented in the previous
section and take the binary time series of coalition con-
figurations. We then emulate the effect of uncorrelated
measurement noise by flipping each bit in the time series
with probability p, yielding the corrupted time series X.
Finally we recalculate Φ on the corrupted time series,
and show the results in Fig. 4. To quantify how fast Φ
changes we calculate the ratio between the corrupted and
the original time series,
η =
Φ[ X, τ ]
Φ[X, τ ]
.
(7)
In order to avoid instabilities as Φ[X, τ ] gets close to
zero, we calculate η only in the region within 0.5 rad of
the centre of the peak, where Φ[X, τ ] is large. The inset
of Fig. 4 shows the mean and standard deviation of η at
different noise levels p.
FIG. 4. Integrated information Φ for different levels of mea-
surement noise p. Inset: (blue) Mean and variance of the ratio
η between Φ of the corrupted and the original time series.
(red) Exponential fit η = exp(−p/(cid:96)), with (cid:96) ≈ 0.04.
FIG. 3. Integrated information Φ and time-delayed mutual
information I(Xt−τ , Xt) for several timescales τ . See text for
details.
clear trend of TDMI diminishing at longer timescales but
becoming progressively more integrated in nature.
It might seem counterintuitive to the reader that both
TDMI and Φ converge to a non-zero value for arbitrarily
high τ . However, recall that the system is deterministic,
so there is no reason to expect TDMI to vanish even
in the τ → ∞ limit. Perfect knowledge of all of the
oscillators' phases at any given time step is enough to
reconstruct the whole history of the system. The reason
why we see a decreasing TDMI is because the informa-
tional state we chose (the coalition configuration) is not
descriptive enough to capture the evolution of the system
perfectly. That is, this is not a result of stochasticity, but
of degeneracy.
Put another way, the TDMI of the whole system re-
mains non-zero in the τ → ∞ limit because we are dealing
with a causally closed system. Again, if we considered
the totality of the system's state (i.e. the phases θi) then
TDMI would be maximal and constant for any τ .
In
contrast, when considering the TDMI of any partition M
we start dealing with a causally open system, since one
partition is affected by the other. This effectively intro-
duces stochasticity in our observations Mt, which does
cause TDMI of the partition to vanish when τ → ∞. This
explains that the TDMI of the whole system converges
to a non-zero value for large τ while the TDMI of any
partition fades to zero, leaving a positive Φ.
Finally, it is interesting to combine the insights from the
dynamical and information-theoretic analyses. Inspecting
Figs. 1 and 2 we see that the peak in Φ is much narrower
than the peaks in λ and Hc. While some values of β do
give rise to non-trivial dynamics, it is only at the centre
of the critical region that these dynamics give rise to
integration. A certain degree of internal variability is
necessary to establish integrated information, but not all
configurations with high internal variability lead to a high
00.10.20.30.40.50.01.02.0Φ[X,τ]τ=1τ=10τ=10000.10.20.30.40.50.02.04.06.0PhaselagβI(Xt−τ,Xt)00.10.20.30.40.500.511.52PhaselagβΦ[X,τ]p=0.00p=0.02p=0.04p=0.0600.050.10.150.200.51pη 5
We find that Φ monotonically decays with p, reflecting
the gradual loss of the precise spatiotemporal patterns
characteristic of the system. The distortion has a greater
effect on time series with greater Φ, but preserves the
dominant peak in β ≈ 0.15. The inset shows that both
the mean and variance of η decay as a clean exponential
with p. Φ is highly sensitive to noise and undergoes a
rapid decline, as a measurement noise of 5% wipes out
70% of the perceived integrated information of the system.
IV. CONCLUSION
We have presented a community-structured network
of Kuramoto oscillators and discussed their collective
behaviour in terms of metastability [4] and integrated
information [9]. We showed that the system undergoes
a phase transition whose critical region presents a sharp,
clear peak of integrated information Φ that coincides with
a strong increase in metastability. To our knowledge, this
is the first description of a dynamical system in which
the three major complexity indicators of criticality, high
metastability, and high integrated information all appear.
The resulting confluence of two major research directions
in complexity science suggests that this is a system that
merits further study.
In the context of the present model, the high internal
variability of the system's components enables system-
wide interaction, which in turn leads to high Φ. As we
have seen, the system presents a region of high metasta-
bility, but notably it is only within an even narrower band
that we find strong integrated information. Moreover, as
we have also seen, shuffling the time series data preserves
the peak of metastability, despite the fact that the result is
meaningless. By contrast, Φ only peaks when the relevant
temporal structure is present in the data. In this way we
provide evidence that complex dynamics – as quantified
by the metastability index λ – are a necessary but not
sufficient condition for complex information processing –
as quantified by integrated information Φ.
Dynamical and information-theoretic measures provide
different lenses through which we can understand a sys-
tem, and offer complementary views of its behaviour.
Our findings support the claim that Φ, despite having
some theoretical drawbacks [11], is a valuable tool for
understanding complex spatial and temporal behaviour
in dynamical systems, particularly when combined with
other analysis techniques.
[1] A. Pikovsky, M. Rosenblum, and J. Kurths, Synchro-
nization: A Universal Concept in Nonlinear Sciences
(Cambridge University Press, Cambridge, 2001) p. 432.
[2] Y. Kuramoto, Chemical Oscillations, Waves and Turbu-
lence (Dover Publications, 1984) p. 164.
[3] M. J. Panaggio and D. M. Abrams, Nonlinearity 28, R67
(2015), arXiv:1403.6204.
[4] M. Shanahan, Chaos 20, 013108 (2010), arXiv:0908.3881.
[5] J. Cabral, E. Hugues, O. Sporns, and G. Deco, NeuroIm-
age 57, 130 (2011).
[6] P. J. Hellyer, G. Scott, M. Shanahan, D. J. Sharp, and
R. Leech, The Journal of Neuroscience 35, 9050 (2015).
[7] M. Schartner, A. Seth, Q. Noirhomme, M. Boly, M.-A.
Bruno, S. Laureys, and A. B. Barrett, PloS One 10,
e0133532 (2015).
[8] M. Shanahan, Physical Review E 78, 041924 (2008).
[9] D. Balduzzi and G. Tononi, PLoS Computational Biology
4, e1000091 (2008).
[10] A. B. Barrett and A. K. Seth, PLoS Computational Biol-
ogy 7, e1001052 (2011).
[11] V. Griffith, arXiv:1401.0978.
[12] M. Tegmark, arXiv:1601.02626.
[13] D. M. Abrams, R. Mirollo, S. H. Strogatz, and D. A.
Wiley, Physical Review Letters 101, 084103 (2008).
[14] J.-R. King, J. D. Sitt, F. Faugeras, B. Rohaut, I. El
Karoui, L. Cohen, L. Naccache, and S. Dehaene, Current
Biology 23, 1914 (2013).
[15] M. Oizumi, L. Albantakis, and G. Tononi, PLoS Compu-
tational Biology 10, e1003588 (2014).
[16] G. Tononi, Archives Italiennes de Biologie 150, 56 (2012).
[17] A. B. Barrett, arXiv:1411.2832.
|
1608.08040 | 2 | 1608 | 2017-03-08T18:14:34 | Rules and mechanisms for efficient two-stage learning in neural circuits | [
"q-bio.NC",
"physics.bio-ph"
] | Trial-and-error learning requires evaluating variable actions and reinforcing successful variants. In songbirds, vocal exploration is induced by LMAN, the output of a basal ganglia-circuit that also contributes a corrective bias to the vocal output. This bias is gradually consolidated in RA, a motor cortex analogue downstream of LMAN. We develop a new model of such two-stage learning. Using stochastic gradient descent, we derive how the activity in 'tutor' circuits (e.g., LMAN) should match plasticity mechanisms in 'student' circuits (e.g., RA) to achieve efficient learning. We further describe a reinforcement learning framework through which the tutor can build its teaching signal. We show that mismatches between the tutor signal and the plasticity mechanism can impair learning. Applied to birdsong, our results predict the temporal structure of the corrective bias from LMAN given a plasticity rule in RA. Our framework can be applied predictively to other paired brain areas showing two-stage learning. | q-bio.NC | q-bio |
Rules and mechanisms for efficient two-stage learning in neural
circuits
Tiberiu Te¸sileanu1,2, Bence Olveczky3, and Vijay Balasubramanian1,2
1Initiative for the Theoretical Sciences, The Graduate Center, CUNY, New York, NY 10016
2David Rittenhouse Laboratories, University of Pennsylvania, Philadelphia, PA 19104
3Department of Organismic and Evolutionary Biology and Center for Brain Science, Harvard
University, Cambridge, MA 02138
November 9, 2018
Abstract
Trial-and-error learning requires evaluating variable actions and reinforcing successful variants. In
songbirds, vocal exploration is induced by LMAN, the output of a basal ganglia-circuit that also contributes
a corrective bias to the vocal output. This bias is gradually consolidated in RA, a motor cortex analogue
downstream of LMAN. We develop a new model of such two-stage learning. Using stochastic gradient
descent, we derive how the activity in 'tutor' circuits (e.g., LMAN) should match plasticity mechanisms
in 'student' circuits (e.g., RA) to achieve efficient learning. We further describe a reinforcement learning
framework through which the tutor can build its teaching signal. We show that mismatches between the
tutor signal and the plasticity mechanism can impair learning. Applied to birdsong, our results predict the
temporal structure of the corrective bias from LMAN given a plasticity rule in RA. Our framework can be
applied predictively to other paired brain areas showing two-stage learning.
1
Introduction
Two-stage learning has been described in a variety of different contexts and neural circuits. During hippocampal
memory consolidation, recent memories, that are dependent on the hippocampus, are transferred to the
neocortex for long-term storage (Frankland and Bontempi 2005). Similarly, the rat motor cortex provides
essential input to sub-cortical circuits during skill learning, but then becomes dispensable for executing certain
skills (Kawai et al. 2015). A paradigmatic example of two-stage learning occurs in songbirds learning their
courtship songs (Andalman and Fee 2009; Turner and Desmurget 2010; Warren et al. 2011). Zebra finches,
commonly used in birdsong research, learn their song from their fathers as juveniles and keep the same song
for life (Immelmann 1969).
The birdsong circuit has been extensively studied; see Figure 1A for an outline. Area HVC is a timebase
circuit, with projection neurons that fire sparse spike bursts in precise synchrony with the song (Hahnloser,
Kozhevnikov, and Fee 2002; Lynch et al. 2016; Picardo et al. 2016). A population of neurons from HVC
projects to the robust nucleus of the arcopallium (RA), a pre-motor area, which then projects to motor
neurons controlling respiratory and syringeal muscles (Leonardo and Fee 2005; Simpson and Vicario 1990; Yu
and Margoliash 1996). A second input to RA comes from the lateral magnocellular nucleus of the anterior
nidopallium (LMAN). Unlike HVC and RA activity patterns, LMAN spiking is highly variable across different
renditions of the song (Kao, Wright, and Doupe 2008; Olveczky, Andalman, and Fee 2005). LMAN is the
output of the anterior forebrain pathway, a circuit involving the song-specialized basal ganglia (Perkel 2004).
Because of the variability in its activity patterns, it was thought that LMAN's role was simply to inject
variability into the song ( Olveczky, Andalman, and Fee 2005). The resulting vocal experimentation would
enable reinforcement-based learning. For this reason, prior models tended to treat LMAN as a pure Poisson
noise generator, and assume that a reward signal is received directly in RA (Fiete, Fee, and Seung 2007). More
recent evidence, however, suggests that the reward signal reaches Area X, the song-specialized basal ganglia,
1
rather than RA (Gadagkar et al. 2016; Hoffmann et al. 2016; Kubikova and Kost´al 2010). Taken together with
the fact that LMAN firing patterns are not uniformly random, but rather contain a corrective bias guiding
plasticity in RA (Andalman and Fee 2009; Warren et al. 2011), this suggests that we should rethink our models
of song acquisition.
Here we build a general model of two-stage learning where one neural circuit "tutors" another. We develop
a formalism for determining how the teaching signal should be adapted to a specific plasticity rule, to best
instruct a student circuit to improve its performance at each learning step. We develop analytical results in
a rate based model, and show through simulations that the general findings carry over to realistic spiking
neurons. Applied to the vocal control circuit of songbirds, our model reproduces the observed changes in the
spiking statistics of RA neurons as juvenile birds learn their song. Our framework also predicts how the LMAN
signal should be adapted to properties of RA synapses. This prediction can be tested in future experiments.
Our approach separates the mechanistic question of how learning is implemented from what the resulting
learning rules are. We nevertheless demonstrate that a simple reinforcement learning algorithm suffices to
implement the learning rule we propose. Our framework makes general predictions for how instructive signals
are matched to plasticity rules whenever information is transferred between different brain regions.
Figure 1: Relation between the song system in zebra finches and our model. A. Diagram of the major brain
regions involved in birdsong. B. Conceptual model inspired by the birdsong system. The line from output
to tutor is dashed because the reinforcement signal can reach the tutor either directly or, as in songbirds,
indirectly. C. Plasticity rule measured in bird RA (measurement done in slice). When an HVC burst leads
an LMAN burst by about 100 ms, the HVC -- RA synapse is strengthened, while coincident firing leads to
suppression. Figure adapted from (Mehaffey and Doupe 2015). D. Plasticity rule in our model that mimics
the Mehaffey and Doupe (2015) rule.
2
motorHVCRALMANArea XDLMconductorstudenttutoroutputreinforcementABCD 2 Results
2.1 Model
We considered a model for information transfer that is composed of three sub-circuits: a conductor, a student,
and a tutor (see Figure 1B). The conductor provides input to the student in the form of temporally precise
patterns. The goal of learning is for the student to convert this input to a predefined output pattern. The tutor
provides a signal that guides plasticity at the conductor -- student synapses. For simplicity, we assumed that the
conductor always presents the input patterns in the same order, and without repetitions. This allowed us to
use the time t to label input patterns, making it easier to analyze the on-line learning rules that we studied.
This model of learning is based on the logic implemented by the vocal circuits of the songbird (Figure 1A).
Relating this to the songbird, the conductor is HVC, the student is RA, and the tutor is LMAN. The song can
be viewed as a mapping between clock-like HVC activity patterns and muscle-related RA outputs. The goal of
learning is to find a mapping that reproduces the tutor song.
Birdsong provides interesting insights into the role of variability in tutor signals. If we focus solely on
information transfer, the tutor output need not be variable; it can deterministically provide the best instructive
signal to guide the student. This, however, would require the tutor to have a detailed model of the student.
More realistically, the tutor might only have access to a scalar representation of how successful the student
rendition of the desired output is, perhaps in the form of a reward signal. A tutor in this case has to solve
the so-called 'credit assignment problem' -- it needs to identify which student neurons are responsible for the
reward. A standard way to achieve this is to inject variability in the student output and reinforcing the firing
of neurons that precede reward (see for example (Fiete, Fee, and Seung 2007) in the birdsong context). Thus,
in our model, the tutor has a dual role of providing both an instructive signal and variability, as in birdsong.
We described the output of our model using a vector ya(t) where a indexed the various output channels
(Figure 2A). In the context of motor control a might index the muscle to be controlled, or, more abstractly,
different features of the motor output, such as pitch and amplitude in the case of birdsong. The output ya(t)
was a function of the activity of the student neurons sj(t). The student neurons were in turn driven by the
activity of the conductor neurons ci(t). The student also received tutor signals to guide plasticity; in the
songbird, the guiding signals for each RA neuron come from several LMAN neurons (Canady et al. 1988;
Garst-Orozco, Babadi, and Olveczky 2014; Herrmann and Arnold 1991). In our model, we summarized the net
input from the tutor to the jth student neuron as a single function gj(t).
We started with a rate-based implementation of the model (Figure 2A) that was analytically tractable but
averaged over tutor variability. We further took the neurons to be in a linear operating regime (Figure 2A)
away from the threshold and saturation present in real neurons. We then relaxed these conditions and tested
our results in spiking networks with initial parameters selected to imitate measured firing patterns in juvenile
birds prior to song learning. The student circuit in both the rate-based and spiking models included a global
inhibitory signal that helped to suppress excess activity driven by ongoing conductor and tutor input. Such
recurrent inhibition is present in area RA of the bird (Spiro, Dalva, and Mooney 1999). In the spiking model
we implemented the suppression as an activity-dependent inhibition, while for the analytic calculations we
used a constant negative bias for the student neurons.
2.2 Learning in a rate-based model
Learning in our model was enabled by plasticity at the conductor -- student synapses that was modulated by
signals from tutor neurons (Figure 2B). Many different forms of such hetero-synaptic plasticity have been
observed. For example, in rate-based synaptic plasticity high tutor firing rates lead to synaptic potentiation
and low tutor firing rates lead to depression (Chistiakova, Bannon, et al. 2014; Chistiakova and Volgushev
2009). In timing-dependent rules, such as the one recently measured by Mehaffey and Doupe (2015) in slices of
zebra finch RA (see Figure 1C), the relative arrival times of spike bursts from different input pathways set
the sign of synaptic change. To model learning that lies between these rate and timing-based extremes, we
introduced a class of plasticity rules governed by two parameters α and β (see also Methods and Figure 2B):
(cid:21)
,
)/τ1 −
β
τ2
e
−(t−t(cid:48)
)/τ2
(1)
dWij
dt
ci(t) =
0
= ηci(t)(cid:0)gj(t) − θ(cid:1) ,
(cid:20) α
(cid:90) t
(cid:48)
(cid:48)
dt
ci(t
)
−(t−t(cid:48)
e
τ1
3
Figure 2: Schematic representation of our rate-based model. A. Conductor neurons fire precisely-timed bursts,
similar to HVC neurons in songbirds. Conductor and tutor activities, c(t) and g(t), provide excitation to
student neurons, which integrate these inputs and respond linearly, with activity s(t). Student neurons also
receive a constant inhibitory input, xinh. The output neurons linearly combine the activities from groups of
student neurons using weights Maj. The linearity assumptions were made for mathematical convenience but
are not essential for our qualitative results (see Appendix). B. The conductor -- student synaptic weights Wij
are updated based on a plasticity rule that depends on two parameters, α and β, and two timescales, τ1 and
τ2 (see eq. (1) and Methods). The tutor signal enters this rule as a deviation from a constant threshold θ.
The figure shows how synaptic weights change (∆W ) for a student neuron that receives a tutor burst and
a conductor burst separated by a short lag. Two different choices of plasticity parameters are illustrated in
the case when the threshold θ = 0. C. The amount of mismatch between the system's output and the target
output is quantified using a loss (error) function. The figure sketches the loss landscape obtained by varying
the synaptic weights Wij and calculating the loss function in each case (only two of the weight axes are shown).
The blue dot shows the lowest value of the loss function, corresponding to the best match between the motor
output and the target, while the orange dot shows the starting point. The dashed line shows how learning
would proceed in a gradient descent approach, where the weights change in the direction of steepest descent in
the loss landscape.
4
conductorstudent∑tutoroutput-400400-400400tutorconductorerrorABClosslandscapesynapticplasticity(loss function)convolution∑∑∑∑∑t (ms)t (ms)(ms)(ms) where Wij is the weight of the synapse from the ith conductor to the jth student neuron, η is a learning
rate, θ is a threshold on the firing rate of tutor neurons, and τ1 and τ2 are timescales associated with the
plasticity. This is similar to an STDP rule, except that the dependence on postsynaptic activity was replaced
by dependence on the input from the tutor. Thus plasticity acts heterosynaptically, with activation of the
tutor -- student synapse controlling the change in the conductor -- student synaptic weight. The timescales τ1
and τ2, as well as the coefficients α and β, can be thought of as effective parameters describing the plasticity
observed in student neurons. As such, they do not necessarily have a simple correspondence in terms of the
biochemistry of the plasticity mechanism, and the framework we describe here is not specifically tied to such
an interpretation.
If we set α or β to zero in our rule, eq. (1), the sign of the synaptic change is determined solely by the
firing rate of the tutor gj(t) as compared to a threshold, reproducing the rate rules observed in experiments.
When α/β ≈ 1, if the conductor leads the tutor, potentiation occurs, while coincident signals lead to depression
(Figure 2B), which mimics the empirical findings from (Mehaffey and Doupe 2015). For general α and β, the
sign of plasticity is controlled by both the firing rate of the tutor relative to the baseline, and by the relative
timing of tutor and conductor. The overall scale of the parameters α and β can be absorbed into the learning
rate η and so we set α − β = 1 in all our simulations without loss of generality (see Methods). Note that if α
and β are both large, it can be that α − β = 1 and α/β ≈ 1 also, as needed to realize the Mehaffey and Doupe
(2015) curve.
We can ask how the conductor -- student weights Wij (Figure 2A) should change in order to best improve
the output ya(t). We first need a loss function L that quantifies the distance between the current output ya(t)
and the target ¯ya(t) (Figure 2C). We used a quadratic loss function, but other choices can also be incorporated
into our framework (see Appendix). Learning should change the synaptic weights so that the loss function is
minimized, leading to a good rendition of the targeted output. This can be achieved by changing the synaptic
weights in the direction of steepest descent of the loss function (Figure 2C).
We used the synaptic plasticity rule from eq. (1) to calculate the overall change of the weights, ∆Wij, over
the course of the motor program. This is a function of the time course of the tutor signal, gj(t). Not every
choice for the tutor signal leads to motor output changes that best improve the match to the target. Imposing
the condition that these changes follow the gradient descent procedure described above, we derived the tutor
signal that was best matched to the student plasticity rule (detailed derivation in Methods). The result is that
the best tutor for driving gradient descent learning must keep track of the motor error
(cid:88)
j(t) =
Maj(ya(t) − ¯ya(t))
(2)
(3)
integrated over the recent past
gj(t) = θ −
a
ζ
1
(cid:90) t
α − β
τtutor
0
(cid:48)
j(t
)e
−(t−t(cid:48)
(cid:48)
)/τtutor dt
,
where Maj are the weights describing the linear relationship between student activities and motor outputs
(Figure 2A) and ζ is a learning rate. Moreover, for effective learning, the timescale τtutor appearing in eq. (3),
which quantifies the timescale on which error information is integrated into the tutor signal, should be related
to the synaptic plasticity parameters according to
τtutor = τ
∗
tutor ≡
is the optimal timescale for the error integration.
τ
∗
tutor , where
ατ1 − βτ2
α − β
(4)
In short, motor learning with a heterosynaptic plasticity rule requires convolving the motor error with a
kernel whose timescale is related to the structure of the plasticity rule, but is otherwise independent of the
motor program.1 As explained in more detail in Methods, this result is derived in an approximation that
assumes that the tutor signal does not vary significantly over timescales of the order of the student timescales
τ1 and τ2. Given eq. (4), this implies that we are assuming τtutor (cid:29) τ1,2. This is a reasonable approximation
because variations in the tutor signal that are much faster than the student timescales τ1,2 have little effect on
learning since the plasticity rule (1) blurs conductor inputs over these timescales.
1We thank the referees for suggesting this way of describing our results.
5
2.3 Matched vs. unmatched learning
Our rate-based model predicts that when the timescale on which error information is integrated into the tutor
signal (τtutor) is matched to the student plasticity rule as described above, learning will proceed efficiently.
A mismatched tutor should slow or disrupt convergence to the desired output. To test this, we numerically
simulated the birdsong circuit using the linear model from Figure 2A with a motor output ya filtered to more
realistically reflect muscle response times (see Methods). We selected plasticity rules as described in eq. (1) and
Figure 2B and picked a target output pattern to learn. The target was chosen to resemble recordings of air-sac
pressure from singing zebra finches in terms of smoothness and characteristic timescales (Veit, Aronov, and Fee
2011), but was otherwise arbitrary. In our simulations, the output typically involved two different channels,
each with its own target, but for brevity, in figures we typically showed the output from only one of these.
For our analytical calculations, we made a series of assumptions and approximations meant to enhance
tractability, such as linearity of the model and a focus on the regime τtutor (cid:29) τ1,2. These constraints can
be lifted in our simulations, and indeed below we test our numerical model in regimes that go beyond the
approximations made in our derivation. In many cases, we found that the basic findings regarding tutor -- student
matching from our analytical model remain true even when some of the assumptions we used to derive it no
longer hold.
We tested tutors that were matched or mismatched to the plasticity rule to see how effectively they
instructed the student. Figure 3A and online Video 1 show convergence with a matched tutor when the sign of
plasticity is determined by the tutor's firing rate. We see that the student output rapidly converged to the
target. Figure 3B and online Video 2 show convergence with a matched tutor when the sign of plasticity is
largely determined by the relative timing of the tutor signal and the student output. We see again that the
student converged steadily to the desired output, but at a somewhat slower rate than in Figure 3A.
To test the effects of mismatch between tutor and student, we used tutors with timescales that did not
match eq. (4). All student plasticity rules had the same effective time constants τ1 and τ2, but different
parameters α and β (see eq. (1)), subject to the constraint α − β = 1 described in section 2.2. Different tutors
had different memory time scales τtutor (eq. (3)). Figures 3C and 3D demonstrate that learning was more
rapid for well-matched tutor-student pairs (the diagonal neighborhood, where τtutor ≈ τ∗
tutor). When the tutor
error integration timescale was shorter than the matched value in eq. (4), τtutor < τ∗
tutor, learning was often
completely disrupted (many pairs below the diagonal in Figures 3C and 3D). When the tutor error integration
timescale was longer than the matched value in eq. (4), τtutor > τ∗
tutor learning was slowed down. Figure 3C
also shows that a certain amount of mismatch between the tutor error integration timescale τtutor and the
matched timescale τ∗
tutor implied by the student plasticity rule is tolerated by the system. Interestingly, the
diagonal band over which learning is effective in Figure 3C is roughly of constant width -- note that the scale on
both axes is logarithmic, so that this means that the tutor error integration timescale τtutor has to be within a
constant factor of the optimal timescale τ∗
tutor for good learning. We also see that the breakdown in learning is
more abrupt when τtutor < τ∗
tutor than in the opposite regime.
An interesting feature of the results from Figures 3C, 3D is that the difference in performance between
matched and mismatched pairs becomes less pronounced for timescales shorter than about 100 ms. This is due
to the fact that the plasticity rule (eq. (1)) implicitly smooths over timescales of the order of τ1,2, which in our
simulations were equal to τ1 = 80 ms, τ2 = 40 ms. Thus, variations of the tutor signal on shorter timescales
have little effect on learning. Using different values for the effective timescales τ1,2 describing the plasticity
rule can increase or decrease the range of parameters over which learning is robust against tutor -- student
mismatches (see Appendix).
2.4 Robust learning with nonlinearities
In the model above, firing rates for the tutor were allowed to grow as large as necessary to implement the most
efficient learning. However, the firing rates of realistic neurons typically saturate at some fixed bound. To
test the effects of this nonlinearity in the tutor, we passed the ideal tutor activity (3) through a sigmoidal
nonlinearity,
(cid:90) t
6
gj(t) = θ − ρ tanh
α − β
τtutor
0
ζ
1
(cid:48)
j(t
−(t−t(cid:48)
)e
(cid:48)
)/τtutor dt
.
(5)
where 2ρ is the range of firing rates. We typically chose θ = ρ = 80 Hz to constrain the rates to the range
0 -- 160 Hz (Garst-Orozco, Babadi, and Olveczky 2014; Olveczky, Andalman, and Fee 2005). Learning slowed
Figure 3: Learning with matched or mismatched tutors in rate-based simulations. A. Error trace showing
how the average motor error evolved with the number of repetitions of the motor program for a rate-based
(α = 0) plasticity rule paired with a matching tutor. (See online Video 1.) B. The error trace and final motor
output shown for a timing-based learning rule matched by a tutor with a long integration timescale. (See
online Video 2.) In both A and B the inset shows the final motor output for one of the two output channels
(thick orange line) compared to the target output for that channel (dotted black line). The output on the first
rendition and at two other stages of learning indicated by orange arrows on the error trace are also shown as
thin orange lines. C. Effects of mismatch between student and tutor on reproduction accuracy. The heatmap
shows the final reproduction error of the motor output after 1000 learning cycles in a rate-based simulation
where a student with parameters α, β, τ1, and τ2 was paired with a tutor with memory timescale τtutor. On
the y axis, τ1 and τ2 were kept fixed at 80 ms and 40 ms, respectively, while α and β were varied (subject to
the constraint α − β = 1; see text). Different choices of α and β lead to different optimal timescales τ∗
tutor
according to eq. (4). The diagonal elements correspond to matched tutor and student, τtutor = τ∗
tutor. Note
that the color scale is logarithmic. D. Error evolution curves as a function of the mismatch between student
and tutor. Each plot shows how the error in the motor program changed during 1000 learning cycles for the
same conditions as those shown in the heatmap. The region shaded in light pink shows simulations where the
mismatch between student and tutor led to a deteriorating instead of improving performance during learning.
Video 1: Evolution of motor output during learning in a rate-based simulation using a rate-based (α = 0)
plasticity rule paired with a matching tutor. This video relates to Figure 3A.
Video 2: Evolution of motor output during learning in a rate-based simulation using a timing-based (α ≈ β)
plasticity rule paired with a matching tutor. This video relates to Figure 3B.
7
102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β0.501.002.005.0010.00050100150200250repetitions02468101214errorα=7.0,β=6.0,τtutor=320.0time010203040506070outputtargetoutput050100150200250repetitions02468101214errorα=0.0,β=−1.0,τtutor=40.0time01020304050607080outputtargetoutput0600ABCD0600worsebetter down with this change (Figure 4A and online Video 3) as a result of the tutor firing rates saturating when
the mismatch between the motor output and the target output was large. However, the accuracy of the final
rendition was not affected by saturation in the tutor (Figure 4A, inset). An interesting effect occurred when
the firing rate constraint was imposed on a matched tutor with a long memory timescale. When this happened
and the motor error was large, the tutor signal saturated and stopped growing in relation to the motor error
before the end of the motor program. In the extreme case of very long integration timescales, learning became
sequential: early features in the output were learned first, before later features were addressed, as in Figure 4B
and online Video 4. This is reminiscent of the learning rule described in (Memmesheimer et al. 2014).
Figure 4: Effects of adding a constraint on the tutor firing rate to the simulations. A. Learning was slowed
down by the firing rate constraint, but the accuracy of the final rendition stayed the same (inset, shown here for
one of two simulated output channels). Here α = 0, β = −1, and τtutor = τ∗
tutor = 40 ms. (See online Video 3.)
B. Sequential learning occurred when the firing rate constraint was imposed on a matched tutor with a long
memory scale. The plots show the evolution of the motor output for one of the two channels that were used in
the simulation. Here α = 24, β = 23, and τtutor = τ∗
tutor = 1000 ms. (See online Video 4.)
Video 3: Effects of adding a constraint on tutor firing rates on the evolution of motor output during learning
in a rate-based simulation. The plasticity rule here was rate-based (α = 0). This video relates to Figure 4A.
Video 4: Evolution of the motor output showing sequential learning in a rate-based simulation, which occurs
when the firing rate constraint is imposed on a tutor with a long memory timescale. This video relates to
Figure 4B.
Nonlinearities can similarly affect the activities of student neurons. Our model can be readily extended
to describe efficient learning even in this case. The key result is that for efficient learning to occur, the
synaptic plasticity rule should depend not just on the tutor and conductor, but also on the activity of the
postsynaptic student neurons (details in Appendix). Such dependence on postsynaptic activity is commonly
seen in experiments (Chistiakova, Bannon, et al. 2014; Chistiakova and Volgushev 2009).
The relation between student neuron activations sj(t) and motor outputs ya(t) (Figure 2A) is in general
also nonlinear. Compared to the linear assumption that we used, the effect of a monotonic nonlinearity,
8
0600time020406080Stage1targetoutput0600time020406080Stage17targetoutput0600time020406080Stage34targetoutput050100150200250repetition05101520errortimeoutputtargetnoconstraintwithconstraintAB0600 ya = Na((cid:80)
j Majsj), with Na an increasing function, is similar to modifying the loss function L, and does not
significantly change our results (see Appendix). We also checked that imposing a rectification constraint that
conductor -- student weights Wij must be positive does not modify our results either (see Appendix). This shows
that our model continues to work with biologically realistic synapses that cannot change sign from excitatory
to inhibitory during learning.
2.5 Spiking neurons and birdsong
To apply our model to vocal learning in birds, we extended our analysis to networks of spiking neurons.
Juvenile songbirds produce a "babble" that converges through learning to an adult song strongly resembling
the tutor song. This is reflected in the song-aligned spiking patterns in pre-motor area RA, which become
more stereotyped and cluster in shorter, better-defined bursts as the bird matures (Figure 5A). We tested
whether our model could reproduce key statistics of spiking in RA over the course of song learning. In this
context, our theory of efficient learning, derived in a rate-based scenario, predicts a specific relation between
the teaching signal embedded in LMAN firing patterns, and the plasticity rule implemented in RA. We tested
whether these predictions continued to hold in the spiking context.
Following the experiments of Hahnloser, Kozhevnikov, and Fee (2002), we modeled each neuron in HVC
(the conductor) as firing one short, precisely timed burst of 5-6 spikes at a single moment in the motor program.
Thus the population of HVC neurons produced a precise timebase for the song. LMAN (tutor) neurons are
known to have highly variable firing patterns that facilitate experimentation, but also contain a corrective
bias (Andalman and Fee 2009). Thus we modeled LMAN as producing inhomogeneous Poisson spike trains
with a time-dependent firing rate given by eq. (5) in our model. Although biologically there are several LMAN
neurons projecting to each RA neuron, we again simplified by "summing" the LMAN inputs into a single,
effective tutor neuron, similarly to the approach in (Fiete, Fee, and Seung 2007). The LMAN-RA synapses
were modeled in a current-based approach as a mixture of AMPA and NMDA receptors, following the songbird
data (Garst-Orozco, Babadi, and Olveczky 2014; Stark and Perkel 1999). The initial weights for all synapses
were tuned to produce RA firing patterns resembling juvenile birds ( Olveczky, Otchy, et al. 2011), subject
to constraints from direct measurements in slice recordings (Garst-Orozco, Babadi, and Olveczky 2014) (see
Methods for details, and Figure 5B for a comparison between neural recordings and spiking in our model). In
contrast to the constant inhibitory bias that we used in our rate-based simulations, for the spiking simulations
we chose an activity-dependent global inhibition for RA neurons. We also tested that a constant bias produced
similar results (see Appendix).
Synaptic strength updates followed the same two-timescale dynamics that was used in the rate-based models
(Figure 2B). The firing rates ci(t) and gj(t) that appear in the plasticity equation were calculated in the spiking
model by filtering the spike trains from conductor and tutor neurons with exponential kernels. The synaptic
weights were constrained to be non-negative. (See Methods for details.)
As long as the tutor error integration timescale was not too large, learning proceeded effectively when
the tutor error integration timescale and the student plasticity rule were matched (see Figure 5C and online
Video 5), with mismatches slowing down or abolishing learning, just as in our rate-based study (compare
Figure 5D with Figure 3C). The rate of learning and the accuracy of the trained state were lower in the
spiking model compared to the rate-based model. The lower accuracy arises because the tutor neurons fire
stochastically, unlike the deterministic neurons used in the rate-based simulations. The stochastic nature of the
tutor firing also led to a decrease in learning accuracy as the tutor error integration timescale τtutor increased
(Figure 5D). This happens through two related effects: (1) the signal-to-noise ratio in the tutor guiding signal
decreases as τtutor increases once the tutor error integration timescale is longer than the duration T of the
motor program (see Appendix); and (2) the fluctuations in the conductor -- student weights lead to some weights
getting clamped at 0 due to the positivity constraint, which leads to the motor program overshooting the target
(see Appendix). The latter effect can be reduced by either allowing for negative weights, or changing the motor
output to a push-pull architecture in which some student neurons enhance the output while others inhibit it.
The signal-to-noise ratio effect can be attenuated by increasing the gain of the tutor signal, which inhibits
early learning, but improves the quality of the guiding signal in the latter stages of the learning process. It is
also worth emphasizing that these effects only become relevant once the tutor error integration timescale τtutor
becomes significantly longer than the duration of the motor program, T , which for a birdsong motif would be
around 1 second.
Spiking in our model tends to be a little more regular than that in the recordings (compare Figure 5A
9
Figure 5: Results from simulations in spiking neural networks. A. Spike patterns recorded from zebra finch
RA during song production, for a juvenile (top) and an adult (bottom). Each color corresponds to a single
neuron, and the song-aligned spikes for six renditions of the song are shown. Adapted from ( Olveczky, Otchy,
et al. 2011). B. Spike patterns from model student neurons in our simulations, for the untrained (top) and
trained (bottom) models. The training used α = 1, β = 0, and τtutor = 80 ms, and ran for 600 iterations of the
song. Each model neuron corresponds to a different output channel of the simulation. In this case, the targets
for each channel were chosen to roughly approximate the time course observed in the neural recordings. C.
Progression of reproduction error in the spiking simulation as a function of the number of repetitions for the
same conditions as in panel B. The inset shows the accuracy of reproduction in the trained model for one of
the output channels. (See online Video 5.) D. Effects of mismatch between student and tutor on reproduction
accuracy in the spiking model. The heatmap shows the final reproduction error of the motor output after 1000
learning cycles in a spiking simulation where a student with parameters α, β, τ1, and τ2 was paired with a
tutor with memory timescale τtutor. On the y axis, τ1 and τ2 were kept fixed at 80 ms and 40 ms, respectively,
while α and β were varied (subject to the constraint α − β = 1; see section 2.2). Different choices of α and β
lead to different optimal timescales τ∗
tutor according to eq. (4). The diagonal elements correspond to matched
tutor and student, τtutor = τ∗
tutor. Note that the color scale is logarithmic.
Video 5: Evolution of motor output during learning in a spiking simulation. The plasticity rule parameters were
α = 1, β = 0, and the tutor had a matching timescale τtutor = τ∗
tutor = 80 ms. This video relates to Figure 5C.
10
102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β0.501.002.005.0010.000100200300400500600repetitions02468101214errorα=1.0,β=0.0,τtutor=80.0time0102030405060708090outputtargetoutput06000100200300400500600t(ms)0100200300400500600juvenileadultjuvenileadultABCD01002003004005000100200300400500t(ms)worsebetter and Figure 5B). This could be due to sources of noise that are present in the brain which we did not model.
One detail that our model does not capture is the fact that many LMAN spikes occur in bursts, while in
our simulation LMAN firing is Poisson. Bursts are more likely to produce spikes in downstream RA neurons
particularly because of the NMDA dynamics, and thus a bursty LMAN will be more effective at injecting
variability into RA (Kojima, Kao, and Doupe 2013). Small inaccuracies in aligning the recorded spikes to the
song are also likely to contribute apparent variability between renditions in the experiment. Indeed, some of
the variability in Figure 5A looks like it could be due to time warping and global time shifts that were not
fully corrected.
2.6 Robust learning with credit assignment errors
The calculation of the tutor output in our rule involved estimating the motor error j from eq. (2). This
required knowledge of the assignment between student activities and motor output, which in our model was
represented by the matrix Maj (Figure 2A). In our simulations, we typically chose an assignment in which
each student neuron contributed to a single output channel, mimicking the empirical findings for neurons
in bird RA. Mathematically, this implies that each column of Maj contained a single non-zero element. In
Figure 6A, we show what happened in the rate-based model when the tutor incorrectly assigned a certain
fraction of the neurons to the wrong output. Specifically, we considered two output channels, y1 and y2,
with half of the student neurons contributing only to y1 and the other half contributing only to y2. We then
scrambled a fraction ρ of this assignment when calculating the motor error, so that the tutor effectively had an
imperfect knowledge of the student -- output relation. Figure 6A shows that learning is robust to this kind of
mis-assignment even for fairly large values of the error fraction ρ up to about 40%, but quickly deteriorates as
this fraction approaches 50%.
Due to environmental factors that affect development of different individuals in different ways, it is unlikely
that the student -- output mapping can be innate. As such, the tutor circuit must learn the mapping. Indeed,
it is known that LMAN in the bird receives an indirect evaluation signal via Area X, which might be used
to effect this learning (Andalman and Fee 2009; Gadagkar et al. 2016; Hoffmann et al. 2016; Kubikova and
Kost´al 2010). One way in which this can be achieved is through a reinforcement paradigm. We thus considered
a learning rule where the tutor circuit receives a reward signal that enables it to infer the student -- output
mapping. In general the output of the tutor circuit should depend on an integral of the motor error, as in
eq. (3), to best instruct the student. For simplicity, we start with the memory-less case, τtutor = 0, in which
only the instantaneous value of the motor error is reflected in the tutor signal; we then show how to generalize
this for τtutor > 0.
As before, we took the tutor neurons to fire Poisson spikes with time-dependent rates fj(t), which were
initialized arbitrarily. Because of stochastic fluctuations, the actual tutor activity on any given trial, gj(t),
differs somewhat from the average, ¯gj(t). Denoting the difference by ξj(t) = gj(t) − ¯gj(t), the update rule for
the tutor firing rates was given by
(6)
where ηtutor is a learning rate, R(t) is the instantaneous reward signal, and ¯R is its average over recent renditions
of the motor program. In our implementation, ¯R is obtained by convolving R(t) with an exponential kernel
(timescale = 1 second). The reward R(tmax) at the end of one rendition becomes the baseline at the start
of the next rendition R(0). The baseline ¯gj(t) of the tutor activity is calculated by averaging over recent
renditions of the song with exponentially decaying weights (one e-fold of decay for every 5 renditions). Further
implementation details are available in our code at https://github.com/ttesileanu/twostagelearning.
∆fj(t) = ηtutor(R(t) − ¯R)ξj(t) ,
The intuition behind this rule is that, whenever a fluctuation in the tutor activity leads to better-than-
average reward (R(t) > ¯R), the tutor firing rate changes in the direction of the fluctuation for subsequent trials,
"freezing in" the improvement. Conversely, the firing rate moves away from the directions in which fluctuations
tend to reduce the reward.
To test our learning rule, we ran simulations using this reinforcement strategy and found that learning again
converges to an accurate rendition of the target output (Figure 6B, inset and online Video 6). The number
of repetitions needed for training is greatly increased compared to the case in which the credit assignment is
assumed known by the tutor circuit (compare Figure 6B to Figure 5C). This is because the tutor needs to use
many training rounds for experimentation before it can guide the conductor -- student plasticity. The rate of
learning in our model is similar to the songbird (i.e., order 10 000 repetitions for learning, given that a zebra
11
Figure 6: Credit assignment and reinforcement learning. A. Effects of credit mis-assignment on learning
in a rate-based simulation. Here, the system learned output sequences for two independent channels. The
student -- output weights Maj were chosen so that the tutor wrongly assigned a fraction of student neurons to
an output channel different from the one it actually mapped to. The graph shows how the accuracy of the
motor output after 1000 learning steps depended on the fraction of mis-assigned credit. B. Learning curve
and trained motor output (inset) for one of the channels showing two-stage reinforcement-based learning for
the memory-less tutor (τtutor = 0). The accuracy of the trained model is as good as in the case where the
tutor was assumed to have a perfect model of the student -- output relation. However, the speed of learning
is reduced. (See online Video 6.) C. Learning curve and trained motor output (inset) for one of the output
channels showing two-stage reinforcement-based learning when the tutor circuit needs to integrate information
about the motor error on a certain timescale. Again, learning was slow, but the accuracy of the trained state
was unchanged. (See online Video 7.) D. Evolution of the average number of HVC inputs per RA neuron with
learning in a reinforcement example. Synapses were considered pruned if they admitted a current smaller than
1 nA after a pre-synaptic spike in our simulations.
Video 6: Evolution of motor output during learning in a spiking simulation with a reinforcement-based tutor.
Here the tutor was memory-less (τtutor = 0). This video relates to Figure 6B.
Video 7: Evolution of motor output during learning in a spiking simulation with a reinforcement-based tutor.
Here the tutor needed to integrate information about the motor error on a timescale τtutor = 440 ms. This
video relates to Figure 6C.
12
0200040006000800010000repetitions050100150200conductor inputsperstudent neuron0.00.10.20.30.40.5fractionmismatch0246810finalerror02000400060008000repetitions02468101214errorα=10.0,β=9.0,τtutor=440.0time020406080outputtargetoutput02000400060008000repetitions02468101214errorα=1.0,β=0.0,τtutor=0.0time020406080outputtargetoutput0600ABCD0600 finch typically sings about 1000 repetitions of its song each day, and takes about one month to fully develop
adult song).
Because of the extra training time needed for the tutor to adapt its signal, the motor output in our
reward-based simulations tends to initially overshoot the target (leading to the kink in the error at around
2000 repetitions in Figure 6B). Interestingly, the subsequent reduction in output that leads to convergence
of the motor program, combined with the positivity constraint on the synaptic strengths, leads to many
conductor -- student connections being pruned (Figure 6D). This mirrors experiments on songbirds, where the
number of connections between HVC and RA first increases with learning and then decreases (Garst-Orozco,
Babadi, and Olveczky 2014).
The reinforcement rule described above responds only to instantaneous values of the reward signal and
tutor firing rate fluctuations. In general, effective learning requires that the tutor keep a memory trace of its
activity over a timescale τtutor > 0, as in eq. (4). To achieve this in the reinforcement paradigm, we can use a
simple generalization of eq. (6) where the update rule is filtered over the tutor memory timescale:
(cid:90) t
∆fj(t) = ηtutor
1
τtutor
(cid:48)
dt
(R(t
(cid:48)
(cid:48)
) − ¯R)ξj(t
)e
−(t−t(cid:48)
)/τtutor .
(7)
We tested that this rule leads to effective learning when paired with the corresponding student, i.e., one for
which eq. (4) is obeyed (Figure 6C and online Video 7).
The reinforcement rules proposed here are related to the learning rules from (Fiete, Fee, and Seung 2007;
Fiete and Seung 2006) and (Farries and Fairhall 2007). However, those models focused on learning in a single
pass, instead of the two-stage architecture that we studied. In particular, in (Fiete, Fee, and Seung 2007), area
LMAN was assumed to generate pure Poisson noise and reinforcement learning took place at the HVC -- RA
synapses. In our model, which is in better agreement with recent evidence regarding the roles of RA and
LMAN in birdsong (Andalman and Fee 2009), reinforcement learning first takes place in the anterior forebrain
pathway (AFP), for which LMAN is the output. A reward-independent heterosynaptic plasticity rule then
solidifies the information in RA.
In our simulations, tutor neurons fire Poisson spikes with specific time-dependent rates which change during
learning. The timecourse of the firing rates in each repetition must then be stored somewhere in the brain.
In fact, in the songbird, there are indirect projections from HVC to LMAN, going through the basal ganglia
(Area X) and the dorso-lateral division of the medial thalamus (DLM) in the anterior forebrain pathway
(Figure 1A) (Perkel 2004). These synapses could store the required time-dependence of the tutor firing rates.
In addition, the same synapses can provide the timebase input that would ensure synchrony between LMAN
firing and RA output, as necessary for learning. Our reinforcement learning rule for the tutor area, eq. (6), can
be viewed as an effective model for plasticity in the projections between HVC, Area X, DLM, and LMAN,
as in (Fee and Goldberg 2011). In this picture, the indirect HVC -- LMAN connections behave somewhat like
the "hedonistic synapses" from (Seung 2003), though we use a simpler synaptic model here. Implementing
the integral from eq. (7) would require further recurrent circuitry in LMAN which is beyond the scope of this
paper, but would be interesting to investigate in future work.
3 Discussion
We built a two-stage model of learning in which one area (the student) learns to perform a sequence of actions
under guidance from a tutor area. This architecture is inspired by the song system of zebra finches, where area
LMAN provides a corrective bias to the song that is then consolidated in the HVC -- RA synapses. Using an
approach rooted in the efficient coding literature, we showed analytically that, in a simple model, the tutor
output that is most likely to lead to effective learning by the student involves an integral over the recent
magnitude of the motor error. We found that efficiency requires that the timescale for this integral should be
related to the synaptic plasticity rule used by the student. Using simulations, we tested our findings in more
general settings. In particular, we demonstrated that tutor-student matching is important for learning in a
spiking-neuron model constructed to reproduce spiking patterns similar to those measured in zebra finches.
Learning in this model changes the spiking statistics of student neurons in realistic ways, for example, by
producing more bursty, stereotyped firing events as learning progresses. Finally, we showed how the tutor can
build its error-correcting signal by means of reinforcement learning.
13
If the birdsong system supports efficient learning, our model can predict the temporal structure of the firing
patterns of RA-projecting LMAN neurons, given the plasticity rule implemented at the HVC -- RA synapses.
These predictions can be directly tested by recordings from LMAN neurons in singing birds, assuming that a
good measure of motor error is available, and that we can estimate how the neurons contribute to this error.
Moreover, recordings from a tutor circuit, such as LMAN, could be combined with a measure of motor error
to infer the plasticity rule in a downstream student circuit, such as RA. This could be compared with direct
measurements of the plasticity rule obtained in slice. Conversely, knowledge of the student plasticity rule could
be used to predict the time-dependence of tutor firing rates. According to our model, the firing rate should
reflect the integral of the motor error with the timescale predicted by the model. A different approach would
be to artificially tutor RA by stimulating LMAN neurons electrically or optogenetically. We would predict
that if the tutor signal is delivered appropriately (e.g., in conjunction with a particular syllable (Tumer and
Brainard 2007)), then the premotor bias produced by the stimulation should become incorporated into the
motor pathway faster when the timescale of the artificial LMAN signal is properly matched to the RA synaptic
plasticity rule.
Our model can be applied more generally to other systems in the brain exhibiting two-stage learning,
such as motor learning in mammals. If the plasticity mechanisms in these systems are different from those in
songbirds, our predictions for the structure of the guiding signal will vary correspondingly. This would allow
a further test of our model of "efficient learning" in the brain. It is worth pointing out that our model was
derived assuming a certain hierarchy among the timescales that model the student plasticity and the tutor
signal. A mismatch between the model predictions and observations could also imply a breakdown of these
approximations, rather than failure of the hypothesis that the particular system under study evolved to support
efficient learning. Of course our analysis could be extended by relaxing these assumptions, for example by
keeping more terms in the Taylor expansion that we used in our derivation of the matched tutor signal.
Applied to birdsong, our model is best seen as a mechanism for learning song syllables. The ordering of
syllables in song motifs seems to have a second level of control within HVC and perhaps beyond (Basista et al.
2014; Hamaguchi, Tanaka, and Mooney 2016). Songs can also be distorted by warping their timebase through
changes in HVC firing without alterations of the HVC -- RA connectivity (Ali et al. 2013). In view of these
phenomena, it would be interesting to incorporate our model into a larger hierarchical framework in which the
sequencing and temporal structure of the syllables are also learned. A model of transitions between syllables
can be found in (Doya and Sejnowski 2000), where the authors use a "weight perturbation" optimization
scheme in which each HVC -- RA synaptic weight is perturbed individually. We did not follow this approach
because there is no plausible mechanism for LMAN to provide separate guidance to each HVC -- RA synapse; in
particular, there are not enough LMAN neurons (Fiete, Fee, and Seung 2007).
In this paper we assumed a two-stage architecture for learning, inspired by birdsong. An interesting question
is whether and under what conditions such an architecture is more effective than a single-step model. Possibly,
having two stages is better when a single tutor area is responsible for training several different dedicated
controllers, as is likely the case in motor learning. It would then be beneficial to have an area that can learn
arbitrary behaviors, perhaps at the cost of using more resources and having slower reaction times, along with
the ability to transfer these behaviors into low-level circuitry that is only capable of producing stereotyped
motor programs. The question then arises whether having more than two levels in this hierarchy could be
useful, what the other levels might do, and whether such hierarchical learning systems are implemented in the
brain.
Acknowledgments
We would like to thank Serena Bradde for fruitful discussions during the early stages of this work. We also thank
Xuexin Wei and Christopher Glaze for useful discussions. We are grateful to Timothy Otchy for providing us
with some of the data we used in this paper. During this work VB was supported by NSF grant PHY-1066293
at the Aspen Center for Physics and by NSF Physics of Living Systems grant PHY-1058202. TT was supported
by the Swartz Foundation.
14
A Methods
A.1 Equations for rate-based model
The basic equations we used for describing our rate-based model (Figure 2A) are the following:
ya(t) =
Majsj(t) ,
sj(t) =
Wijci(t) + wgj(t) − xinh .
j
(cid:88)
(cid:88)
(cid:88)
i
(cid:90) t
(A.1)
(A.2)
In simulations, we further filtered the output using an exponential kernel,
ya(t) =
Maj
(cid:48)
sj(t
) e
−(t−t(cid:48)
(cid:48)
)/τout dt
,
j
0
with a timescale τout that we typically set to 25 ms. The smoothing produces more realistic outputs by
mimicking the relatively slow reaction time of real muscles, and stabilizes learning by filtering out high-
frequency components of the motor output. The latter interfere with learning because of the delay between the
effect of conductor activity on synaptic strengths vs. motor output. This delay is of the order τ1,2 − τout (see
the plasticity rule below).
The conductor activity in the rate-based model is modeled after songbird HVC (Hahnloser, Kozhevnikov,
and Fee 2002): each neuron fires a single burst during the motor program. Each burst corresponds to a sharp
increase of the firing rate ci(t) from 0 to a constant value, and then a decrease 10 ms later. The activities of
the different neurons are spread out to tile the whole duration of the output program. Other choices for the
conductor activity also work, provided no patterns are repeated (see Appendix).
A.2 Mathematical description of plasticity rule
In our model the rate of change of the synaptic weights obeys a rule that depends on a filtered version of the
conductor signal (see Figure 2B). This is expressed mathematically as
where η is a learning rate and ci = K ∗ ci, with the star representing convolution and K being a filtering kernel.
We considered a linear combination of two exponential kernels with timescales τ1 and τ2,
dWij
dt
= η ci(t) (gj(t) − θ) ,
(A.3)
with Ki(t) given by
Ki(t) =
K(t) = αK1(t) − βK2(t) ,
(cid:40)
τ
−1
i e−t/τi
0
for t ≥ 0,
else.
(A.4)
(A.5)
Different choices for the kernels give similar results (see Appendix). The overall scale of α and β can be
absorbed into the learning rate η in eq. (A.3). In our simulations, we fix α − β = 1 and keep the learning rate
constant as we change the plasticity rule (see eq. (3)).
In the spiking simulations with and without reinforcement learning in the tutor circuit, the firing rates ci(t)
and gj(t) were estimated by filtering spike trains with exponential kernels whose timescales were in the range
5 ms -- 40 ms. The reinforcement studies typically required longer timescales for stability, possibly because of
delays between conductor activity and reward signals.
A.3 Derivation of the matching tutor signal
To find the tutor signal that provides the most effective teaching for the student, we first calculate how much
synaptic weights change according to our plasticity rule, eq. (A.3). Then we require that this change matches
the gradient descent direction. We have
ci(t)(gj(t) − θ) dt .
(A.6)
(cid:90) T
∆Wij =
0
dWij
dt
(cid:90) T
dt = η
0
15
Because of the linearity assumptions in our model, it is sufficient to focus on a case in which each conductor
neuron, i, fires a single short burst, at a time ti. We write this as ci(t) = δ(t − ti), and so
(cid:90) T
(cid:90) T
∆Wij =
0
dWij
dt
dt = η
0
K(t − ti)(gj(t) − θ) dt ,
(A.7)
where we used the definition of ci(t). If the time constants τ1, τ2 are short compared to the timescale on which
the tutor input gj(t) varies, only the values of gj(t) around time ti will contribute to the integral. If we further
assume that T (cid:29) ti, we can use a Taylor expansion of gj(t) around t = ti to perform the calculation:
(cid:90) ∞
∆Wij ≈ η
ti
= η(gj(ti) − θ)
= η(gj(ti) − θ)
(cid:48)
K(t) dt + ηg
K(t − ti)(cid:0)gj(ti) − θ + (t − ti)g
j(ti)(cid:1) dt
(cid:90) ∞
(cid:0)αK1(t) − βK2(t)(cid:1) dt + ηg
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
∆Wij = η(cid:2)(α − β) (gj(ti) − θ) + (ατ1 − βτ2)g
(cid:48)
j(ti)
(cid:48)
j(ti)
tK(t) dt
0
0
0
0
Doing the integrals involving the exponential kernels K1 and K2, we get
t(cid:0)αK1(t) − βK2(t)(cid:1) dt .
j(ti)(cid:3) .
(cid:48)
(A.8)
(A.9)
We would like this synaptic change to optimally reduce a measure of mismatch between the output and
the desired target as measured by a loss function. A generic smooth loss function L(ya(t), ¯ya(t)) can be
quadratically approximated when ya is sufficiently close to the target ¯ya(t). With this in mind, we consider a
quadratic loss
(cid:90) T
(cid:88)
a
0
(cid:2)ya(t) − ¯ya(t)(cid:3)2
L =
1
2
dt .
(A.10)
The loss function would decrease monotonically during learning if synaptic weights changed in proportion to
the negative gradient of L:
where γ is a learning rate. This implies
∆Wij = −γ
∆Wij = −γ
∂L
∂Wij
,
(cid:90) T
(cid:88)
a
0
(cid:2)ya(t) − ¯ya(t)(cid:3) ci(t) .
Maj
(A.11)
(A.12)
(A.13)
Using again ci(t) = δ(t − ti), we obtain
∆Wij = −γj(ti) ,
where we used the notation from eq. (2) for the motor error at student neuron j.
We now set (A.9) and (A.13) equal to each other. If the conductor fires densely in time, we need the
equality to hold for all times, and we thus get a differential equation for the tutor signal gj(t). This identifies
the tutor signal that leads to gradient descent learning as a function of the motor error j(t), eq. (3) (with the
notation ζ = γ/η).
A.4 Spiking simulations
We used spiking models that were based on leaky integrate-and-fire neurons with current-based dynamics
for the synaptic inputs. The magnitude of synaptic potentials generated by the conductor -- student synapses
was independent of the membrane potential, approximating AMPA receptor dynamics, while the synaptic
inputs from the tutor to the student were based on a mixture of AMPA and NMDA dynamics. Specifically, the
16
equations describing the dynamics of the spiking model were:
(cid:1)
= (VR − Vj) + R(cid:0)I AMPA
(cid:88)
− Vinh ,
δ(t − tconductor #i
(cid:88)
I AMPA
j
τAMPA
(cid:88)
+ I NMDA
= −
Wij
+
k
k
i
j
j
δ(t − ttutor
k
) ,
τm
dVj
dt
dI AMPA
j
dt
dI NMDA
j
dt
(cid:88)
k
) + (1 − r)w
+ rwG(Vj)
k
Sj(t) ,
= −
I NMDA
j
τNMDA
ginh
Nstudent
Sj
τinh
+
(cid:88)
(cid:88)
j
k
Vinh =
dSj
dt
= −
(cid:20)
δ(t − tstudent
k
) ,
(cid:21)−1
G(V ) =
1 +
[Mg]
3.57 mM
exp(−V /16.13 mV)
.
(except during refractory period)
δ(t − ttutor
k
) ,
(A.14)
Here Vj is the membrane potential of the jth student neuron and VR is the resting potential, as well as the
potential to which the membrane was reset after a spike. Spikes were registered whenever the membrane
potential went above a threshold Vth, after which a refractory period τref ensued. Apart from excitatory
AMPA and NMDA inputs modeled by the I AMPA
variables in our model, we also included a global
inhibitory signal Vinh which is proportional to the overall activity of student neurons averaged over a timescale
τinh. The averaging is performed using the auxiliary variables Sj which are convolutions of student spike
trains with an exponential kernel. These can be thought of as a simple model for the activities of inhibitory
interneurons in the student.
and I NMDA
j
j
Table 1 gives the values of the parameters we used in the simulations. These values were chosen to match
the firing statistics of neurons in bird RA, as described below.
Parameter
Symbol
Value
Parameter
Symbol
Value
No. of conductor neurons
Reset potential
Threshold potential
Membrane time constant
Refractory period
VR
Vth
τm
τref
AMPA time constant
τAMPA
6.3 ms
300
No. of student neurons
−72.3 mV Input resistance
−48.6 mV Strength of inhibition
24.5 ms
1.1 ms
Fraction NMDA receptors
Strength of synapses from
tutor
No. of conductor synapses
per student neuron
NMDA time constant
τNMDA
81.5 ms Mean strength of synapses
Time constant for global in-
hibition
Conductor firing rate during
bursts
τinh
20 ms
632 Hz
from conductor
Standard
of
conductor -- student weights
deviation
Table 1: Values for parameters used in the spiking simulations.
R
ginh
r
w
80
353 MΩ
1.80 mV
0.9
100 nA
148
32.6 nA
17.4 nA
The voltage dynamics for conductor and tutor neurons was not simulated explicitly. Instead, each conductor
neuron was assumed to fire a burst at a fixed time during the simulation. The onset of each burst had additive
timing jitter of ±0.3 ms and each spike in the burst had a jitter of ±0.2 ms. This modeled the uncertainty in
spike times that is observed in in vivo recordings in birdsong (Hahnloser, Kozhevnikov, and Fee 2002). Tutor
neurons fired Poisson spikes with a time-dependent firing rate that was set as described in the main text.
The initial connectivity between conductor and student neurons was chosen to be sparse (see Table 1).
The initial distribution of synaptic weights was log-normal, matching experimentally measured values for
zebra finches (Garst-Orozco, Babadi, and Olveczky 2014). Since these measurements are done in the slice,
the absolute number of HVC synapses per RA neuron is likely to have been underestimated. The number
17
of conductor -- student synapses we start with in our simulations is thus chosen to be higher than the value
reported in that paper (see Table 1), and is allowed to change during learning. We checked that the learning
paradigm described here is robust to substantial changes in these parameters, but we have chosen values that
are faithful to birdsong experiments and which are thus able to imitate the RA spiking statistics during song.
The synapses projecting onto each student neuron from the tutor have a weight that is fixed during our
simulations reflecting the finding in (Garst-Orozco, Babadi, and Olveczky 2014) that the average strength of
LMAN -- RA synapses for zebra finches does not change with age. There is some evidence that individual LMAN --
RA synapses undergo plasticity concurrently with the HVC -- RA synapses (Mehaffey and Doupe 2015) but we
did not seek to model this effect. There are also developmental changes in the kinetics of NMDA-mediated
synaptic currents in both HVC -- RA and LMAN -- RA synapses which we do not model (Stark and Perkel 1999).
These, however, happen early in development, and thus are unlikely to have an effect on song crystallization,
which is what our model focuses on. Stark and Perkel (1999) also observed changes in the relative contribution
of NMDA to AMPA responses in the HVC -- RA synapses. We do not incorporate such effects in our model
since we do not explicitly model the dynamics of HVC neurons in this paper. However, this is an interesting
avenue for future work, especially since there is evidence that area HVC can also contribute to learning, in
particular in relation to the temporal structure of song (Ali et al. 2013).
A.5 Matching spiking statistics with experimental data
We used an optimization technique to choose parameters to maximize the similarity between the statistics
of spiking in our simulations and the firing statistics observed in neural recordings from the songbird. The
comparison was based on several descriptive statistics: the average firing rate; the coefficient of variation and
skewness of the distribution of inter-spike intervals; the frequency and average duration of bursts; and the
firing rate during bursts. For calculating these statistics, bursts were defined to start if the firing rate went
above 80 Hz and last until the rate decreased below 40 Hz.
To carry out such optimizations in the stochastic context of our simulations, we used an evolutionary
algorithm -- the covariance matrix adaptation evolution strategy (CMA-ES) (Hansen 2006). The objective
and the observed statistics xobs
function was based on the relative error between the simulation statistics xsim
,
i
i
(cid:34)(cid:88)
(cid:18) xsim
i
i − 1
xobs
i
(cid:19)2(cid:35)1/2
error =
.
(A.15)
Equal weight was placed on optimizing the firing statistics in the juvenile (based on a recording from a 43 dph
bird) and optimizing firing in the adult (based on a recording from a 160 dph bird). In this optimization there
was no learning between the juvenile and adult stages. We simply required that the number of HVC synapses
per RA neuron, and the mean and standard deviation of the corresponding synaptic weights were in the ranges
seen in the juvenile and adult by Garst-Orozco, Babadi, and Olveczky (2014). The optimization was carried
out in Python (RRID:SCR_008394), using code from https://www.lri.fr/~hansen/cmaes_inmatlab.html.
The results fixed the parameter choices in Table 1 which were then used to study our learning paradigm. While
these choices are important for achieving firing statistics that are similar to those seen in recordings from the
bird, our learning paradigm is robust to large variations in the parameters in Table 1.
A.6 Software and data
We used custom-built Python (RRID:SCR_008394) code for simulations and data analysis. The software and
data that we used can be accessed online on GitHub (RRID:SCR_002630) at https://github.com/ttesileanu/
twostagelearning.
18
Suppose further that we use a general loss function,
(B.1)
(B.2)
B Appendix
B.1 Effect of nonlinearities
We can generalize the model from eq. (A.1) by using a nonlinear transfer function from student activities to
motor output, and a nonlinear activation function for student neurons:
(cid:17)
(cid:17)
.
(cid:1) dt .
(cid:12)(cid:12)(cid:12)(cid:12)ya(t)−¯ya(t)
∂L
∂ya
j
(cid:16)(cid:88)
(cid:16)(cid:88)
(cid:90) T
i
(cid:0)
(cid:90) T
(cid:88)
0
a
(cid:88)
a
ya(t) = Na
Majsj(t)
,
sj(t) = F
Wijci(t) + wgj(t) − xinh
L =
0 L
{ya(t) − ¯ya(t)}
∆Wij = −γ
MajN
(cid:48)
aF
(cid:48)
ci(t)
(cid:80)
j(t) =
MajN
(cid:48)
a
∂L
∂ya
(cid:12)(cid:12)(cid:12)(cid:12)ya(t)−¯ya(t)
Carrying out the same argument as that from section A.3, the gradient descent condition, eq. (A.11), implies
The departure from the quadratic loss function, L (cid:54)= 1
Na, have the effect of redefining the motor error,
2
a(ya(t) − ¯ya(t))2, and the nonlinearities in the output,
.
(B.3)
.
(B.4)
A proper loss function will be such that the derivatives ∂L/∂ya vanish when ya(t) = ¯ya(t), and so the motor
error j as defined here is zero when the rendition is perfect, as expected. If we use a tutor that ignores the
nonlinearities in a nonlinear system, i.e., if we use eq. (2) instead of eq. (B.4) to calculate the tutor signal
that is plugged into eq. (3), we still expect successful learning provided that N(cid:48)
a > 0 and that L is itself an
increasing function of ya − ¯ya (see section B.2). This is because replacing eq. (B.4) with eq. (2) would affect
the magnitude of the motor error without significantly changing its direction. In more complicated scenarios,
if the transfer function to the output is not monotonic, there is the potential that using eq. (2) would push
the system away from convergence instead of towards it. In such a case, an adaptive mechanism, such as the
reinforcement rules from eqns. (6) or (7) can be used to adapt to the local values of the derivatives N(cid:48)
a and
∂L/∂ya.
Finally, the nonlinear activation function F introduces a dependence on the student output sj(t) in eq. (B.3),
since F (cid:48) is evaluated at F −1(sj(t)). To obtain a good match between the student and the tutor in this context,
we can modify the student plasticity rule (A.3) by adding a dependence on the postsynaptic activity,
dWij
dt
= η ci(t) (gj(t) − θ) F
(cid:48)
(F
−1(sj(t))) .
(B.5)
In general, synaptic plasticity has been observed to indeed depend on postsynaptic activity (Chistiakova,
Bannon, et al. 2014; Chistiakova and Volgushev 2009). Our derivation suggests that the effectiveness of learning
could be improved by tuning this dependence of synaptic change on postsynaptic activity to the activation
function of postsynaptic neurons, according to eq. (B.5). It would be interesting to check whether such tuning
occurs in real neurons.
B.2 Effect of different output functions
In the main text, we assumed a linear mapping between student activities and motor output. Moreover, we
assumed a myotopic organization, in which each student neuron projected to a single muscle, leading to a
student -- output assignment matrix Maj in which each column had a single non-zero entry. We also assumed
19
that student neurons only contributed additively to the outputs, with no inhibitory activity. Here we show
that our results hold for other choices of student -- output mappings.
For example, assume a push-pull architecture, in which half of the student neurons controlling one output
are excitatory and half are inhibitory. This can be used to decouple the overall firing rate in the student
from the magnitude of the outputs. Learning works just as effectively as in the case of the purely additive
student -- output mapping when using matched tutors, Appendix Figures 1A and 1B. The consequences of
mismatching student and tutor circuits are also not significantly changed, Appendix Figures 1C and 1D.
We can also consider nonlinear mappings between the student activity and the final output. If there is a
monotonic output nonlinearity, as in eq. (B.1) with N(cid:48)
a > 0, the tutor signal derived for the linear case, eq. (3),
can still achieve convergence, though at a slower rate and with a somewhat lower accuracy (see Appendix
Figure 1E for the case of a sigmoidal nonlinearity). For non-monotonic nonlinearities, the direction from which
the optimum is approached can be crucial, as learning can get stuck in local minima of the loss function.2
Studying this might provide an interesting avenue to test whether learning in songbirds is based on a gradient
descent-type rule or on a more sophisticated optimization technique.
B.3 Different inhibition models
In the spiking model, we used an activity-dependent inhibitory signal that was proportional to the average
student activity. Using a constant inhibition instead, Vinh = constant, does not significantly change the results:
see Appendix Figure 1F for an example.
B.4 Effect of changing plasticity kernels
In the main text, we used exponential kernels with τ1 = 80 ms and τ2 = 40 ms for the smoothing of the
conductor signal that enters the synaptic plasticity rule, eq. (A.3). We can generalize this in two ways: we can
use different timescales τ1, τ2, or we can use a different functional form for the kernels. (Note that in the main
text we showed the effects of varying the parameters α and β in the plasticity rule, while the timescales τ1 and
τ2 were kept fixed.)
The values for the timescales τ1,2 were chosen to roughly match the shape of the plasticity curve measured
in slices of zebra finch RA (Mehaffey and Doupe 2015) (see Figures 1C, 1D). The main predictions of our
model, that learning is most effective when the tutor signal is matched to the student plasticity rule, and that
large mismatches between tutor and student lead to impaired learning, hold well when the student timescales
change: see Appendix Figure 2A for the case when τ1 = 20 ms and τ2 = 10 ms. In the main text we saw
that the negative effects of tutor -- student mismatch diminish for timescales that are shorter than ∼ τ1,2. In
Appendix Figure 2A, the range of timescales where a precise matching is not essential becomes very small
because the student timescales are short.
Another generalization of our plasticity rule can be obtained by changing the functional form of the kernels
used to smooth the conductor input. As an example, suppose K2 is kept exponential, while K1 is replaced by
(cid:40) 1
¯K1(t) =
te−t/¯τ1
for t ≥ 0,
else.
(B.6)
¯τ 2
1
0
An example of learning using an STDP rule based on kernels ¯K1 and K2 where ¯τ1 = τ2 is shown in Appendix
Figure 2B. The matching tutor has the same form as before, eq. (3) with timescale τtutor = τ∗
tutor given by
eq. (4), but with τ1 = 2¯τ1 = 2τ2. We can see that learning is as effective as in the case of purely exponential
kernels.
B.5 More general conductor patterns
In the main text, we have focused on a conductor whose activity matches that observed in area HVC of
songbirds (Hahnloser, Kozhevnikov, and Fee 2002): each neuron fires a single burst during the motor program.
Our model, however, is not restricted to this case. We generated alternative conductor patterns by using
arbitrarily-placed bursts of activity, as in Appendix Figure 3A. The model converges to a good rendition of the
2We thank Josh Gold for this observation.
20
Appendix figure 1: Robustness of learning. A. Error trace showing how average motor error evolves with
repetitions of the motor program for rate-based plasticity paired with a matching tutor, when the student --
output mapping has a push-pull architecture. The inset shows the final motor output (thick red line) compared
to the target output (dotted black line). The output on the first rendition and at two other stages of learning
are also shown. B. The error trace and final motor output shown for timing-based plasticity matched by a
tutor with a long integration timescale. C. Effects of mismatch between student and tutor on reproduction
accuracy when using a push-pull architecture for the student -- output mapping. The heatmap shows the final
reproduction error of the motor output after 1000 learning cycles when a student with plasticity parameters
α and β is paired with a tutor with memory timescale τtutor. Here τ1 = 80 ms and τ2 = 40 ms. D. Error
evolution curves as a function of the mismatch between student and tutor. Each plot shows how the error in
the motor program changes during 1000 learning cycles for the same conditions as those shown in the heatmap.
The region shaded in light pink shows simulations where the mismatch between student and tutor leads to a
deteriorating instead of improving performance during learning. E. Convergence in the rate-based model with
a linear-nonlinear controller that uses a sigmoidal nonlinearity. F. Convergence in the spiking model when
inhibition is constant instead of activity-dependent (Vinh = constant).
21
102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β0.501.002.005.0010.00050100150200250repetitions02468101214errorα=7.0,β=6.0,τtutor=320.0time01020304050607080outputtargetoutput050100150200250repetitions02468101214errorα=0.0,β=−1.0,τtutor=40.0time010203040506070outputtargetoutput0600ABCD0100200300400500repetitions02468101214errorα=24.0,β=23.0,τtutor=1000.0time01020304050607080outputtargetoutput050100150200repetitions02468101214errorα=1.0,β=0.0,τtutor=80.0time01020304050607080outputtargetoutputEF060006000600worsebetter Appendix figure 2: Effect of changing conductor smoothing kernels in the plasticity rule. (A) Matrix showing
learning accuracy when using different timescales for the student plasticity rule. Each entry in the heatmap
shows the average rendition error after 1000 learning steps when pairing a tutor with timescale τtutor with a
non-matched student. Here the kernels are exponential, with timescales τ1 = 20 ms, τ2 = 10 ms. (B) Evolution
of motor error with learning using kernels ∼ e−t/τ and ∼ te−t/τ , instead of the two exponentials used in the
main text. The tutor signal is as before, eq. (3). The inset shows the final output for the trained model, for
one of the two output channels. Learning is as effective and fast as before.
target program, Appendix Figure 3B. Learning is harder in this case because many conductor neurons can
be active at the same time, and the weight updates affect not only the output of the system at the current
position in the motor program, but also at all the other positions where the conductor neurons fire. This is
in contrast to the HVC-like conductor, where each neuron fires at a single point in the motor program, and
thus the effect of weight updates is better localized. More generally, simulations show that the sparser the
conductor firing, the faster the convergence (data not shown). The accuracy of the final rendition of the motor
program (Appendix Figure 3B, inset) is also not as good as before.
B.6 Edge effects
In our derivation of the matching tutor rule, we assumed that the system has enough time to integrate all the
synaptic weight changes from eq. (A.3). However, some of these changes occur tens or hundreds of milliseconds
after the inputs that generated them, due to the timescales used in the plasticity kernel. Since our simulations
are only run for a finite amount of time, there will in general be edge effects, where periods of the motor
program towards the end of the simulations will have difficulty converging. To offset such numerical issues, we
ran the simulations for a few hundred milliseconds longer than the duration of the motor program, and ignored
the data from this extra period. Our simulations typically run for 600 ms, and the time reserved for relaxation
after the end of the program was set to 1200 ms. The long relaxation time was chosen to allow for cases where
the tutor was chosen to have a very long memory timescale.
B.7 Parameter optimization for reproducing juvenile and adult spiking statistics
We set the parameters in our simulations to reproduce spiking statistics from recordings in zebra finch RA as
closely as possible. Appendix Figure 4 shows how the distribution of summary statistics obtained from 50 runs
of the simulation compares to the distributions calculated from recordings in birds at various developmental
stages. Each plot shows a standard box and whisker plot superimposed over a kernel-density estimate of the
distribution of a given summary statistic, either over simulation runs or over recordings from birds at various
stages of song learning. We ran two sets of simulations, one for a bird with juvenile-like connectivity between
HVC and RA, and one with adult-like connectivity (see Methods). In these simulations there was no learning
to match the timecourse of songs -- the goal was simply to identify parameters that lead to birdsong-like firing
statistics.
22
102040801603206401280256051201024020480τtutor102040801603206401280256051201024020480τ∗tutor=ατ1−βτ2α−β0.501.002.005.0010.00050100150200repetitions02468101214errorα=24.0,β=23.0,τtutor=1000.0time010203040506070outputtargetoutput0600ABworsebetter Appendix figure 3: Learning with arbitrary conductor activity. A. Typical activity of conductor neurons. 20
of the 100 neurons included in the simulation are shown. The activity pattern is chosen so that about 10%
of the neurons are active at any given time. The pattern is chosen randomly but is fixed during learning.
Each conductor burst lasts 30 ms. B. Convergence curve and final rendition of the motor program (in inset).
Learning included two output channels but the final output is shown for only one of them.
The qualitative match between our simulations and recordings is good, but the simulations are less variable
than the measurements. This may be due to sources of variability that we have ignored -- for example, all our
simulated neurons had exactly the same membrane time constants, refractory periods, and threshold potentials,
which is not the case for real neurons. Another reason might be that in our simulations, all the runs were
performed for the same network, while the measurements are from different cells in different birds.
B.8 Effect of spiking stochasticity on learning
As pointed out in the main text, learning is affected in the spiking simulations when the tutor error integration
timescale τtutor becomes very long. More specifically, two distinct effects occur. First, the fluctuations in the
motor output increase, leading to a poorer match to the shape of the target motor program. And second, the
whole output gets shifted up, towards higher muscle activation values. Both of these effects can be traced back
to the stochasticity of the tutor signal.
In the spiking simulations, tutor neurons are assumed to fire Poisson spikes following a time-dependent
firing rate that obeys eq. (5). By the nature of the Poisson process, the tutor output in this case will contain
fluctuations around the mean, g(t) ∼ ¯g(t) + ξ(t). Recall that the scale of g(t) is set by the threshold θ and
thus so is the scale of the variability ξ(t).
As long as the tutor error integration timescale is not very long, g(t) roughly corresponds to a smoothed
version of the motor error (t) (cf. eq. (5)). However, as τtutor grows past the duration T of the motor program,
the exponential term in eq. (5) becomes essentially constant, leading to a tutor signal ¯g(t) whose departures
from the center value θ decrease in proportion to the timescale τtutor. As far as the student is concerned, the
relevant signal is g(t) − θ (eq. (1)), and thus, when τtutor > T , the signal-to-noise ratio in the tutor guiding
signal starts to decrease as 1/τtutor. This ultimately leads to a very noisy rendition of the target program. One
way to improve this would be to increase the gain factor ζ that controls the relation between the motor error
and the tutor signal (see eq. (5)). This improves the ability of the system to converge onto its target in the
late stages of learning. In the early stages of learning, however, this could lead to saturation problems. One
way to fix this would be to use a variable gain factor ζ that ensures the whole range of tutor firing rates is used
without generating too much saturation. This would be an interesting avenue for future research.
Reducing the fluctuations in the tutor signal also decreases the fluctuations in the conductor -- student
synaptic weights, which leads to fewer weights being clamped at 0 because of the positivity constraint. This
reduces the shift between the learned motor program and the target. As mentioned in the main text, another
approach to reducing or eliminating this shift is to allow for negative weights or (more realistically) to use a
push-pull mechanism, in which the activity of some student neurons acts to increase muscle output, while the
23
050010001500200025003000repetitions02468101214errorα=5.0,β=4.0,τtutor=240.0time01020304050607080outputtargetoutput06000100200300400500600time(ms)135791113151719conductorneuronindexAB Appendix figure 4: Violin plots showing how the spiking statistics from our simulation compared to the
statistics obtained from neural recordings. Each violin shows a kernel-density estimate of the distribution that
a particular summary statistic had in either several runs of a simulation, or in several recordings from behaving
birds. The circle and the box within each violin show the median and the interquartile range.
24
simjuvenile40-6565-9090-115115-140140-165simadult020406080100120140160Averagefiringrate(Hz)simjuvenile40-6565-9090-115115-140140-165simadult0123456CVofISIsimjuvenile40-6565-9090-115115-140140-165simadult0510152025SkewnessofISIsimjuvenile40-6565-9090-115115-140140-165simadult024681012Burstfrequency(Hz)simjuvenile40-6565-9090-115115-140140-165simadult0.000.050.100.150.20Averageburstlength(s)simjuvenile40-6565-9090-115115-140140-165simadult050100150200250300350400450Firingrateduringbursts(Hz) activity of other student neurons acts as an inhibition on muscle output.
C Plasticity parameter values
In the heatmaps that appear in many of the figures in the main text and in the supplementary information,
we kept the timescales τ1 and τ2 constant while varying α and β to modify the student plasticity rule. Since
the overall scale of α and β is inconsequential as it can be absorbed into the learning rate (as explained in
section 2.2), we imposed the further constraint α − β = 1. This implies that we effectively focused on a
one-parameter family of student plasticity rule, as identified by the value of α (and the corresponding value for
β = α − 1). In the figures, we expressed this instead in terms of the timescale of the optimally-matching tutor,
τ∗
tutor, as defined in eq. (4).
Below we give the explicit values of α and β that we used for each row in the heatmaps. These can be
calculated by solving for α in eq. (4), using β = α − 1, and assume that τ1 = 80 ms and τ2 = 40 ms.
τ∗
tutor
α
β
10 −0.75 −1.75
−1.5
20
−1.0
40
0.0
80
2.0
160
6.0
320
640
14.0
30.0
1280
62.0
2560
126.0
5120
10240
254.0
510.0
20480
−0.5
0.0
1.0
3.0
7.0
15.0
31.0
63.0
127.0
255.0
511.0
References
[1] Farhan Ali et al. "The basal ganglia is necessary for learning spectral, but not temporal, features of
birdsong". In: Neuron 80.2 (Oct. 2013), pp. 494 -- 506.
[2] Aaron S. Andalman and Michale S. Fee. "A basal ganglia-forebrain circuit in the songbird biases motor
output to avoid vocal errors." In: Proceedings of the National Academy of Sciences of the United States
of America 106.30 (July 2009), pp. 12518 -- 23.
[3] Mark J. Basista et al. "Independent Premotor Encoding of the Sequence and Structure of Birdsong in
Avian Cortex". In: Journal of Neuroscience 34.50 (Dec. 2014), pp. 16821 -- 16834.
[4] R. A. Canady et al. "Effect of testosterone on input received by an identified neuron type of the canary
song system: a Golgi/electron microscopy/degeneration study". In: The Journal of Neuroscience 8.10
(1988), pp. 3770 -- 3784.
[5] Marina Chistiakova, Nicholas M. Bannon, et al. "Heterosynaptic Plasticity: Multiple Mechanisms and
Multiple Roles". In: The Neuroscientist : a review journal bringing neurobiology, neurology and psychiatry
April (2014).
[6] Marina Chistiakova and Maxim Volgushev. "Heterosynaptic plasticity in the neocortex". In: Experimental
Brain Research 199.3-4 (2009), pp. 377 -- 390.
[7] Kenji Doya and Terrence J. Sejnowski. "A computational model of avian song learning". In: The new
cognitive neurosciences. Ed. by Michael S. Gazzaniga. The MIT Press, 2000, pp. 469 -- 482. isbn: 0262071959.
arXiv: arXiv:1011.1669v3.
[8] Michael A. Farries and Adrienne L. Fairhall. "Reinforcement Learning With Modulated Spike Timing-
Dependent Synaptic Plasticity". In: Journal of neurophysiology 98.6 (2007), pp. 3648 -- 3665.
[9] Michale S. Fee and Jesse H. Goldberg. "A hypothesis for basal ganglia-dependent reinforcement learning
in the songbird". In: Neuroscience 198 (2011), pp. 152 -- 170.
25
[10]
[11]
Ila R. Fiete, Michale S. Fee, and H. Sebastian Seung. "Model of birdsong learning based on gradient
estimation by dynamic perturbation of neural conductances". In: Journal of neurophysiology 98.4 (Oct.
2007), pp. 2038 -- 57.
Ila R. Fiete and H. Sebastian Seung. "Gradient learning in spiking neural networks by dynamic per-
turbation of conductances". In: Physical Review Letters 97.4 (2006), p. 048104. arXiv: 0601028v1
[arXiv:q-bio].
[12] Paul W. Frankland and Bruno Bontempi. "The organization of recent and remote memories". In: Nature
reviews. Neuroscience 6.2 (2005), pp. 119 -- 30.
[13] Vikram Gadagkar et al. "Dopamine neurons encode performance error in singing birds". In: Science
354.6317 (2016), pp. 1278 -- 1282. eprint: http://science.sciencemag.org/content/354/6317/1278.
full.pdf.
[14] Jonathan Garst-Orozco, Baktash Babadi, and Bence P. Olveczky. "A neural circuit mechanism for
regulating motor variability during skill learning". In: eLife 3 (2014), e03697.
[15] Richard H. R. Hahnloser, Alexay A. Kozhevnikov, and Michale S. Fee. "An ultra-sparse code underlies
the generation of neural sequences in a songbird". In: Nature 419.1989 (2002), pp. 796 -- 797.
[16] Kosuke Hamaguchi, Masashi Tanaka, and Richard Mooney. "A Distributed Recurrent Network Contributes
to Temporally Precise Vocalizations". In: Neuron 91 (2016), pp. 680 -- 693.
[17] Nikolaus Hansen. "The CMA Evolution Strategy: A Comparing Review". In: Towards a new evolutionary
computation. Advances on estimation of distribution algorithms. 2006, pp. 75 -- 102.
[18] Kathrin Herrmann and Arthur P. Arnold. "The development of afferent projections to the robust
archistriatal nucleus in male zebra finches: a quantitative electron microscopic study". In: The Journal of
Neuroscience 11.7 (1991), pp. 2063 -- 2074.
[19] Lukas A. Hoffmann et al. "Dopaminergic Contributions to Vocal Learning". In: The Journal of Neuro-
science 36.7 (2016), pp. 2176 -- 2189.
[20] K Immelmann. "Song development in the zebra finch and other estrildid finches". In: Bird vocalizations.
Ed. by R. A. Hinde. 1969, pp. 61 -- 74.
[21] Mimi H. Kao, Brian D. Wright, and Allison J. Doupe. "Neurons in a Forebrain Nucleus Required for
Vocal Plasticity Rapidly Switch between Precise Firing and Variable Bursting Depending on Social
Context". In: The Journal of Neuroscience 28.49 (2008), pp. 13232 -- 13247.
[22] Risa Kawai et al. "Motor Cortex Is Required for Learning but Not for Executing a Motor Skill". In:
Neuron 86.3 (2015), pp. 800 -- 812.
[23] Satoshi Kojima, Mimi H. Kao, and Allison J. Doupe. "Task-related "cortical" bursting depends critically
on basal ganglia input and is linked to vocal plasticity". In: PNAS 110.12 (2013), pp. 4756 -- 4761.
[24] L'ubica Kubikova and L'ubor Kost´al. "Dopaminergic system in birdsong learning and maintenance". In:
Journal of Chemical Neuroanatomy 39 (2010), pp. 112 -- 123.
[25] Anthony Leonardo and Michale S. Fee. "Ensemble Coding of Vocal Control in Birdsong". In: The Journal
of Neuroscience 25.3 (2005), pp. 652 -- 661.
[26] Galen F. Lynch et al. "Rhythmic Continuous-Time Coding in the Songbird Analog of Vocal Motor
Cortex". In: Neuron 90.4 (2016), pp. 877 -- 892.
[27] W. Hamish Mehaffey and Allison J. Doupe. "Naturalistic stimulation drives opposing heterosynaptic
plasticity at two inputs to songbird cortex". In: Nature Neuroscience 18.9 (2015), pp. 1272 -- 1280.
[28] Raoul-Martin Memmesheimer et al. "Learning Precisely Timed Spikes". In: Neuron 82 (2014), pp. 1 -- 14.
[29] Bence P. Olveczky, Aaron S. Andalman, and Michale S. Fee. "Vocal experimentation in the juvenile
songbird requires a basal ganglia circuit." In: PLoS biology 3.5 (May 2005), e153.
[30] Bence P. Olveczky, Timothy M. Otchy, et al. "Changes in the neural control of a complex motor sequence
during learning". In: Journal of Neurophysiology 106.1 (2011), pp. 386 -- 397.
[31] David J. Perkel. "Origin of the Anterior Forebrain Pathway". In: Annals of the New York Academy of
Sciences 1016 (2004), pp. 736 -- 748.
26
[32] Michel A. Picardo et al. "Population-Level Representation of a Temporal Sequence Underlying Song
Production in the Zebra Finch". In: Neuron 90.4 (2016), pp. 866 -- 876.
[33] H. Sebastian Seung. "Learning in Spiking Neural Networks by Reinforcement of Stochastic Synaptic
Transmission". In: 40 (2003), pp. 1063 -- 1073.
[34] H. Blair Simpson and David S. Vicario. "Brain Pathways for Learned and Unlearned Vocalizations Differ
in Zebra Finches". In: The Journal of Neuroscience 10.5 (1990), pp. 1541 -- 1556.
[35] John E. Spiro, Matthew B. Dalva, and Richard Mooney. "Long-Range Inhibition Within the Zebra
Finch Song Nucleus RA Can Coordinate the Firing of Multiple Projection Neurons". In: Journal of
neurophysiology 81 (1999), pp. 3007 -- 3020.
[36] Laura L. Stark and David J. Perkel. "Two-stage, input-specific synaptic maturation in a nucleus essential
for vocal production in the zebra finch". In: The Journal of Neuroscience 19.20 (1999), pp. 9107 -- 9116.
[37] Evren C. Tumer and Michael S. Brainard. "Performance variability enables adaptive plasticity of
'crystallized' adult birdsong". In: Nature 450.December (2007), pp. 1240 -- 1244.
[38] Robert S. Turner and Michel Desmurget. "Basal ganglia contributions to motor control: a vigorous tutor".
In: Current Opinion in Neurobiology 20.6 (2010), pp. 704 -- 716.
[39] Lena Veit, Dmitriy Aronov, and Michale S. Fee. "Learning to breathe and sing: development of respiratory-
vocal coordination in young songbirds". In: Journal of Neurophysiology 106.4 (2011), pp. 1747 -- 1765.
eprint: http://jn.physiology.org/content/106/4/1747.full.pdf.
[40] Timothy L. Warren et al. "Mechanisms and time course of vocal learning and consolidation in the adult
songbird". In: Journal of Neurophysiology 106 (2011), pp. 1806 -- 1821.
[41] Albert C. Yu and Daniel Margoliash. "Temporal hierarchical control of singing in birds". In: Science
273.5283 (1996), pp. 1871 -- 1875.
27
|
1010.4145 | 2 | 1010 | 2010-10-29T14:29:24 | Voltage clamp analysis of nonlinear dendritic propertie in prepositus hypoglossi neurons | [
"q-bio.NC"
] | The nonlinear properties of the dendrites in prepositus hypoglossi neurons are involved in maintenance of eye position. The biophysical properties of these neurons are essential for the operation of the vestibular neural integrator that converts a head velocity signal to one that controls eye position. A novel method named QSA (quadratic sinusoidal analysis) for voltage clamped neurons was used to quantify nonlinear responses that are dominated by dendrites. The voltage clamp currents were measured at harmonic and interactive frequencies using specific stimulation frequencies, which act as frequency probes of the intrinsic nonlinear neuronal behavior. These responses to paired frequencies form a matrix that can be reduced by eigendecomposition to provide a very compact piecewise quadratic analysis at different membrane potentials that otherwise is usually described by complex differential equations involving a large numbers of parameters and dendritic compartments. Moreover, the QSA matrix can be interpolated to capture most of the nonlinear neuronal behavior like a Volterra kernel. The interpolated quadratic functions of the two major prepositus hypoglossi neurons, namely type B and D, are strikingly different. A major part of the nonlinear responses is due to the persistent sodium conductance, which appears to be essential for sustained nonlinear effects induced by NMDA activation and thus would be critical for the operation of the neural integrator. Finally, the dominance of the nonlinear responses by the dendrites supports the hypothesis that persistent sodium conductance channels and NMDA receptors act synergistically to dynamically control the influence of individual synaptic inputs on network behavior. | q-bio.NC | q-bio |
Voltage clamp analysis of nonlinear
dendritic properties in prepositus
hypoglossi neurons
Christophe Magnani, Daniel Eugène, Erwin Idoux and L.E. Moore*
CESEM - UMR8194 - CNRS - Université Paris Descartes
Address : 45 rue des Saints-Pères, 75270 PARIS, FRANCE - Tel.:33-(0)142863398,
Fax:33-(0)142863399
Keywords: Electrophysiology; Vestibular neural integrator; Membrane potential;
Impedance; Quadratic analysis; Persistent sodium conductance; NMDA receptors
*Corresponding author
Abstract
The nonlinear properties of the dendrites in prepositus hypoglossi neurons are in-
volved in maintenance of eye position. The biophysical properties of these neurons
are essential for the operation of the vestibular neural integrator that converts a
head velocity signal to one that controls eye position. A novel method named QSA
(quadratic sinusoidal analysis) for voltage clamped neurons was used to quantify
nonlinear responses that are dominated by dendrites. The voltage clamp currents
were measured at harmonic and interactive frequencies using specific stimulation
frequencies, which act as frequency probes of the intrinsic nonlinear neuronal be-
havior. These responses to paired frequencies form a matrix that can be reduced
by eigendecomposition to provide a very compact piecewise quadratic analysis at
different membrane potentials that otherwise is usually described by complex dif-
ferential equations involving a large numbers of parameters and dendritic com-
partments. Moreover, the QSA matrix can be interpolated to capture most of
the nonlinear neuronal behavior like a Volterra kernel. The interpolated quadratic
functions of the two major prepositus hypoglossi neurons, namely type B and D,
are strikingly different. A major part of the nonlinear responses is due to the per-
1
2
sistent sodium conductance, which appears to be essential for sustained nonlinear
effects induced by NMDA activation and thus would be critical for the operation
of the neural integrator. Finally, the dominance of the nonlinear responses by the
dendrites supports the hypothesis that persistent sodium conductance channels
and NMDA receptors act synergistically to dynamically control the influence of
individual synaptic inputs on network behavior.
Introduction
Mathematical models based on the experimentally measured biophysical proper-
ties of neurons generally consist of complicated sets of differential equations derived
from the historical Hodgkin and Huxley (1952) model. Extending the HH formal-
ism to branching neurons requires a large number of parameters that must be
determined to obtain a realistic neuronal model. The techniques previously em-
ployed to measure these parameters involve either linear admittance (or impedance)
measurements or ad hoc extrapolations from voltage clamp experiments with poor
space clamp control. Thus, it is important to consider more refined theories from
nonlinear analysis, such as nonlinear dynamics of neurons (Gutkin and Ermen-
trout, 1998; Izhikevich, 2002) or nonlinear system identification (Marmarelis and
Naka, 1973; French, 1976; Victor and Shapley, 1980; Boyd et al., 1983; Schetzen,
2006). A goal of nonlinear analysis is not just a refinement of the linear systems
approach, but the development of a fundamental insight into how neurons process
information.
In a seminal paper, Fitzhugh derived the equations of the nonlinear response for
a single sinusoidal voltage clamp (FitzHugh, 1983). This approach has been ex-
tended to the quadratic response for a multi-sinusoidal voltage clamp and been
developed as a matrix theory termed quadratic sinusoidal analysis (QSA) (Mag-
nani and Moore, 2010). The proposed experimental approach is essentially based
on QSA, requiring that the stimulus amplitudes evoke mainly linear and quadratic
responses.
In addition, QSA provides the mathematical tools for a model inde-
pendent analysis of quadratic nonlinearities and provides an innovative way to
quantitatively describe real neurons and their models. The measurement of non-
linearities in neurons under normal physiological conditions is clearly important in
order to understand how they process synaptic inputs, which typically evoke 5-10
mV post-synaptic nonlinear responses.
Two types of neurons of the rat prepositus hypoglossi nucleus (PHN) were inves-
tigated. Both types clearly manifest nonlinearities at multiple subthreshold step
levels. The type D neurons are known to show marked spontaneous, voltage depen-
dent and irregular oscillatory properties. By contrast, type B neurons, the majority
3
in this nucleus, are non-oscillatory and have regular spontaneous activity that is
highly dependent on a significant persistent sodium (gNaP) conductance (Vervaeke
et al., 2006). In this paper, the novel QSA method has been used to investigate the
quadratic response to time varying voltage clamped stimuli and establish a quan-
titative characterization of the nonlinear behavior in order to understand neuronal
responses elicited by normal physiological synaptic inputs.
It will be shown that at physiological levels of stimulation, neurons and their mod-
els can generate significant responses at harmonic and interactive frequencies that
are not present in the input signal. Thus, the nonlinear frequency responses contain
more frequencies over a wider frequency band than the input signal. As a conse-
quence they provide significant amplification at dynamically changing membrane
potentials. The use of stimuli with multiple input frequencies allows one to probe
neuronal function and characterize it by a matrix of quadratic interactions, namely
the QSA matrix. It is then possible to extract information about active membrane
properties from this matrix by eigendecomposition. Finally, biologically realistic
simulations have been implemented using neuronal models based on vestibular neu-
ronal experimental data. These simulations suggest that the nonlinear responses
in voltage clamp are dominated by active dendritic structures.
Materials and Methods
Whole-cell patch-clamp recordings and statistical
analysis
This paper is both a theoretical and experimental nonlinear approach to neuronal
function that adds to previous steady state linear analyses (Fishman et al., 1977;
Murphey et al., 1995). It provides a quantitative assessment of quadratic responses
of both data recorded from individual neurons and their corresponding biophysical
models. Experiments were carried out on male Wistar rats (25- to 52-days-old)
supplied by Centre d'Elevage Roger Janvier (Le Genest Saint Isle, France). All ef-
forts were made to minimize animal suffering as well as the number of animals used.
All experiments followed the guidelines on the ethical use of animals from the Eu-
ropean Communities Council Directive of 24 November 1986 (86/609/EEC). Brain
dissections were performed as described elsewhere (Idoux et al., 2008). Briefly, af-
ter decapitation under deep anesthesia, the brain was quickly removed and placed
in ice-cold, phosphate/bicarbonate-buffered artificial cerebro-spinal fluid (ACSF),
which included (in mM) 225 sucrose, 5 KCl, 1 NaH2PO4, 26 NaHCO3, 0.25 CaCl2,
1.3 MgCl2, 11 glucose and was bubbled with 95% O2-5% CO2 (pH 7.4). Four or
five 250 µm thick, coronal slices containing the PHN were cut from the brainstem
with a microslicer (Leica, Rueil-Malmaison, France) and transferred into an incu-
4
bating vial filled with a regular ACSF containing (in mM) 124 NaCl, 5 KCl, 1
NaH2PO4, 26 NaHCO3, 2.5 CaCl2, 1.3 MgCl2, 11 glucose and bubbled with 95%
O2 and 5% CO2 (pH 7.4). Slices were then placed one at a time in the recording
chamber maintained at 32-34°C, where the slice was superfused with regular ACSF
at a constant flow rate of 3 mL min-1.
Patch-clamp pipettes were pulled from borosilicate glass tubing to a resistance of 5-
8 MΩ. The control internal solution contained (in mM) 140 K-gluconate, 2 MgCl2,
10 HEPES, 0.1 EGTA, 4 Na2ATP, and 0.4 Na2GTP (adjusted to pH 7.3 with
KOH). The junction potential for this internal solution was not subtracted for the
potential measurements or the model simulations. PHN neurons were visualized
with a Nomarski optic microscope under infrared illumination. Recordings were
made with an Axoclamp 2B amplifier (Axon Instruments, Union City, CA, USA)
or a Multiclamp 700B (Molecular Devices, Sunnyvale, CA, USA). The spontaneous
discharge was first recorded in the current-clamp mode for 8 to 10 minutes once
a stable level had been reached and the recorded PHN neuron was determined as
B or D type (see Idoux et al., 2008). PHN neurons that had resting membrane
potential more negative than -50 mV and a spike amplitude > 45 mV were selected
for the voltage clamp experiments. Types D and B neurons from the prepositus
hypoglossi nucleus were measured for different stimulus amplitudes and membrane
potentials. Based on the criteria for time invariance discussed in the Rationale
section, five type D and six type B neurons were selected for detailed analysis.
All measurements were made with stimuli applied for twice the duration used in
the analysis. Only the last half of the record was used to assure that a steady
state condition was reached. At voltage clamp potentials near threshold, transient
currents due to uncontrolled action potentials occasionally occurred in the non-
analyzed initial part of the recording, however they were completely inactivated
by the maintained depolarization with no firing during the latter analyzed part of
the record. In some experiments 25-50 µM NMDA (Sigma, St Quentin Fallavier,
France) was applied in the presence or absence of 2 µM TTX (Tocris, Bristol, UK).
The data acquisition was done with a PC-compatible computer running Win-
dows XP, using MATLAB scripts (MATLAB 7.0, MathWorks, Natick, MA, USA).
Recordings were low-pass filtered at 2 kHz and digitized at 5 kHz (BNC-2090
+ PCI-6052E, National Instruments, Austin, TX, USA). The OneSidedPValue (p-
value) of MeanTest[x-y] for paired differences, mean values and standard deviations
(±SD) were calculated with the HypothesisTesting package of MATHEMATICA
7.0, (Wolfram Research, Champain, IL, USA).
5
Rationale
PHN neurons were analyzed with QSA, which specifically selects harmonic and
intermodulation frequencies, described as follows.
If a double sinusoidal input
has frequencies f1 and f2 then the linear response will have exactly the same
frequencies f1 and f2. However, the quadratic response will include additional
harmonics 2f1 and 2f2 as well as intermodulation products f1 + f2 and f1 − f2.
This principle can be generalized to a multi-sinusoidal input in which case the
quadratic response will include double of each input frequency as well as sum
and difference of each pair of distinct input frequencies. A quadratic response
can generate frequency overlaps when distinct combinations of input frequencies
generate the same output frequency. For instance, a multi-sinusoidal input with
frequencies 1, 2, 3, 4 (in Hertz) would generate many frequency overlaps such as
1 = 2 − 1 or 2 + 3 = 1 + 4 and so on. In presence of frequency overlaps, it is not
possible to unambiguously measure the nonlinear frequency interactions. In the
previous example, the measurement at 5 Hz is ambiguous because one is unable
to distinguish between the contributions of 2 + 3 and 1 + 4. In order to avoid this
problem, the QSA was used with a flexible algorithm generating incommensurable
frequencies. This approach is based on a practical measurement technique, namely
harmonic probing on Volterra kernels (Boyd et al., 1983; Victor and Shapley, 1980).
Since harmonic and intermodulation responses also exist for nonlinearities of higher
degrees, for example third order intermodulation products f1 + f2 + f3, it is impor-
tant to ensure that the neurons mainly manifest quadratic nonlinearities, otherwise
the results would be significantly contaminated. For this, a necessary, but not suf-
ficient condition, consists of using relatively small stimulus amplitudes in such a
way that only linear and quadratic responses are significant. This approach was
used in our previous piecewise linear analysis (Murphey et al., 1995), hence the
term piecewise quadratic analysis can be used to describe the QSA extension. The
influence of the input amplitude on the harmonic response has been investigated
previously (see Moore et al., 1980; FitzHugh, 1983). In order to ensure that the
stimulus amplitudes were sufficient to overcome spontaneous noise and avoid signif-
icant higher order responses, several algorithms described elsewhere (Magnani and
Moore, 2010) have been implemented in MATLAB to verify that the experimental
traces are time invariant for both linear and quadratic outputs, and that the signal
can be adequately reconstructed by quadratic analysis.
It will be shown that the oscillatory type D neurons of the PHN have quadratic
responses over a range of subthreshold membrane potentials, namely they con-
vert limited amplitude and bandwidth input signals to wider bandwidth and more
complex output responses as mentioned above for nonlinear responses. Under nor-
mal physiological conditions of current clamp, a depolarization activates gNaP
6
conductance that in turn increases the impedance and consequently reduces the
electrotonic length. Type B neurons show similar effects and in addition, they
have a significant increase in the electrotonic length at hyperpolarized membrane
potentials because of the activation of a hyperpolarization activated conductance
(Erchova et al., 2004).
Voltage clamp experiments were done to partially control the oscillatory and bistable
responses of PHN neurons in order to analyze the nonlinear membrane properties
of both the somatic and dendritic regions of these neurons. Indeed, an important
advantage of a quantitative voltage clamp analysis of central neurons is the ex-
ploitation of the space clamp problem as a way to separate somatic and dendritic
responses, namely the voltage clamp current measured from the voltage controlled
soma is generally dominated by the unclamped voltage responses of the dendrites
(Moore et al., 1999). Thus, this current can be used as a measure of both linear
and nonlinear dendritic potential responses while the somatic membrane potential
is voltage clamped. These additional currents can flow because of a potential dif-
ference between the soma and the rest of the dendrite. In fact, these additional
currents reflect the behavior of the dendritic membrane potential that can be taken
into account with multi-compartmental models.
Previous voltage clamp measurements have been done using signals of small am-
plitude to obtain steady state linear responses at different membrane potentials.
These measurements were done over a potential range to obtain a quantitative
description of the voltage dependent conductances and have allowed the construc-
tion of neuronal multi-compartmental models of both type B and D PHN neurons
(Idoux et al., 2008). These models allow an estimation of both somatic and den-
dritic membrane properties from somatic voltage clamp experiments that probe
the soma and all regions of the dendritic structure.
When the stimulus amplitudes are sufficiently small to elicit linear responses, both
the voltage clamp and the current clamp generate equivalent linear results, however
this is not the case for nonlinear responses. The present series of voltage clamp
experiments is a quadratic extension of the steady state linear analysis. Voltage
clamped neurons show two kinds of nonlinearities : first, space clamped somatic
ionic currents, and second, ionic currents in the soma due to an unclamped den-
dritic membrane. However, in a similar current clamp experiment, the nonlinear
behaviors measured from the soma are caused by the voltage responses of both
the somatic and the dendritic membranes. Due to this asymmetry between voltage
clamp and current clamp, there is no obvious way to predict the voltage clamp non-
linear response from the current clamp nonlinear response nor the converse. It is
well known in the linear case that the admittance from voltage clamp, Y (f ) =
,
I(f )
V (f )
, that is Y = Z−1
is the inverse of the impedance from current clamp Z (f ) =
V (f )
I(f )
7
V (f1+f2)
I(f1) I(f2)
where I or V refers to Fourier transforms of I and V, respectively. In the nonlin-
ear case, with nonoverlapping frequencies, the quadratic response for the voltage
clamp is defined by QSA as Bvc (f1, f2) = γf1,f2
and for the current clamp
I(f1+f2)
V (f1) V (f2)
by Bcc (f1, f2) = γf1,f2
, where γf1,f2
is a symmetry factor (Magnani and
Moore, 2010). In general Bvc (f1, f2) and Bcc (f1, f2) are not reciprocally equivalent
because of the asymmetry discussed above. Hence, this is an important concep-
tual difference between linear and nonlinear analysis, which also plays a role in
interpretation of current versus voltage clamp experiments.
The terms, V (f1 + f2) and I (f1 + f2) where f1 and f2 are either positive or neg-
ative frequencies refer to all complex values of V and I at the interactive and
harmonic quadratic frequencies. The quadratic response function can then be rep-
resented as magnitude and phase plots versus f1 and f2.
Matrix reduction
Since Bvc (f1, f2) is difficult to interpret, it is convenient to reduce it to a diagonal
matrix through eigendecomposition methods. Similar methods have been used in
quantitative neuronal analyses, for example singular value decomposition (Lewis
et al., 2002) or principal component analysis (Haas et al., 2007). To make this
reduction, it is important to note that the matrix Qvc obtained by row flipping
Qvc (f1, f2) = Bvc (−f1, f2) is actually Hermitian. As a consequence, Qvc can be
reduced to a diagonal matrix D such that Qvc = P ∗DP where P is a unitary ma-
trix and P ∗ its complex conjugate transpose. The unitary matrix P contains no
information about the magnitude and can be viewed as a kind of generalization of
the phase. The magnitude is entirely encoded in the diagonal matrix D. The ele-
ments of D are called eigenvalues. Each column of the matrix P ∗ is a special vector
called an eigenvector, whose coordinates are expressed relatively to the stimulus
frequencies. The amplitude of each eigenvalue indicates the relative contribution
of the corresponding eigenvector to the quadratic response. Practically, the eigen-
decomposition can be interpreted as a reduction of the quadratic neuronal function
to a set of quadratic filters in which eigenvalues play the role of amplitudes.
The responses at the stimulus frequencies are shown in Figure 1A as interpolated
black points. The color coded points illustrate the smaller amplitude current re-
sponses at the quadratic frequencies. At low frequencies both the linear and non-
linear responses are quite comparable despite the fact that the standard deviation
of the imposed voltage clamp stimulus was only 2.85 mV. In Figure 1B, each cell
(i, j) of the QSA matrix encodes the magnitude of the corresponding quadratic
interaction (fi, fj).
Informally, the QSA matrix can be viewed as a quadratic
generalization of the admittance, thus it is an intrinsic characterization of the
8
measured neuronal response. The interpolations were performed by the MATLAB
command GRIDDATA (linear method) in order to represent the responses in 3D
color plots over a continuous range of frequencies. As a characteristic function,
the QSA matrix is independent of the stimulus amplitude, however the current re-
sponses become proportionally insignificant as the stimulus approaches zero. The
responses of higher orders are insignificant for stimulation amplitudes below ±3
mV.
The QSA matrix, Qvc, is the core mathematical object which encodes the total
effect of the pairwise interactions at the interactive frequencies. Each matrix cell is
located at the intersection of two input frequencies. In this way, the one dimensional
admittance function Y (f ) is generalized to a two dimensional quadratic function
Qvc (f1, f2). The QSA along with linear analysis can be used to reconstruct the
signal using Y and Qvc. Figure 1B shows that the maximum amplitude for the
interactive frequencies occurs at the intersection of 2 Hz and 10.4 Hz (f1 + f2 ).
The eigendecomposition of these data strongly suggests that depolarized type D
neurons are dominated by a single eigenvalue as illustrated in Figure 1C. Even when
two eigenvalues are required to adequately describe the response, the eigendecom-
position provides a remarkably compact representation of the nonlinear response
that otherwise can only be quantitatively described by complex differential equa-
tions involving large numbers of dendritic compartments. It would appear that the
neuronal function becomes more complex, in the sense of information processing,
as the number of significant eigenvalues increases.
Matrix summation
The previous matrix reduction has the great advantage of being reversible, such
that from P and D one can exactly recover Qvc. There exists a coarser simplifi-
cation by summing the QSA matrix by columns to obtain a vector indexed by the
stimulation frequencies. The summation is defined as follows :
(cid:88)
R (j) =
Qvc (i, j)
i
The values of R are illustrated in Figure 1D. The advantage of the R functions
is that they can be presented as classical Bode plots. Moreover, each R function
can be intuitively interpreted as a measure of the influence of each individual
stimulation frequency on the nonlinear responses involving its interaction with all
other stimulation frequencies. Hence, matrix summation is especially well suited to
superimpose and compare with the piecewise quadratic analysis for different steady
state responses.
Figure 2A shows a reconstruction of the measured currents with the first and second
order responses. The contribution of the nonlinear frequencies can be quantitatively
9
Fig. 1: Voltage clamp measurement at -50 mV on a type D neuron.
(A) Amplitude of the evoked linear current responses (black points with interpolation line) and
nonlinear frequency components (colored points). The stimulation show an increasing amplitude
as the frequency increases. The nonlinear interactions are shown as individual points for fi + fj
in red, fi − fj in blue, 2fi in green where fi and fj are positive.
(B) Three dimensional plot of the magnitudes of the interpolated QSA matrix for a continuous
range of frequencies. Each axis is indexed by 8 negative followed by 8 positive frequencies, which
are shown as a 16x16 square matrix.
(C) Eigenvalues of the QSA matrix, illustrating the dominance of one eigenvalue. The abscissa
is shown with numbers indicating each of the 2N eigenvalues by decreasing magnitudes.
(D) Summation R (j) showing the influence of each stimulation frequency in the quadratic
responses.
The voltage command was a multsinusoidal stimulation, which had a standard deviation of 2.85
mV. For all plots, the frequency components were computed with the MATLAB command FFT
divided by the number of points. The current was measured in nA and the voltage in mV. The
stimulation was constructed from the nonoverlapping set (0.2, 0.8, 2, 3.4, 5.8, 10.4, 13.4, 17.8)
in Hz.
10
Fig. 2: Nonlinear and linear responses of the voltage clamped type D neuron of Figure 1 .
(A) Superposition of the averaged data with the first (green) and second order (red) responses
to describe the total neuronal response (blue).
(B) Superposition of only first order responses for four traces.
(C) Superposition of only second order responses for four traces which provides an indication of
the synchrony and reproducibility of the second order components in the data for significantly
large amplitudes.
(D) Linear impedance interpolated from the stimulating frequencies.
11
expressed by comparing the ratios of the spectral energy of the linear versus linear
+ quadratic responses to the total spectral energy for the entire range of frequen-
cies. Since the maximum stimulus frequency was 17.8 Hz, a frequency range of 36
Hz > 2 x 17.8 was selected for the total sum. Clearly, the quadratic reconstruction
is much more accurate than the linear one, which is confirmed by evaluating the
signal energy, 61% for the linear analysis against 96% for the quadratic analysis.
The discrepancy between the second order ratio and 100% shows how well a second
order approximation fits the total response, as well as an indication of the presence
of other higher order responses for the frequency range selected. Higher order re-
sponses clearly are more prominent with a depolarization due to the augmentation
of the nonlinear behavior of both second and higher order responses.
Since the second order responses have relatively low amplitudes it is essential to
ensure that the contributions of other noise sources, such as synaptic events or
membrane ion channel fluctuations, do not significantly contribute to the observed
responses at the harmonic and interacting frequencies. The synchrony and repro-
ducibility of the data is shown by superimposing first order (Figure 2B, delta1) and
second order (Figure 2C, delta2) computed responses for four sequential measure-
ments. A time invariance correlation function was used to determine an optimal
stimulus amplitude, which is large enough to overcome the spontaneous noise of
the neuron and not too great to evoke significant higher order responses (Magnani
and Moore, 2010). Finally, the usual linear impedance, Z (f ) =
, is shown in
Figure 2D.
V (f )
I(f )
Results
Type D neurons
It is difficult to get accurate measurements of the quadratic responses in current
clamp for the type D neurons, mainly due to uncontrolled spontaneous oscillations.
Thus, voltage clamp experiments were done to control the oscillations and measure
the negative current associated with the persistent sodium conductance gNaP. The
nonlinear responses evoked by 5-10 mV voltage clamp stimuli can also be blocked
(not shown) by riluzole as described in an earlier steady state piecewise linear
analysis of gNaP (Idoux et al., 2008).
Figure 3A shows the increased linear impedance magnitude evoked by progressive
depolarizations. Since gNaP can be seen as a negative conductance, its activation
by depolarized membrane potentials reduces the total conductance of the cell,
which leads to an increase of impedance. The nonlinear responses are indicated by
eigenvalues (Figure 3B) and R summation (Figure 3C), whose magnitudes clearly
12
increase with depolarization. For this neuron, the spectral energy analysis led to
ratios 99% and 100% at -60 mV, and 77% and 98% at -50 mV for the linear versus
linear + quadratic responses, respectively.
Type B neurons
It has been shown previously that type B neurons have a prominent gNaP, which
often leads to net inward current carried by Na+ for a limited range of voltage
clamped depolarized membrane potentials (Idoux et al., 2008). Voltage clamped
data in Figure 4 illustrate that type B nonlinear responses are significantly en-
hanced by depolarization and in addition, often show a resonant enhancement of
the impedance. The impedance shows a maximum at an intermediate depolariza-
tion and a shift to a higher resonance frequency with further depolarization. In
contrast, Figures 4B and C show that both the eigenvalues and the R summation
values increase monotonically with depolarized membrane potentials. Interestingly,
the number of significant eigenvalues required to describe the nonlinear response
is generally two or more, unlike the single eigenvalue usually needed for type D
neurons. For this neuron, the spectral energy analysis led to ratios 99% and 100%
at -60 mV, and 82% and 96% at -41 mV for the linear versus linear + quadratic
responses, respectively.
Effect of gNaP and NMDA activation
Nonlinear responses are likely to be enhanced by the activation of dendritic NMDA
receptors, which would occur during synaptic activity of the neural integrator net-
work. NMDA activation could trigger dendritic bistable responses that contributes
to maintaining a particular firing rate after an input impulse. Figures 5B and C
show that the addition of NMDA clearly enhances nonlinear response of a type
D neuron consistent with the trigger hypothesis. The mean value of R and the
maximum eigenvalue for all neuronal types increased significantly with the addi-
tion of NMDA. TTX reduces the nonlinear effects of NMDA to control values or
less, however it is still capable of inducing potential oscillations in current clamp
(not shown). Control (with gNaP, without NMDA) versus NMDA + TTX treated
neurons (without gNaP, with NMDA) did not show a significant p-value. Thus,
NMDA and gNaP are synergistic in their action, where the combined effects are
generally greater than either alone. Since the normal physiological activation of
NMDA receptors is due to transient synaptic currents, the non-inactivation of gNaP
and membrane potential bistability contribute to the maintenance of a depolarized
potential. In conclusion gNaP appears to be essential for sustained nonlinear ef-
fects induced by NMDA activation and thus would be critical for the operation of
13
Fig. 3: Voltage clamp experiments on a PHN type D neuron show the effect of an activated gNaP by
comparing both the linear and quadratic voltage clamp responses at -70, -60, -55, -50 mV (blue,
red, green, and black respectively).
(A) Linear impedance, however it should be noted that the stimulus is a controlled voltage that
leads to a current response. The depolarization markedly increases the impedance.
(B) Eigenvalues of the QSA matrix. The abscissa is shown with numbers indicating each of the
2N eigenvalues by decreasing magnitudes.
(C) Magnitude of R summations of the corresponding QSA matrix. They also increase during
a depolarization.
Statistically in five neurons the maximum eigenvalue increased for a change of membrane poten-
tial of -60 to -50 mV from 0.002±0.001 to 0.010±0.001 with p-value=0.00004. The mean value
of R±SD increased from 0.005±0.003 to 0.023±0.004 with p-value=0.0003.
14
Fig. 4: Voltage clamped type B neuron at -75, -60, -46, -41 mV (blue, red, green, black, respectively).
(A) Linear impedance showing resonance induced by the depolarization.
(B) Multiple eigenvalues increasing with depolarization in contrast to a single dominant eigen-
value observed for type D neurons. The abscissa numerically labels the eigenvalues by decreasing
magnitude.
(C) Magnitude of R summations.
These plots show that nonlinear responses are present at both hyperpolarized and depolarized
membrane potentials.
Statistically in six neurons, for a positive membrane potential change of 5 mV in the range -60
to -40 mV, the maximum value of the eigenvalue increased from 0.004±0.003 to 0.014±0.009
with p-value=0.008 and the mean value of R±SD increased from 0.009±0.006 to 0.03±0.02 with
p-value=0.01.
15
the neural integrator.
Type D and B neuronal model simulations
In order to provide an interpretation of these experimental results, numerical sim-
ulations have been done using previously published models of types D and B for
rat PHN neurons (Idoux et al., 2008). These models have been constructed from
a piecewise linear analysis to fit parameters of nonlinear differential equations in
voltage clamp. Unless otherwise indicated, both models had three uniformly dis-
tributed voltage dependent conductances, gK, gNaP and gH with a soma and eight
dendritic compartments. Typically the gH of type D neurons is quite small and
could be neglected.
Simulations done with the published average parameter values were consistent with
the voltage clamp data, however some appropriate modifications of the average
parameters were done for the comparison with the data for the individual neurons
of Figures 3 and 4. Four potentials are shown to cover the range observed in the
experiments. The type D data and model generally show a dominant eigenvalue
(see -50 mV in Table 1 and Figure 6).
The simulations of the type B model in Figure 7 show behavior similar to the data
of Figure 4, showing two or more significant eigenvalues compared to an essentially
single dominant eigenvalue for the type D model (Figure 6). However Tables 1
and 2 indicate that the number of significant eigenvalues for both neuronal types
is dependent on both the membrane potential and specific parameter values. For
example, at -40 mV with or without NMDA, Figure 5 B shows a type D neuron
with multiple significant eigenvalues (also, see Table 1).
In addition, the monotonic increase in the R summation of the type B model non-
linear responses with depolarization contrasts with the peaking of linear impedance
increase similar to that observed in the type B neuron of Figure 4. In general the
impedance increase with depolarization is caused by the activation of the gNaP neg-
ative conductance that balances the other positive conductances. An impedance
maximum, often with a resonance, occurs because the impedance decreases with
further depolarization due to an increased gK.
The nonlinear responses of type B model simulations, in contrast to type D, show
that the quadratic response is greatly reduced in magnitude at -60 mV, in part
because the type B model has a greater density of voltage dependent gNaP con-
ductances on a relatively more compact dendritic structure than found for type
D neurons. In addition, the nonlinear responses in the type B model, as in the
data of Figure 4, are enhanced at hyperpolarized potentials directly due to the gH
conductance.
The use of nonoverlapping frequencies is required for the construction of QSA
16
Fig. 5: Effect of NMDA (red) and NMDA+TTX (green) compared to control (blue) on a voltage
clamped type D neuron at -40 mV.
(A) Effect of 25 µM NMDA on linear impedance. The impedance decreases with NMDA (red)
and increases with 2 µM TTX (green) compared to the control (blue)
(B) The NMDA plots (red) illustrate that the nonlinear responses are increased by the presence
of NMDA. Moreover, TTX (green), which blocks gNaP, when added to NMDA shows decreased
eigenvalues. The abscissa numerically labels the eigenvalues by decreasing magnitude.
(C) Effect of NMDA and NMDA+TTX on R summation. The R values are in inverse order to
the impedance magnitudes.
Statistically, pooled data from five neurons (three of type B and two of type D) in the presence
versus absence of 25-50 µM NMDA over the range -66 to -40 mV, showed a significant difference
in the mean values of R summation (p-value = 0.026). Similarly the maximum eigenvalues were
significantly different with a p-value of 0.035. TTX reduces the nonlinear effects of NMDA to
control values or less (p-value = 0.09). An insignificant p-value of 0.2 was found comparing
control with NMDA + TTX treated neurons.
17
Fig. 6: Model of voltage clamped type D PHN neuron. Simulations were done at -70, -60, -55, -50 mV
(blue, red, green, black, respectively).
(A) Linear impedance. The depolarization markedly increases the linear impedance.
(B) Eigenvalues of the QSA matrix. The abscissa numerically labels the eigenvalues by decreas-
ing magnitude.
(C) Magnitude of R summations of the corresponding QSA matrix.
The following parameter values were used: csoma = 20.5 pF; aratio = 3.77; elength = 0.54;
gleak = 1.2 nS(1.37 nS); vleak = -53 mV; gk = 1.18 nS; vn = -32 mV; sn = 0.06(mV)−1 (0.05
(mV)−1); tn = 10 msec; vk = -87 mV; gH = 0; sm = 0.06 (mV)−1; tm = 150 µsec; vNa = 77
mV; vm = -38 mV (-35 mV); gNaP= 0.9 nS (0.6 nS). Parameter values shown in parentheses
are default values defined in Idoux et al., 2008.
18
Fig. 7: Model of voltage clamped type B neuron. Simulations were done at -76, -60, -46, -41 mV (blue,
red, green, black, respectively).
(A) Linear impedance. The magnitude shows a resonant maximum at -46 mV.
(B) Eigenvalues of the QSA matrix. The abscissa numerically labels the eigenvalues by decreas-
ing magnitude.
(C) Magnitude of R summations of the corresponding QSA matrix.
The parameter values were as follows: csoma = 0.0000995 µF (0.0000265); aratio = 2.6 (2.85);
elength = 0.31 (0.37); gleak = 0.001297 µS (0.000878); vleak = -56.61 µS (-54.9); gk = 0.005555
µS (0.0024); vn = -38.7 mV (-35); sn = 0.07715/mV (0.05); tn = 0.66 sec (.09); vk = -87 mV;
gH = 0.00644 µS (0.00591); vq = -79.7 mV (-63.9); sq = -0.0667/mV (-0.0647); tq = 1.75 sec
(8.43); vk2 = -43 mV; gNaP = 0.00147 µS (0.00131); vm = -35 mV; sm = 0.05/mV (.056); tm
= 0.00015 sec; vnmda = 77 mV; Parameter values shown in parentheses are default values from
Idoux et al., 2008.
19
soma-Na dendrite-Na Vm Rmean
µ1
µ2
µ2/µ1(%)
1
1
0
0.01
1
1
1
0
1
0.01
-40
-50
-50
-50
-50
0.0162 +0.0160
0.0095
0.0011 +0.0011
0.0075
0.0012
-0.0049
-0.0095 +0.0015
-0.0002
-0.0074 +0.0012
-0.0012 +0.0003
31
16
18
16
25
Tab. 1: Type D model simulations. The voltage clamped somatic membrane potential is given along
with presence or absence of conductances indicated by 1 or 0. The value of gNaP in either
the soma or all dendritic compartments was reduced to 1% indicated by 0.01 in the labeled
columns. The mean values of R summation and the two largest eigenvalues µ1 and µ2 are given
for each simulation. The last column is µ2 represented as the percentage of the absolute µ1
value. A positive µ1 occurs when gK has a greater effect than gNaP.
matrices from experimental measurements. Nevertheless, it is possible to do a
coarse interpolation for other frequencies as plotted in Figure 8. The interpolated
QSA matrices for type D and B neurons show striking differences, which also occur
in their mathematical models. Figures 8A and 1B show local peaks in the QSA plots
at two membrane potentials for the same type D neuron, which are comparable
to the type D model simulation in Figure 8B. By contrast the type B neuron and
model (Figure 8C, D) show a prominent low frequency peak distinctly different
than observed in type D neurons. The QSA matrix clearly provides significantly
more information about the neuronal behavior than linear parameters and has the
possiblilty of being extended to all frequencies of interest.
In order to test the hypothesis that the dendritic conductances are mainly respon-
sible for the nonlinear responses, simulations were done with gNaP reduced in the
soma and unchanged in the dendrite or the converse as indicated in Tables 1 and
2.
In type D neurons, decreasing the soma gNaP has a lesser effect (79%) on
the R summation compared with decreasing the dendritic gNaP (13%). Simula-
tions of type B model neurons showed a lesser effect for a gNap reduction in the
more compact dendrite (34%). Such simulations support the hypothesis that the
nonlinear responses measured under voltage clamp conditions are dominated by
dendritic responses, while the linear responses are determined by both the soma
and dendrite. Since the somatic region has good voltage space clamp control, one
would expect a dominant linear current response from the somatic conductances
at all membrane potentials, however the significantly larger dendritic membrane
area is not space clamped leading to uncontrolled voltage excursions and greater
nonlinearites. In conclusion, these simulations suggest that nonlinear responses in
voltage clamped neurons are dominated by active dendritic structures, when their
electrotonic lengths are above 0.3.
soma-Na dendrite-Na Vm Rmean
0.0140
0.00652
0.00104 +0.0018
0.00467
0.00221
-0.0141 +0.0027
-0.0063 +0.0018
-0.0008
-0.0044 +0.0020
-0.0019 +0.0015
1
1
0
0.01
1
µ1
µ2
1
1
0
1
0.01
-40
-50
-50
-50
-50
20
µ2/µ1(%)
19
29
44
45
79
Tab. 2: Type B model simulations as in Table 1. The mean values of R summation and the two largest
eigenvalues µ1 and µ2 are given for each simulation. When gNaP in the dendrite is reduced,
the negative µ1 decreases and becomes as large as the unchanged positive µ2 (79%), which
suggests structurally different quadratic filters for type B compared to type D models.
Fig. 8: Interpolated QSA plots of type D and B neurons. (A) type D neuron of Figure 3 at -50 mV.
(B) Type D model of Figure 6 at -50 mV. (C) Type B neuron of Figure 4 at -46 mV. (D) Type
B model of Figure 7 at -46 mV.
21
Discussion
The neurons of the prepositus hypoglossi nucleus (PHN) provide a useful system to
investigate nonlinear behavior, such as persistent activity to maintain eye position.
The oscillatory character of some of these neurons is similar to that observed in the
stellate neurons in layer IV of entorhinal cortex (Haas and White, 2002; Schreiber
et al., 2004), which are also involved in the processing of orientational information.
PHN neurons are part of the brainstem that receives head velocity signals and
integrates them to control eye position for the stabilization of an image at the center
of the visual field during head rotation. This specific processing is called neural
integration (Aksay et al., 2007) due to an analogy with integration in mathematical
calculus.
In order to understand how the PHN neural network can perform neural integra-
tion, it is important to understand the biophysical properties of individual neurons
involved in the circuitry. Single neurons of the PHN show oscillatory and bistable
nonlinear properties that are likely to be involved in the operation of the neural
integrator. Since models based on recurrent excitation, even including lateral in-
hibition, are not sufficiently robust, a number of theoretical papers have suggested
that the nonlinear properties of individual neurons are essential for the network be-
havior of the neural integrator (Koulakov et al., 2002; Goldman et al., 2003). Thus,
the finding that oscillatory nonlinear behavior is clearly present in the neurons in-
volved in the eye movement circuitry (Idoux et al., 2006) lends strong support
to these theoretical notions. Dendritic characteristics (Johnston and Narayanan,
2008) are potentially critical for the function of these neurons, which in addition
have also shown to be different for the two main classes of PHN neurons (Idoux et
al., 2008).
Thus, type D oscillatory neurons have bistability properties, which are consistent
with neural integrator models that rely on remote dendritic processing (Goldman et
al., 2003; Idoux et al., 2008). The experiments and analysis in this paper strongly
support the hypothesis that type D neurons have persistent sodium channels in
their dendrites, which would promote remote bistable potentials shifts leading to
persistent activity. Type B neurons have an even higher density of persistent
sodium channels and their passive electrotonic length is less than type D. As a
consequence, the type B neurons at moderate depolarizations would have a more
uniform potential throughout the less isolated dendritic tree.
The gNaP conductance in the dendrites can easily be activated by the synaptic
stimulation of the NMDA receptors likely to occur during normal physiological
activity. In addition, NMDA activation enhances total nonlinearities, which are
shown for a type D neuron and also observed in type B neurons (see Figure 5). In
current clamp conditions, both types of neurons show marked potential oscillations
22
in the presence of 25-50 µM NMDA (Idoux et al., 2006). Clearly the level of NMDA
activation expected during synaptic activation would be much less, however it is
likely to be sufficient to evoke significant nonlinear responses due to the gNaP in
both type B and D neurons. In this regard, the neural integrator model of Koulakov
et al. (2002), depends on NMDA synapses that activate gNaP dependent bistable
states in dendritic compartments.
The quantitative measurement of the biophysical properties of intact neurons is
seriously compromised by the inherent inability to voltage clamp the electrotonic
structure of the dendritic tree. In general whole cell measurements are restricted
to patch clamp electrodes placed in the soma, which makes it difficult to infer
the remote properties of the dendrites. Previous piecewise linear analyses have
permitted the development of realistic neuronal models, however it has been diffi-
cult to separate the properties of the dendrites from the soma unless patch clamp
electrodes can be placed in the dendrites. The new approach described in this pa-
per takes advantage of the space clamp problem associated with voltage clamping
neurons. The quadratic analysis has shown that it is possible to characterize the
nonlinear behavior of the uncontrolled dendritic membrane voltage responses while
maintaining voltage clamp control of the somatic membrane. When the dendritic
electrotonic structure is remote with relatively large surface areas compared to the
soma, the nonlinear behavior of the dendritic membrane dominates the soma. This
is especially the case in a voltage clamp experiment because of the lack of dendritic
potential control.
It was found that responses in the range of ±(5-10) mV could be well described
by quadratic nonlinearities suggesting that nonlinearities of higher degrees only
add marginal improvement. Thus, the quadratic response is likely to sufficiently
capture most of the nonlinear behavior of neuronal systems except for extremely
large synaptic inputs. The quadratic functions are quite sensitive to the mean
membrane potential and appear to be valid for a range of sinusoidal inputs. This
behavior extends significantly the validity of the quantitative quadratic description.
The quadratic functions are computed on particular sets of nonoverlapping frequen-
cies, which can be interpolated over a continuous range of frequencies as illustrated
in Figures 1B and 8. Thus, they provide a significantly concise description of the
neuronal behavior and could potentially be used as computational devices that
would be independent of nonlinear differential equations. Practically, this could be
an alternative approach to large scale neural network simulations.
The estimation of the parameters of both the voltage dependent conductances and
the electrotonic structure have shown quantitative differences between type B and
D neurons (Idoux et al., 2008). In addition, the nonlinear analysis in this paper
suggests that the number of significant eigenvalues is greater for the type B versus
the type D neurons and their individual models. Thus, the measured nonlinear-
23
ities seems to be structurally different between type B and D, namely the two
corresponding types of quadratic functions are intrinsically different (Figure 8). In
general the dominant eigenvalue was negative, which is related to the negative slope
conductance due to gNaP. Type D model simulations at large depolarizations (see
Table 1 at -40 mV) show that the maximum eigenvalue (µ1) is positive, consistent
with a positive slope conductance due to an increased outward potassium current.
In contrast, the greater gNaP of the type B model maintains a negative µ1 at -40
mV (Table 2). Both type D and B models show positive µ1 values if gNaP is totally
removed from the soma and dendrite.
In conclusion, the work described in this paper provides a novel way to concisely
quantify the fundamental nonlinearities underlying of individual neurons. By allow-
ing rigorous comparison of any neuronal model with the behavior of real neurons,
it makes possible to show that nonlinear responses in voltage clamp are dominated
by the active dendritic structure. A determination of the molecular basis of the
eigenvalue analysis should provide a better understanding of how neurons use their
remarkable nonlinear properties in information processing.
References
Aksay E, Olasagasti I, Mensh BD, Baker R, Goldman MS, Tank DW (2007) Func-
tional dissection of circuitry in a neural integrator. Nat Neurosci 10:494 -- 504.
Boyd S, Tang Y, Chua L (1983) Measuring volterra kernels. IEEE Transactions
on Circuits and Systems 30:571 -- 577.
Erchova I, Kreck G, Heinemann U, Herz AVM (2004) Dynamics of rat entorhinal
cortex layer ii and iii cells: characteristics of membrane potential resonance at
rest predict oscillation properties near threshold. J Physiol 560:89 -- 110.
Fishman HM, Poussart DJ, Moore LE, Siebenga E (1977) K+ conduction descrip-
tion from the low frequency impedance and admittance of squid axon. J Membr
Biol 32:255 -- 290.
FitzHugh R (1983) Sinusoidal voltage clamp of the hodgkin-huxley model. Biophys
J 42:11 -- 16.
French AS (1976) Practical nonlinear system analysis by wiener kernel estimation
in the frequency domain. Biological Cybernetics 24:111 -- 119.
Goldman MS, Levine JH, Major G, Tank DW, Seung HS (2003) Robust persis-
tent neural activity in a model integrator with multiple hysteretic dendrites per
neuron. Cereb Cortex 13:1185 -- 1195.
24
Gutkin BS, Ermentrout GB (1998) Dynamics of membrane excitability determine
interspike interval variability: a link between spike generation mechanisms and
cortical spike train statistics. Neural Comput 10:1047 -- 1065.
Haas JS, Dorval AD, White JA (2007) Contributions of ih to feature selectivity in
layer ii stellate cells of the entorhinal cortex. J Comput Neurosci 22:161 -- 171.
Haas JS, White JA (2002) Frequency selectivity of layer ii stellate cells in the
medial entorhinal cortex. J Neurophysiol 88:2422 -- 2429.
Hodgkin AL, Huxley AF (1952) A quantitative description of membrane current
and its application to conduction and excitation in nerve. J Physiol 117:500 -- 544.
Idoux E, Eugene D, Chambaz A, Magnani C, White JA, Moore LE (2008) Con-
trol of neuronal persistent activity by voltage-dependent dendritic properties. J
Neurophysiol 100:1278 -- 1286.
Idoux E, Serafin M, Fort P, Vidal PP, Beraneck M, Vibert N, Muehlethaler M,
Moore L (2006) Oscillatory and intrinsic membrane properties of guinea pig
nucleus prepositus hypoglossi neurons in vitro. J Neurophysiol 96:175 -- 196.
Izhikevich EM (2002) Resonance and selective communication via bursts in neurons
having subthreshold oscillations. Biosystems 67:95 -- 102.
Johnston D, Narayanan R (2008) Active dendrites: colorful wings of the mysterious
butterflies. Trends Neurosci 31:309 -- 316.
Koulakov AA, Raghavachari S, Kepecs A, Lisman JE (2002) Model for a robust
neural integrator. Nat Neurosci 5:775 -- 782.
Lewis ER, Henry KR, Yamada WM (2002) Tuning and timing of excitation and
inhibition in primary auditory nerve fibers. Hear Res 171:13 -- 31.
Magnani C, Moore LE (2010) Quadratic sinusoidal analysis of voltage clamped
neurons. arXiv:1007.0858v1.
Marmarelis PZ, Naka KI (1973) Nonlinear analysis and synthesis of receptive-field
responses in the catfish retina. i. horizontal cell leads to ganglion cell chain. J
Neurophysiol 36:605 -- 618.
Moore LE, Chub N, Tabak J, O'Donovan M (1999) Nmda-induced dendritic
J Neu-
oscillations during a soma voltage clamp of chick spinal neurons.
rosci 19:8271 -- 8280.
25
Moore LE, Fishman HM, Poussart DJ (1980) Small-signal analysis of k+ conduc-
tion in squid axons. J Membr Biol 54:157 -- 164.
Murphey CR, Moore LE, Buchanan JT (1995) Quantitative analysis of electrotonic
structure and membrane properties of nmda-activated lamprey spinal neurons.
Neural Comput 7:486 -- 506.
Schetzen M (2006) The Volterra and Wiener Theories of Nonlinear Systems Krieger
Publishing Company.
Schreiber S, Erchova I, Heinemann U, Herz AVM (2004) Subthreshold resonance
explains the frequency-dependent integration of periodic as well as random stim-
uli in the entorhinal cortex. J Neurophysiol 92:408 -- 415.
Vervaeke K, Hu H, Graham LJ, Storm JF (2006) Contrasting effects of the persis-
tent na+ current on neuronal excitability and spike timing. Neuron 49:257 -- 270.
Victor J, Shapley R (1980) A method of nonlinear analysis in the frequency domain.
Biophys J 29:459 -- 483.
|
1701.01311 | 3 | 1701 | 2017-01-09T14:21:38 | Extracting the Groupwise Core Structural Connectivity Network: Bridging Statistical and Graph-Theoretical Approaches | [
"q-bio.NC",
"cs.DM"
] | Finding the common structural brain connectivity network for a given population is an open problem, crucial for current neuro-science. Recent evidence suggests there's a tightly connected network shared between humans. Obtaining this network will, among many advantages , allow us to focus cognitive and clinical analyses on common connections, thus increasing their statistical power. In turn, knowledge about the common network will facilitate novel analyses to understand the structure-function relationship in the brain. In this work, we present a new algorithm for computing the core structural connectivity network of a subject sample combining graph theory and statistics. Our algorithm works in accordance with novel evidence on brain topology. We analyze the problem theoretically and prove its complexity. Using 309 subjects, we show its advantages when used as a feature selection for connectivity analysis on populations, outperforming the current approaches. | q-bio.NC | q-bio | Extracting the Groupwise Core Structural
Connectivity Network: Bridging Statistical and
Graph-Theoretical Approaches
Nahuel Lascano1,2, Guillermo Gallardo1, Rachid Deriche1, Dorian Mazauric3,
and Demian Wassermann1
1 Athena EPI, Universit´e Cote d'Azur, Inria, France
2 Computer Science Department, FCEyN, Universidad de Buenos Aires, Argentina
3 ABS EPI, Universit´e Cote d'Azur, Inria, France
Abstract. Finding the common structural brain connectivity network
for a given population is an open problem, crucial for current neuro-
science. Recent evidence suggests there's a tightly connected network
shared between humans. Obtaining this network will, among many ad-
vantages, allow us to focus cognitive and clinical analyses on common
connections, thus increasing their statistical power. In turn, knowledge
about the common network will facilitate novel analyses to understand
the structure-function relationship in the brain.
In this work, we present a new algorithm for computing the core struc-
tural connectivity network of a subject sample combining graph theory
and statistics. Our algorithm works in accordance with novel evidence
on brain topology. We analyze the problem theoretically and prove its
complexity. Using 309 subjects, we show its advantages when used as a
feature selection for connectivity analysis on populations, outperforming
the current approaches.
Keywords: Group-wise connectome, core graph problem, brain connec-
tivity, diffusion MRI
1
Introduction
Isolating the common brain connectivity network from a population is a main
problem in current neuroscience [3,7,10]. Recent evidence suggests that there's a
common and densely connected brain connectome across humans [2]. In this work
we present a new approach for selecting these common connections, combining
recent topological hypotheses [2] and current methods [7,10].
Finding the common brain connectome across subjects has the potential
to increase our understanding of the relationship between function and struc-
ture in the brain. This relationship is one of the main open questions in neu-
roscience [3,5]. Moreover, knowledge about the most common connections in a
population will facilitate clinical and cognitive Diffusion MRI analyses by re-
ducing the number of surveyed connections, increasing the statistical power of
those analyses. Finding the common connectome will also allow us to increase
7
1
0
2
n
a
J
9
]
.
C
N
o
i
b
-
q
[
3
v
1
1
3
1
0
.
1
0
7
1
:
v
i
X
r
a
2
N. Lascano et al.
our knowledge about the brain structure by comparing core networks across
different populations.
We formalize the problem of selecting the common connections combining
graph theory and statistics. Then, we prove that the problem is NP-Hard and
propose a polynomial-time algorithm to find approximate solutions. To do this,
we develop an exact polynomial-time algorithm for a relaxed version of the
problem and prove the algorithm's correctness and complexity.
Currently, the most used algorithm to extract a population's core structural
connectivity network (CSNC) [7] uses an statistical approach: first, compute a
connectivity matrix for each subject; then, analize each connection separately
with a hypothesis test, using as null hypothesis that that edge is not present in
the population; finally, construct a binary graph with the edges for which the
null hypothesis was rejected. The main problem of Gong et al.'s [7] algorithm
is that the resulting graph can be a set of disconnected subgraphs. Moreover,
recent studies have shown that the brain has a core network tightly connected
and a sparsely connected outer one [2]. In other words, this approach ignores
the resulting network's topology. Performing statistical analyses in a feature set
chosen by hypothesis testing incurs in the double dipping problem [8].
A newer approach to solve the CSNC problem, designed by Wassermann et
al. [10], uses graph theory to get a connected CSCN: first, compute a binary
connectivity graph for each subject using a threshold; for each possible connec-
tion compute the "cost" of including or excluding it from the common graph by
evaluating in how many subjects that connection is present; finally, construct
the binary graph with all the edges that is "cheaper" to include than to ex-
clude and connect the resulting graph if it's disconnected, using the minimum
possible cost. This algorithm guarantees that the resulting graph is connected,
but the connection binarization discards significant information for the resulting
common network. In other words, it discards information of the probability of
each connection being in the brain. This is problematic because the resulting
graph may include edges for which tractography assigned a very low existence
probability across subjects. Also, the outer part of the brain, the connections
which do not result in the core network, should also be sparsely connected [2],
which this algorithm does not enforce.
In this work we propose, for the first time, a polynomial-time algorithm to ob-
tain the CSCN of a population addressing the issues listed above. Our algorithm
combines the recent graph-theoretical approach [10] with the statistical aware-
ness of the most popular one [7]. We start by formalizing the problem, which
allow us to prove that it's NP-Hard. Then, we propose a first algorithm that
solves a relaxed version of the problem in an exact way, giving the best possible
core graph for our formalization. Then, we adapt it to guarantee a connected
result, agreeing with recent evidence on structural connectivity network topol-
ogy [2, e.g.]. Finally, we validate our approach using 300 subjects from the HCP
database and comparing the performance of the networks obtained by our new
approach, Wassermann et al.'s [10] and Gong et al.'s [7] predicting connectivity
values from handedness in the core network.
Extracting the Groupwise Core Structural Connectivity Network
3
2 Definitions, Problems and Contributions
We want to develop a new algorithm to extract the core structural connectivity
network, a problem that implies working with different brains. Thus, the first
thing we need to do is unify them into a common connectivity model. This allows
us to model all brains with graphs in which each node represents a cortical or sub-
cortical region, and each edge represents a white matter connection between two
regions. We choose the Desikan parcellation [4] to uniformize the brain cortical
and sub-cortical regions across subjects.
To compute the connectivity matrices we use a probabilistic tractography al-
gorithm, which outputs one matrix per subject. The resulting matrices represent
the existence probability of a connection across parcels in each subject [5]. As
these are symmetric, we interpret the matrices as weighted undirected graphs,
sharing the node set across subjects.
Formally, we represent a sample of N brain structural networks by N com-
plete weighted graphs G1 = (V, E, w1), . . . , GN = (V, E, wN ) with a common
node set V . We call G1, . . . , GN the sample graphs. Each graph Gi corresponds
to a subject. Each vertex v ∈ V represents a cortical or sub-cortical region.
Each edge e ∈ E = V × V represents a white matter bundle connecting two
regions. Finally, the weight wi(e) is the connection probability for the edge e in
the subject i obtained through tractography:
w1(e), w2(e), . . . , wN (e) ∈ [0, 1] ∀e ∈ E.
(1)
Note that all graphs have the same ordered node set and all of them are com-
plete: an edge weight, or connection probability, wi(e) of 0 represents an absent
connection. Using this formalization we express the general core structural con-
nectivity network problem as follows: find a core graph G∗ = (V ∗, E∗) densely
connected such that G∗ keeps the more relevant connections E∗ ⊆ E in the
sample and discards the less relevant ones, for some definition of relevance and
density. For simplicity, once we select E∗ we can define V ∗ as
V ∗ = {v ∈ V : ∃u ∈ V, (u, v) ∈ E∗} ,
(2)
the set of nodes that the edges in E∗ cover. Then, we can reduce the problem of
finding G∗ to find E∗ alone.
We want a formalization of relevance that represents the probability that
a connection is present across subjects. Thus, we choose to model the group-
wise relevance, w∗(e) as the mean existence probability across subjects, factored
by the standard deviation of these probabilities. In other words, w∗(e) is the
number of standard deviations that the mean existing probability of connection
e is larger than 0. We use w∗(e) as a statistical measure of edge presence across
the population. Formally,
where w(e) (cid:44) N(cid:88)
wi(e)
N
, s(e) (cid:44)
i=1
w∗(e) (cid:44) w(e)
s(e)
(cid:16)
(cid:118)(cid:117)(cid:117)(cid:117)(cid:116) N(cid:88)
i=1
(cid:17)2
w(e) − wi(e)
N
.
(3)
(cid:80)
e∈E∗ w∗(e)
E∗
(cid:80)
e∈E\E∗ w∗(e)
E∗
.
.
(4)
(5)
α(w∗, E∗) (cid:44)
β(w∗, E∗) (cid:44)
4
N. Lascano et al.
Note that w∗(e) is the statistic of a hypothesis z-test which assumes a media
of 0 for the population weight of e. We choose the z-statistic because of the
normal distribution's properties, e.g. linearity, even if other distributions, such
as Beta distribution, may be more appropriate for modeling the probability. In
any case, note that for the purpose of our contribution w∗ can be any function
E → R which grows with the relevance of the edges in the sample.
We also want a formalization that represents the density of the core subgraph.
We use the relationship between the number of edges and the total statistical
relevance w∗ that those edges sum:
As we want also a sparse outer subgraph, we also define its density:
Now we can express our objective informally as: choose E∗ such that α(w∗, E∗)
(Eq. 4) is large and β(w∗, E∗) (Eq. 5) small. In accordance to recent evidence
on the core network, we also want G∗ to be connected. Here, connected means
that for every pair of vertices u, v in V ∗ there is a path of edges in E∗ from u
to v.
Let E c be the family of sets of edges that induce a connected graph. We now
formalize the problem of finding this common graph G∗ in two different ways.
• The optimization version consists in computing:
f (w∗, E∗) = λα(w∗, E∗) − (1 − λ)β(w∗, E∗)
(6)
max
E∗∈E c
The parameter λ (between 0 and 1) can be adjusted to weight the density of
the inner and the outer network. Note that if λ = 1, the solution to (6) only
considers the density of the core network, and if λ = 0, it only considers the
edges excluded of the core network.
• Given A and B, the decision version consists in finding E∗ ⊆ E c such that:
α(w∗, E∗) ≥ A
β(w∗, E∗) ≤ B
(7)
Having formalized the Core Structural Connectivity Network into an opti-
mization and a decision problem, we proceed with one of our main theoretical
contributions: proving that the problem is NP-Complete.
2.1 CSCN Problem's NP-Completeness
We have formalized the problem of the Core Structural Connectivity Network
taking into account the density and connectedness of the core subgraph and the
sparsity of the outer one. We will now prove that, with this formalization, the
problem is NP-Complete.
Extracting the Groupwise Core Structural Connectivity Network
5
Definition 1 (Core Structural Connectivity Network problem). Given
G1 = (V, E, w1), G2 = (V, E, w2), . . . , GN = (V, E, wN ) weighted graphs (the
sample graphs) with a common node set, a complete edges set (E = V × V ) and
w1(e), w2(e), . . . , wN (e) ∈ R≥0 ∀e ∈ E weights of their edges, and given A, B
real numbers, find G∗ = (V ∗, E∗) connected graph (the core graph) such that
α(w∗, E∗) ≥ A
β(w∗, E∗) ≤ B
for α and β as defined in Eq. (4) and Eq .(5).
Here we prove that the Core Structural Connectivity Network problem, called
CSCN problem, is NP-complete. In our reduction, we use the Steiner Tree prob-
lem [6], called ST problem in the following. Given an edge-weighted graph
G(cid:48) = (V (cid:48), E(cid:48), w), a subset S ⊆ V (cid:48) of nodes, and a real k ≥ 0, ST problem consists
(cid:80)
in determining if there exists a connected subgraph H such that S ⊆ V (H) and
e∈E(H) w(e) ≤ k. The decision version of ST problem is NP-complete even if
Instance of ST problem. Consider any edge-weighted graph G(cid:48) = (V (cid:48), E(cid:48), w)
2 for every e ∈ E(cid:48). Given k ≥ 0, ST problem consists in de-
such that w(e) = 1
termining if there exists a connected subgraph H such that S ⊆ V (H) and
E(S) ≤ 2k. Without loss of generality, we assume that E(cid:48) ≥ 2k and that G(cid:48)
is connected.
all weights are equal [6].
Reduction. We construct the instance of CSCN problem as follows. Let
s = S and let t ≥ 1 be any positive integer. Let G = (V, E, w∗) defined as
follows. Let V = V (cid:48) ∪ {vi,j 1 ≤ i ≤ s, 1 ≤ j ≤ t} and E = V × V . Let
S = {u1, . . . , us}. For every i, j, 1 ≤ i ≤ s, 1 ≤ j ≤ t, w∗
vi,j ,u = 0
for every u ∈ V \ {ui}. Furthermore, for every e ∈ E(cid:48), set w∗(e) = w(e) = 1
2 ,
and for every u, u(cid:48) ∈ V (cid:48) such that {u, u(cid:48)} /∈ E(cid:48), then set w∗
e = 0. Finally, we set
A = s.t+k
Lemma 1. If E∗ < s.t + 2k, then any solution for CSCN problem is not
admissible because β(w∗, E∗) > B.
Proof. Suppose that E∗ < s.t + 2k. In order to minimize(cid:80)
s.t+2k and B =
= 1 and w∗
2 (E(cid:48)−2k)
vi,j ,ui
s.t+2k
.
1
e∈E\E∗ w∗(e), E∗
must contain {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t} if E∗ ≥ s.t. (Otherwise we select
a subset of this set of edges.) Indeed, by construction of G, we have w∗
= 1
for every i, j, 1 ≤ i ≤ s, 1 ≤ j ≤ t. Then, if E∗ − s.t > 0, E∗ must contain
E∗ − s.t edges of E(cid:48), that is edges of E of weight 1
2 each. Recall that there are
exactly s.t edges of weight 1, and the other edges have weight 0 or 1
2 .
e∈E\E∗ w∗(e) =
There are two cases. First, suppose that E∗ ≥ s.t. We get that(cid:80)
2 (E(cid:48) − (E∗ − s.t)). Since E∗ − s.t < 2k, then we get that(cid:80)
2 (E(cid:48) − (E∗ − s.t)) > 1
2 (E(cid:48)−(E∗−s.t))
(cid:80)
e∈E\E∗ w∗(e) =
2 (E(cid:48) − 2k). Furthermore, since E∗ < s.t + 2k,
. Thus, we proved that β(w∗, E∗) =
2 (E(cid:48)−2k)
E∗
vi,j ,ui
s.t+2k
>
1
1
1
1
we get that
e∈E\E∗ w∗(e)
E∗
> B.
6
N. Lascano et al.
Second, suppose that E∗ < s.t. We get that(cid:80)
e∈E\E∗ w∗(e) = s.t − E∗ +
E(cid:48)
2 (E(cid:48) − (E∗ − s.t)), we obtain the result by the
2 . Since s.t − E∗ +
Finally, if E∗ < s.t + 2k, then there is no admissible solution for CSCN
(cid:3)
arguments described for the first case.
E(cid:48)
2 > 1
problem.
Lemma 2. If E∗ > s.t + 2k, then any solution for CSCN problem is not
admissible because α(w∗, E∗) < A.
Proof. Suppose that E∗ > s.t + 2k. In order to maximize (cid:80)
e∈E∗ w∗(e), E∗
must contain {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t} and E∗ − s.t edges of E(cid:48).
= 1 for every i, j, 1 ≤ i ≤ s,
Indeed, by construction of G, we have w∗
2 for every e ∈ E(cid:48). Recall that there are
1 ≤ j ≤ t. Furthermore, w∗(e) = 1
exactly s.t edges of weight 1, and the other edges have weight 0 or 1
2 . We get
that α(w∗, E∗) = s.t+ 1
s.t+2k = A. Indeed, the average weight is
lower when there are more edges of weight 1
2 (the number of edges of weight 1
is the same in both ratios).
Finally, if E∗ > s.t + 2k, then there is no admissible solution for CSCN
(cid:3)
2 (E∗−s.t)
E∗
< s.t+k
vi,j ,ui
problem.
By Lemma 1 and Lemma 2, we get the following corollary.
Corollary 1. Any solution for CSCN problem is such that E∗ = s.t + 2k.
We prove in Lemma 3 and in Lemma 4 that there is an admissible solution
for CSCN problem if and only if there is an admissible solution for ST problem.
Lemma 3. If there is an admissible solution for ST problem, then there is an
admissible solution for CSCN problem.
e∈E(H) w(e) = 1
graph such that S ⊆ V (H) and(cid:80)
Proof. Suppose there is an admissible solution for ST problem. We prove that
there is an admissible solution for CSCN problem. Let H be a connected sub-
2E(H) ≤ k. If E(H) = 2k,
then set E∗ = E(H) ∪ {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t}. If E(H) < 2k, then
set E∗ = E(H) ∪ {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t} ∪ F , where F ⊆ E(cid:48) such that
F (cid:54)= E(H) = ∅, w∗(e) = 1
2 for every e ∈ F , and such that the graph induced by
E(H) ∪ F is connected. The last condition comes from Corollary 1 in order to
get the right number of edges in E∗. This condition is always possible to satisfy
because G(cid:48) is connected.
The graph induced by E∗ is connected. Indeed, H is an admissible solution for
ST problem, E(H)∪ F is connected by construction, and {{vi,j, ui} 1 ≤ j ≤ t}
is a set of edges all adjacent to ui ∈ S for every i, 1 ≤ i ≤ s.
Furthermore, we get
α(w∗, E∗) =
(cid:80)
e∈E∗ w∗(e)
E∗
=
s.t + k
s.t + 2k
= A
Extracting the Groupwise Core Structural Connectivity Network
7
and
β(w∗, E∗) =
(cid:80)
e∈E\E∗ w∗(e)
E∗
Finally, we proved that there is an admissible solution for CSCN problem. (cid:3)
1
2 (E(cid:48) − 2k)
s.t + 2k
=
= B.
Lemma 4. If there is an admissible solution for CSCN problem, then there is
an admissible solution for ST problem.
1
2 (E(cid:48)−2k)
s.t+2k = B.
Proof. Suppose there is an admissible solution for CSCN problem. We prove
that there is an admissible solution for ST problem. Let E∗ ⊆ E be such that
the graph induced by E∗ is connected, and such that α(w∗, E∗) ≥ s.t+k
s.t+2k = A
and β(w∗, E∗) ≤=
We first prove that {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t} ⊆ E∗. By Corollary 1,
e∈E∗ w∗(e) ≥
s.t + k. By construction of G, the set of edges of weight 1 is {{vi,j, ui} 1 ≤ i ≤
2} = {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t}.
s, 1 ≤ j ≤ t}, that is {e ∈ E w∗(e) > 1
We get that {{vi,j, ui} 1 ≤ i ≤ s, 1 ≤ j ≤ t} ⊆ E∗ because, otherwise, we
we know that E∗ = s.t + 2k. Thus, it necessarily means that(cid:80)
would have(cid:80)
we would have(cid:80)
e∈E∗ w∗(e) < s.t + k.
e∈E∗ w∗(e) < s.t + k.
Furthermore, every e ∈ E∗ ∩ E(cid:48) is such that w∗(e) = 1
2 . Indeed, otherwise,
Finally, since E∗ is an admissible solution for CSCN problem, then it means
that the graph induced by the set of edges E∗ ∩ E(cid:48) is connected and is such that
for every ui, 1 ≤ i ≤ s, then there is an edge e ∈ E∗ ∩ E(cid:48) that is adjacent to
ui. By the previous remark, every edge in E∗ ∩ E(cid:48) has weight 1
2 . Thus, it means
that there is E∗ ∩ E(cid:48) is an admissible solution for ST problem considering the
(cid:3)
graph G(cid:48). Indeed E∗ ∩ E(cid:48) = 2k and so(cid:80)
e∈E∗∩E(cid:48) w(e) = k.
We are now able to prove the NP-completeness of CSCN problem.
Theorem 1. CSCN problem is NP-complete.
Proof. The reduction is clearly polynomial. Furthermore, Lemma 3 and Lemma 4
prove the equivalence between CSCN problem and ST problem. Since the deci-
sion version of ST problem is NP-complete even if all weights are equal [6], then
we obtain the NP-completeness of the decision version of CSCN problem. (cid:3)
In Theorem 1 we have proved that CSCN problem is NP-complete. Hence, to
be able to solve it in reasonable time we need a relaxation to make it tractable
or an approximate algorithm for the complete version. In this article we will
propose both.
2.2 Relaxation of the CSCN problem
We proved in the previous section that the connectivity constraint is the main
reason of the difficulty of the problem. Without it, it becomes tractable. So
we solve, in this section, a relaxed version of problem without the connectivity
constraint. Then, we use this solution to approximate the full problem.
8
N. Lascano et al.
Theorem 2. The decision version of CSCN problem without the connectivity
constraint is in P.
Algorithm 1 Maximum edges
Compute w∗(e) for each e ∈ E
Sort(E)
for each e ∈ E do
E∗ ← E∗ ∪ e
if α(w∗, E∗) > A and β(w∗, E∗) < B then
return T rue
end if
end for
return F alse
(cid:46) sorts edges by w∗ non-increasingly
Proof. Algorithm 1, in each step i, defines E∗ as the i maximum weighted edges
and tries to use that to fulfill the constraints.
Assume that there exists an E∗ that fulfills the constraints. There are two
cases: 1) E∗ has the E∗ maximum weighted edges, 2) there are ej ∈ E∗, ek ∈
E \ E∗ such that w∗(ek) ≥ w∗(ej). In 1), Algorithm 1 will find E∗. In 2), let
E(cid:48) = (E∗ ∪ {ek}) \ {ej} another subset of E. Then
α(w∗, E(cid:48)) =
e∈E(cid:48) w∗(e)
E(cid:48)
=
e∈E(cid:48) w∗(e)
E∗
≥
e∈E∗ w∗(e)
E∗
= α(w∗, E∗) ≥ A
because the edges in E∗ are the same as the ones in E(cid:48) except from one that has
a larger weight. For the same reason,
(cid:80)
(cid:80)
(cid:80)
(cid:80)
≤
(cid:80)
e∈E\E(cid:48) w∗(e)
(cid:80)
e∈E\E(cid:48) w∗(e)
E(cid:48)
=
E∗
β(w∗, E(cid:48)) =
e∈E\E∗ w∗(e)
E∗
= β(w∗, E∗) ≤ B.
Thus, we found a new subset of E that stills fulfills the constraints. We can do
the same process with E(cid:48) (replace an edge with another one of larger weight)
iteratively, always getting subsets that fulfills the constraints, until we cannot do
this anymore. At that point we will have a subset that has only the maximum
E∗ edges and fulfills the constraints. Thus, algorithm 1 will find this subset.
We now need to prove algorithm 1 runs in polynomial time in the size of E.
The first operation, computing w∗(e) for each e, implies computing the mean
and standard deviation for each edge across the population, which can be done
in O(N ) per edge (where N is the size of the population). This is O(N ∗E) for
all the edges. The second step, sorting, can be done in O(E log E).
The main loop runs at most E times, and in each loop it adds an edge to
E∗, computes α and β and performs two comparisons. The comparisons can be
done in constant time, as the addition to E∗ if we use a linked list of edges to
represent it. To compute α and β it is needed to iterate once again E (the part
Extracting the Groupwise Core Structural Connectivity Network
9
in E∗ for α, the part in E \ E∗ for β) adding the weights together and then
performing two divisions. This can be done in linear time in the size of E, and
even quicker (constant time) if we optimize it by keeping the values of α and β
across loops and updating them with the weight of the edge that changed sets.
Then, algorithm 1 solves the CSCN problem in O(max(E2,E ∗ N ) or in
O(E ∗ N ) if a little optimization is used.
(cid:3)
2.3 Heuristic approach
In Section 2.2 we developed Algorithm 1 to solve the problem of finding the Core
Structural Connectivity Network in polynomial time. However, this algorithm
does not guarantee a connected result. We solve the original problem, presented
in Section 2, by first applying Algorithm 1 and then modifying the resulting
core graph G∗ to guarantee its connectedness. This results in an approximate
solution for the full problem computable in polynomial time.
To extend G∗ into a connected graph we add the necessary edges while de-
creasing the minimum possible the objective function f defined in Eq. 6. For
this, we use the same approach that Wassermann et al. [10]. Namely, we make a
multigraph Gcc with the connected components of G∗ as nodes, complete it with
all the possible edges between those connected components, and run a Maximum
Spanning Tree algorithm. This selects the edges needed to produce a connected
subgraph with the maximum possible weight. For the full details, see Wasser-
mann et al. [10]. This way we get a connected subgraph close to the best possible
subgraph, which we obtained using Algorithm 1.
3 Experiments and Results
Fig. 1. Core structural connectivity network computed by our approach. On the left,
we show the adjacency matrix for λ = 0.5, where 48.19% of the connections were
included in the CSCN. In the central and right panels, we show the resulting CSCN
for λ = 0.9 and 0.99 respectively. The percentage of included connections in the CSCN
is 5.99% and 1.27% respectively.
We formalized the CSCN problem in section 2 and designed an algorithm to
solve it in section 2.3. Now we will asses the performance of our method. For
LRLR 10
N. Lascano et al.
this, we compare it with the most used [7] and with the recent one [10] in the
task of connectivity prediction performance.
We use a subset of the HCP500 dataset [9]: all subjects aged 21-40 with
complete dMRI protocol, totaling 309. We compute the weighted connectivity
matrices between the cortical regions defined by the Desikan atlas [4] as done
by Sotiropoulos et al. [9]. Examples of CSCN exctracted with our algorithm at
different λ levels are shown in Fig. 1, which was generated using Nilearn [1].
3.1 Consistency of the Extracted Graph
To compare the stability across different algorithms for CSCN, we use an analysis
based on Wassermann et al. [10]: we randomly take 500 subsets of 100 subjects
each and computed the core graphs for all subsets. We then compute the number
of unstable connections: connections that present in at least one core graph but
not in all of them. Finally, in Table 1 we report each algorithm's stability:
stability of the algorithm (cid:44) 1 − #{unstable connections}
#{total connections}
.
This measure quantifies the CSCN consistency across subsamples. Due to the
homogeneity of our sample, we expect the CSCNs obtained by an algorithm to
be similar across subsamples. Hence, a stabler algorithm is preferable.
3.2 Predicting Handedness-Specific Connectivity
We evaluate performance of the methods by using the generated core graphs as
a feature selection for handedness specific connectivity. We use a nested Leave-
3 -Out procedure: the outer loop performs model selection on 1
1
3 of the subjects
using the core graph algorithm and the inner loop performs model fitting and
prediction using the selected features.
Table 1. Stability of the algorithms and amount of features selected by linear regression
in the core graph relating the weights with handedness. Our procedure gets more
features selected than Gong et al. [7] and Wassermann et al. [10], showing better
statistical power. It's also more stable than Gong et al., showing improved consistency.
Algorithm
Gong et al. 2009
Wassermann et al. 2016
Our approach (λ = 0.50)
Features (mean) Features (std) Stability
0.066
0.415
1.042
0.256
0.723
1.269
0.364
0.644
0.528
Specifically, we first take 1
3 subjects randomly and compute the core graph for
those subjects using the three different algorithms. Then we add the weights for
the selected edges for each subject, and select the features F that are more deter-
minant of handedness using a linear least-squares regression and the Bonferroni
Extracting the Groupwise Core Structural Connectivity Network
11
Fig. 2. Performance of core network as feature selection for a linear model for handed-
ness specific connectivity. We evaluate model prediction (left) and fit (right) for Gong
et al. [7] in green, Wassermann et al. [10] in blue and ours, in red. We show the his-
tograms from our nested Leave- 1
3 -Out experiment. In both measures, our approach has
more frequent lower values than Gong et al., showing a better performance.
correction for multiple hypotheses. This experiment is repeated 500 times. We
quantify the amount of features that are selected after each procedure, which
indicates how useful is the core graph algorithm for selecting the edges related
to handedness. We show the results in Table 1.
To evaluate the prediction, we randomly take 1
2 of the remaining subjects
and fit a linear model on the features F to predict connectivity weights using
the handedness of each subject. Finally, we predict the values of the features
F from the handedness in the subjects left out. We quantify the quality of the
linear model fitting Akaike Information Criterion (AIC) and of the prediction
performance with the mean squared error (MSE) of the prediction. For both
measures a lower value indicates better performance. The outer loop is performed
500 times and the inner loop 100 times per outer loop, which totals 50,000
experiments. We show the results of this experiments in Fig. 2.
4 Discussion and Conclusion
We presented for the first time a polynomial algorithm to extract the core struc-
tural connectivity network of a population combining a graph-theoretical ap-
proach with statistic relevance of the connections, observing the recent evidence
of the structural network topology.
Our results show that our algorithm outperforms, in the prediction experi-
ment, the most used technique [7] as well as latest approaches [10]. In Table 1
we can see that our algorithm preserves, in average, more connections correlated
with the handedness of the subjects. We can also see that despite being less
012345Value1e40.00.20.40.60.81.0Frecuency1e4MSE25002000150010005000Value0123456Frecuency1e3AICWassermann et al. 2016Gong et al. 2009Our approach (λ = 0.50) 12
N. Lascano et al.
stable than Wassermann et al.'s it is stabler than Gong et al.'s. Finally, Fig. 2
shows that, in the handedness prediction experiment, our method outperforms
Gong et al.'s and Wassermann et al's: the number of cases with lower AIC and
MSE is larger in our case. Hence, our CSCN is better as linear model relating
connectivity with handedness in terms of model fitting and prediction.
In terms of theoretical contributions, we formalized the problem, proved its
difficulty and gave a novel algorithm for dealing with it. We then validated our
approach by showing its power as feature selector for getting connections related
to handedness with 300 real subjects' data. The experiment shows our method
performs better than the currently available. Moreover, our method avoids the
double dipping problem by not choosing the feature set with hypothesis testing.
Acknowledgements Authors acknowledge funding from ERC Advanced Grant
agreement No 694665 : CoBCoM - Computational Brain Connectivity Mapping
References
1. Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi,
J., Gramfort, A., Thirion, B., Varoquaux, G.: Machine learning for neuroimaging
with scikit-learn. Frontiers in neuroinformatics 8(February), 14 (2014)
2. Bassett, D.S., Porter, M.A., Wymbs, N.F., Grafton, S.T., Carlson, J.M., Mucha,
P.J.: Robust detection of dynamic community structure in networks. Chaos 23(1)
(2013)
3. Bullmore, E.T., Sporns, O., Solla, S.A.: Complex brain networks: graph theoretical
analysis of structural and functional systems. Nature reviews. Neuroscience 10(3),
186–98 (2009), http://www.ncbi.nlm.nih.gov/pubmed/19190637
4. Desikan, R.S., S´egonne, 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.: An automated labeling system for subdividing the human cerebral cortex on
MRI scans into gyral based regions of interest. NeuroImage 31(3), 968–980 (2006)
5. Donahue, C.J., Sotiropoulos, S.N., Jbabdi, S., Hernandez-Fernandez, M., Behrens,
T.E., Dyrby, T.B., Coalson, T., Kennedy, H., Knoblauch, K., Van Essen, D.C.,
Glasser, M.F.: Using Diffusion Tractography to Predict Cortical Connection
Strength and Distance: A Quantitative Comparison with Tracers in the Monkey.
Journal of Neuroscience 36(25), 6758–6770 (jun 2016), http://www.jneurosci.
org/cgi/doi/10.1523/JNEUROSCI.0493-16.2016
6. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman & Co., New York, NY, USA (1979)
7. Gong, G., He, Y., Concha, L., Lebel, C., Gross, D.W., Evans, A.C., Beaulieu, C.:
Mapping anatomical connectivity patterns of human cerebral cortex using in vivo
diffusion tensor imaging tractography. Cerebral Cortex 19(3), 524–536 (2009)
8. Kriegeskorte, N., Simmons, W.K., Bellgowan, P.S.F., Baker, C.I.: Circular analysis
in systems neuroscience: the dangers of double dipping. Nature Neuroscience (2009)
9. Sotiropoulos, S.N., Jbabdi, S., Xu, J., Andersson, J.L., Moeller, S., Auerbach, E.J.,
Glasser, M.F., Hernandez, M., Sapiro, G., Jenkinson, M., Feinberg, D.A., Yacoub,
E., Lenglet, C., Van Essen, D.C., Ugurbil, K., Behrens, T.E.: Advances in diffusion
MRI acquisition and processing in the Human Connectome Project. NeuroImage
80(3), 125–143 (oct 2013)
10. Wassermann, D., Mazauric, D., Gallardo Diez, G.A., Deriche, R.: Extracting
the Core Structural Connectivity Network: Guaranteeing Network Connectedness
Through a Graph-Theoretical Approach. In: MICCAI 2016 (2016)
|
1908.03514 | 1 | 1908 | 2019-08-09T16:13:12 | A practical guide to methodological considerations in the controllability of structural brain networks | [
"q-bio.NC",
"q-bio.QM"
] | Predicting how the brain can be driven to specific states by means of internal or external control requires a fundamental understanding of the relationship between neural connectivity and activity. Network control theory is a powerful tool from the physical and engineering sciences that can provide insights regarding that relationship; it formalizes the study of how the dynamics of a complex system can arise from its underlying structure of interconnected units. Given the recent use of network control theory in neuroscience, it is now timely to offer a practical guide to methodological considerations in the controllability of structural brain networks. Here we provide a systematic overview of the framework, examine the impact of modeling choices on frequently studied control metrics, and suggest potentially useful theoretical extensions. We ground our discussions, numerical demonstrations, and theoretical advances in a dataset of high-resolution diffusion imaging with 730 diffusion directions acquired over approximately 1 hour of scanning from ten healthy young adults. Following a didactic introduction of the theory, we probe how a selection of modeling choices affects four common statistics: average controllability, modal controllability, minimum control energy, and optimal control energy. Next, we extend the current state of the art in two ways: first, by developing an alternative measure of structural connectivity that accounts for radial propagation of activity through abutting tissue, and second, by defining a complementary metric quantifying the complexity of the energy landscape of a system. We close with specific modeling recommendations and a discussion of methodological constraints. | q-bio.NC | q-bio |
A practical guide to methodological considerations in the controllability
of structural brain networks
Teresa M. Karrera,b, Jason Z. Kimb,*, Jennifer Stisoc,*, Ari E. Kahnc, Fabio Pasqualettid, Ute
Habela,e,f, and Danielle S. Bassettb,g,h,i,j,k,l
aDepartment of Psychiatry, Psychotherapy and Psychosomatics, Faculty of Medicine, RWTH Aachen, Germany.
bDepartment of Bioengineering, School of Engineering & Applied Science, University of Pennsylvania, Philadelphia, PA 19104, USA.
cDepartment of Neuroscience, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104, USA.
dDepartment of Mechanical Engineering, University of California, Riverside, CA 92521, USA.
eJARA - Translational Brain Medicine, Aachen, Germany.
fInstitute of Neuroscience and Medicine: JARA-Institute Brain Structure Function Relationship (INM 10), Research Center Jlich,
gDepartment of Physics and Astronomy, College of Arts & Sciences, University of Pennsylvania, Philadelphia, PA 19104, USA.
hDepartment of Neurology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104, USA.
iDepartment of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104, USA.
jDepartment of Electrical and Systems Engineering, School of Engineering & Applied Science, University of Pennsylvania,
Jlich, Germany.
Philadelphia, PA 19104, USA.
kSanta Fe Institute, Santa Fe, NM 87501, USA.
lTo whom correspondence should be addressed: [email protected]
*These two authors contributed equally.
Abstract
Predicting how the brain can be driven to specific states by means of internal or external control
requires a fundamental understanding of the relationship between neural connectivity and activity.
Network control theory is a powerful tool from the physical and engineering sciences that can provide
insights regarding that relationship; it formalizes the study of how the dynamics of a complex system
can arise from its underlying structure of interconnected units. Given the recent use of network control
theory in neuroscience, it is now timely to offer a practical guide to methodological considerations in the
controllability of structural brain networks. Here we provide a systematic overview of the framework,
examine the impact of modeling choices on frequently studied control metrics, and suggest potentially
useful theoretical extensions. We ground our discussions, numerical demonstrations, and theoretical
advances in a dataset of high-resolution diffusion imaging with 730 diffusion directions acquired over
approximately 1 hour of scanning from ten healthy young adults. Following a didactic introduction
of the theory, we probe how a selection of modeling choices affects four common statistics: average
controllability, modal controllability, minimum control energy, and optimal control energy. Next, we
extend the current state of the art in two ways: first, by developing an alternative measure of structural
connectivity that accounts for radial propagation of activity through abutting tissue, and second, by
defining a complementary metric quantifying the complexity of the energy landscape of a system. We
close with specific modeling recommendations and a discussion of methodological constraints. Our
hope is that this accessible account will inspire the neuroimaging community to more fully exploit
the potential of network control theory in tackling pressing questions in cognitive, developmental, and
clinical neuroscience.
Keywords: network neuroscience, control theory, structural connectivity, diffusion imaging.
1
1
Introduction
The brain is a complex system of interconnected units that dynamically transitions through diverse ac-
tivation states supporting cognitive function [1]. Understanding the mechanisms and processes that give
rise to these trajectories through state space is crucial for intervening in disease to restore cognitive
functioning [2]. One relevant factor enabling such rich neural dynamics is the network architecture of the
underlying structural substrate [3 -- 5]. Yet, the exact mechanisms by which the physical architecture of
the brain both supports and constrains its function remain largely unknown [6 -- 8].
Recent advances in network control theory offer a formal means to study how the temporal dynam-
ics of a complex system emerges from its underlying network structure [9 -- 11]. Applying this theory to
the brain requires that one first builds a network model in which brain regions (nodes) are anatomically
connected to one another (edges) [12, 13]. The state of the brain network system is then reflected in the
pattern of neurophysiological activity across network nodes, and state trajectories represent the temporal
sequence of brain states that the system traverses [14, 15]. With definitions of the network and its state
in hand, we can consider the problem of network controllability, which in essence amounts to asking how
the system can be driven to specific target states by means of internal or external control input [16]. In
the context of the brain, such input can intuitively take the form of electrical stimulation [17 -- 21], task
modulation [22 -- 24], or other perturbations from the world or from different portions of the body [25, 26].
Practically, network control theory and its associated toolkit enables us to study the general role of brain
regions in controlling neural dynamics in diverse scales and species [27 -- 30], and in both health [31, 32]
and disease [33, 34] or injury [35]. Moreover, the approach can be used to determine the patterns of input
required to induce specific state transitions necessary for behavior [18, 22, 35, 36].
Network control theory offers three primary advantages over traditional approaches to the study of
brain network function. First, the multi-modal nature of the theoretical framework explicitly enforces
a simultaneous study of brain structure and function, in contrast to approaches that characterize each
separately and then assess statistical covariance. Second, network control theory exceeds the often purely
descriptive approach of network science [37 -- 39] by building a generative model parameterized by both a
network's spatial features and its temporal features [40]. The model then offers predictions of the brain's
response to both endogenous and exogenous input signals. In the case of the former, the model could
hypothetically prove useful in understanding how the brain enacts cognitive control to reach task-relevant
cognitive states [22, 24, 31]. In the context of the latter, the model could similarly prove useful in in-
forming neuromodulation for the treatment of neurological and psychiatric disorders [40]. Third, initial
studies applying network control theory in neuroscience demonstrate that network controllability is a
useful marker of brain dynamics, from quantifying the capacity of different brain regions to alter whole-
brain dynamics [31], over demonstrating that this capacity grows with development [41], to finding that
controllability is linked to executive functioning [22, 24, 42]. Moreover, applications of the theory to data
collected during invasive neuromodulation regimens demonstrate the utility of the theory in predicting
response to electrical stimulation in practice [18, 21] and in theory [17].
In light of the promising applicability of network control theory in neuroscience, we wish to provide
a systematic overview of how the framework can be used to study the controllability of neural dynamics.
This primer is constructed so as to offer neuroscientists some basic intuitions regarding the foundational
concepts, and to guide them through the necessary prerequisites and considerations. For a more technical
2
introduction that nevertheless remains heavily motivated by neuroscience, we refer the interested reader
to [43, 44]; and for further information about the underlying mathematics (which remains agnostic to
the application domain), we refer the reader to [9, 11]. Because the application of network control theory
can be formulated in several ways, we systematically probe how diverging theoretical assumptions and
possible modeling choices influence controllability metrics and the estimated energy of state transitions.
For example, we consider discrete and continuous time systems, methods for system stabilization, the
time horizon for control, and the set of control nodes. We complement these studies with specific rec-
ommendations for best practices, which depend in no small part upon the nature of the neuroscientific
question being investigated. To further stimulate research in this exciting field, we suggest a few useful
extensions of the theoretical framework, such as alternative estimates of structural connectivity and a
complementary metric that quantifies the complexity of the energy landscape.
2 Theoretical framework
2.1 Network control theory
The core of the theoretical framework is the structural network of neurons (or larger neural units) in
the brain that allows the activity of a brain region to diffuse and change the activity of connected brain
regions (Fig. 1A). Here, we introduce a mathematical model that describes the natural dynamics of a
complex linear system (Fig. 1B). Formally, the temporal evolution of network activity is modeled as a
linear function of its connectivity:
x = Ax(t),
(1)
where x(t) is a vector of size N × 1 that represents the state of the system. Here we operationalize the
system's state to reflect the magnitude of the neurophysiological activity of the N brain regions at a
single point in time. Over time, x(t) denotes the state trajectory, which is the temporal sequence of
states or activity patterns that is traversed by the system. The adjacency matrix A is of size N × N ,
and denotes the relationships between the system elements. Here, we operationalize that relation as the
structural connectivity between each pair of brain regions.
Next, we extend this model to account for controlled dynamics, which occur when the brain is induced
to deviate from its natural trajectory by the injection of internal or external input signals (Fig. 1C). In
this case, the temporal dynamics of a system additionally depends on the control energy injected into a
set of nodes across time
x = Ax(t) + Bκuκ(t).
(2)
Here, Bκ is a matrix of size N × m that denotes the set of m control nodes or brain regions into which we
wish to inject inputs. This matrix consists of m indicator vectors, each having set only the i-th element to
1, corresponding to a control node. If we control all brain regions, Bκ corresponds to the N × N identity
matrix with ones on the diagonal and zeros elsewhere. If we control only a single brain region i, Bκ reduces
to a single N × 1 vector with a one in the i-th element and zeros elsewhere. The term uκ(t) is a vector
of m functions of size m × 1 denoting the control input, which is the amount of input injected into each
of the m control nodes at each time point t. Over time, uκ(t) denotes the injected control input over time.
For the interested reader, we wish to provide a few mathematical intuitions that might facilitate a deeper
understanding of the presented concepts. By Eq. 1, the structure of the network determines its dynamic
evolution over time. Mathematically, the structural connectivity matrix A serves as linear operator that
3
maps each state, x, to the rate of change from that state, x. This linear transformation can be described
in terms of the evolutionary modes of the system consisting of the N eigenvectors of A and their associ-
ated eigenvalues (Fig. 1D). Each eigenvector of A can be imagined as an axis of the linear transformation
which remains invariant over time. Thus, the eigenvectors reflect directions in the state space along which
the system independently moves, each characterized by a specific pattern of brain region activity. Each
eigenvalue, in turn, determines the rate of growth or decay along its associated eigenvector; that is, each
eigenvalue determines how slow or fast the system grows or decays in the direction defined by the eigen-
vector. Thus, the eigenvalues control the temporal persistence of the set of supported modes of activity.
Especially for the interpretation of results, it is important to keep in mind that the dynamic model
is relatively simple and relies on the assumptions of linearity, time invariance, and freedom from noise.
Linearity implies that the system evolves linearly over time which is not an accurate reflection of extended
dynamics in most neural processes. However, it has been shown that non-linear dynamics can be locally
approximated by linear dynamics [45, 46]. Time invariance implies that the systems response does not
depend on the time point because both the structural network A and the control set Bκ are constant
over time. This assumption likely holds true for short time scales but could be challenged by long-term
structural reorganization, which has been observed across development and adulthood [41, 47, 48]. Free-
dom from noise implies that all properties of signal propagation are accounted for deterministically by
the model. Yet, noise is a feature of neural signals at both small [49 -- 51] and large time scales [52, 53].
Nevertheless, it is customary and reasonable when first developing a mathematical model of a complex
system to consider the salient features of the model that do not depend on noise [44, 54, 55].
2.2 Prerequisites
The core of the dynamic model is the network structure that enables activity changes within a particular
brain region to diffuse and induce state changes in connected regions. Thus, the first step is to build a
structural connectivity network by defining the weighted adjacency matrix A (Fig. 1A). The structural
network of human and nonhuman animals can be modeled using a range of spatial scales of neural units
and physical links between them [13]. Here, we focus on the construction of the human connectome
which requires (i) a brain parcellation that defines the N nodes of the network and (ii) diffusion imaging
data that define the strength of structural connectivity Ai,j between two brain regions i and j. To avoid
self-loops in the system, we set the diagonal of A to zero. As a further practical note, the sparse nature
of human connectomes typically does not require any thresholding of the matrix.
The next step is to pick a time-system that best reflects the neural dynamics under study. Here, we
consider two options: discrete and continuous. A discrete-time system assumes that the system evolves
in discrete time steps whereas a continuous-time system models continuously changing dynamics. Neural
processes can often not be clearly assigned to one of these categories. The spatial scale of the neural
unit under study is one factor that could guide this choice [44]. Smaller spatial scales such as single
neurons could be modeled as either firing or not, and the accompanying uniform delays in neural activity
could be better approximated by discrete-time systems. Larger neural units such as brain regions usually
contain more neurons. The ensuing more heterogeneous timing of neural state changes might be bet-
ter represented by continuous-time systems. Depending on the choice, the modeled dynamics can differ
substantially because of their distinct mathematical implementation. More concretely, discrete-time dy-
namics rely on difference equations whereas continuous-time systems are based on differential equations.
Note that we exclusively present formulas for continuous-time systems in the main text; the discrete-time
4
Figure 1: Schematics of network control theory and relevant concepts. (A) Structural brain network construction.
Brain atlas and diffusion imaging data define the nodes and edges of the structural connectivity matrix A. (B) Natural
dynamics of the brain. The temporal evolution of brain states, such as the magnitude of neurophysiological activity across
brain regions, is modeled as a linear function of brain structure. The area under the curve illustrates the impulse response
and thus, average controllability of brain region #4. (C) Controlled dynamics. The state trajectory additionally depends
on control input injected into the system. The control input matrix Bκ determines the nodes into which a control signal
uκ is injected (yellow flash) over time. The area under the curve of the control energy signals corresponds to the control
energy required by the given state transition.
(D) Activity modes of a system. The structural connectivity matrix A
can be decomposed into N eigenvectors and eigenvalues that determine the system's dynamics. Eigenvectors determine the
supported modes of activity; eigenvalues determine the rate of decline of their associated mode. Brain region i's controllability
vi,j of mode j corresponds to a projection of the jth eigenvector onto the dimension spanned by brain region i. (E) Control
energies and controllability metrics. (Left) Control energy for specific state transitions. Here we illustrate the minimum
control energy required to drive the brain from a specific initial state to a specific target state using a particular control node
set. The optimal control energy additionally constrains the size of the state trajectory. (Right) Control strategies potentially
examining all possible state transitions (dashed arrows). Average controllability has been previously described as a brain
region's ability to control nearby states that require little energy. Modal controllability has been previously described as a
brain region's ability to control distant states that require more energy. (F) Controllability metrics and control energies can
be relevant on an individual and regional level. To examine both levels separately, we will summarize statistics across brain
regions and individuals, respectively. For consistency between two parameter choices such as discrete- and continuous-time
systems, we will calculate the Pearson correlation between individual (regional) values extracted from one parameter choice
and those extracted from the second parameter choice. G) Complexity of the energy landscape. The landscape of possible
minimum control energy trajectories is determined by the eigenvalues of the inverse of the controllability Gramian Wκ,T .
We used the variability of the eigenvalues to quantify the heterogeneity of the energy landscape. Abbreviations: IQR,
interquartile range.
versions can be found in the Supplementary Formulas.
5
The third step is to choose a method to stabilize the system to avoid its infinite growth over time.
Because extremely large brain states are neurobiologically implausible, we normalize the system such
that it either approaches the largest supported mode of activity or goes to zero over time:
Anorm =
A
λ(A)max + c
− I
(3)
Here, I denotes the identity matrix of size N × N , and λ(A)max denotes the largest eigenvalue of the
system. To normalize the system, we must specify the parameter c, which determines the rate of stabi-
lization of the system. If c = 0, the largest mode of activity is stable and all other modes decay; thus,
the system approaches the largest mode over time. If c > 0, such as the commonly used choice c = 1, all
modes decay; thus, the system goes to zero over time. As will become clear in the next section, the latter
variant can be especially useful for the computation of average controllability in infinite time as well as
modal controllability due to its mathematical definition.
2.3 Optimal control energy
To quantify the degree of controllability of a network, we consider an optimal control problem to steer
the network from a specific initial state x(0) = x0 to a specific target state x(T ) = xT over the time
horizon T while minimizing a combination of both the length of the state trajectory and the required
control energy [35, 56, 57]. Formally, we consider the problem
u(t)∗
κ = argmin
uκ
J(uκ) = argmin
uκ
0
((xT − x(t))(cid:62)(xT − x(t)) + ρuκ(t)(cid:62)uκ(t))dt,
(4)
where the parameter ρ determines the relative weighting between the costs associated with the length
of the state trajectory and input energy. We use the cost function J(u(t)∗
κ) to find the unique optimal
control input u(t)∗
κ which allows us to calculate the optimal control energy (Fig. 1E) required by a single
brain region i (Fig. 1C):
(cid:90) T
(cid:90) T
i (t)(cid:107)2
(cid:90) T
(cid:107)u∗
0
2dt,
E∗
i =
N(cid:88)
i=1
(5)
(6)
and in total
E∗ =
E∗
i =
u∗
κ(t)(cid:62)u∗
κ(t)dt.
0
To calculate optimal control energy, we must specify an initial brain state x0 and a target brain state xT
by assigning each brain region an initial and target activity level. If available, we can extract regional
activity values directly from functional neuroimaging data such as electrocorticography or magnetic res-
onance imaging [18, 58, 59], or we can use model-based estimates of task-related activation such as β
values from a general linear model [42]. Otherwise, we can also model brain states by artificially defining
a subset of nodes to be active, such as brain regions belonging to the same cognitive system [22, 35, 36].
Additionally, we must specify the control set Bκ, a set of brain regions into which we wish to inject
signals. Theoretically, this choice can vary from controlling a single region to controlling the full brain.
The choice of small- to medium-sized control sets, however, can lead to large numerical instabilities that
accumulate and bias the results. As a rule of thumb, it is advisable to ensure that the numerical error
does not exceed 10−6. To reduce the numerical error of the calculation, we can also define a relaxed
control set Bκ by allowing large control input to control regions and small, random inputs to all other
brain regions [18]. We must also specify the time horizon T over which the control input is effective. For
6
pragmatic reasons such as the potential translation to real external brain stimulation, the time horizon is
usually set to finite time. Note that the time horizon is measured in arbitrary units even if brain states
are defined by functional imaging data. Finally, we must specify the time step dt, which should be small
enough to sufficiently approximate the dynamics [35]; a reasonable choice is dt = 0.001.
The cost function J is motivated by the fact that biological systems might constrain the features of
the traversed states, such as their type, diversity, or magnitude. Transitioning through states not too far
away from the target state is supposed to avoid extremely large and thus neurobiologically implausible
brain state transitions. In the case where no specific assumptions are made on the relative importance of
the two constraints and where both the distance and energy values are of a comparable scale, an equal
weighting of ρ = 1 is a reasonable choice. Depending on our neurobiological assumptions, we can also
define alternative cost functions and potentially restrict them to a subset of brain regions [22].
2.4 Minimum control energy
A specific and commonly used subform of optimal control energy is obtained by letting ρ → ∞ in (4), so
that the cost function J accounts only for the energy of the control input to steer the network from an
initial state x(0) = x0 to a target state x(T ) = xT . Thus, we call this metric minimum control energy
(Fig. 1E). To compute the minimum control energy for a given network, it is convenient to define the
controllability Gramian as
(cid:90) T
0
Wκ,T =
eAtBκB(cid:62)
κ eA(cid:62)tdt.
(7)
The eigenvalues of Wκ,T can be used to answer several questions regarding the controllability of a net-
work. First, if the smallest eigenvalue of Wκ,T is zero, then the network is not controllable. That is, there
exist final states xT that cannot be reached by any control input, independent of its energy. Second, the
magnitude of the smallest eigenvalue of Wκ,T is inversely proportional to the largest energy needed to
reach a final state. That is, there exists a final state xT that can be reached only using inputs whose
energy is at least proportional to the inverse of the smallest eigenvalue of Wκ,T . The foundational papers
[31, 60] have shown that brain networks are controllable from any single region; that is, the smallest
eigenvalue of Wκ,T is greater than zero. However, brain networks require very large control energy; that
is, the smallest eigenvalue of Wκ,T can be extremely small. It should also be noted that the computation
of the smallest eigenvalue of Wκ,T tends to be numerically difficult, which motivates the next metric.
2.5 Average controllability
Apart from examining specific state transitions, the theoretical framework also allows us to ask questions
regarding the general role of brain regions in controlling neural dynamics. A third metric is obtained by
measuring the average input energy required to drive the system via a specified set of control nodes to
all possible target states xT with unit norm [61, 62]. Following the above discussion, this metric equals
T race(W −1
κ,T ) where the inverse of the controllability Gramian is a map from target states to control
energy. To avoid numerical difficulties when controlling only a few nodes in very large systems, the met-
ric is often approximated as T race(Wκ,T ). Using this last form, our definition of average controllability
measures the ability of a network to amplify and spread control inputs, rather than being associated
with the problem of steering the network state from x0 to xT . More concretely, average controllability
quantifies the energy of the impulse response of a system, which describes how a system naturally evolves
over time from some initial condition [9]. Starting from an exclusive activation of the specified control
regions, we observe the brain's natural response (Fig. 1B). The larger and more variable this natural
7
response, the more states can be reached with low energy input by controlling this specific set of brain
regions.
In prior work, average controllability was also intuitively described as the ability of a set of
control nodes to drive the system to easily reachable, nearby states (Fig. 1E) [31].
To calculate average controllability, we must specify the time horizon T , which is the time period over
which we wish to observe the impulse response of the system. Note that the units of the time horizon
depend on the units of A. To observe the complete impulse response, we often assume infinite time. Fur-
thermore, we must determine the control set Bκ, which is the set of brain regions into which control input
can be injected. Even if the control set can comprise multiple, and even all nodes, average controllability
is often examined for individual brain regions to enable comparison to another single-node metric: most
commonly, modal controllability.
2.6 Modal controllability
Lastly, we introduce the metric modal controllability, which was previously described as the ability of a
single node to drive the system to distant, more difficult-to-reach states (Fig. 1E) [31]. The controllability
metric is obtained directly from the eigenvalues and eigenvectors of the network weighted adjacency
matrix. In particular, we use
N(cid:88)
φi =
(1 − (eλj (A)))v2
ij,
(8)
j=1
as a scaled summary of node i's ability to control all N modes of the network [11]. To calculate modal
controllability, we are not required to specify any parameters except the symmetric adjacency matrix A.
This metric capitalizes on information housed in the modes of A, as summarized in the eigenvalues λj
and the matrix of normalized eigenvectors V = [vi,j]. Entry vi,j is a measure of the controllability of
mode λj(A) from node i that geometrically corresponds to projecting node i onto the eigenvector j (Fig.
1D) [9, 63]. According to this heuristic, the larger the magnitude of the projection, the higher the ability
of node i to control mode j. The metric summarizes this notion across all modes, and then scales them by
their rate of decline as determined by the eigenvalues. This weighting emphasizes especially fast decaying
modes which might on average be more difficult to control because the injected control energy only has a
short-term impact. We note that modal controllability is exclusively formalized for symmetric matrices
whereas all of the other definitions that we present can be naturally extended to directed networks.
For completeness, we note that boundary controllability is another controllability metric used in the
literature and measures the ability of a brain region to integrate information between network commu-
nities [11, 31]. However, because the metric is less commonly used, we will not discuss it further in this
work.
3 Materials and methods
3.1 Acquisition of diffusion imaging data
High resolution anatomical brain images were collected from 10 healthy young adults (23.9 ± 3.6 years;
20-31 years; 70% female). The participants underwent a 53:24 minute diffusion spectrum imaging (DSI)
scan with 730 diffusion directions (maximum b-value = 5010s/mm2, 21 b = 0 images, TR = 4300ms,
TE = 102ms, matrix size = 144×144, field of view = 260 × 260mm2, slice number = 87, resolution =
1.8×1.8×1.8mm3, multi-band acceleration factor = 3). Additionally, T1-weighted images were obtained
8
using an MPRAGE sequence (TR = 2500ms, TE = 2.18ms, flip angle = 7 degrees, slice number = 208,
slice thickness = 0.9mm). Both scans were acquired on a Siemens Magnetom Prisma 3 Tesla scanner with
a 64-channel head coil. The study was approved by the Institutional Review Board of the University of
Pennsylvania and all participants provided informed consent in writing.
3.2 Preprocessing of diffusion imaging data
As previously described in more detail [30], the individual DSI scans were skull-stripped, realigned, and
motion-corrected using an improved average b=0 reference image. The preprocessing was implemented
in nipype [64] using the Advanced Normalization Tools (ANTs, [65]) for image registration. We quanti-
fied the diffusion at different orientations in each voxel using the generalized q-sampling reconstruction
method [66] in DSI Studio (dsi-studio.labsolver.org). Based on the derived quantitative anisotropy values,
we performed deterministic tractography across the whole-brain [67]. For each participant, we generated
1,000,000 streamlines with a maximum length of 500mm [68] and a maximum turning angle of 35 degrees
[69].
3.3 Construction of structural brain networks
Based on the diffusion imaging data, we constructed a structural brain network for each participant.
Consistent with previous work [31, 35, 36, 41], we defined nodes of the network as brain regions according
to the 234-node Lausanne atlas (excluding brainstem) [70]. For this purpose, the Lausanne parcels were
dilated by 4mm so that the parcels reached down into the white matter enough to ensure accurate sam-
pling of underlying fibers. In the process of dilation, some voxels were assigned to two or more regions
of interest; to eradicate this redundancy, we assigned each voxel to the mode of its neighbors [71]. After
warping the parcellation into the subject's diffusion space, we quantified the edges of the network as total
streamline count connecting a pair of brain regions, corrected for their volume. Overall, we constructed
a 233×233 sparse, weighted, and undirected adjacency matrix for each participant with the number of
interregional streamlines representing structural connectivity.
3.4 Mapping to cognitive systems
To define neurobiologically meaningful brain states, we capitalized on an established functional brain atlas
[72]. By clustering the resting state functional magnetic resonance imaging data of 1000 healthy adults,
Yeo et al.
identified seven cognitive systems, each consisting of a set of distributed brain regions that
are functionally coupled [72]. The functional parcellation comprises visual (VIS), somatomotor (SOM),
dorsal attention (DOR), ventral attention (VEN), limbic (LIM), frontoparietal control (FPC), and default
mode (DM) systems. To link the functional and anatomical atlases, we mapped each brain region to the
cognitive system with the highest spatial overlap as reported previously [22, 48]. More concretely, each
Lausanne parcel was assigned to the cognitive system that was most frequently associated with its voxels
as defined by the purity index. Subcortical regions were summarized in an eighth, subcortical system (SC).
3.5 Probing different modeling choices
We used the structural connectivity matrices of our sample to probe the impact of several modeling
choices on average and modal controllability, and on minimum and optimal control energy. In our anal-
yses, we systematically varied one parameter at a time while keeping all other parameters constant.
Constant modeling choices were guided by the modeling choices most commonly used in the literature
[31, 35, 36, 41]. Concretely, we employed a simplified noise-free linear continuous-time and time-invariant
network model, stabilized using c = 1. When estimating average controllability, we set the time horizon
9
T to infinite time. When estimating control energies, we used T = 3, approximated by 1000 time steps.
When the system matrix A is stable, the controllability Gramian equation converges as T approaches
infinity. In this case, the Gramian can be computed algebraically by solving the Lyapunov equation.
Motivated by the questions most relevant to each approach, we calculated average and modal controlla-
bility for each brain region based on single-node control sets, whereas control energies were based on full
brain control. To examine control energies for transitions between previously defined functional systems,
we simulated state transitions from an initially active default mode system to the activation of six differ-
ent cognitive systems representing the target states [72]. For each specific brain state, regions belonging
to the activated cognitive system were set to one, whereas all other brain regions were set to zero. Except
for the section on full versus partial control, we averaged across the examined state transitions. For op-
timal control energy, we set the relative energy weight ρ = 1. In the restricted set of state transitions we
investigated, minimum and optimal control energy yielded highly similar results. To avoid redundancy,
we report the results on optimal control energy in the Supplementary Results (SFig. 1-6). Nevertheless,
we point out deviating results of optimal control energy in the main text.
3.6 Examining metrics on an individual and regional level
Since network control theory can be utilized to examine controllability differences in both individuals and
brain regions, we separately studied the metrics on an individual and regional level. For this purpose,
we summarized average controllability, modal controllability, and minimum control energy across either
brain regions or individuals to subsequently investigate individual and regional values, respectively (Fig.
1F). To estimate the consistency of a metric across parameter choices on an individual (regional) level, we
first summarized the metric across brain regions (individuals) and then computed the Pearson correlation
between the individual (regional) values obtained with one modeling choice and the individual (regional)
values obtained with the second modeling choice; for example, we compare discrete- and continuous-time
systems, and we compare two different parameter choices of time horizon T .
3.7 Construction of spatial adjacency network
In addition to diffusing along white matter fibers, neural signals could potentially also diffuse between
spatially adjacent brain regions. In other words, physical contact between two regions can be seen as a
form of structural connectivity. To examine this complementary measure of structural connectivity, we
generated brain networks, S, based on the amount of shared neighborhood between two brain regions. We
defined the edges of S as the number of face-touching voxels between two parcels of the Lausanne-atlas
warped into subject space. In addition to studying each structural matrix separately, we also exploit
the combined information of both measures by constructing the matrix AS as an average of A and S.
Because both diffusion and adjacency measures are expressed in arbitrary units and the actual scaling
might impact controllability metrics, we scaled S and AS to the range of A. We tested the effects of
the structural connectivity types and their binarized version in a repeated measures ANOVA with two
within-subject factors. To ensure that the effects of matrix type and binarization were not exclusively
based on different edge weight distributions [73], we verified the results by sampling edge weights of S
and AS from the distribution of A while preserving their rank order.
3.8 Controllability of fast and slow dynamics
We capitalized on the concept of modal controllability to probe the ability of a brain region to control
a specific set of temporal dynamics such as fast and slow modes [18, 74]. Instead of summarizing across
10
all modes that a system supports, we restricted the calculation of modal controllability to a subset of
fastest (slowest) modes. We define transient (persistent) modal controllability as the ability of a brain
region to control fast (slow) modes. The temporal dynamics of modes are determined by the magnitude
of their eigenvalues. In continuous-time systems, large (small) eigenvalues relate to quickly (slowly) de-
caying modes. The lack of a formal definition of fast and slow dynamics requires the choice of a threshold
that specifies the subset of modes (Fig. 1D). We systematically probed the influence of threshold on
a brain region's ability to control different temporal dynamics by calculating transient and persistent
modal controllability using the 10%, 20%, 30%, 40%, and 50% fastest and slowest modes. To disentangle
these overlapping control tasks, we additionally summarized the ability of each brain region to control
a specific interval of modes using the unscaled eigenvector matrix V . For this purpose, we separated V
into both 10 intervals and 2 intervals; this separation enabled a comparison to persistent and transient
modal controllability based on a cut-off of 10% and 50%, respectively.
3.9 Definition of complexity of the energy landscape
The control trajectories from any initial state to any target state span the energy landscape of a dynamic
system. The heterogeneity of the minimum control energy landscape is determined by the eigenvalues
of the inverse of the controllability Gramian [9]. Note that we assume independent control from all
brain regions because the inverse of the Gramian is often ill-conditioned for small control sets [31]. We
capitalized on the variability of the eigenvalues to quantify the complexity of the minimum control energy
landscape of a brain network, that is how the magnitude of the minimum control energy varies across
all possible state transitions (Fig. 1G). To account for the observed skewness of the distribution, we
adopted the interquartile range as a measure of variability. Formally, we define the complexity of the
energy landscape as the difference between the 75th and 25th percentile of the eigenvalue distribution of
the inverted controllability Gramian
Cκ,T = P75(λW −1
κ,T
) − P25(λW −1
κ,T
).
(9)
We calculated the complexity of the energy landscape for each participant based on an infinite-time
controllability Gramian. Then, we tested the complexity of the energy landscape of the brain network
against three null models preserving distinct network characteristics. The topological null model preserved
degree and strength distribution by iteratively switching connections between randomly selected edge
pairs and subsequently associating the connections with the empirically observed edge weights [75]. The
spatial null model preserved the relationship between Euclidean distance on the edge weights by adding
the initially removed distance effects to the randomly rewired graph [76]. The combined null model
preserved both the strength distribution and spatial embedding of the brain networks by approximating
the observed strength distributions and effects of Euclidean distance on the edge weights [76]. Overall,
we generated 1000 random instantiations of each null model.
4 Results
In the application of network control theory, we can rely on different neurobiological assumptions that are
reflected in our modeling decisions. We begin with an examination of the impact of different modeling
choices, before investigating several proposed model extensions.
11
4.1 Consistency across time systems
When examining how the brains architecture gives rise to its complex dynamics by means of network
control theory, one of the first modeling steps represents the type of the dynamic model. We can either
assume that the neural dynamics evolve in discrete time steps or continuously. In light of potentially
distinct dynamics of discrete- and continuous-time systems, we initially examined the consistency of min-
imum control energy, average controllability, and modal controllability across time systems. For this
purpose, we calculated the Pearson correlation of each metric between discrete- and continuous-time sys-
tems, separately summarized across brain regions and individuals, and -- if applicable -- for different time
horizons T (Fig. 2A). Average controllability showed a high consistency across time systems (individual
level: rmin = 0.80, p = 5 × 10−3; regional level rmin = 0.99, p = 5 × 10−221), particularly for time
horizons close to zero or infinity. Likewise, modal controllability demonstrated a high consistency across
time systems (individual level: r = 0.99, p = 3 × 10−11; regional level r = 1.0, p = 2 × 10−16). Minimum
control energy, however, was less consistent across time systems (individual level: rmin = 0.77, p = 0.01;
regional level rmin = 0.33, p = 2 × 10−7), particularly for short time horizons. The observed results
are in line with theoretical considerations that suggest a convergence of discrete- and continuous-time
systems for infinite time. Overall, the consistency across discrete- and continuous-time systems was high
but depended on the metric, the observation level, and the chosen time horizon.
Figure 2: Consistency of metrics across time. (A) Consistency of average controllability and minimum control energy
across time systems. Pearson correlation coefficient between a given metric estimated for discrete- versus continuous-time
systems across a range of time horizons T . (Left) Average controllability; (Right) minimum control energy. (B) Consistency
of average controllability and minimum control energy in a continuous-time system for any two choices of time horizon. Heat
maps depict correlation matrix of different time horizons. Each heat map entry corresponds to the Pearson correlation of
a metric based on two different time horizon choices. (Top) Pearson correlation of individual metrics averaged across brain
regions. (Bottom) Pearson correlation of regional metrics averaged across participants. From these results, we deduce that
average controllability and minimum control energy differ qualitatively for discrete- versus continuous-time systems when
comparing estimates from short time horizons versus from longer time horizons.
12
4.2 Consistency across time horizons
Network control theory might lend itself particularly well to evaluate how local perturbations of the brain,
for instance elicited by deep brain stimulation or transcranial magnetic stimulation, affect whole brain
dynamics. In such a setting we might be interested in assessing different temporal scales of brain stimu-
lation such as the effect of stimulation in the short term or in the long term. This question prompts the
examination of the time horizon of the injected signal as another early modeling decision. We addressed
this question by quantifying the Pearson correlation between values estimated for one time horizon T and
for another time horizon T (cid:48), separately averaged across brain regions or across individuals (Fig. 2B). We
first noted that the time horizon affected the scaling of the metrics (SFig. 1A). More specifically, aver-
age controllability monotonically increased in magnitude with larger time horizons because we observed
the impulse response of the system for a longer time interval. Minimum control energy monotonically
decreased with larger time horizons; this relation is intuitive when we consider the fact that longer time
horizons allow the system to capitalize on its own natural dynamics, thereby demanding less exogenous
control input. In contrast, optimal control energy first rapidly decreased and then slightly increased with
larger time horizons (SFig. 1A). The increasing amount of optimal control energy might be required to
additionally constrain the distance of traversed brain states over longer time horizons. In general, we
found a high consistency between the metrics across a wide range of examined time horizons. However,
smaller time horizons demonstrated a different control regime in which average controllability (individual
level: rmin = −0.56, p = 0.09; regional level (rmin = 0.88, p = 7 × 10−78) and minimum control energy
(individual level: rmin = 0.28, p = 0.44; regional level (rmin = −0.71, p = 6 × 10−37) were partly anti-
correlated with the corresponding metrics in larger time horizons. In sum, short time horizons induced
an alternative control regime in average controllability and minimum control energy compared to longer
time horizons.
Impact of normalization
4.3
The normalization step represents another modeling decision that is related to time. For mathematical
reasons, we often assume the neural dynamics to diminish and stabilize over time. Neurobiological con-
siderations determine the degree of normalization; that is, how fast or slow we assume the neural system
to stabilize. To investigate the effect of normalization on controllability metrics and control energies, we
calculated average controllability, modal controllability, and minimum control energy for different choices
of the normalization parameter c. At both individual and regional levels, we first observed that with
increasing c, average controllability decreased whereas modal controllability and minimum control energy
increased (SFig. 2). Next, we investigated the consistency of the metrics across different manners of
normalization by quantifying the Pearson correlation between metrics for two choices of the normaliza-
tion parameter c, separately summarized across brain regions (Fig. 3A) and individuals (SFig. 3A). In
both cases, we observed two different control regimes depending on small (c = 0.1 to c = 102; Fig. 3A)
and large (c = 104 to c = 106; Fig. 3C) normalization parameters. Within each regime, the results were
highly consistent independent of the normalization parameter c. Between both regimes, however, the con-
sistency in average controllability (individual level: rmin = −0.19, p = 0.61; regional level: rmin = 0.86,
p = 1× 10−69;), modal controllability (individual level: rmin = 0.29, p = 0.41; regional level: rmin = 0.99,
p = 7× 10−320), and minimum control energy (individual level: rmin = 0.87, p = 2× 10−3; regional level:
rmin = 0.81, p = 6 × 10−56) was reduced. Hypothesizing that this alternative control regime might be
due to a faster stabilization of the system, we quantified the Spearman correlation between c and the
decay rate of the slowest mode. We indeed found that an increase of the normalization parameter led to
a faster stabilization of the system (r = −1.0, p = 0). Taken together, a faster stabilization of the system
13
introduced an alternative control regime that particularly affected controllability metrics.
Figure 3: Different control regimes depending on normalization. Consistency of (Left) average controllability,
(Middle) modal controllability, and (Right) minimum control energy for different choices of the normalization parameter c.
(A) Heat maps depict correlation matrices of different normalization parameters. Each heat map entry corresponds to the
Pearson correlation between a metric calculated using one normalization parameter and the same metric calculated using a
second normalization parameter. Pearson correlation of individual metrics summarized across brain regions. Normalizing
the system such that it stabilizes faster introduces a different control regime. (B) Regional controllability metrics and control
energy projected onto the brain surface using a normalization parameter c = 1. (C) Analogously, the same metrics plotted
onto the brain surface using a normalization parameter c = 105. Together, panels (B) and (C) illustrate the fact that distinct
choices for the normalization parameter can induce distinct control regimes. Note: metric values are ranked for visualization
purposes only.
14
4.4 Impact of control set size
In the study of the effects of brain stimulation on brain activity, we can also ask how many and which
brain regions we should control in order to drive the system to a, for instance healthy, state. More
concretely, we could compare the effects of targeting a specific neural circuit to the effects of whole-brain
stimulation. This motivates the examination of a final modeling choice: the number of controlled brain
regions. To probe the effect of control set size on minimum control energy, we began by generating ran-
dom control sets of varying number of active brain regions ranging from single-node to full-brain control.
We then proceeded by testing the impact on minimum control energy and the numerical error in six brain
state transitions: from activation of the default mode to activation of six canonical cognitive systems as
defined by Yeo et al. [72]. Importantly, the numerical error was reasonably small (< 1 × 10−6) when we
controlled at least 28.3% to 29.6% of brain regions (NV IS = 66, NSOM = 67, NDOR = 67, NV EN = 68,
NLIM = 69, NF P C = 68), increasing our confidence in the results. We observed that minimum control
energy and the numerical error decreased exponentially with increasing control set size (SFig. 4A). In-
tuitively, the control of a larger number of brain regions required less control energy. The exponential
relationship between control energy and control node set can also be mathematically derived [11].
Next, we were interested in how control and state trajectories differ in partial- compared to full-brain
control sets. We calculated the minimum control energy trajectory and the distance between the state
trajectory and the target state for the same six state transitions controlling all versus randomly drawn
sets of 150 brain regions. In full-brain control, we observed an exponential increase in energy (Fig. 4A)
and an approximately linear decrease in the distance between current and target state (Fig. 4B) across
the control horizon. When we controlled only a part of the brain, control and state trajectories differed
considerably. For instance, instead of taking the direct route through the state space, the system tra-
versed more distant states before it reached the target state. Theoretical work has indeed shown that
such non-local trajectories generally emerge if only a subset of nodes is controlled [77].
Finally, we wished to study the effect of distance between initial and target state on minimum con-
trol energy. Because the set of state transitions that we studied lacked sufficient variability in these
distances, we additionally simulated trajectories from a zero-activity initial state to random target states
with a varying size of brain regions activated. We found a monotonic increase of minimum control energy
with increasing distance between initial and target states (Fig. 4C). When employing a partial control
set, a subset of the random state transitions required massive amounts of control energy. A further
exploration revealed that these hardly controllable state transitions involved an activation of two weakly
connected limbic regions that were not part of the random control set. Similarly, the six state transitions
likely required less control on average because the activation of densely connected cognitive systems is
an easier control task than the activation of randomly chosen regions in target states of equal distance.
The findings in optimal control energy were highly similar, even if the exact control and state trajectories
were different (SFig. 5). Overall, state and control trajectories differed substantially depending on which
brain regions were allowed to receive energy input.
4.5 Relation of metrics
How easily a brain network can be steered to different states and the amount of energy required to achieve
a specific state transition could prove to be informative markers for pathology in or injury to the central
nervous system. The selection of the metric put to the test primarily depends on the specific research
question. To elucidate the empirical associations between controllability metrics and control energies,
15
Figure 4: Minimum control energy in full and partial control sets. Minimum control energy for state transitions from
the default mode system to six target cognitive systems (colored lines and triangles). (Top) Results from simulations using
a whole-brain control set including all 233 brain regions. (Bottom) Results from simulations using a control set consisting
of random subsets of 150 brain regions. (A) Minimum control energy across the control trajectory differs quantitatively
between full and partial brain control.
(B) Euclidean distance between the current state and the target state across
the control trajectory differs between full and partial brain control. (C) Minimum control energy increases with larger
Euclidean distance between initial and target states. Blue dots depict state transitions from a zero-activity brain state to
states comprised of a varying number of randomly activated brain regions. All values are averaged across participants. Lines
and ribbons represent the best fit to the data and the 95% confidence interval, respectively. Abbreviations: x0 = initial
state, xT = target state, x(t) = state at time t, VIS = visual, SOM = somatomotor, DOR = dorsal attention, VEN =
ventral attention, LIM = limbic, and FPC = frontoparietal control.
we measured the Pearson correlation between each pair of metrics, summarized separately across brain
regions (Fig. 5A) and individuals (Fig. 5B and Fig. 5C). In line with previous research [31], average and
modal controllability showed a positive association on the individual level (r = 0.5, p = 0.14), but asignifi-
cant large negative association on the regional level (r = −0.87, p = 1×10−74). In the six state transitions
we studied, we found a significant large positive correlation between minimum and optimal control energy
both on the individual level (r = 0.99, p = 1× 10−7) and on the regional level (r = 0.97, p = 1× 10−136).
Consistent with the mathematical constraints of their definitions, minimum control energy was signifi-
cantly lower than optimal control energy on both the individual (meanmin = 43.8, meanopt = 56.0, V = 0,
p = 2 × 10−3) and regional (meanmin = 0.19, meanopt = 0.24, V = 2, p = 6 × 10−40) levels. Finally,
we investigated how controllability metrics and control energies are related. On both levels of examina-
tion, average controllability demonstrated small negative correlations with control energies (ranging from
16
r = −0.29, p = 6 × 10−6 to r = −0.03, p = 0.93), whereas modal controllability showed slightly larger,
positive associations with control energies (ranging from r = 0.23, p = 3 × 10−4 to r = 0.45, p = 0.19).
The size of the correlations estimated with empirical data supported the scarcity of a clear theoretical
link between the concepts. Yet, the directions of the effects was consistent with the general notion that
high average controllability implies low control energy, whereas modal controllability is linked to higher
control energies.
Figure 5: Relation between controllability metrics and control energies. (A) The correlation matrix of individual
average controllability, modal controllability, minimum control energy, and optimal control energy, summarized across brain
regions. Pearson correlation coefficients (upper matrix triangle) additionally encoded via color and size of circles (lower
matrix triangle). (B) Analogously, the correlation matrix of regional controllability metrics and control energies averaged
across participants. (C) Regional controllability metrics and control energies projected onto the brain surface. Note that
the metric values are ranked for visualization purposes only. Collectively, these panels illustrate the negative association
between regional average and modal controllability, and the high consistency between minimum and optimal control energy.
Abbreviations: AVG = average controllability, MOD = modal controllability, MIN = minimum control energy, and OPT
= optimal control energy. Asterisks indicate significance on a Bonferroni-corrected α-level of 0.05 (*), 0.01 (**), and 0.001
(***).
4.6 Structural connectivity measures
After systematically examining the impact of diverse modeling choices, we wished to provide several,
potentially useful extensions of the theoretical framework. We begin with a consideration of the archi-
tecture of the brain which represents the core of network control theory. Thus, it is particularly relevant
how we define the inter-connections between brain regions. Typically used DTI data do not take into
17
account the fact that the signal can theoretically diffuse via physical contact between two brain regions.
To evaluate the consequences of different forms of the adjacency matrix reflecting different modes of signal
propagation in the brain, we additionally built structural connectivity networks based on the amount of
shared neighborhood between two brain regions. Then, we calculated controllability metrics and control
energies for the two alternative measures of structural connectivity, their combination, and their bina-
rized versions (Fig. 6). We first examined the similarity in controllability of structural networks based
on diffusion imaging (A) and based on spatial adjacency (S). Between A and S, we found small- to
medium-sized Pearson correlations in average controllability (individual level: r = 0.02, p = 0.95; regional
level: r = −0.01, p = 0.92), modal controllability (individual level: r = −0.15, p = 0.67; regional level:
r = 0.41, p = 10 × 10−11), and minimum control energy (individual level: r = 0.36, p = 0.31; regional
level: r = 0.64, p = 2× 10−16). Thus, the two measures of structural connectivity provide complementary
information. Next, we quantified the effect of binarization, matrix type (A vs. S), and their combination
AS, on controllability metrics and control energy. Repeated measures ANOVAs revealed significant main
effects of matrix type and binarization on a Bonferroni-corrected level of α = 0.01 except for the effect of
matrix type on average controllability (individual level: F = 5.98, p = 5 × 10−3; regional level F = 1.30,
p = 0.27).
To ensure that these results were not exclusively due to different edge weight distributions, we veri-
fied these results using S and AS based on the same edge weight distribution as A. When we examined
the effects in more detail, we observed that the binarization reduced the absolute values and variance of
average controllability on both the regional and individual level, whereas modal controllability displayed
a reverse effect. This pattern of results is in line with findings that less connected brain regions exhibit
lower average controllability but higher modal controllability [36]. Similarly, individual minimum control
energy was increased for binary matrices compared to fully weighted matrices; this result is consistent
with previous evidence demonstrating that control nodes with more homogeneous edge weights require
larger control energy [30]. Overall, the binarization of the structural connectivity matrix substantially
reduced the variance of controllability metrics but not minimum control energy, suggesting that the edge
weights carry valuable information especially for controllability metrics.
4.7 Persistent and transient modal controllability
Many neuroscientific endeavors focus on the speed of neural dynamics. Network control theory allows
us to explicitly study whether a brain region is capable of controlling fast and slowly changing activity
modes by means of transient and persistent modal controllability. However, there is no clear definition of
which activity modes are considered as fast or slow. Thus, we wished to further inspect how the definition
of fast and slow temporal dynamics affects transient and persistent modal controllability. We began with
the calculation of both metrics across various thresholds for determining which modes were considered
to be transient versus persistent. First, we observed that with increasing threshold the magnitude of
both transient and persistent modal controllability increased because the number of summed modes was
expanded. As expected, we further noted that transient and persistent modal controllability based on a
threshold of 0.5 summed up to modal controllability. The initially positive Pearson correlation between
transient and persistent modal controllability of brain regions reduced and turned into a negative associ-
ation with increasing thresholds (r0.1 = 0.82, p0.1 = 7 × 10−59; r0.2 = 0.78, p0.2 = 7 × 10−49; r0.3 = 0.65,
p0.3 = 1 × 10−29; r0.4 = −0.20, p0.4 = 3 × 10−3; r0.5 = −0.99, p0.5 = 4 × 10−187) (Fig. 7A). Notably,
for small thresholds such as 0.1, a subset of brain regions was found to be capable of controlling both
fast and slow temporal dynamics (Fig. 7B). While controlling for the size of each cognitive system, we
18
Figure 6: Structural connectivity measures. (A) Average controllability, (B) modal controllability, and (C) minimum
control energy for different measures of structural connectivity. The network encoded in A is based on streamline counts
between two brain regions from diffusion imaging. The network encoded in S is based on the extent of spatial adjacency
between two brain regions from T1-weighted images. The network encoded in AS is an average of A and S. Additionally, we
consider binary versions of the three networks, and refer to them as bA, bS, and bAS, respectively. (Top) Box plots depict
individual controllability metrics and control energy summarized across brain regions. Diamonds represent individuals.
(Bottom) Violin plots depict regional metrics averaged across participants. Collectively, these panels illustrate the fact that
the two structural connectivity measures provide complementary information that is retained by their combination.
found that these brain regions belonged primarily to the subcortex (36%) and VIS (22%) systems, but
also VEN (12%), DOR (9%), SOM (8%), DM (8%), and FPC (5%) systems. For large thresholds such
as 0.5, brain regions seem to be either able to control fast dynamics (39% SC, 14% DOR, 12% VIS, 11%
DM, 8% FPC, 6% SOM, 5% VEN, and 5% LIM systems) or slow dynamics (31% FPC, 26% subcortex,
22% SOM, 10% DOR, 7% DM, and 3% VIS systems), but not both.
To explore this ambiguous relationship in more detail, we disentangled the overlapping thresholds by
considering the unscaled controllability matrix V , and then by summarizing the modes into 10 intervals
versus 2 intervals (Fig. 7C, top versus bottom). Interestingly, this investigation into the controllability of
separate mode intervals also supported the notion that similar brain regions were capable of controlling
fast and slow dynamics in the strict definition of these control tasks (10 intervals) but not in the broader
definition of these control tasks (2 intervals).
Importantly, we note that these results do not extend
to discrete-time systems because the definition of modes that are considered as fast versus slow differs
substantially between time systems. Overall, the ability of a brain region to control fast and slow modes
largely depended on the definition of the control tasks.
19
Figure 7: Impact of threshold on persistent and transient modal controllability. Regional controllability of fast and
slow modes for two exemplary thresholds. (Top) Persistent and transient modal controllability defined as a brain region's
ability to control the 10% slowest and 10% fastest modes, respectively. (Bottom) Analogously, persistent and transient
modal controllability based on a threshold of 50% of the modes. (A) Scatter plots show the relationship between transient
and persistent modal controllability of brain regions averaged across participants.
(B) Transient and persistent modal
controllability projected onto the brain surface. Note that metric values are ranked for visualization purposes only. (C)
Heat maps depict each node's ability to control a specific interval of modes, ranging from the fastest (1) to the slowest (10
and 2 respectively) modes. For this purpose, we summarized the unscaled controllability matrix V into 10 and 2 intervals
respectively. When we aggregated the modes into 10 intervals, similar brain regions were capable of controlling both the
slowest and fastest group of modes. When we, however, aggregated the modes into 2 intervals, brain regions were able
to control either fast or slow modes. Thus, the ability of a brain region to control fast and slow modes depended on the
definition of the specific control task. Abbreviations: ctrb = controllability, VIS = visual, SOM = somatomotor, DOR =
dorsal attention, VEN = ventral attention, LIM = limbic, FPC = frontoparietal control, DM = default mode network, and
SC = subcortical.
4.8 Complexity of energy landscape
Finally, we sought to extend the types of research question we can address with the set of currently
available controllability and energy metrics. For this purpose, we developed and validated a complemen-
tary metric that measures the heterogeneity of all possible minimum control energy trajectories. The
complexity of the energy landscape allows us to quantify the similarity or dissimilarity of all possible
state transitions in respect to their required amount of control energy. Based on the variability of the
eigenvalues of the controllability Gramian, we quantified the complexity of the minimum control energy
landscape in each individual. Probing the consistency of the complexity of the energy landscape across
time systems, we observed a large positive Pearson correlation between discrete- and continuous-time
systems (r = 0.87, p = 1 × 10−3). We further examined the complementarity of the complexity of the
energy landscape by calculating the Pearson correlation between the complexity measure and the other
20
established control metrics defined earlier. We found a small negative association between complexity
and average controllability (r = −0.15, p = 0.68), a large negative association with modal controllability
(r = −0.67, p = 0.04), and a medium negative association with minimum control energy (r = −0.40,
p = 0.26). Next, we validated the complexity of the energy landscape of the brain against three null
models, preserving either the strength distribution or the spatial embedding, or both. Brain networks
showed a significantly lower complexity of the energy landscape than the topological null model (W = 65,
p = 8 × 10−8), the spatial null model (W = 0, p = 5 × 10−8), and the combined null model (W = 2498,
p = 6× 10−3), as quantified by a Wilcoxon test (Fig. 8). Interestingly, the combination of topological and
spatial characteristics seemed to partially explain the brain's higher homogeneity of the energy landscape.
We found consistent evidence in discrete-time systems (SFig. 7). Overall, the complexity of the energy
landscape of the brain was complementary to other controllability metrics and low compared to several
null models.
Figure 8: Complexity of the energy landscape of the human brain. Heterogeneity of the minimum control energy
landscape of individual participants (dark blue diamonds) as compared to three null models preserving different characteris-
tics of brain networks. The complexity of the energy landscape was quantified by the variability of the eigenvalue distribution
of the controllability Gramian. Null model distributions (box plots) were estimated by randomly rewiring each brain net-
work 100 times. Spatial null models (blue box plots) preserved the relationship between edge weight and Euclidean distance.
Topological null models (yellow box plots) preserved degree and strength distributions. Combined null models (green box
plots) preserved both strength distribution and spatial embedding. Dashed lines indicate complexity of the energy landscape
of brain networks and null models averaged across individuals. The combination of topological and spatial characteristics
partially explains the homogeneous energy landscape of the brain.
5 Discussion
Network control theory is an emerging field in neuroscience that has the potential to yield promising
insights into structure-function relationships in health and disease. Here, we provided an overview of
the theoretical framework by illustrating the underlying model of neural dynamics and commonly stud-
ied controllability concepts. Based on the structural brain networks from ultra high-resolution diffusion
imaging data (730 diffusion directions) of 10 healthy adults, we calculated average and modal control-
lability as well as minimum and optimal control energy. We then systematically probed the impact of
21
different modeling choices, specifically the choice of time system, time horizon, normalization, and size
of the control set, on these metrics. We further suggested potentially useful model extensions such as an
alternative measure of structural connectivity accounting for propagation of signals through gray matter
to abutting regions, and a complementary metric quantifying the complexity of the energy landscape of
brain networks.
Specific modeling recommendations. Based on our systematic examination of different modeling
choices, we derived several specific recommendations. First, we observed a generally high consistency be-
tween the behavior of discrete- and continuous-time systems, which depended on the metric, observation
level, and time horizon. Classifying the neural dynamics under study as clearly discrete- or continuous-
time is often challenging. Unless an investigator has a clear justification for choosing one time system
over another, we recommend to verify the obtained results in the alternative time-system to allow for a
better generality of the findings and inferences drawn therefrom. Second, we demonstrated that short
time horizons led to an alternative time system compared to longer time horizons. The arbitrary units of
the time scale further challenge the decision of which time horizon to choose. If there exists no concrete
justification for the choice of time horizon, we recommend to validate the obtained findings using several
different time horizons. Third, we found that a fast stabilization of the system induced a substantially
different control scenario. Again if there are no concrete neurobiological variables that can be used to
constrain one's choice, we suggest that a slow stabilization could be a plausible representation of most
neural dynamics, allowing for a broader range of dynamics. Since the influence of the normalization pa-
rameter c depends on the largest eigenvalues, the same c can have different stabilization effects in different
brain networks. To ensure consistency across studies, we suggest to make c dependent on λ(A)max, for
instance by c = 0.01·λ(A)max. Finally, we observed that the composition and size of the control region
set substantially influenced state and control trajectories. The decision critically depends on the individ-
ual research question and hence, should be well informed by theoretical or practical considerations. From
a methodological perspective, it is important to control a sufficiently large number of brain regions to
robustly estimate control energies. In sum, these recommendations could guide more informed modeling
choices in future applications of network control theory to pressing questions in cognitive, developmental,
and clinical neuroscience.
The role of time in network controllability.
In our examination of different modeling choices,
we found that both a short time horizon and a fast stabilization of the system induced an alternative
control regime. We suggest a common mechanism underlying both time-related observations. Whereas
the injected control input has time to diffuse along inter-connections between brain regions over longer
time horizons, it might be possible that this diffusion process is constrained over short time horizons.
Instead, a different control regime could come into effect in which the injected input primarily controls
each brain region independently rather than capitalizing on their interconnections. This finding suggests
that time might play a more important role in the controllability of structural brain networks than is
commonly assumed. Thus, it could be interesting to further investigate the factor of time, for instance
by linking control to real-time measures of brain function [18, 58]. Another potentially fruitful venture
could be to determine optimal control horizons by capitalizing on the natural dynamics of the system or
by changing inter-connections in more advanced dynamic models [78]. Such methods emphasizing the
role of time could help to develop minimal clinical interventions such as neuromodulation [79], which is
immediately relevant for the control of seizures in epilepsy [80 -- 84]. The temporal nature of control is
also potentially relevant for further refining brain-machine interfaces [85, 86].
22
Future directions for proposed model extensions. Moreover, the present work provides several
potentially useful extensions of network control theory. We first developed and validated a complemen-
tary measure of structural connectivity motivated by the fact that brain networks based on diffusion
imaging data disregard the potential for neural signals to diffuse between spatially adjacent brain re-
gions. We demonstrated that this alternative structural connectivity measure based on the amount of
shared neighborhood between two brain regions was complementary to the tractography version. We fur-
ther showed that their combination introduced more inter-individual variability in controllability metrics,
motivating future efforts to employ this approach in studies of individual differences. An important next
step is to test whether structural brain networks based on both diffusion imaging and spatial adjacency
outperform networks purely based on diffusion imaging data by better accounting for the observed neural
dynamics [45, 46]. Additionally, we examined the ability of the brain to control slow and fast dynamics.
We found that the capability of a brain region to control different fast modes depended on the specific
definition of the control task and was not consistent between time-systems. Neuroscientists interested in
the speed of neural changes such as different frequency bands [87, 88] should be careful in justifying their
choice of time system and the threshold which defines slow versus fast modes.
Lastly, we wished to extend the existing set of controllability metrics. For this purpose, we developed and
validated a new metric that quantifies the complexity of the energy landscape of a given brain network.
In other words, the metric measures how heterogeneous all possible state transitions are in the control
energy that they require. We showed that the brain exhibited a more homogeneous energy landscape
compared to two different null models. We found that both the brain networks' strength distribution and
spatial embedding partially explained this observation, which is in line with previous findings connecting
local and global network characteristics to network controllability [30, 31, 36]. The requirement of a sim-
ilar amount of energy to enable diverse state transitions implies that brain architecture supports diverse
transitions, which in turn could explain the complex functional dynamics consistently observed in neural
systems. A crucial next step is to test the practical utility of this new metric by linking it to development,
cognition, and psychiatric disorders. Taken together, the proposed model extensions hopefully stimulate
and enrich future research.
Expanding horizons of network control theory. New developments in network control theory
are constantly expanding the horizons of research questions that can be tackled with the associated tools.
Many (but perhaps not all) of these developments could be helpful in the study of the mind and brain.
Efforts have recently revealed a relation between controllability and symmetry [89 -- 91], which could prove
useful in determining the impact of bilateral and other symmetries on neural dynamics. The field has
begun considering multiobjective functions, tradeoffs, and constraints in control [92, 93], in addition
to probing a system's potential for control via local topological information [94]. As the field of neu-
roscience moves more concertedly towards multimodal approaches, efforts in the control of multilayer
networks could prove particularly useful [95], as could methods for detecting control nodes across scales
[96 -- 98]. For some questions, advances in the control of nonlinear systems could prove effective [99 -- 101],
including applications of Ising models [102, 103] and considerations related to the dynamics of neural
mass models [104]. Finally, moving beyond network controllability, recent work expanding system identi-
fication methods to identify specific form of nonlinear dynamics present in brain is particularly promising
[105 -- 107].
23
Methodological considerations. Several methodological aspects could potentially constrain the in-
terpretability of our results. First, we capitalized on high-resolution diffusion weighted imaging data for
the construction of structural connectivity networks. Associated tractography algorithms are still limited
in their capacity to reliably track fiber bundles, particularly long-range connections [108, 109], in terms
of their origin, exact direction, and intersection [110]. Nevertheless, diffusion weighted imaging serves as
the state-of-the art method to study white matter architecture in humans and therefore, tractography
algorithms are continuously being refined [111]. Second, our dynamic model of neural processes relied
on several simplifying assumptions including linearity and time-invariance. However, such basic models
often provide a good starting-point to approximate higher-order dynamics [45, 46] and can subsequently
be adapted to contain more complex features such as non-linearity [101, 112] and time-dependence [78].
Third, it was beyond the scope of this paper to examine the impact of modeling choices in all of their
theoretically possible combinations.
Instead, we systematically varied one modeling choice at a time
while keeping all other choices constant. Thus, the obtained results might not automatically generalize
to left-out choices, for example in the presence of higher order interactions. A further limitation of our
study is the investigation of control energies in a restricted set of six state transitions. In addition, the
initial and target brains states were constructed in a controlled, yet unnatural way by capitalizing on
the artificial activation of brain regions belonging to the same cognitive system. For greater ecological
validity, future studies could instead rely on real brain states as measured by functional neuroimaging
[18, 58]. Lastly, we wish to point out that the high consistency between minimum and optimal control
energy could also be due to different scales of distance and energy costs. To avoid such effects, future
efforts could develop an optimal energy algorithm that balances both constraints equally independent of
their scale.
Conclusions. Our systematic overview of network control theory and possible modeling choices aimed to
facilitate a deeper understanding and better evaluation of network control theory applications in neuro-
science. Future work can potentially benefit from our specific recommendations and the proposed model
extensions. Overall, this work hopefully inspires the neuroscience community to fully exploit the poten-
tial of network control theory on multiple spatial scales ranging from single neurons to brain regions.
Ultimately, such endeavors could advance our understanding of how the architecture of the brain gives
rise to complex neural dynamics.
6 Conflict of Interest
The authors declare no competing interests.
7 Acknowledgements
TMK was supported by the International Research Training Group (IRTG 2150) of the German Research
Foundation and by a full doctoral scholarship of the German National Academic foundation. We gratefully
acknowledge Eli J. Cornblath, Xiaosong He, Urs Braun, and Lorenzo Caciagli for useful discussions. DSB
acknowledges support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan
Foundation, the ISI Foundation, the Paul Allen Foundation, the Army Research Laboratory (W911NF-10-
2-0022), the Army Research Office (Bassett-W911NF-14-1-0679, DCIST- W911NF-17-2-0181, Grafton-
W911NF-16-1-0474), the Office of Naval Research, the National Institute of Mental Health (2-R01-DC-
24
009209-11, R01-MH112847, R01-MH107235, R21-M MH-106799), the National Institute of Child Health
and Human Development (1R01HD086888-01), National Institute of Neurological Disorders and Stroke
(R01 NS099348), and the National Science Foundation (BCS-1441502, BCS-1430087, NSF PHY-1554488
and BCS-1631550). The content is solely the responsibility of the authors and does not necessarily
represent the official views of any of the funding agencies.
8 Author Contributions
DSB, JS, and TMK designed the study. AEK preprocessed the data. JZK contributed analytic solutions.
TMK and JZK wrote code. TMK analyzed the data and wrote the manuscript. All authors edited the
manuscript approved the final version.
25
References
[1] G Deco, V K Jirsa, and A R McIntosh. Emerging concepts for the dynamical organization of
resting-state activity in the brain. Nature Reviews Neuroscience, 12(1):43, 2011.
[2] G Deco and M L Kringelbach. Great expectations: using whole-brain computational connectomics
for understanding neuropsychiatric disorders. Neuron, 84(5):892 -- 905, 2014.
[3] A M Hermundstad, K S Brown, D S Bassett, and J M Carlson. Learning, memory, and the role of
neural network architecture. PLoS computational biology, 7(6):e1002063, 2011.
[4] A M Hermundstad, D S Bassett, K S Brown, E M Aminoff, D Clewett, S Freeman, A Frithsen,
A Johnson, C M Tipper, M B Miller, et al. Structural foundations of resting-state and task-based
functional connectivity in the human brain. Proceedings of the National Academy of Sciences, 110
(15):6169 -- 6174, 2013.
[5] A M Hermundstad, K S Brown, D S Bassett, E M Aminoff, A Frithsen, A Johnson, C M Tipper,
M B Miller, S T Grafton, and J M Carlson. Structurally-constrained relationships between cognitive
states in the human brain. PLoS computational biology, 10(5):e1003591, 2014.
[6] K Rajan, C D Harvey, and D W Tank. Recurrent network models of sequence generation and
memory. Neuron, 90(1):128 -- 142, 2016.
[7] N Levy, D Horn, I Meilijson, and E Ruppin. Distributed synchrony in a hebbian cell assembly of
spiking neurons. Advances in Neural Information Processing Systems, 11, 1999.
[8] I R Fiete, W Senn, C ZH Wang, and R HR Hahnloser. Spike-time-dependent plasticity and het-
erosynaptic competition organize networks to produce long scale-free sequences of neural activity.
Neuron, 65(4):563 -- 576, 2010.
[9] T Kailath. Linear systems, volume 156. Prentice-Hall Englewood Cliffs, NJ, 1980.
[10] Y-Y Liu, J-J Slotine, and journal=nature volume=473 number=7346 pages=167 year=2011 pub-
lisher=Nature Publishing Group Barab´asi, A-L. Controllability of complex networks.
[11] F Pasqualetti, S Zampieri, and F Bullo. Controllability metrics, limitations and algorithms for
complex networks. IEEE Transactions on Control of Network Systems, 1(1):40 -- 52, 2014.
[12] E Bullmore and O Sporns. Complex brain networks: graph theoretical analysis of structural and
functional systems. Nature reviews neuroscience, 10(3):186, 2009.
[13] D S Bassett, P Zurn, and J I Gold. On the nature and use of models in network neuroscience. Nat
Rev Neurosci, 19(9):566 -- 578, 2018.
[14] K V Shenoy, M T Kaufman, M Sahani, and M M Churchland. A dynamical systems view of motor
preparation: implications for neural prosthetic system design. In Progress in brain research, volume
192, pages 33 -- 58. Elsevier, 2011.
[15] W J Freeman. Characterization of state transitions in spatially distributed, chaotic, nonlinear,
dynamical systems in cerebral cortex. Integrative Physiological and Behavioral Science, 29(3):294 --
306, 1994.
26
[16] R E Kalman, Y-C Ho, and K S Narendra. Controllability of linear dynamical systems, contributions
to differential equations 1 (1963), 189 -- 213. MR, 155070(27):5012.
[17] S F Muldoon, F Pasqualetti, S Gu, M Cieslak, S T Grafton, J M Vettel, and D S Bassett.
Stimulation-based control of dynamic brain networks. PLoS Comput Biol, 12(9):e1005076, 2016.
[18] J Stiso, A N Khambhati, T Menara, A E Kahn, J M Stein, S R Das, R Gorniak, J Tracy, B Litt,
K A Davis, et al. White matter network architecture guides direct electrical stimulation through
optimal state transitions. arXiv preprint arXiv:1805.01260, 2018.
[19] E A Solomon, J E Kragel, R Gross, B Lega, M R Sperling, G Worrell, S A Sheth, K A Zaghloul,
B C Jobst, J M Stein, S Das, R Gorniak, C S Inman, S Seger, D S Rizzuto, and M J Kahana. Medial
temporal lobe functional connectivity predicts stimulation-induced theta power. Nat Commun, 9
(1):4437, 2018.
[20] J D Medaglia, D Y Harvey, N White, A Kelkar, J Zimmerman, D S Bassett, and R H Hamilton.
Network controllability in the inferior frontal gyrus relates to controlled language variability and
susceptibility to tms. Journal of Neuroscience, pages 0092 -- 17, 2018.
[21] A N Khambhati, A E Kahn, J Costantini, Y Ezzyat, E A Solomon, R E Gross, B C Jobst, S A
Sheth, K A Zaghloul, G Worrell, S Seger, B C Lega, S Weiss, M R Sperling, R Gorniak, S R
Das, J M Stein, D S Rizzuto, M J Kahana, T H Lucas, K A Davis, J I Tracy, and D S Bassett.
Functional control of electrophysiological network architecture using direct neurostimulation in
humans. Network Neuroscience, Epub Ahead of Pring, 2019.
[22] Z Cui, J Stiso, G L Baum, J Z Kim, D R Roalf, R F Betzel, Shi Gu, Z Lu, C H Xia, R Ciric, et al.
Optimization of energy state transition trajectory supports the development of executive function
during youth. bioRxiv, page 424929, 2018.
[23] S F Muldoon, J Costantini, W R S Webber, R Lesser, and D S Bassett. Locally stable brain states
predict suppression of epileptic activity by enhanced cognitive effort. Neuroimage Clin, 18:599 -- 607,
2018.
[24] E J Cornblath, E Tang, G L Baum, T M Moore, A Adebimpe, D R Roalf, R C Gur, R E Gur,
F Pasqualetti, T D Satterthwaite, et al. Sex differences in network controllability as a predictor of
executive function in youth. NeuroImage, 188:122 -- 134, 2019.
[25] R P Dum, D J Levinthal, and P L Strick. Motor, cognitive, and affective areas of the cerebral
cortex influence the adrenal medulla. Proc Natl Acad Sci U S A, 113(35):9922 -- 9927, 2016.
[26] M S Goyal, S Venkatesh, J Milbrandt, J I Gordon, and M E Raichle. Feeding the brain and
nurturing the mind: Linking nutrition and the gut microbiota to brain development. Proc Natl
Acad Sci U S A, 112(46):14105 -- 14112, 2015.
[27] G Yan, P E Vrtes, E K Towlson, Y L Chew, D S Walker, W R Schafer, and A L Barabsi. Network
control principles predict neuron function in the Caenorhabditis elegans connectome. Nature, 550
(7677):519 -- 523, 2017.
[28] E K Towlson, P E V´ertes, G Yan, Y L Chew, D S Walker, W R Schafer, and A Barab´asi. Caenorhab-
ditis elegans and the network control frameworkfaqs. Philosophical Transactions of the Royal Society
B: Biological Sciences, 373(1758):20170372, 2018.
27
[29] L Wiles, S Gu, F Pasqualetti, B Parvesse, D Gabrieli, D S Bassett, and D F Meaney. Autaptic
connections shift network excitability and bursting. Sci Rep, 7:44006, 2017.
[30] J Z Kim, J M Soffer, A E Kahn, J M Vettel, F Pasqualetti, and D S Bassett. Role of graph
architecture in controlling dynamical networks with applications to neural systems. Nature physics,
14(1):91, 2018.
[31] S Gu, F Pasqualetti, M Cieslak, Q K Telesford, B Y Alfred, A E Kahn, J D Medaglia, J M Vettel,
M B Miller, S T Grafton, et al. Controllability of structural brain networks. Nature communications,
6:8414, 2015.
[32] J M Shine, M Breakspear, P T Bell, K A Ehgoetz Martens, R Shine, O Koyejo, O Sporns, and R A
Poldrack. Human cognition involves the dynamic integration of neural activity and neuromodula-
tory systems. Nat Neurosci, 22(2):289 -- 296, 2019.
[33] J Jeganathan, A Perry, D S Bassett, G Roberts, P B Mitchell, and M Breakspear. Fronto-limbic
dysconnectivity leads to impaired brain network controllability in young people with bipolar disorder
and those at high genetic risk. Neuroimage Clin, 19:71 -- 81, 2018.
[34] B C Bernhardt, F Fadaie, M Liu, B Caldairou, S Gu, E Jefferies, J Smallwood, D S Bassett,
A Bernasconi, and N Bernasconi. Temporal lobe epilepsy: Hippocampal pathology modulates
connectome topology and controllability. Neurology, 92(19):e2209 -- e2220, 2019.
[35] S Gu, R F Betzel, M G Mattar, M Cieslak, P R Delio, S T Grafton, F Pasqualetti, and D S Bassett.
Optimal trajectories of brain state transitions. Neuroimage, 148:305 -- 317, 2017.
[36] R F Betzel, S Gu, J D Medaglia, F Pasqualetti, and D S Bassett. Optimally controlling the human
connectome: the role of network topology. Scientific reports, 6:30770, 2016.
[37] D S Bassett and E T Bullmore. Human brain networks in health and disease. Current opinion in
neurology, 22(4):340, 2009.
[38] J D Medaglia, M-E Lynall, and D S Bassett. Cognitive network neuroscience. Journal of cognitive
neuroscience, 27(8):1471 -- 1491, 2015.
[39] M P Van den Heuvel, E T Bullmore, and O Sporns. Comparative connectomics. Trends in cognitive
sciences, 20(5):345 -- 361, 2016.
[40] U Braun, A Schaefer, R F Betzel, H Tost, A Meyer-Lindenberg, and D S Bassett. From maps to
multi-dimensional network mechanisms of mental disorders. Neuron, 97(1):14 -- 31, 2018.
[41] E Tang, C Giusti, G L Baum, S Gu, E Pollock, A E Kahn, D R Roalf, T M Moore, K Ruparel,
R C Gur, et al. Developmental increases in white matter network controllability support a growing
diversity of brain dynamics. Nature Communications, 8(1):1252, 2017.
[42] U Braun, A Harneit, G Pergola, T Menara, A Schaefer, R F Betzel, Z Zang, J I Schweiger,
K Schwarz, J Chen, G Blasi, A Bertolino, D Durstewitz, F Pasqualetti, E Schwarz, A Meyer-
Lindenberg, D S Bassett, and H Tost. Brain state stability during working memory is explained
by network control theory, modulated by dopamine d1/d2 receptor function, and diminished in
schizophrenia. arxiv, Epub ahead of print, 2019.
28
[43] E Tang and D S Bassett. Colloquium: Control of dynamics in brain networks. Reviews of Modern
Physics, 90(3):031003, 2018.
[44] J Z Kim and D S Bassett. Linear dynamics & control of brain networks.
arXiv preprint
arXiv:1902.03309, 2019.
[45] R F Gal´an. On how network architecture determines the dominant patterns of spontaneous neural
activity. PloS one, 3(5):e2148, 2008.
[46] CJ Honey, O Sporns, Leila Cammoun, Xavier Gigandet, Jean-Philippe Thiran, Reto Meuli, and
Patric Hagmann. Predicting human resting-state functional connectivity from structural connec-
tivity. Proceedings of the National Academy of Sciences, 106(6):2035 -- 2040, 2009.
[47] A May. Experience-dependent structural plasticity in the adult human brain. Trends in cognitive
sciences, 15(10):475 -- 482, 2011.
[48] G L Baum, R Ciric, D R Roalf, R F Betzel, T M Moore, R T Shinohara, A E Kahn, S N Vandekar,
P E Rupert, M Quarmley, P A Cook, M A Elliott, K Ruparel, R E Gur, R C Gur, D S Bassett, and
T D Satterthwaite. Modular segregation of structural brain networks supports the development of
executive function in youth. Curr Biol, 27(11):1561 -- 1572.e8, 2017.
[49] B A Brinkman, A I Weber, F Rieke, and E Shea-Brown. How do efficient coding strategies depend
on origins of noise in neural circuits? PLoS Comput Biol, 12(10):e1005150, 2016.
[50] W F Mlynarski and A M Hermundstad. Adaptive coding for dynamic sensory inference. Elife,
e32055, 2018.
[51] Y Gai, B Doiron, and J Rinzel. Slope-based stochastic resonance: how noise enables phasic neurons
to encode slow signals. PLoS Comput Biol, 6(6), 2010.
[52] D D Garrett, A R McIntosh, and C L Grady. Brain signal variability is parametrically modifiable.
Cereb Cortex, 24(11):2931 -- 2940, 2014.
[53] M Breakspear and A R McIntosh. Networks, noise and models: reconceptualizing the brain as a
complex, distributed system. Neuroimage, 58(2):293 -- 295, 2011.
[54] L L Gollo, C Mirasso, O Sporns, and M Breakspear. Mechanisms of zero-lag synchronization in
cortical motifs. PLoS Comput Biol, 10(4):e1003548, 2014.
[55] P N Taylor, Y Wang, M Goodfellow, J Dauwels, F Moeller, U Stephani, and G Baier. A computa-
tional study of stimulus driven epileptic seizure abatement. PLoS One, 9(12):e114316, 2014.
[56] V G Boltyanskii, R V Gamkrelidze, and L S Pontryagin. The theory of optimal processes. i.
the maximum principle. Technical report, TRW SPACE TECHNOLOGY LABS LOS ANGELES
CALIF, 1960.
[57] J P Hespanha. Linear systems theory. Princeton university press, 2018.
[58] E J Cornblath, A Ashourvan, J Z Kim, R F Betzel, R Ciric, G L Baum, X He, K Ruparel, T M
Moore, R C Gur, et al. Context-dependent architecture of brain state dynamics is explained by
white matter connectivity and theories of network control. arXiv preprint arXiv:1809.02849, 2018.
29
[59] D Zller, C Sandini, M Schaer, S Eliez, D S Bassett, and D Van De Ville. Structural control energy
of resting-state functional brain states reveals inefficient brain dynamics in psychosis vulnerability.
bioRxiv, 703561, 2019.
[60] T Menara, D Bassett, and F Pasqualetti. Structural controllability of symmetric networks. IEEE
Transactions on Automatic Control, 2018.
[61] B Marx, D Koenig, and Didier Georges. Optimal sensor and actuator location for descriptor systems
using generalized gramians and balanced realizations. In Proceedings of the 2004 American Control
Conference, volume 3, pages 2729 -- 2734. IEEE, 2004.
[62] H R Shaker and M Tahavori. Optimal sensor and actuator location for unstable systems. Journal
of Vibration and Control, 19(12):1915 -- 1920, 2013.
[63] AMA Hamdan and AH Nayfeh. Measures of modal controllability and observability for first-and
second-order linear systems. Journal of guidance, control, and dynamics, 12(3):421 -- 428, 1989.
[64] K Gorgolewski, C D Burns, C Madison, D Clark, Y O Halchenko, M L Waskom, and S S Ghosh.
Nipype: a flexible, lightweight and extensible neuroimaging data processing framework in python.
Front Neuroinform, 5, 08 2011. ISSN 1662-5196.
[65] B B Avants, N J Tustison, G Song, P A Cook, A Klein, and J C Gee. A reproducible evaluation
of ants similarity metric performance in brain image registration. Neuroimage, 54(3):2033 -- 2044,
2011.
[66] F-C Yeh, V J Wedeen, and W-Y I Tseng. Generalized q-sampling imaging. IEEE transactions on
medical imaging, 29(9):1626 -- 1635, 2010.
[67] F-C Yeh, T D Verstynen, Y Wang, J C Fern´andez-Miranda, and W-Y I Tseng. Deterministic
diffusion fiber tracking improved by quantitative anisotropy. PloS one, 8(11):e80713, 2013.
[68] M Cieslak and ST Grafton. Local termination pattern analysis: a tool for comparing white matter
morphology. Brain imaging and behavior, 8(2):292 -- 299, 2014.
[69] D S Bassett, J A Brown, V Deshpande, J M Carlson, and S T Grafton. Conserved and variable
architecture of human white matter connectivity. Neuroimage, 54(2):1262 -- 1279, 2011.
[70] L Cammoun, X Gigandet, D Meskaldji, J P Thiran, O Sporns, K Q Do, P Maeder, R Meuli,
and P Hagmann. Mapping the human connectome at multiple scales with diffusion spectrum mri.
Journal of neuroscience methods, 203(2):386 -- 397, 2012.
[71] A Daducci, S Gerhard, A Griffa, A Lemkaddem, L Cammoun, X Gigandet, R Meuli, P Hagmann,
and J-P Thiran. The connectome mapper: an open-source processing pipeline to map connectomes
with mri. PloS one, 7(12):e48121, 2012.
[72] B. T. Yeo, F. M. Krienen, J. Sepulcre, M. R. Sabuncu, D. Lashkari, M. Hollinshead, J. L. Roffman,
Smoller J. W., L. Zllei, J. R. Polimeni, B. Fischl, ... Liu, H., and R. L. Buckner. The organization
of the human cerebral cortex estimated by intrinsic functional connectivity. Journal of neurophys-
iology, 106(3):1125 -- 65, 2011.
30
[73] E Wu-Yan, R F Betzel, E Tang, S Gu, F Pasqualetti, and D S Bassett. Benchmarking measures of
network controllability on canonical graph models. Journal of Nonlinear Science, pages 1 -- 39, 2018.
[74] E Tang, G L Baum, D R Roalf, T D Satterthwaite, F Pasqualetti, and D S Bassett. The control of
brain network dynamics across diverse scales of space and time. arXiv preprint arXiv:1901.07536,
2019.
[75] M Rubinov and O Sporns. Complex network measures of brain connectivity: uses and interpreta-
tions. Neuroimage, 52(3):1059 -- 1069, 2010.
[76] J A Roberts, A Perry, A R Lord, G Roberts, P B Mitchell, R E Smith, F Calamante, and M Break-
spear. The contribution of geometry to the human connectome. Neuroimage, 124:379 -- 393, 2016.
[77] J Sun and A E Motter. Controllability transition and nonlocality in network control. Physical
review letters, 110(20):208701, 2013.
[78] A Li, S P Cornelius, Y-Y Liu, L Wang, and A-L Barab´asi. The fundamental advantages of temporal
networks. Science, 358(6366):1042 -- 1046, 2017.
[79] P Afshar, X Wei, M Lazarewicz, R Gupta, G Molnar, and T Denison. Advancing neuromodulation
using a dynamic control framework. Conf Proc IEEE Eng Med Biol Soc, 2011:671 -- 674, 2011.
[80] S Ching, E N Brown, and M A Kramer. Distributed control in a mean-field cortical network model:
implications for seizure suppression. Phys Rev E Stat Nonlin Soft Matter Phys, 86(2 Pt 1):021920,
2012.
[81] S Stanslaski, J Giftakis, P Stypulkowski, D Carlson, P Afshar, P Cong, and T Denison. Emerging
technology for advancing the treatment of epilepsy using a dynamic control framework. Conf Proc
IEEE Eng Med Biol Soc, 2011:753 -- 756, 2011.
[82] P N Taylor, J Thomas, N Sinha, J Dauwels, M Kaiser, T Thesen, and J Ruths. Optimal control
based seizure abatement using patient derived connectivity. Front Neurosci, 9:202, 2015.
[83] S Olmi, S Petkoski, M Guye, F Bartolomei, and V Jirsa. Controlling seizure propagation in large-
scale brain networks. PLoS Comput Biol, 15(2):e1006805, 2019.
[84] D Ehrens, D Sritharan, and S V Sarma. Closed-loop control of a fragile network: application to
seizure-like dynamics of an epilepsy model. Front Neurosci, 9:58, 2015.
[85] S Gowda, A L Orsborn, and J M Carmena. Parameter estimation for maximizing controllability of
linear brain-machine interfaces. Conf Proc IEEE Eng Med Biol Soc, 2012:1314 -- 1317, 2012.
[86] J Stiso, M-C Corsi, J O Garcia, J M Vettel, F De Vico Fallani, T H Lucas, and D S Bassett. Learn-
ing in brain-computer interface control evidenced by joint decomposition of brain and behavior.
psyArxiv, en69j, 2019.
[87] M Siegel, T H Donner, and A K Engel. Spectral fingerprints of large-scale neuronal interactions.
Nature Reviews Neuroscience, 13(2):121, 2012.
[88] S Baillet. Magnetoencephalography for brain electrophysiology and imaging. Nature neuroscience,
20(3):327, 2017.
31
[89] A J Whalen, S N Brennan, T D Sauer, and S J Schiff. Observability and controllability of nonlinear
networks: The role of symmetry. Phys Rev X, 5:1, 2015.
[90] C Zhao, W X Wang, Y Y Liu, and J J Slotine.
Intrinsic dynamics induce global symmetry in
network controllability. Sci Rep, 5(8422), 2015.
[91] A J Whalen, S N Brennan, T D Sauer, and S J Schiff. Effects of symmetry on the structural
controllability of neural networks: A perspective. Proc Am Control Conf, 2016:5785 -- 5790, 2016.
[92] Y Tang, H Gao, W Du, J Lu, A V Vasilakos, and J Kurths. Robust multiobjective controllability
of complex neuronal networks. IEEE/ACM Trans Comput Biol Bioinform, 13(4):778 -- 791, 2016.
[93] H Keren and S Marom. Controlling neural network responsiveness: tradeoffs and constraints. Front
Neuroeng, 7:11, 2014.
[94] G Li, L Deng, G Xiao, P Tang, C Wen, W Hu, J Pei, L Shi, and H E Stanley. Enabling controlling
complex networks with local topological information. Sci Rep, 8(1):4593, 2018.
[95] G Menichetti, L Dall'Asta, and G Bianconi. Control of multilayer networks. Sci Rep, 6:20706, 2016.
[96] Y Tang, H Gao, and J Kurths. Multiobjective identification of controlling areas in neuronal net-
works. IEEE/ACM Trans Comput Biol Bioinform, 10(3):708 -- 720, 2013.
[97] Y Tang, Z Wang, H Gao, S Swift, and J Kurths. A constrained evolutionary computation method
for detecting controlling regions of cortical networks. IEEE/ACM Trans Comput Biol Bioinform,
9(6):1569 -- 1581, 2012.
[98] Y Tang, H Gao, W Zou, and J Kurths.
Identifying controlling nodes in neuronal networks in
different scales. PLoS One, 7(7):e41375, 2012.
[99] S P Cornelius, W L Kath, and A E Motter. Realistic control of network dynamics. Nat Commun,
4:1942, 2013.
[100] A E Motter. Networkcontrology. Chaos, 25(9):097621, 2015.
[101] J G T Zanudo, G Yang, and R Albert. Structure-based control of complex networks with nonlinear
dynamics. Proceedings of the National Academy of Sciences, 114(28):7234 -- 7239, 2017.
[102] C W Lynn and D D Lee. Statistical mechanics of influence maximization with thermal noise.
European Physics Letters, 117, 2017.
[103] S Gu, M Cieslak, B Baird, S F Muldoon, S T Grafton, F Pasqualetti, and D S Bassett. The energy
landscape of neurophysiological activity implicit in brain network structure. Sci Rep, 8(1):2507,
2018.
[104] X Liu, J Gao, G Wang, and Z W Chen. Controllability analysis of the neural mass model with
dynamic parameters. Neural Comput, 29(2):485 -- 501, 2017.
[105] C O Becker, D S Bassett, and V M Preciado. Large-scale dynamic modeling of task-fMRI signals
via subspace system identification. J Neural Eng, 15(6):066016, 2018.
32
[106] A Ashourvan, S Pequito, M Bertolero, J Z Kim, D S Bassett, and B Litt. A dynamical systems
framework to uncover the drivers of large-scale cortical activity. bioRxiv, 638718, 2019.
[107] G Koppe, J Toutounji, P Kirsch, S Lis, and D Durstewitz. Identifying nonlinear dynamical systems
via generative recurrent neural networks with applications to fMRI. arXiv, 1902:07186, 2019.
[108] C Thomas, Q Y Frank, M O Irfanoglu, P Modi, K S Saleem, D A Leopold, and C Pierpaoli.
Anatomical accuracy of brain connections derived from diffusion mri tractography is inherently
limited. Proceedings of the National Academy of Sciences, 111(46):16574 -- 16579, 2014.
[109] C Reveley, A K Seth, C Pierpaoli, A C Silva, D Yu, R C Saunders, D A Leopold, and Q Y
Frank. Superficial white matter fiber systems impede detection of long-range cortical connections
in diffusion mr tractography. Proceedings of the National Academy of Sciences, 112(21):E2820 --
E2828, 2015.
[110] S Jbabdi and H Johansen-Berg. Tractography: where do we go from here? Brain connectivity, 1
(3):169 -- 183, 2011.
[111] F Pestilli, J D Yeatman, A Rokem, K N Kay, and B A Wandell. Evaluation and statistical inference
for human connectomes. Nature methods, 11(10):1058, 2014.
[112] B Fiedler, A Mochizuki, G Kurosawa, and D Saito. Dynamics and control at feedback vertex sets.
i: Informative and determining nodes in regulatory networks. Journal of Dynamics and Differential
Equations, 25(3):563 -- 604, 2013.
33
|
1501.06890 | 1 | 1501 | 2015-01-27T20:03:50 | Comment on "Human Time-Frequency Acuity Beats the Fourier Uncertainty Principle" | [
"q-bio.NC"
] | In the initial article [Phys. Rev. Lett. 110, 044301 (2013), arXiv:1208.4611] it was claimed that human hearing can beat the Fourier uncertainty principle. In this Comment, we demonstrate that the experiment designed and implemented in the original article was ill-chosen to test Fourier uncertainty in human hearing. | q-bio.NC | q-bio | Comment on "Human Time-Frequency Acuity Beats the Fourier Uncertainty
Principle"
National Research Council of Canada, 100 Sussex Drive, Ottawa, ON, Canada K1A 0R6
G.S. Thekkadath and Michael Spanner
(Dated: October 10, 2018)
In the initial article [Phys. Rev. Lett. 110, 044301 (2013), arXiv:1208.4611] it was claimed
that human hearing can beat the Fourier uncertainty principle. In this Comment, we demonstrate
that the experiment designed and implemented in the original article was ill-chosen to test Fourier
uncertainty in human hearing.
5
1
0
2
n
a
J
7
2
]
.
C
N
o
i
b
-
q
[
1
v
0
9
8
6
0
.
1
0
5
1
:
v
i
X
r
a
The Gabor limit [1],
∆t∆f ≥ 1
4π
,
(1)
refers to the lower bound on the product of the standard
deviations (STD) in time (∆t) and frequency (∆f ) of an
audio signal. This limit is a consequence of the Fourier
uncertainty principle.
In their Letter, Oppenheim and
Magnasco [2] claim that human hearing can surpass this
limit. They design an experiment which establishes psy-
chological limens, δt and δf , and show that their subjects
can discriminate signals that beat a limen-based uncer-
tainty
δtδf ≥ 1
4π
.
(2)
The frequency and time limens used by the authors re-
late to the accuracy with which human participants can
distinguish small frequency and time shifts present in a
sequence of three test pulses. The sequence in question is
referred to as "Task 5" in the paper. It is our view that
their experiment is ill-chosen to test Fourier uncertainty.
Firstly, the ∆t and ∆f that appear in the Gabor limit
must be the STD of time and frequency evaluated over
the whole test signal. This point is made clear in the
derivation of the uncertainty principle that can be found
in the book of Cohen [3] (Sections 3.2 and 3.3). The li-
mens used by the authors, however, are simply ad hoc
parameters that relate to the STD of statistical errors
made by the human participants when tasked with esti-
mating frequency and timing shifts in the test signal, and
are unrelated to the STD of time and frequency evaluated
over the test signal. Therefore, the limen-based inequal-
ity Eq.(2) is in no way related to the actual Gabor limit
Eq.(1), and there is no expectation that the limen-based
inequality should be satisfied.
Secondly, one can straightforwardly use Fourier anal-
ysis itself to "beat" Task 5, which again demonstrates
that Task 5 does not test for violations of Fourier uncer-
tainty since any Fourier-based analysis would necessarily
be limited by the uncertainty principle. One method is
to use a window Fourier Transform (WFT) to construct
a spectrogram given by
(cid:90) ∞
−∞
F (f, t0) =
− (t−t0)2
ei2πf te
2γ2 X(t)dt
(3)
FIG. 1: F (f, t0) for the Gaussian pulse train using the pa-
rameters Dt = 0.015s, Df = 4Hz, and σ = 50 ms.
where X(t) is the pulse sequence and γ is the width of
the window function. The γ is a free parameter that
controls the aspect ratio of the individual signals as
they appear in F (f, t0), and must be chosen to ensure
that the signals do not overlap in F (f, t0). This can
be readily achieved by setting γ equal to, for example,
the temporal variance of the first pulse received in
the pulse sequence.
In Fig.1, we show F (f, t0) for the
sequence of Gaussian pulses used in Ref.[2]. Fig. 1
clearly demonstrates that the function F (f, t0) can be
used to obtain all the frequency and time shifts required
to perform Task 5. When evaluating the WFT integral
using a fast Fourier transform (FFT), the peak positions
can be resolved to (at least) within one grid point in
both frequency (df ) and time (dt). For a sampling
rate R and time range T , the grid spacings in the
FFT are given by dt = 1/R and df = 1/T . Therefore,
the WFT would result in a limen-based uncertainty of
δtδf = dtdf = 1/(RT ). Using R=96 kHz and T =2s,
for example, we get δtδf = dtdf = 5.2083 × 10−6 which
is orders of magnitude smaller than 1/(4π) ≈ 0.0796,
and also orders of magnitude smaller than what was
achieved by the human participants. Furthermore, since
the WFT is a linear transform that can "beat" Task
Time (s)Frequency (Hz)00.511.52350400450500550600650700750800 5, we also conclude that Task 5 does not test for the
necessity of nonlinear transforms in models of human
hearing.
[2] J.N. Oppenheim and M.O. Magnasco, Phys. Rev. Lett.,
110, 044301 (2013).
[3] L. Cohen, Time-Frequency Analysis (Prentice Hall PTR,
Englewood Cliffs, N.J., 1995).
2
[1] D. Gabor, Nature 159, 591 (1947).
|
1703.01357 | 7 | 1703 | 2019-09-27T21:22:14 | A Computational Model of Systems Memory Consolidation and Reconsolidation | [
"q-bio.NC"
] | In the mammalian brain, newly acquired memories depend on the hippocampus for maintenance and recall, but over time the neocortex takes over these functions, rendering memories hippocampus-independent. The process responsible for this transformation is called systems memory consolidation. However, reactivation of a well-consolidated memory can trigger a temporary return to a hippocampus-dependent state, a phenomenon known as systems memory reconsolidation. The neural mechanisms underlying systems memory consolidation and reconsolidation are not well understood. Here, we propose a neural model based on well-documented mechanisms of synaptic plasticity and stability and describe a computational implementation that demonstrates the model's ability to account for a range of findings from the systems consolidation and reconsolidation literature. We derive several predictions from the computational model and suggest experiments that may test its validity. | q-bio.NC | q-bio | Running head: MODELING SYSTEMS RECONSOLIDATION
Draft version 3.23, 2019-09-27.
This paper has not been peer
reviewed. Please do not copy or
cite without author's permission.
A Computational Model of Systems Memory Consolidation and Reconsolidation
Peter Helfer1 and Thomas R. Shultz1
1McGill University, 2001 McGill College, Montreal, QC, Canada H3A 1G1
1
MODELING SYSTEMS RECONSOLIDATION
Acknowledgements
This work was supported by a Discovery Grant to TRS from the Natural Sciences and
Engineering Research Council of Canada (7927-2012-RGPIN). The funding source had no other
role than financial support.
Abstract
In the mammalian brain, newly acquired memories depend on the hippocampus for maintenance
and recall, but over time the neocortex takes over these functions, rendering memories
hippocampus-independent. The process responsible for this transformation is called systems
memory consolidation. However, reactivation of a well-consolidated memory can trigger a
temporary return to a hippocampus-dependent state, a phenomenon known as systems memory
reconsolidation. The neural mechanisms underlying systems memory consolidation and
reconsolidation are not well understood. Here, we propose a neural model based on well-
documented mechanisms of synaptic plasticity and stability and describe a computational
implementation that demonstrates the model's ability to account for a range of findings from the
systems consolidation and reconsolidation literature. We derive several predictions from the
computational model, and suggest experiments that may test its validity.
Keywords: memory reconsolidation; artificial neural network; AMPA receptor exchange; neural
plasticity; computational model
2
MODELING SYSTEMS RECONSOLIDATION
A Computational Model of Systems Memory Consolidation and Reconsolidation
The neural processes that transform memories from short-term to long-term storage are
collectively known as memory consolidation. They include synaptic consolidation, relatively
rapid intra-cellular changes that stabilize synaptic potentiation, and systems consolidation, slower
and larger-scale processes that reorganize and restructure memory traces across brain systems.
Specifically, systems consolidation refers to mechanisms that gradually make memories
independent of the hippocampus, a structure in the medial temporal lobe of the mammalian
brain.
Whereas new memories are susceptible to disruption by a number of different types of
interventions (e.g. electroconvulsive shock, certain pharmaceuticals, surgical procedures, and
interference from new learning), consolidated memories are resistant to these treatments.
However, retrieval of a consolidated memory can trigger a process in which it transiently
becomes vulnerable to such interventions again, but subsequently restabilizes into a consolidated
state. This is known as reconsolidation, and like consolidation it can be observed both at the
synaptic and systems level. While much has been learned about the molecular underpinnings of
synaptic consolidation and reconsolidation, the mechanisms responsible for the systems-level
phenomena remain largely unknown (Asok, Leroy, Rayman, & Kandel, 2019; Hardt & Nadel,
2018).
Here, we present an artificial neural network model that includes connection dynamics
based on mechanisms of synaptic plasticity and demonstrate how these low-level processes can
account for systems consolidation and reconsolidation.
3
MODELING SYSTEMS RECONSOLIDATION
We begin with overviews of synaptic and systems memory consolidation and
reconsolidation, and of previously published computer simulations. Next, we describe our model,
report on simulation results and discuss their implications.
Synaptic Consolidation and Reconsolidation
Synaptic transmission. Neurons generate electrical signals called action potentials (APs)
that travel along nerve fibers (axons) toward synapses where connections are made with other
neurons. When an action potential reaches a synapse, neurotransmitter is released into the
synaptic cleft, a narrow gap between the presynaptic active zone and the post-synaptic density
(PSD), specialized areas of neuronal cell membrane that together make up the synapse. The
neurotransmitter molecules bind to receptor proteins in the PSD, thereby triggering a response in
the postsynaptic neuron. Depending on the type of neurotransmitter and the type of receptor, the
response may be excitatory or inhibitory (making the postsynaptic neuron more or less likely to
generate an AP), or have some other, e.g. regulatory, function. The average size of the excitatory
or inhibitory response that is generated by the arrival of an action potential at a particular
synapse is a measure of synaptic strength, and it depends both on the amount of transmitter
released and on the numbers and types of receptors in the PSD (Citri & Malenka, 2008; Kandel,
Dudai, & Mayford, 2014). Most neuroscientists believe that memories are stored in the strengths
of synapses (Kandel et al., 2014; Sossin, 2008), an idea first articulated by Santiago Ramón y
Cajal in the late 19th century (1894) and known as the synaptic theory of memory (Langille &
Brown, 2018).
Glutamate receptors. The amino acid glutamate is the most abundant neurotransmitter
in the vertebrate nervous system (Platt, 2007). There are several types of glutamate receptors,
among which the α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor (AMPA
4
MODELING SYSTEMS RECONSOLIDATION
receptor or AMPAR) is chiefly responsible for mediating excitatory synaptic transmission (Citri
& Malenka, 2008; Maren, Tocco, Standley, Baudry, & Thompson, 1993), and the N-methyl-D-
aspartate receptor (NMDA receptor or NMDAR) is involved with regulatory functions including
the regulation of synaptic strength (Ben Mamou, Gamache, & Nader, 2006; Citri & Malenka,
2008; D. L. Walker & Davis, 2002).
Long-term potentiation. When a neuron is stimulated strongly enough to make it fire
(generate an AP), participating glutamatergic synapses are strengthened by a process called long-
term potentiation (LTP) (Bliss & Lømo, 1973), which is associated with an increase in the
number of AMPARs inserted in the PSD (Malenka & Bear, 2004). There are different stages of
LTP. Moderately strong stimulation gives rise to early-phase LTP (E-LTP), which lasts for at
most a few hours. More intense stimulation can trigger the induction of late-phase LTP (L-LTP),
which can persist for months or longer (Abraham, 2003).
Many researchers believe that LTP as studied in the laboratory is a cellular model for
learning and memory, i.e. that it replicates the synaptic changes that occur during memory
formation (Abraham, 2003; Bliss & Collingridge, 1993; Kandel, 2000; Lisman, 2017; Pastalkova
et al., 2006), although conclusive proof for this has proved difficult to obtain (Lynch, 2004;
Stevens, 1998).
Consolidation. Induction of L-LTP -- but not E-LTP -- is believed to require RNA
translation (synthesis of new proteins), based on experiments with protein synthesis inhibiting
drugs (PSIs) such as anisomycin or cycloheximide. Infusion of such drugs before or immediately
after stimulation can prevent establishment of L-LTP (Fonseca, Nägerl, & Bonhoeffer, 2006;
Frey, Huang, & Kandel, 1993; Frey, Krug, Reymann, & Matthies, 1988; Huang, Li, & Kandel,
1994). However, once L-LTP has been established, a process that takes on the order of one hour,
5
MODELING SYSTEMS RECONSOLIDATION
it is no longer vulnerable to PSI infusion (Bourtchouladze et al., 1998; Fonseca et al., 2006;
Kandel, 2000).
At the behavioral level, PSI injection within the first hour after training, but not later, has
been shown to cause memory impairments (Davis & Squire, 1984; Judge & Quartermain, 1982;
R. Kim, Moki, & Kida, 2011; Lattal & Abel, 2004), suggesting that the formation of long-term
memory similarly requires protein synthesis (but see Canal and Gold (Canal, Chang, & Gold,
2007), who argue that, at least in the case of anisomycin, the memory impairments may be
caused by other effects of the drug, not by protein synthesis inhibition).
Reconsolidation. Several studies from the 1940s and 1950s demonstrated that
electroconvulsive shock (ECS) could interfere with the establishment of long-term memory in
rodents (Duncan, 1949; Gerard, 1955; Thompson & Dean, 1955) and in humans (Cronholm &
Lagergren, 1959; Kehlet & Lunn, 1951), but only when applied within an hour or two after
acquisition. However, in 1968, Misanin et al. reported that ECS could impair 24-hour-old, i.e.
consolidated, memories of fear conditioning in rats - but only if the convulsive treatment was
immediately preceded by memory "reactivation", i.e. retrieval cued by presentation of the
conditioned stimulus (Misanin, Miller, & Lewis, 1968). Post-reactivation susceptibility to ECS
was also demonstrated by Schneider and Sherman (1968) and Lewis, Mahan and Bregman
(1972). Judge and Quartermain (1982) reported that injection of the protein synthesis inhibitor
anisomycin, which was known to produce memory deficits when administered to mice
immediately after training, could also impair older memories if given 30 minutes or less after
reactivation. Przybyslawski and Sara (1997) showed that the NMDA receptor antagonist MK-
801 could induce memory deficits in rats that had been trained on a maze-running task, if
injected up to 90 minutes after a maze run, but not after 120 minutes. The authors proposed that
6
MODELING SYSTEMS RECONSOLIDATION
reactivation returns a well-established memory to a labile state from which it normally
restabilizes spontaneously, and that this restabilization requires some or all of the same NMDA
receptor-dependent events that are needed for consolidation of new memories. They therefore
referred to the process as memory reconsolidation (Przybyslawski & Sara, 1997), a term first
introduced by Spear (1973). Nader et al. (2000) demonstrated that anisomycin infusion into the
amygdala of rats could disrupt an established fear-conditioning memory if performed
immediately after reactivation, but not six hours later. Taken together, these studies support the
notion that reactivation can render a memory trace that has undergone synaptic consolidation
labile, and that an NMDA-dependent process, likely involving protein synthesis, is required to
subsequently restabilize it. The phenomenon, known as synaptic memory reconsolidation, has
attracted much interest in the wake of the Przybyslawski and Sara (1997) and Nader et al. (2000)
papers, and a large reconsolidation literature now exists (for reviews, see Baldi & Bucherelli,
2015; Besnard, Caboche, & Laroche, 2012; Lee, Nader, & Schiller, 2017; Nader & Einarsson,
2010).
A reconsolidation-like phenomenon has also been observed at the synaptic level: protein
synthesis inhibition normally does not disrupt established (i.e. several hours old) L-LTP, but if
administered together with low-frequency electric stimulation, it can cause depotentiation
(Fonseca et al., 2006; Okubo-Suzuki et al., 2016).
The function of reconsolidation. It is generally believed that the function of post-
retrieval plasticity is to permit memory modification or updating when new information is
encountered (Lee, 2009; Lee et al., 2017; Schiller & Phelps, 2011). Several studies with human
subjects have shown that new training material is more likely to interfere with an established
memory if presented after reactivation of the original memory (Forcato et al., 2007; Hupbach,
7
MODELING SYSTEMS RECONSOLIDATION
Gomez, Hardt, & Nadel, 2007; M. P. Walker, Brakefield, Hobson, & Stickgold, 2003), although
some authors have questioned whether such results really are evidence of a reconsolidation
process (Ecker & Lewandowsky, 2012; Sederberg, Gershman, Polyn, & Norman, 2011). Post-
reactivation memory updating has also been demonstrated in rodents (Jones, Pest, Vargas,
Glisky, & Fellous, 2015; Monfils, Cowansage, Klann, & LeDoux, 2009).
The molecular underpinnings of LTP. The biochemical basis of LTP is not completely
understood, but intense research efforts during the last several decades have begun to throw light
on some of the underlying molecular mechanisms (Baltaci, Mogulkoc, & Baltaci, 2019; Kandel
et al., 2014). One significant discovery is that different types of AMPA receptors are of
importance for the induction and maintenance of the early and late phases of LTP (Hong et al.,
2013; McCormack, Stornetta, & Zhu, 2006; Rao-Ruiz et al., 2011). An AMPA receptor is made
up of four subunits, each of which can be of several different kinds. Depending on its subunit
composition, an AMPA receptor may or may not permit calcium ions to pass through the cell
membrane, and is accordingly designated as calcium-permeable (CP) or calcium-impermeable
(CI) (Henley & Wilkinson, 2013). E-LTP induction is characterized by a rapid increase in the
number of CP-AMPARs, while the establishment of L-LTP requires insertion of CI-AMPARs
(Clem & Huganir, 2010; Hong et al., 2013; Kessels & Malinow, 2009; Plant et al., 2006). This
discovery is significant because the increased CP-AMPAR count is relatively short-lived,
whereas an elevated level of CI-AMPARs can be sustained for a long time (Clem & Huganir,
2010; Hong et al., 2013; Plant et al., 2006). The ability of CI-AMPARs to remain at the PSD for
a long time has been hypothesized to be due to molecular mechanisms that traffic CI-AMPARs
to the potentiated synapse and protect them against removal (Clopath, Ziegler, Vasilaki, Buesing,
& Gerstner, 2008; Helfer & Shultz, 2018; Jalil, Sacktor, & Shouval, 2015; Sacktor, 2011;
8
MODELING SYSTEMS RECONSOLIDATION
Smolen, Baxter, & Byrne, 2012; Zhang, Smolen, Baxter, & Byrne, 2010). Interestingly, memory
retrieval has been shown to trigger a transient reversal to a state with high CP-AMPAR and low
CI-AMPAR counts (Hong et al., 2013; Rao-Ruiz et al., 2011), providing a potential explanation
for post-retrieval synaptic instability. As discussed above, infusion of a protein synthesis
inhibitor can interfere both with the induction of L-LTP and with restabilization of L-LTP after
reactivation. A proposed explanation for this is that proteins required to transport CI-AMPARs
into the synapse, and to maintain them there, need to be synthesized in order for either of these
processes to occur (Sacktor, 2011).
Additional evidence supporting the notion of LTP as a memory model includes other
pharmacological interventions that affect both LTP and memory. For example, inhibition of
NMDA receptors blocks both LTP and fear memory acquisition (Fanselow & Kim, 1994;
Miserendino, Sananes, Melia, & Davis, 1990), inhibition of the protein kinase PKMζ with ZIP
(zeta-inhibitory peptide) disrupts both L-LTP (Ling et al., 2002; Serrano, Yao, & Sacktor, 2005)
and long-term memory (LTM) (Pastalkova et al., 2006; Shema, Hazvi, Sacktor, & Dudai, 2009;
Shema, Sacktor, & Dudai, 2007), and blocking endocytosis (removal from the cell membrane) of
CI-AMPARs rescues both L-LTP and LTM from these effects of ZIP (Migues et al., 2010,
2016). Also, optogenetic stimulation protocols that induce long-term depression (LTD) and LTP
have been shown to inactivate and reactivate, respectively, a conditioned fear response (Nabavi
et al., 2014).
To summarize, LTP is considered a cellular model of memory, because it is compatible
with the synaptic theory of memory, and although a causal relationship has proved difficult to
establish, many results support the notion that the synaptic changes that characterize LTP also
play an important role in the cellular mechanisms of memory.
9
MODELING SYSTEMS RECONSOLIDATION
Systems Consolidation and Reconsolidation
Consolidation. Scoville and Milner (1957) famously documented that bilateral
hippocampal lesions in human patients resulted in profound memory loss for past events
(retrograde amnesia), as well as a nearly complete inability to form new long-term memories
(dense anterograde amnesia). The retrograde amnesia appeared to be graded: it was most severe
for the period shortly before surgery, while older memories were relatively spared. These
findings have been confirmed with other human patients (Dede & Smith, 2016; Manns, Hopkins,
& Squire, 2003; Penfield & Milner, 1958), and reproduced in animal studies with primates
(Squire & Zola-Morgan, 1991; Zola-Morgan & Squire, 1990) and rodents (Anagnostaras, Maren,
& Fanselow, 1999; Winocur, 1990) (although a temporal gradient is not always observed (Lah &
Miller, 2008; Sutherland, Sparks, & Lehmann, 2010)). The observation that recent memories are
more vulnerable to hippocampal damage than older ones has given rise to the notion that the
hippocampus plays a crucial role in the establishment, maintenance and recall of new memories,
but over time memories become hippocampus-independent (Frankland & Bontempi, 2005;
McClelland, McNaughton, & O'Reilly, 1995; Milner, 1989; Nadel & Hardt, 2010; Squire &
Zola-Morgan, 1991). The process responsible for this putative reorganization is called systems
memory consolidation (Dudai, 2004). According to what is called the standard theory of systems
consolidation (Squire & Alvarez, 1995), the hippocampus quickly records information about
events in real time, then repeatedly replays them, perhaps primarily during sleep (Genzel et al.,
2017), thereby driving a more time-consuming process of memory trace creation in the neocortex
(Frankland & Bontempi, 2005; McClelland et al., 1995; Teyler & Rudy, 2007). A possible
explanation for why it takes longer to lay down a memory trace in neocortex is that it is more
sparsely inter-connected than the hippocampus, and that creation of a memory trace therefore
10
MODELING SYSTEMS RECONSOLIDATION
requires axonal growth and synaptogenesis (Alvarez & Squire, 1994; Frankland & Bontempi,
2005). Another possibility, suggested by Teyler and Rudy (2007), is that strong inhibitory
control mechanisms in the neocortex slow down the induction of LTP.
The idea that hippocampal "replay" can trigger reinstatement of activation patterns in
neocortex raises the question of what kind of information is stored in the hippocampus.
McClelland et al., in their "two memory systems" proposal (1995), suggested that the
hippocampus stores "compressed representations" of cortical patterns, that can be decoded to
reconstruct activation patterns in neocortex, but most authors favor a view where the
hippocampus does not store memory content, but rather "indices" (Teyler & Discenna, 1986;
Teyler & Rudy, 2007) or "links" to loci in the neocortex where components of memory traces
reside (Alvarez & Squire, 1994; Frankland & Bontempi, 2005). In these accounts, the
hippocampus plays a dual role: it supports, through replay, the establishment and/or
strengthening of intra-neocortical connections, and it also assist in memory retrieval until those
connections become strong enough to function independently.
Not all researchers agree with the standard theory of systems consolidation. In particular,
proponents of the Multiple-Trace Theory (Nadel & Moscovitch, 1997) and its descendant the
Trace Transformation Hypothesis (Moscovitch, Cabeza, Winocur, & Nadel, 2016; Nadel,
Winocur, Ryan, & Moscovitch, 2007; Winocur, Sekeres, Binns, & Moscovitch, 2013) hold that
whereas semantic memories (knowledge about the world) are consolidated in the manner
described by the standard theory, autobiographical memories (spatial and temporal information)
remain hippocampus-dependent for as long as they persist. According to this view, the loss of
hippocampal support (whether due to lesion/damage or to gradual weakening as memories age)
results in the loss of spatio-temporal information. Empirical support for the notion that spatial
11
MODELING SYSTEMS RECONSOLIDATION
specificity is lost as memories age includes rodent studies by Wiltgen et al. (2010), Gafford,
Parsons, & Helmstetter (2013) and Einarsson et al. (2015). On the other hand, Dede and Smith,
in a comprehensive review of human cases of temporal lobe damage (Dede & Smith, 2016),
conclude that there is no evidence for loss of remote episodic information after hippocampal
damage, contradicting the transformation hypothesis and supporting the standard theory.
While the standard theory thus is not universally accepted (Nadel & Moscovitch, 1997;
Rudy, Biedenkapp, & O'Reilly, 2005; Winocur, Moscovitch, & Bontempi, 2010), it enjoys
widespread support and is compatible with a large body of empirical evidence (Dede & Smith,
2016; Runyan, Moore, & Dash, 2019; Squire, Genzel, Wixted, & Morris, 2015; Wang & Morris,
2010; Wiltgen, Brown, Talton, & Silva, 2004).
Several neocortical regions in the frontal and temporal lobes have been identified as
locations where memories consolidate, including the frontal cortex, posterior and anterior
cingulate and temporal cortex (Bontempi, Laurent-Demir, Destrade, & Jaffard, 1999; Frankland,
Bontempi, Talton, Kaczmarek, & Silva, 2004; Quirk & Mueller, 2008), parietal and retrosplenial
cortex (Maviel, Durkin, Menzaghi, & Bontempi, 2004) (both using a spatial discrimination task
in rodents), and frontal, lateral temporal, and occipital lobes for autobiographical memories in
humans (Bayley, Gold, Hopkins, & Squire, 2005). For reviews, see Dede and Smith (2016) and
Frankland and Bontempi (2005). The anterior cingulate cortex (ACC), part of the prefrontal
cortex, has received particular attention because it has been shown to play an important role for
recall of remote memories in several experimental paradigms including fear conditioning,
conditioned taste aversion, trace eyeblink conditioning and spatial discrimination (Bontempi et
al., 1999; Ding, Teixeira, & Frankland, 2008; Einarsson & Nader, 2012; Einarsson et al., 2015;
12
MODELING SYSTEMS RECONSOLIDATION
Frankland & Bontempi, 2005; Frankland et al., 2004; Maviel et al., 2004; Takehara, Kawahara,
& Kirino, 2003).
Experimental methods. Much information regarding the role of the hippocampus in
systems consolidation comes from lesion experiments with rats or mice, where bilateral
hippocampal lesions are performed at different intervals following training. In a common type of
fear conditioning, a sound or a specific spatial context (conditioned stimulus) is paired with an
electric foot shock or other aversive unconditioned stimulus. Recall is subsequently tested by
presenting the conditioned stimulus alone and measuring the degree of fear response. Findings
from this type of study indicate that hippocampal lesions immediately after training result in
severely impaired recall, but longer delays between conditioning and lesions produce gradually
less severe impairments. The time interval after conditioning during which hippocampal lesions
result in significant memory impairment is called the systems consolidation window. For fear
conditioning in rodents, most studies report a consolidation window of between three and four
weeks (J. Kim, Clark, & Thompson, 1995; J. Kim & Fanselow, 1992; Kitamura et al., 2009;
Land, Bunsey, & Riccio, 2000; Winocur, Moscovitch, & Sekeres, 2013). Longer windows have
been reported for primates: months in monkeys (Zola-Morgan & Squire, 1990), and years in
humans (Scoville & Milner, 1957; Squire & Cohen, 1979).
Systems consolidation has also been investigated by studying the effect of reversible
inactivation of specific brain areas. Studies using pharmaceutical inactivation of the rodent
hippocampus and/or ACC have shown that retrieval of a fear memory is hippocampus-dependent
one or three days after acquisition, but not after 28 or 30 days. At this point it has instead become
dependent on the ACC for retrieval (Einarsson et al., 2015; Frankland et al., 2004; Sierra et al.,
2017; Wiltgen et al., 2010).
13
MODELING SYSTEMS RECONSOLIDATION
Thus results from both lesion and inactivation studies are compatible with the standard
theory, and similar results have also been obtained in studies measuring brain activity during
retrieval of recent and remote memories (Bontempi et al., 1999; Gafford et al., 2013; Nadel &
Hardt, 2010).
Reconsolidation. Whereas synaptic reconsolidation -- the transient post-reactivation
susceptibility of memories to amnestic interventions like electro-convulsive shock or
administration of protein synthesis inhibitors -- has been studied since the 1960s (Judge &
Quartermain, 1982; Lewis et al., 1972; Misanin et al., 1968), systems reconsolidation -- the return
of hippocampus-dependence after reactivation -- was first described by Land et al. in 2000 (Land
et al., 2000), and subsequently by Debiec et al. (2002) and Winocur et al. (2009; 2013). These
researchers found that hippocampal lesions produced amnesia for 30- or 45-day old fear
memories if performed immediately after reactivating the memories by presentation of the
conditioned stimulus. Hippocampal lesions did not produce memory impairments without
preceding reactivation, nor if administered after the reactivated memory was allowed 48 hours to
restabilize after reactivation (Debiec et al., 2002). It thus appears that reactivation renders a
neocortical memory trace unstable and that a functioning hippocampus is needed for its
restabilization.
Inactivation studies have provided additional information about the effect that
reactivation has on remote memories. As noted above, 30-day old fear memories are strongly
dependent on ACC for retrieval. In an elegant study, Einarsson et al. (2015) showed that six
hours after memory reactivation, ACC inactivation no longer impaired retrieval, but after 24
hours ACC-dependence had returned. At the six-hour time point, inactivation of the
14
MODELING SYSTEMS RECONSOLIDATION
hippocampus also did not affect retrieval, but simultaneous activation of both hippocampus and
ACC did block recall.
Together, the findings from post-reactivation lesion and inactivation studies suggest that
reactivation triggers the creation of a short-lived hippocampal memory trace that is able to
support recall and is also required for restabilization of the ACC trace.
An additional interesting finding is that PSI infusion into hippocampus immediately after
reactivation blocks reconsolidation when performed three days (Debiec et al., 2002; 2006), five
days (Rossato, Bevilaqua, Medina, Izquierdo, & Cammarota, 2006), or seven days (Haubrich et
al., 2015; Milekic & Alberini, 2002) after training. These results suggest that recovery from the
destabilization caused by reactivation depends on protein synthesis in the hippocampus.
Researchers differ regarding the maximum memory age at which reactivation can trigger
systems reconsolidation. Whereas the lesion and inactivation studies indicate that memories are
susceptible for at least a month after training (Debiec et al., 2002; Einarsson et al., 2015; Land et
al., 2000; Winocur et al., 2009; Winocur, Sekeres, et al., 2013), the results from post-reactivation
PSI infusion into the hippocampus are less clear. Debiec et al. (Debiec et al., 2002) were able to
demonstrate reconsolidation blockade with anisomycin as late as 45 days after training, whereas
Frankland et al. (Frankland et al., 2006) were unable to show the effect at 36 days. The reason
for this discrepancy is not clear.
An interesting clue to a connection between synaptic and systems reconsolidation is
provided by Ghazal (2016): reactivation of a 1-day-old fear memory triggered a strong reduction
of CI-AMPARs in hippocampus but not in ACC, whereas at 30 days the opposite was true. This
result is consistent with a transition from engagement of the hippocampus to the ACC over a
15
MODELING SYSTEMS RECONSOLIDATION
systems-consolidation timeframe and with synaptic reconsolidation taking place in whichever of
the two systems is engaged in the retrieval.
In summary, retrieval of a new (e.g. 3-day-old) fear memory requires the hippocampus
but not the ACC. Over time, a reversal takes place so that retrieval of a 30-day-old memory
requires the ACC but not the hippocampus. Reactivation of a consolidated memory temporarily
returns it to ACC-independence for retrieval. Systems consolidation (establishment of a
neocortical trace) and systems reconsolidation (restabilization after reactivation-induced
destabilization of the neocortical trace) both require hippocampal involvement.
Model
Based on the findings described in the foregoing, we here present a model of systems
memory consolidation and reconsolidation. Although the role of LTP in learning and memory
remains conjectural, for the purpose of modeling we assume that LTP -- or an LTP-like
mechanism -- is the cellular substrate of memory and investigate to what extent this type of
synaptic mechanism can explain the systems-level phenomena.
Below we provide a conceptual description of the model; the computational
implementation is described in the Methods section.
Synaptic level
• Moderately intense stimulation induces E-LTP, which involves the rapid insertion
of CP-AMPARs. Constitutive processes subsequently remove these within hours.
• More intense stimulation sets in motion L-LTP induction (synaptic consolidation)
which involves a state change in a bistable mechanism (molecular switch). When in
the ON state, this mechanism maintains a high CI-AMPAR count in the synapse;
16
MODELING SYSTEMS RECONSOLIDATION
when the switch is OFF, the CI-AMPAR count drifts towards the basal level
characteristic of the unpotentiated synapse.
• Memory retrieval abruptly removes CI-AMPARs from the synapse and replaces
them with CP-AMPARs, thus returning the synapse to an E-LTP-like state. The
subsequent restoration of L-LTP is protein-synthesis-dependent and requires
NMDAR activity, i.e. neural stimulation.
Systems level
As discussed above, some researchers disagree with the standard theory, in particular
with respect to episodic/autobiographical memories. Nevertheless, there is widespread
agreement that semantic information does undergo systems consolidation. The aim of
the present work is to model those types of memories that do undergo a reorganization
whereby they gradually become hippocampus-independent; the question of whether
episodic/autobiographical memories are processed differently is beyond our present
scope.
• Stimulus presentations trigger patterns of activation in multiple ensembles of
neurons in the neocortex (NC). These active neurons in turn project onto and
activate neurons in the hippocampus (HPC), where a memory trace is quickly
created, providing linkages between the activated NC ensembles. Linkages are also
created through prefrontal cortex, in particular the ACC, but these connections are
initially too weak to support memory retrieval without HPC support.
• Subsequently, the HPC memory trace is spontaneously and repeatedly reactivated
which causes stimulation of these same NC neural ensembles through nerve fibers
projecting back from the HPC to the NC. Over time, the repeated reactivation of the
17
MODELING SYSTEMS RECONSOLIDATION
NC neural ensembles strengthens the ACC linkages, eventually to a point where
they can support retrieval of the memory without assistance from the hippocampus.
• Meanwhile, the HPC trace is gradually weakened by constitutive decay processes
(Hardt, Nader, & Nadel, 2013; Sachser et al., 2016).
•
If a consolidated memory is reactivated by a reminder, then activity in the
neocortical neural ensembles triggers re-establishment of the HPC linkage.
Simultaneously, the retrieval causes transient destabilization of the synapses in the
ACC linkage (synaptic reconsolidation).
• Following reactivation, hippocampal replay stimulates the now destabilized
synapses of the ACC linkage. This activity drives restabilization of these synapses.
Meanwhile, the reactivated HPC trace rapidly decays, leading to a return to ACC-
dependence in 24 hours or less.
Temporal characteristics
The temporal characteristics of systems consolidation and reconsolidation provide
some clues to the underlying mechanisms:
The reconsolidation window is shorter than the consolidation window.
Hippocampal lesion after training produces a memory impairment, more severe the
shorter the delay between conditioning and lesion. Similarly, hippocampal lesion
after reactivation also causes more severe impairment if performed early. However,
the time scales are very different. New memories are vulnerable for at least several
weeks after training (J. Kim & Fanselow, 1992; Squire & Cohen, 1979; Winocur,
1990; Zola-Morgan & Squire, 1990), whereas the post-reactivation window of
vulnerability only lasts for a few days (Debiec et al., 2002). Our model attributes this
18
MODELING SYSTEMS RECONSOLIDATION
difference to the different natures of the processes being interrupted by HPC
lesioning in the two scenarios. In the case of consolidation, what is being interrupted
is the gradual and relatively slow establishment and strengthening of intra-
neocortical connections; thus the earlier the intervention is performed, the weaker the
partially consolidated memory trace will be. Reconsolidation blockade, on the other
hand, interrupts the much faster process of re-stabilization of destabilized neocortical
synapses. The earlier in the reconsolidation window hippocampal lesion is
performed, the fewer synapses will have had time to restabilize, leading to more
severe memory loss.
Memories become transiently ACC-independent after reactivation.
Whereas a consolidated fear memory depends strongly on the ACC for recall,
reactivation triggers a brief period of ACC-independence, such that six hours after
reactivation ACC inactivation has little or no effect on retrieval, but 24 hours after
reactivation full ACC dependence has returned (Einarsson et al., 2015). This finding
suggests that reactivation triggers the creation of a short-lived hippocampal memory
trace that is able to support recall six hours later, but not 24 hours later. For a
memory trace to last for six hours or more, it must have undergone synaptic
consolidation, yet 24 hours is a short lifetime for a consolidated trace -- compare the
situation after initial acquisition, where the hippocampal trace is able to support
recall for at least several days. The mechanism that causes a faster decline of a post-
reactivation hippocampal trace is unknown.
19
MODELING SYSTEMS RECONSOLIDATION
Previously Published Computational Models
Several computational models of systems consolidation have been published. Alvarez
and Squire (1994) described a recurrent artificial neural network (ANN) model with three
memory banks representing, respectively, the medial temporal lobe (MTL - the hippocampus and
adjacent areas) and two areas of neocortex (NC). Learning was implemented by a Hebbian-like
algorithm that adjusted connection weights, with a fast learning rate for the connections between
the MTL and each of the NC areas (MTL-NC connections) and a slower rate for the connections
between the two NC areas (NC-NC connections). After training on a pair of random patterns,
one in each NC area, the model could be presented with either of the two patterns and reproduce
the other with good accuracy thanks to the strong linkage provided by the MTL-NC connections,
but not if the MTL subsystem was disabled. However, if after initial training, MTL units were
repeatedly reactivated in random fashion, then the resulting stimulation through the strong MTL-
NC connections would cause the two trained NC patterns to become active, and Hebbian
learning would strengthen the NC-NC connections to a point where they eventually became able
to support recall without the MTL linkage. The Alvarez and Squire model thus demonstrated the
computational feasibility of the standard theory of systems consolidation.
McClelland et al. (1995) hypothesized that systems consolidation serves to avoid
"catastrophic interference", impairment of older memory traces when new patterns are integrated
too quickly, an effect that had been observed in artificial neural networks. They suggested that
the brain avoids this problem by having the hippocampus quickly capture new patterns, then
repeatedly replaying them to the slow-learning neocortex. A computational model described in
the paper demonstrated that such repeated replay, interleaved with rehearsal of previously
20
MODELING SYSTEMS RECONSOLIDATION
learned patterns, could avoid catastrophic interference. The model used in their simulations
consisted of a traditional three-layer feed-forward network.
Meeter and Murre's TraceLink model (Meeter & Murre, 2005; Murre, 1996) combines
elements from these two earlier studies. Like Alvarez and Squire's (1994) model, TraceLink is a
recurrent network with neocortical and hippocampal layers. Unlike the Alvarez and Squire
model, it is capable of learning multiple patterns in succession, avoiding catastrophic
interference by repeated activation of previously learned material. It differs from the McClelland
et al. (1995) model by including simulation of a hippocampal system which autonomously
presents a random mix of new and old patterns to neocortex. McClelland et al.'s simulation did
not include a hippocampus component; the interleaved set of old and new training patterns was
provided to the network by the programmer.
A paper by Amaral et al. (2008) is primarily concerned with demonstrating that a single
model can accommodate seemingly contradictory results with respect to the transience or
permanence of amnesia resulting from hippocampal lesions at different time points in the course
of systems consolidation. The model consists of two recurrent layers, representing hippocampus
and neocortex and employs the "synaptic reentry reinforcement" (SRR) mechanism, first
described by Wittenberg and Tsien (2002), which can autonomously strengthen a memory trace
over time, and even repair partial damage. The model is able to reproduce the empirical finding
that both early and late hippocampal disruption can produce either permanent or transient
memory impairment, but permanent loss is a more likely outcome after early and/or extensive
hippocampal disruption.
These simulations thus illuminate various aspects of systems consolidation. Our work
builds on these previous models and adds the capability to simulate systems reconsolidation, i.e.
21
MODELING SYSTEMS RECONSOLIDATION
a transient period of hippocampus-dependence following cued retrieval of a consolidated
memory. The key to this capability is a more elaborate connection design than is traditionally
used in neural networks, Specifically, the connections in our model simulate the AMPA receptor
exchanges underlying synaptic consolidation and reconsolidation. Below, we introduce our
computational implementation of the model described above, and show that it is capable of
reproducing both systems consolidation and reconsolidation, including the effects of different
pharmaceutical and surgical interventions.
22
MODELING SYSTEMS RECONSOLIDATION
Methods
Simulation Targets
Table 1 summarizes the empirical findings, as described in the introduction, which our
model aims to reproduce in simulations.
Table 1: Simulation targets
3 PSI during conditioning impairs LTM but not
STM.5,7,8
4 PSI infusion during maintenance does not cause
memory impairment.5,7
5 HPC lesion at 3d causes memory impairment.9,10 HPC lesion at 3 days or less after training causes
Result
1 Retrieval is HPC-dependent and ACC-
independent at 3d.1,2,3,4,5
2 Retrieval is ACC-dependent and HPC-
independent at 30d.1,2,3,4,6
6 HPC lesion at 30d does not cause memory
impairment.9,10,11
7 Reactivation alone does not impair a
consolidated memory.10,11
8 Reactivation + PSI causes memory
impairment.8,11,12,13,14
9 Delayed impairment effect of post-reactivation
PSI infusion in HPC.8,11
10 Reactivation + HPC lesion causes memory
impairment.10,11,15
11 Graded effect of post-reactivation HPC lesion.11
12 Retrieval can be supported by either ACC or
HPC 6h after reactivation.1
Description
Retrieval of a three-day old memory is impaired
by HPC inactivation but unaffected by ACC
inactivation.
Retrieval of a 30-day old memory is impaired by
ACC inactivation but unaffected by HPC
inactivation.
Systemic injection of PSI during training prevents
LTM induction -- but does not impair STM .
Systemic PSI injection does not impair a
consolidated memory.
later recall to be severely impaired compared to non-
lesioned animals.
HPC lesion, when not preceded by memory
reactivation, does not result in subsequently impaired
recall.
Reactivation without subsequent HPC lesion or PSI
infusion does not by itself impair a consolidated
memory.
Reactivation immediately followed by PSI infusion
into HPC causes impairment of a consolidated
memory.
The impairment caused by post-reactivation PSI
infusion in HPC does not manifest in retrieval test
soon after lesion (4h) but only later (24h).
Reactivation immediately followed by HPC lesion
causes impairment of a consolidated memory.
The severity of memory impairment caused by post-
reactivation HPC lesion diminishes with increasing
reactivation-lesion delay.
Six hours after reactivation, neither ACC inactivation
nor HPC inactivation alone impairs retrieval, but
simultaneous inactivation of both does block
retrieval.
1Einarsson et al. (2015) 2Gafford et al. (2013) 3Ghazal (2016) 4de Oliveira Alvares et al. (2012) 5Frey (1997)
6Frankland et al. (2004) 7Davis and Squire (1984) 8Rossato et al. (2006) 9J. Kim and Fanselow (1992) 10Land et al.
(2000) 11Debiec et al. (2002) 12Frankland et al. (2006) 13Haubrich et al. (2015) 14Milekic and Alberini (2002)
15Winocur et al. (2009)
23
MODELING SYSTEMS RECONSOLIDATION
Model Design
Like other artificial neural networks, ours consists of units and connections, where units
are analogs of biological neurons (or ensembles of neurons) and connections model synapses.
(For a primer on artificial neural networks, see e.g. Shultz (2003)).
Topology. We use a recurrent artificial neural network with four regions representing
hippocampus (HPC), anterior cingulate cortex (ACC) and two sensory cortex areas, SC0 and
SC1, to which stimuli are presented. Each region consists of 25 units. The units are linked
together by connections, such that each HPC and ACC unit is bidirectionally connected to all the
units in the other three regions. The set of connections between any two regions, e.g. from HPC
to SC1, is referred to as a tract, see Figure 1.
24
MODELING SYSTEMS RECONSOLIDATION
Figure 1: Network architecture. To reduce clutter only three of the twenty-five units in
each region are shown. Each double-headed arrow represents two independent
connections, one in each direction, between a pair of units. The diagram illustrates the
state after initial acquisition: presentation of the unconditioned stimulus (US) and the
conditioned stimulus (CS) has activated some units in SC0 and SC1 (filled circles) and
fast learning has created strong linkages (bold lines) through the HPC. Linkages through
the slower-learning ACC are still weak.
This topology is an evolution of the one used by Alvarez and Squire in their influential model of
systems memory consolidation (Alvarez & Squire, 1994). Their model had two neocortical
regions with direct linkage between them; we have added the ACC region in order to be able to
simulate interruption of linkage between the SC1 and SC2 regions by reversible inactivation of
either the ACC and/or HPC regions (or both).
As noted in the introduction, a number of neocortical regions other than the ACC have
been identified as loci for remote memory traces. For simplicity, the model only includes the
ACC as a target region for systems consolidation.
Similarly, the model's connectivity between the hippocampus and neocortical regions is
simplified for reasons of computational feasibility. In a real brain, the hippocampal complex is
comprised of many intricately interconnected subareas that receive input from and project back
to not only sensory cortices, but to a widespread network of different structures in hierarchies of
association areas. For an overview, see Lavenex and Amaral (2000).
The algorithms for activation, learning, and retrieval are adapted from Meeter and
Murre's TraceLink model of systems memory consolidation (Meeter & Murre, 2005), with minor
modifications that will be detailed in the following.
25
MODELING SYSTEMS RECONSOLIDATION
The model description that follows makes reference to a number of configurable
parameters. These are documented in tables 2-4.
Units. The units are bistable and stochastic. A unit's activity level is either 1 (active) or 0
(inactive). The probability that a unit is active (the activation function) is an asymmetric sigmoid
function of its net input:
(
P isActive
)
j
=
1
(
actK net
−
inhib
r
j
)
1
+
e−
Where
j
P isActive is the probability that unit j is active, and
j
)
(
(1)
jnet is the net input to unit j,
calculated as the sum of the activity levels of units connected to unit j, weighted by inbound
connection strengths:
net
j
= ∑
i
a s
i
ij
Here,
ia is the activation level of unit i and
ijs is the strength of the connection from unit i to
(2)
unit j. The actK symbol in Equation 1 is a configurable parameter that controls the slope of the
sigmoid function. The
inhib variable in Equation 1 is a region-specific dynamic inhibition level,
r
which will be described below.
Figure 2 shows two examples of activation functions with actK values of 2.0 and 10.0,
respectively, both with an inhibition level of 5.0. The probability of activation grows from near
zero when net input is low to near one when net input is high, with a value of 0.5 at the current
inhibition level. An increase in inhibition shifts the curve to the right, with the effect that greater
net input is required for activation. This tends to results in fewer active units in the region.
Greater actK values produce curves with steeper, more abrupt transitions, corresponding to a
26
MODELING SYSTEMS RECONSOLIDATION
more deterministic behavior. All simulations reported here use actK = 2.0.
Figure 2: Activation functions. Two examples are shown, one with actK = 2.0 (solid) and
one with actK = 10.0 (dashed). Both are plotted with inhibr = 5.0.
The use of the activation function, including the inhibition mechanism, will be further described
below.
Connections. The connections are abstract models of glutamatergic synapses,
characterized by four variables:
• psdSize: the maximum number of AMPARs that can be inserted, analogous to the
number of receptor "slots" in the PSD.
• numCpAmpars: the number of currently inserted CP-AMPARs
• numCiAmpars: the number of currently inserted CI-AMPARs
•
isPotentiated: a Boolean attribute that models the bistable nature of L-LTP
This combination of binary and continuously variable attributes makes it possible to
model synapses where potentiation behaves like a bistable switch (Bortolotto, Bashir, Davies, &
Collingridge, 1994; Jalil et al., 2015; Westmark et al., 2010), yet synaptic strength is variable
with many distinct levels (Bartol et al., 2015).
27
MODELING SYSTEMS RECONSOLIDATION
As will be described below, stimulation causes an increase of the psdSize attribute,
allowing more AMPARs to be inserted. A connection's weight at any moment is proportional to
its total number of inserted AMPARs (numCpAmpars + numCiAmpars).
Simulation
A simulation consists of a sequence of time steps. Various interventions may be
scheduled for any time point during the simulation, and in addition several background processes
execute at each time step. The scheduled event types are training, reactivation, HPC lesion, HPC
inactivation and ACC inactivation. The background processes are consolidation, AMPAR
trafficking and random depotentiation. In addition, a retrieval test can be executed at any time.
The different interventions and background processes will be described below, but first we
introduce the concept of a learning cycle.
Learning cycle. Learning takes place when stimuli are presented, when a memory is
reactivated, and during memory consolidation. In each of these instances, learning is achieved by
the execution of a learning cycle in every connection that satisfies the Hebbian condition that the
two units that it connects are both active. A learning cycle consists of three steps: PSD growth,
CP-AMPAR influx, and (possibly) L-LTP induction.
1. PSD growth. The network learns activation patterns by a Hebbian learning rule (Hebb,
1949) that increases the psdSize attribute of connections between simultaneously activated units,
asymptotically towards a maximum value:
psdSize
ij
←
psdSize
ij
+
µ
T
(
maxPsdSize
−
psdSize
ij
)
(3)
where
psdSize is the PSD size of the connection between units i and j, maxPsdSize is a
ij
configurable parameter, and
Tµ is a learning rate specific to the tract that connection ij belongs
to. This learning rule differs from that of TraceLink in three respects: (1) We use an asymptotic
28
MODELING SYSTEMS RECONSOLIDATION
function where TraceLink used a linear function with hard limits; this is an esthetic choice; (2) In
TraceLink, the learning rule applies directly to connection weights; in our model, learning affects
PSD size, i.e. the maximum weight that the connection can attain; actual weight is defined by the
number of inserted AMPARs, which may at times be lower than psdSize; (3) TraceLink's
learning rule supported "unlearning" (learning at a negative rate) to model interference, a
phenomenon that we are not addressing in the present model.
To simulate different intensities of stimulation, the learning rule is applied a variable
number of times, specified by the parameter numStimCycles. The stimulation resulting from
spontaneous hippocampal activation during consolidation is modeled as much less intense than
the stimulation due to stimulus presentation.
2. CP-AMPAR influx. PSD growth is followed by an increase in the number of
CP-AMPARs such that the total AMPAR count becomes equal to the connection's PSD size:
numCpAmpars
ij
←
psdSize
ij
−
numCiAmpars
ij
(4)
This models the rapid CP-AMPAR influx during E-LTP induction.
3. L-LTP induction. Finally, probabilistic induction of L-LTP in a connection is
implemented by turning on its isPotentiated attribute with a probability that depends on the
intensity of the stimulation:
prob
←
f
(
numStimCyc
les
,
stimTh
resh
)
(5)
where numStimCycles is as described above, stimThresh defines the value of numStimCycles that
produces a 0.5 probability of L-LTP induction, and f is an asymmetric sigmoid function:
(
f stim thresh
,
)
=
1
stim thresh
−
)
1
+
(
e−
(6)
29
MODELING SYSTEMS RECONSOLIDATION
Learning cycles are executed (a) when stimuli are presented for training, (b) at memory
retrieval (reactivation), and (c) when patterns are spontaneously activated by the memory
consolidation process. Descriptions of these mechanisms follow:
Training. To train an association, patterns (subsets) of units are activated in the SC0 and
SC1 regions to represent an unconditioned stimulus, US, and a conditioned stimulus, CS,
respectively. Each stimulus pattern consists of
k
*r
numUnits units, where
r
numUnits is the
r
number of units in region r and
rk is a region-specific parameter that specifies the fraction of the
region's units to activate in a pattern. (This is described further in the section about inhibition,
below.) The network then randomly selects and activates
rk linkage units in each of the HPC and
ACC regions and executes a learning cycle in all connections that connect two simultaneously
active units, i.e. connections that are in the Hebbian condition. The learning rate,
Tµ , is defined
to be relatively high in HPC tracts, allowing rapid creation of linkages strong enough to support
recall. The ACC learning rate is lower, hence linkages through the ACC region are not strong
enough to independently support recall immediately after training.
Retrieval. To test recall of a trained pattern, the CS units are activated in SC0, and the
network is cycled by repeated application of the activation function in all units a number of times
specified by the parameter numSettleCycles. The units are updated synchronously: new
activation values are first computed for all units, based on their current net inputs. The new
values are then applied to all the units.
The activity pattern that the SC1 region eventually settles into is compared to the
associated US pattern to calculate a retrieval test score.
Inhibition. To ensure fast settling of the network, we use a variation of an inhibition
mechanism described by Meeter and Murre for their TraceLink architecture, designed to
30
MODELING SYSTEMS RECONSOLIDATION
(roughly) model the working of inhibitory neurons (Meeter & Murre, 2005). Each region has a
configurable target fraction of active units,
rk . Following each update cycle during the settling
phase of memory retrieval (and during consolidation, see below), each region's inhibition level,
inhib (see Equation 1), is adjusted by an amount proportional to the current deviation of the
r
number of active units in the region from the target value defined by the region's
rk value:
inhib
r
←
inhib
r
+
numActive
r
r
k
−
k
r
*
numUnits
r
*
inhibIncr
inhib
r
←
(
min maxInhib max minInhib inhib
r
(
(
,
)))
(7)
(8)
where
numActive is the number of active units in region r and
r
numUnits is the total number of
r
units in region r. The configurable inhibIncr parameter controls the rate of adjustment of the
inhibition level. The assignment in Equation 8 restricts
inhib to a configurable range of values
r
[
]
minInhib maxInhib .
,
Systems consolidation. At every simulation time step a randomly selected trained pattern
is activated in HPC, after which the entire network is cycled in the same manner as for recall test
(but without stimulus presentation). Whatever pattern the ACC and SC regions settle into is then
reinforced by application of the learning rule. Because the network is more likely to settle into
trained patterns than other random states, this will tend to strengthen CS-US linkages through the
ACC, eventually making recall of trained patterns HPC-independent.
AMPAR trafficking. At each simulation time step, AMPAR trafficking is simulated in
all connections as follows:
CP-AMPARs: the numCpAmpar value declines asymptotically towards zero, simulating the
limited dwell time of CP-AMPARs at the synapse:
31
MODELING SYSTEMS RECONSOLIDATION
numCpAmpars
ij
←
numCpAmpars
ij
−
cpAmparRemovalRate
*(
numCpAmpars minNumCpAmpars
−
ij
where
numCpAmpars is the current number of CP-AMPARs at connection ij, and
ij
cpAmparRemovalRate and minNumCpAmpars are configurable parameters.
(8)
)
CI-AMPARs: if a connection's isPotentiated attribute is true and its source and destination units
are both active (Hebbian condition), then numCiAmpars grows asymptotically towards the
number of available slots in the connection (psdSize), simulating activity-dependent trafficking
of CI-AMPARs into a potentiated synapse:
numCiAmpars
ij
← (
min numCiAmpars
ij
+
ciAmparInsertionRate psdSize
ij
,
)
(9)
In unpotentiated connections,
numCiAmpars declines exponentially, simulating a depotentiated
ij
synapse's gradual return to baseline:
numCiAmpars
ij
←
numCiAmpars
ij
−
ciAmparRemovalRate
*(
numCiAmpars minNumCiAmpars
−
)
(10)
minNumCiAmpars, ciAmparInsertionRate and ciAmparRemovalRate are configurable
parameters.
Depotentiation. Potentiated connections are subject to random constitutive
depotentiation at a rate controlled by the tract-specific parameter
baseDepotProb ,which has a
t
higher value for HPC tracts than for intra-neocortical tracts, to account for the observed faster
decline of hippocampal memory traces.
Reactivation. Reactivation is modeled as an unreinforced CS presentation, i.e. a cued
retrieval. The CS pattern is activated in the SC0 region, the network is settled as described for
training, and an AMPAR exchange is simulated in all connections situated between
simultaneously active units (i.e. connections in the Hebbian condition):
numCiAmpars is
ij
reduced to its minimum value minNumCiAmpars, and
numCpAmpars is set to its maximum
ij
32
MODELING SYSTEMS RECONSOLIDATION
allowed value
numCiAmpars
ij
psdSize
ij
−
. This puts those consolidated ACC linkage
connections that were activated by memory retrieval in an unstable E-LTP-like state, modeling
the post-reactivation instability documented in empirical studies. A random set of
rk HPC
linkage units is then activated, in the same manner as during initial acquisition, and a round of
Hebbian learning takes place.
As noted in the introduction, the hippocampal engagement after reactivation is much
briefer (less than 24h) than after initial training, when HPC can support recall for several days.
Although the mechanism underlying this faster disengagement is unknown, in our model we
simulate the phenomenon by increasing the depotentiation probability in connections that are
activated following memory retrieval:
depotProb
ij
←
maxDepotProb
t
(11)
where
maxDepotProb is a tract-specific parameter that has a higher value in HPC tracts than in
t
neocortical tract. Subsequently, the depotentiation rate returns asymptotically to its base value:
depotProb
←
depotProb depotProbDecayRate
−
*(
depotProb baseDepotProb
−
)
(12)
where depotProbDecayRate is a global parameter.
Hippocampal lesion. Hippocampal lesion is simulated by disconnecting the HPC region
from the simulation.
PSI infusion. Infusion of PSI into a region -- HPC or ACC -- is simulated by disabling
potentiation and CI-AMPAR insertion for nine hours of simulation time, corresponding to the
amount of time that the protein synthesis inhibitor anisomycin remains active in brain tissue
(Wanisch & Wotjak, 2008).
Inactivation. Reversible inactivation of HPC or ACC is modeled by transiently disabling
activation of all units in the HPC or ACC region, respectively.
33
MODELING SYSTEMS RECONSOLIDATION
Simulation environment
The model is implemented as a C++ program and all simulations were executed on an
Intel i5-2400 computer running the Debian Linux 8.4 operating system.
Model fitting
Simulations are controlled by the parameters indicated in the following tables. Parameter
values were manually chosen to make simulation time courses approximate those reported in the
referenced literature.
Table 2: Global parameters
Parameter
actK
minPsdSize
maxPsdSize
numSettleCycles
trainNumStimCycles
consNumStimCycles
minInhib
maxInhib
inhibIncr
Value
2.0
10
100
20
50
1
2.5
10
0.05
Description
Steepness of the activation function
Min value for a connection's psdSize
Max value for a connection's psdSize
Number of times the network is cycled for settling
Number of times the learning rule is applied at
training
Number of times the learning rule is applied at
consolidation
Minimum inhibition level in a region
Maximum inhibition level in a region
Inhibition level adjustment rate
Table 3: Region-specific parameters.
HPC ACC SC0/1 Description
25
0.2
Number of units
Target fraction of active units
25
0.2
25
0.2
Parameter
numUnitsr
kr
34
MODELING SYSTEMS RECONSOLIDATION
Table 4: Tract-specific parameters. The "HPC" columns indicates values for the tracts that
connect HPC with the other regions, and the "ACC" column for the tracts that connect ACC with
SC0 and SC1.
Name
µt
psdDecayRate01h
cpAmparRemovalRate01h
ciAmparInsertionRate01h
ciAmparRemovalRate01h
HPC ACC Description
0.004 Learning rate
0.08
0.01
0.01
0.1
0.1
2.0
2.0
0.015 0.015 CI-AMPAR removal rate in unpotentiated
PSD shrink rate when unpopulated
Constitutive removal rate for CP-AMPARs
CI-AMPAR insertion rate in potentiated synapse
baseDepotProb01h
maxDepotProb01h
depotProbDecayRate01h
minNumCpAmpars
minNumCiAmpars
0.002 0.0
0.0
0.05
0.03
0.03
0
0
2
2
synapse
Probability of spontaneous depotentiation
Prob. of depotentiation after reactivation
depotProb decay rate towards baseDepotProb
Num CP-AMPARs in an unpotentiated conn.
Num CI-AMPARs in an unpotentiated conn.
Results
After training the network with a CS-US association, recall is tested by presenting the CS
in the SC0 region, settling the system, and comparing the resulting activation pattern in the SC1
region with the trained US (a score of 1.0 indicates a perfect match). The descriptions below
compare the recall scores in simulation runs where an intervention (lesion or inactivation) is
performed with simulation runs without the intervention (labeled "baseline" in the diagrams).
The score values in all diagrams are means of 100 simulation runs. Error bars indicate standard
deviation.
Consolidation and reconsolidation windows. HPC lesions produce memory deficits when
performed in the consolidation or reconsolidation windows, but not otherwise. See Figure 3
and Figure 4. These simulations reproduce findings 5-7 and 10-11 of Table 1.
35
MODELING SYSTEMS RECONSOLIDATION
Figure 3: Consolidation window. Simulated HPC lesions produce severe impairment
when performed shortly after training, but not later. Recall tests are performed 7 days
after lesioning, e.g. the data point corresponding to lesions performed 5 days after
conditioning reflects recall tests executed on day 12. This simulates delays used in
behavioral experiments to allow the animals to recover from surgery before testing. The
baseline data points reflect recall tests at the corresponding time points, but without
preceding lesions.
36
MODELING SYSTEMS RECONSOLIDATION
Figure 4: Reconsolidation window: Simulated HPC lesions produce severe impairment
when performed shortly after reactivation, but not later. Recall tests are performed 7 days
after lesioning, i.e. the data points corresponding to lesions performed 0-23 hours after
reactivation reflect recall tests executed on day 7 after reactivation, etc. The baseline data
points correspond to recall tests at the corresponding time points, but without preceding
lesions.
37
MODELING SYSTEMS RECONSOLIDATION
Effect of systemic PSI infusion before training. Systemic PSI infusion before conditioning
impairs formation of long-term memory but does not impair short-term memory, see Figure
5. This simulation reproduces finding 3 of Table 1.
Figure 5: The effect of systemic PSI infusion before training. Short-term memory (1h) is
unaffected, but long-term memory (24h) is severely impaired.
38
MODELING SYSTEMS RECONSOLIDATION
Effect of PSI on consolidated memory. Systemic PSI infusion during maintenance, i.e. after
completed systems consolidation, does not impair subsequent memory retrieval, see Figure 6.
This simulation reproduces finding 4 of Table 1.
Figure 6: Effect of PSI on consolidated memory. Systemic PSI infusion is simulated on
day 30 after conditioning. The diagram shows recall performance before and after the
intervention. PSI infusion in the maintenance phase does not affect recall performance.
PSI infusion in HPC after reactivation. Post-reactivation PSI infusion in HPC does not
cause immediate memory loss. Rather, the recall impairment develops over a period of more
39
MODELING SYSTEMS RECONSOLIDATION
than 4 but less than 48 hours (Debiec et al., 2002), see Figure 7. This simulation reproduces
findings 8 and 9 of Table 1.
Figure 7: PSI infusion in HPC after reactivation. PSI is infused immediately after
reactivation; recall tests are performed four hours later or 48 hours later. The impairment
only manifests at the later time point.
40
MODELING SYSTEMS RECONSOLIDATION
HPC/ACC-dependence for recall. Consolidation transforms memories from being HPC-
dependent to being ACC-dependent for recall, see Figure 8. These simulations reproduce
findings 1 and 2 of Table 1.
Figure 8: HPC/ACC-dependence for recall. HPC inactivation impairs recall 3 days after
training, but not at 30 days. ACC inactivation does not affect recall 3 days after training,
but causes severe impairment at 30 days.
Temporary ACC-independence after reactivation. Reactivation creates a transient HPC
linkage which temporarily returns the memory to ACC-independence, see Figure 9. These
simulations reproduce findings 12 and 13 of Table 1.
41
MODELING SYSTEMS RECONSOLIDATION
Figure 9: Temporary ACC-independence after reactivation. Before reactivation ACC
inactivation severely impairs recall of a consolidated memory. Six hours after
reactivation neither ACC inactivation nor HPC activation produces significant
impairment. At 24h after reactivation, ACC dependence has returned.
Discussion
We have presented an artificial neural network model of systems memory consolidation
and reconsolidation that accounts for a broad range of findings from the empirical literature,
including those from studies employing hippocampal lesions as well as ones using reversible
inactivation of the hippocampus or anterior cingulate cortex. At the core of the model is a new
connection design in which variable stability arises from simulation of receptor exchanges that
have been observed in glutamatergic synapses.
42
MODELING SYSTEMS RECONSOLIDATION
Reconsolidation
It is worth noting that although the term "reconsolidation" suggests a recapitulation of
consolidation, the model reflects our view that the two processes are quite different. Whereas
systems consolidation is a gradual strengthening of the intra-neocortical connections that link
together the components of a memory, systems reconsolidation consists in the restoration of
stability in such synapses following reactivation-induced destabilization. HPC lesions in the
"consolidation window" (the first couple of weeks after training) and in the "reconsolidation
window" (the first day or two after reactivation) both result in memory deficits -- but for different
reasons: HPC lesions in the consolidation window prematurely interrupts the hippocampal replay
that drives establishment and strengthening of neocortical linkages, leaving a weak memory trace
there. Loss of HPC in the reconsolidation window, in contrast, deprives neocortical neurons of
the HPC stimulation that is needed to restabilize synapses that have been destabilized by
reactivation-induced AMPAR exchange. Without such stimulation, the destabilized neocortical
trace decays.
Our model is the first to demonstrate results analogous to both systems consolidation and
systems reconsolidation in an artificial neural network, and also to reproduce the effects of a
number of experimental interventions described in the empirical literature. This was made
possible by the use of a connection model with a plasticity mechanism inspired by biological
synapses.
Limitations
Although the model is capable of reproducing a number of empirical results, it should be
kept in mind that it is a restricted model with many simplifying assumptions. Thus, while the
model only includes synaptic potentiation as a substrate for the creation of memory traces in
43
MODELING SYSTEMS RECONSOLIDATION
systems consolidation and reconsolidation, it seems likely that other mechanisms are also at play,
perhaps including axonal growth, synaptogenesis, and possibly neurogenesis (Frankland &
Bontempi, 2005). Similarly, the simple receptor trafficking in the network connections is a crude
model of synaptic potentiation, which is known to involve a large number of molecular
processes, most of which are not well understood (Asok et al., 2019). We chose to model
AMPAR trafficking because of the strong evidence of its importance for induction and
maintenance of synaptic potentiation, as discussed in the introduction. Despite its limitations, it
is interesting to find that a model as simple as this is able to capture a significant set of diverse
findings from the consolidation/reconsolidation literature.
Predictions
In closing, we list a number of predictions that have been derived from the model and
may be used to test its validity in future experiments.
1. In the model, post-reactivation susceptibility to HPC lesion is due to the CI-AMPAR to CP-
AMPAR exchange that reactivation triggers in the consolidated connections in the ACC:
subsequent CI-AMPAR restoration depends on HPC activity. A prediction that can be drawn
from the model is therefore, that if reactivation is prevented from triggering AMPA receptor
exchange in the ACC, then HPC lesion in the reconsolidation window will not impair recall.
This could be tested by infusing a drug like GluA23Y into the ACC before reactivation.
GluA23Y is a synthetic peptide that prevents endocytosis (removal) of CI-AMPARs from the
synapse. In contrast, GluA23Y should not be able to prevent the amnestic effect of
hippocampal lesion during the consolidation window, because in this case the impairment is
not due to depotentiation but to interrupted consolidation.
44
MODELING SYSTEMS RECONSOLIDATION
2. The amnestic effect of post-reactivation PSI infusion into HPC is also abolished if
destabilization of ACC connections is blocked. The reasoning is the same as above: If the
ACC connections are not destabilized, then they do not need to be restabilized, and thus
depriving them of hippocampal stimulation should have no amnestic effect. So the prediction
is that GluA23Y infusion in ACC before reactivation would abolish the amnestic effect of
post-reactivation PSI infusion in hippocampus.
3. Destabilization can also be prevented by selectively inhibiting GluN2B-containing NMDA
receptors (Ben Mamou et al., 2006; Milton et al., 2013), cannabinoid receptor 1, or L-type
voltage gated calcium channels (R. Kim et al., 2011). The model predicts that inhibiting any
of these receptors or channels in ACC during reactivation would reduce or eliminate the
amnestic effect of post-reactivation hippocampal lesion or PSI injection.
4. The model attributes the short duration of post-reactivation ACC-independence to
accelerated depotentiation of linkage synapses in hippocampus. Blocking AMPAR
endocytosis in hippocampus, e.g. by local GluA23Y infusion, should extend the durability of
reactivation-induced hippocampal memory traces, and therefore lengthen the duration of
post-reactivation ACC-independence.
5. In the model, post-reactivation hippocampal lesion or PSI infusion causes memory
impairments by depriving neocortical linkage synapses of the stimulation they need to
restabilize. If this is correct, then other means of blocking such stimulation should have the
same amnestic effect. For example, we predict that prolonged (several days) reversible
inactivation of hippocampus after reactivation will impair subsequent recall of the reactivated
memory.
45
MODELING SYSTEMS RECONSOLIDATION
6. In the model, all that is needed for restabilization of a reactivated memory trace is repeated
activation of its ACC linkage. This suggests that it may be possible to compensate for an
inactivated hippocampus (as in prediction 5) by triggering reactivations externally, i.e. by
reminders. This idea is related to a result reported by Lehmann et al. (2009), where repeated
conditioning sessions were shown to significantly speed up the development of
hippocampus-independence.
Data availability statement
All computer program files are available from the ModelDB database, accession number
TBD (http://modeldb.yale.edu/TBD).
References
Abraham, W. C. (2003). How Long Will Long-Term Potentiation Last? Philosophical
Transactions: Biological Sciences, 358(1432), 735 -- 744.
Ackley, D., Hinton, G., & Sejnowski, T. (1985). A Learning Algorithm for Boltzmann
Machines. Cognitive Science, 9(1), 147 -- 169.
Alvarez, P., & Squire, L. R. (1994). Memory and the medial temporal-lobe: A simple network
model. Proceedings of the National Academy of Sciences of the United States of America,
91(15), 7041 -- 7045. https://doi.org/10.1073/pnas.91.15.7041
Amaral, O. B., Osan, R., Roesler, R., & Tort, A. B. L. (2008). A synaptic reinforcement‐based
model for transient amnesia following disruptions of memory consolidation and
reconsolidation. Hippocampus, 18(6), 584 -- 601. https://doi.org/10.1002/hipo.20420
Anagnostaras, S. G., Maren, S., & Fanselow, M. S. (1999). Temporally graded retrograde
amnesia of contextual fear after hippocampal damage in rats: Within-subjects
examination. Journal of Neuroscience, 19(3), 1106 -- 1114.
46
MODELING SYSTEMS RECONSOLIDATION
Asok, A., Leroy, F., Rayman, J. B., & Kandel, E. R. (2019). Molecular Mechanisms of the
Memory Trace. Trends in Neurosciences, 42(1), 14 -- 22.
https://doi.org/10.1016/j.tins.2018.10.005
Baldi, E., & Bucherelli, C. (2015). Brain sites involved in fear memory reconsolidation and
extinction of rodents. Neuroscience and Biobehavioral Reviews, 53, 160 -- 190.
https://doi.org/10.1016/j.neubiorev.2015.04.003
Baltaci, S. B., Mogulkoc, R., & Baltaci, A. K. (2019). Molecular Mechanisms of Early and Late
LTP. Neurochemical Research, 44(2), 281 -- 296. https://doi.org/10.1007/s11064-018-
2695-4
Bartol, T. M., Bromer, C., Kinney, J. P., Chirillo, M. A., Bourne, J. N., Harris, K. M., &
Sejnowski, T. J. (2015). Hippocampal spine head sizes are highly precise. BioRxiv,
016329. https://doi.org/10.1101/016329
Bayley, P. J., Gold, J. J., Hopkins, R. O., & Squire, L. R. (2005). The neuroanatomy of remote
memory. Neuron, 46(5), 799 -- 810. https://doi.org/10.1016/j.neuron.2005.04.034
Ben Mamou, C., Gamache, K., & Nader, K. (2006). NMDA receptors are critical for unleashing
consolidated auditory fear memories. Nature Neuroscience, 9(10), 1237 -- 1239.
https://doi.org/10.1038/nn1778
Besnard, A., Caboche, J., & Laroche, S. (2012). Reconsolidation of memory: A decade of
debate. Progress in Neurobiology, 99(1), 61 -- 80.
https://doi.org/10.1016/j.pneurobio.2012.07.002
Bliss, T., & Collingridge, G. (1993). A Synaptic Model of Memory -- Long-Term Potentiation in
the Hippocampus. Nature, 361(6407), 31 -- 39. https://doi.org/10.1038/361031a0
47
MODELING SYSTEMS RECONSOLIDATION
Bliss, T., & Lømo, T. (1973). Long-Lasting Potentiation of Synaptic Transmission in Dentate
Area of Anesthetized Rabbit Following Stimulation of Perforant Path. Journal of
Physiology-London, 232(2), 331 -- 356.
Bontempi, B., Laurent-Demir, C., Destrade, C., & Jaffard, R. (1999). Time-dependent
reorganization of brain circuitry underlying long-term memory storage. Nature,
400(6745), 671 -- 675. https://doi.org/10.1038/23270
Bortolotto, Z., Bashir, Z., Davies, C., & Collingridge, G. (1994). A Molecular Switch Activated
by Metabotropic Glutamate Receptors Regulates Induction of Long-Term Potentiation.
Nature, 368(6473), 740 -- 743. https://doi.org/10.1038/368740a0
Bourtchouladze, R., Abel, T., Berman, N., Gordon, R., Lapidus, K., & Kandel, E. R. (1998).
Different training procedures recruit either one or two critical periods for contextual
memory consolidation, each of which requires protein synthesis and PKA. Learning &
Memory, 5(4 -- 5), 365 -- 374.
Cajal, S. R. Y. (1894). The Croonian Lecture: La Fine Structure des Centres Nerveux.
Proceedings of the Royal Society of London, 55(331 -- 335), 444 -- 468.
https://doi.org/10.1098/rspl.1894.0063
Canal, C. E., Chang, Q., & Gold, P. E. (2007). Amnesia produced by altered release of
neurotransmitters after intraamygdala injections of a protein synthesis inhibitor.
Proceedings of the National Academy of Sciences of the United States of America,
104(30), 12500 -- 12505. https://doi.org/10.1073/pnas.0705195104
Citri, A., & Malenka, R. C. (2008). Synaptic plasticity: Multiple forms, functions, and
mechanisms. Neuropsychopharmacology, 33(1), 18 -- 41.
https://doi.org/10.1038/sj.npp.1301559
48
MODELING SYSTEMS RECONSOLIDATION
Clem, R. L., & Huganir, R. L. (2010). Calcium-Permeable AMPA Receptor Dynamics Mediate
Fear Memory Erasure. Science, 330(6007), 1108 -- 1112.
https://doi.org/10.1126/science.1195298
Clopath, C., Ziegler, L., Vasilaki, E., Buesing, L., & Gerstner, W. (2008). Tag-Trigger-
Consolidation: A Model of Early and Late Long-Term-Potentiation and Depression. Plos
Computational Biology, 4(12), e1000248. https://doi.org/10.1371/journal.pcbi.1000248
Cronholm, B., & Lagergren, A. (1959). Memory Disturbances After Electroconvulsive-Therapy.
Acta Psychiatrica Et Neurologica, 34(3), 283 -- 310. https://doi.org/10.1111/j.1600-
0447.1959.tb07580.x
Davis, H. P., & Squire, L. R. (1984). Protein synthesis and memory: A review. Psychological
Bulletin, 96(3), 518 -- 559.
de Oliveira Alvares, L., Einarsson, E. Ö., Santana, F., Crestani, A. P., Haubrich, J., Cassini, L.
F., … Quillfeldt, J. A. (2012). Periodically reactivated context memory retains its
precision and dependence on the hippocampus. Hippocampus, 22(5), 1092 -- 1095.
https://doi.org/10.1002/hipo.20983
Debiec, J., LeDoux, J. E., & Nader, K. (2002). Cellular and systems reconsolidation in the
hippocampus. Neuron, 36(3), 527 -- 538.
Dede, A. J. O., & Smith, C. N. (2016). The Functional and Structural Neuroanatomy of Systems
Consolidation for Autobiographical and Semantic Memory. In Current Topics in
Behavioral Neurosciences. SpringerLink (pp. 1 -- 32).
https://doi.org/10.1007/7854_2016_452
49
MODELING SYSTEMS RECONSOLIDATION
Ding, H. K., Teixeira, C. M., & Frankland, P. W. (2008). Inactivation of the anterior cingulate
cortex blocks expression of remote, but not recent, conditioned taste aversion memory.
Learning & Memory, 15(5), 290 -- 293. https://doi.org/10.1101/lm.905008
Dudai, Y. (2004). The neurobiology of consolidations, or, how stable is the engram? Annual
Review of Psychology, 55, 51 -- 86.
https://doi.org/10.1146/annurev.psych.55.090902.142050
Duncan, C. (1949). The Retroactive Effect of Electroshock on Learning. Journal of Comparative
and Physiological Psychology, 42(1), 32 -- 44. https://doi.org/10.1037/h0058173
Ecker, U. K. H., & Lewandowsky, S. (2012). Computational constraints in cognitive theories of
forgetting. Frontiers in Cognition, 3, 400. https://doi.org/10.3389/fpsyg.2012.00400
Einarsson, E. Ö., & Nader, K. (2012). Involvement of the anterior cingulate cortex in formation,
consolidation, and reconsolidation of recent and remote contextual fear memory.
Learning & Memory, 19(10), 449 -- 452. https://doi.org/10.1101/lm.027227.112
Einarsson, E. Ö., Pors, J., & Nader, K. (2015). Systems Reconsolidation Reveals a Selective
Role for the Anterior Cingulate Cortex in Generalized Contextual Fear Memory
Expression. Neuropsychopharmacology, 40(2), 480 -- 487.
https://doi.org/10.1038/npp.2014.197
Fanselow, M., & Kim, J. (1994). Acquisition of Contextual Pavlovian Fear Conditioning Is
Blocked by Application of an Nmda Receptor Antagonist D,l-2-Amino-5-
Phosphonovaleric Acid to the Basolateral Amygdala. Behavioral Neuroscience, 108(1),
210 -- 212. https://doi.org/10.1037//0735-7044.108.1.210
50
MODELING SYSTEMS RECONSOLIDATION
Fonseca, R., Nägerl, U. V., & Bonhoeffer, T. (2006). Neuronal activity determines the protein
synthesis dependence of long-term potentiation. Nature Neuroscience, 9(4), 478 -- 480.
https://doi.org/10.1038/nn1667
Forcato, C., Burgos, V. L., Argibay, P. F., Molina, V. A., Pedreira, M. E., & Maldonado, H.
(2007). Reconsolidation of declarative memory in humans. Learning & Memory, 14(4),
295 -- 303. https://doi.org/10.1101/lm.486107
Frankland, P. W., & Bontempi, B. (2005). The organization of recent and remote memories.
Nature Reviews Neuroscience, 6(2), 119 -- 130. https://doi.org/10.1038/nrn1607
Frankland, P. W., Bontempi, B., Talton, L. E., Kaczmarek, L., & Silva, A. J. (2004). The
involvement of the anterior cingulate cortex in remote contextual fear memory. Science,
304(5672), 881 -- 883. https://doi.org/10.1126/science.1094804
Frankland, P. W., Ding, H.-K., Takahashi, E., Suzuki, A., Kida, S., & Silva, A. J. (2006).
Stability of recent and remote contextual fear memory. Learning & Memory, 13(4), 451 --
457. https://doi.org/10.1101/lm.183406
Frey, U. (1997). Chapter 6 - Cellular Mechanisms Of Long-Term Potentiation: Late
Maintenance. In J. W. Donahoe & V. Packard Dorsel (Eds.), Advances in Psychology
(pp. 105 -- 128). https://doi.org/10.1016/S0166-4115(97)80092-1
Frey, U., Huang, Y., & Kandel, E. (1993). Effects of Camp Simulate a Late-Stage of Ltp in
Hippocampal Ca1 Neurons. Science, 260(5114), 1661 -- 1664.
https://doi.org/10.1126/science.8389057
Frey, U., Krug, M., Reymann, K. G., & Matthies, H. (1988). Anisomycin, an inhibitor of protein
synthesis, blocks late phases of LTP phenomena in the hippocampal CA1 region in vitro.
Brain Research, 452(1 -- 2), 57 -- 65.
51
MODELING SYSTEMS RECONSOLIDATION
Gafford, G. M., Parsons, R. G., & Helmstetter, F. J. (2013). Memory accuracy predicts
hippocampal mTOR pathway activation following retrieval of contextual fear memory.
Hippocampus, 23(9), 842 -- 847. https://doi.org/10.1002/hipo.22140
Genzel, L., Rossato, J. I., Jacobse, J., Grieves, R. M., Spooner, P. A., Battaglia, F. P., … Morris,
R. G. M. (2017). The Yin and Yang of Memory Consolidation: Hippocampal and
Neocortical. Plos Biology, 15(1), e200531. https://doi.org/10.1371/journal.pbio.2000531
Gerard, R. (1955). Biological Roots of Psychiatry. Science, 122(3162), 225 -- 230.
https://doi.org/10.1126/science.122.3162.225
Ghazal, P. (2016). AMPA receptor trafficking in recent vs. Remote memory. Brain Research,
1650, 203 -- 207. https://doi.org/10.1016/j.brainres.2016.09.010
Hardt, O., & Nadel, L. (2018). Systems consolidation revisited, but not revised: The promise and
limits of optogenetics in the study of memory. Neuroscience Letters, 680, 54 -- 59.
https://doi.org/10.1016/j.neulet.2017.11.062
Hardt, O., Nader, K., & Nadel, L. (2013). Decay happens: The role of active forgetting in
memory. Trends in Cognitive Sciences, 17(3), 111 -- 120.
https://doi.org/10.1016/j.tics.2013.01.001
Haubrich, J., Crestani, A. P., Cassini, L. F., Santana, F., Sierra, R. O., de Oliveira Alvares, L., &
Quillfeldt, J. A. (2015). Reconsolidation Allows Fear Memory to Be Updated to a Less
Aversive Level through the Incorporation of Appetitive Information.
Neuropsychopharmacology, 40(2), 315 -- 326. https://doi.org/10.1038/npp.2014.174
Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Approach. John
Wiley & Sons.
52
MODELING SYSTEMS RECONSOLIDATION
Helfer, P., & Shultz, T. R. (2018). Coupled feedback loops maintain synaptic long-term
potentiation: A computational model of PKMzeta synthesis and AMPA receptor
trafficking. PLOS Computational Biology, 14(5), e1006147.
https://doi.org/10.1371/journal.pcbi.1006147
Henley, J. M., & Wilkinson, K. A. (2013). AMPA receptor trafficking and the mechanisms
underlying synaptic plasticity and cognitive aging. Dialogues in Clinical Neuroscience,
15(1), 11 -- 27.
Hong, I., Kim, J., Kim, J., Lee, S., Ko, H.-G., Nader, K., … Choi, S. (2013). AMPA receptor
exchange underlies transient memory destabilization on retrieval. Proceedings of the
National Academy of Sciences of the United States of America, 110(20), 8218 -- 8223.
https://doi.org/10.1073/pnas.1305235110
Huang, Y., Li, X., & Kandel, E. (1994). Camp Contributes to Messy Fiber Ltp by Initiating Both
a Covalently Mediated Early Phase and Macromolecular Synthesis-Dependent Late-
Phase. Cell, 79(1), 69 -- 79. https://doi.org/10.1016/0092-8674(94)90401-4
Hupbach, A., Gomez, R., Hardt, O., & Nadel, L. (2007). Reconsolidation of episodic memories:
A subtle reminder triggers integration of new information. Learning & Memory, 14(1 -- 2),
47 -- 53. https://doi.org/10.1101/lm.365707
Jalil, S. J., Sacktor, T. C., & Shouval, H. Z. (2015). Atypical PKCs in memory maintenance: The
roles of feedback and redundancy. Learning & Memory (Cold Spring Harbor, N.Y.),
22(7), 344 -- 353. https://doi.org/10.1101/lm.038844.115
Jones, B. J., Pest, S. M., Vargas, I. M., Glisky, E. L., & Fellous, J.-M. (2015). Contextual
reminders fail to trigger memory reconsolidation in aged rats and aged humans.
Neurobiology of Learning and Memory, 120. https://doi.org/10.1016/j.nlm.2015.02.003
53
MODELING SYSTEMS RECONSOLIDATION
Judge, M., & Quartermain, D. (1982). Characteristics of Retrograde-Amnesia Following
Reactivation of Memory. Physiology & Behavior, 28(4), 585 -- 590.
https://doi.org/10.1016/0031-9384(82)90034-8
Kandel, E. R. (2000). Cellular mechanisms of learning and the biological basis of individuality.
In E. R. Kandel, J. H. Schwartz, & T. M. Jessel (Eds.), Principles of neural science (4th
ed., pp. 1247 -- 1279). New York: McGraw Hill.
Kandel, E. R., Dudai, Y., & Mayford, M. R. (2014). The Molecular and Systems Biology of
Memory. Cell, 157(1), 163 -- 186. https://doi.org/10.1016/j.cell.2014.03.001
Kehlet, H., & Lunn, V. (1951). Retrograde Amnesia Following Electroshock Therapy. Nordisk
Medicin, 45(19), 716 -- 719.
Kessels, H. W., & Malinow, R. (2009). Synaptic AMPA Receptor Plasticity and Behavior.
Neuron, 61(3), 340 -- 350. https://doi.org/10.1016/j.neuron.2009.01.015
Kim, J., Clark, R., & Thompson, R. (1995). Hippocampectomy Impairs the Memory of Recently,
but Not Remotely, Acquired Trace Eyeblink Conditioned-Responses. Behavioral
Neuroscience, 109(2), 195 -- 203. https://doi.org/10.1037/0735-7044.109.2.195
Kim, J., & Fanselow, M. (1992). Modality-specific retrograde-amnesia of fear. Science,
256(5057), 675 -- 677. https://doi.org/10.1126/science.1585183
Kim, R., Moki, R., & Kida, S. (2011). Molecular mechanisms for the destabilization and
restabilization of reactivated spatial memory in the Morris water maze. Molecular Brain,
4, 9. https://doi.org/10.1186/1756-6606-4-9
Kitamura, T., Saitoh, Y., Takashima, N., Murayama, A., Niibori, Y., Ageta, H., … Inokuchi, K.
(2009). Adult Neurogenesis Modulates the Hippocampus-Dependent Period of
54
MODELING SYSTEMS RECONSOLIDATION
Associative Fear Memory. Cell, 139(4), 814 -- 827.
https://doi.org/10.1016/j.cell.2009.10.020
Lah, S., & Miller, L. (2008). Effects of Temporal Lobe Lesions on Retrograde Memory: A
Critical Review. Neuropsychology Review, 18(1), 24. https://doi.org/10.1007/s11065-
008-9053-2
Land, C., Bunsey, M., & Riccio, D. C. (2000). Anomalous properties of hippocampal lesion-
induced retrograde amnesia. Psychobiology, 28(4), 476 -- 485.
Langille, J. J., & Brown, R. E. (2018). The Synaptic Theory of Memory: A Historical Survey and
Reconciliation of Recent Opposition. Frontiers in Systems Neuroscience, 12.
https://doi.org/10.3389/fnsys.2018.00052
Lattal, K. M., & Abel, T. (2004). Behavioral impairments caused by injections of the protein
synthesis inhibitor anisomycin after contextual retrieval reverse with time. Proceedings
of the National Academy of Sciences of the United States of America, 101(13), 4667 --
4672. https://doi.org/10.1073/pnas.0306546101
Lavenex, P., & Amaral, D. G. (2000). Hippocampal-neocortical interaction: A hierarchy of
associativity. Hippocampus, 10(4), 420 -- 430. https://doi.org/10.1002/1098-
1063(2000)10:4<420::AID-HIPO8>3.0.CO;2-5
Lee, J. L. C. (2009). Reconsolidation: Maintaining memory relevance. Trends in Neurosciences,
32, 413 -- 420. https://doi.org/10.1016/j.tins.2009.05.002
Lee, J. L. C., Nader, K., & Schiller, D. (2017). An Update on Memory Reconsolidation
Updating. Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2017.04.006
55
MODELING SYSTEMS RECONSOLIDATION
Lehmann, H., Sparks, F. T., Spanswick, S. C., Hadikin, C., McDonald, R. J., & Sutherland, R. J.
(2009). Making context memories independent of the hippocampus. Learning & Memory,
16(7), 417 -- 420. https://doi.org/10.1101/lm.1385409
Lewis, D., Mahan, J., & Bregman, N. (1972). Cue-Dependent Amnesia in Rats. Journal of
Comparative and Physiological Psychology, 81(2), 243-.
https://doi.org/10.1037/h0033524
Ling, D. S. F., Benardo, L. S., Serrano, P. A., Blace, N., Kelly, M. T., Crary, John. F., & Sacktor,
T. C. (2002). Protein kinase M zeta is necessary and sufficient for LTP maintenance.
Nature Neuroscience, 5(4), 295 -- 296. https://doi.org/10.1038/nn829
Lisman, J. (2017). Criteria for identifying the molecular basis of the engram (CaMKII,
PKMzeta). Molecular Brain, 10, 55. https://doi.org/10.1186/s13041-017-0337-4
Lynch, M. A. (2004). Long-term potentiation and memory. Physiological Reviews, 84(1), 87 --
136. https://doi.org/10.1152/physrev.00014.2003
Malenka, R. C., & Bear, M. F. (2004). LTP and LTD: An embarrassment of riches. Neuron,
44(1), 5 -- 21. https://doi.org/10.1016/j.neuron.2004.09.012
Manns, J. R., Hopkins, R. O., & Squire, L. R. (2003). Semantic memory and the human
hippocampus. Neuron, 38(1), 127 -- 133. https://doi.org/10.1016/S0896-6273(03)00146-6
Maren, S., Tocco, G., Standley, S., Baudry, M., & Thompson, R. (1993). Postsynaptic Factors in
the Expression of Long-Term Potentiation (ltp) -- Increased Glutamate-Receptor Binding
Following Ltp Induction in-Vivo. Proceedings of the National Academy of Sciences of
the United States of America, 90(20), 9654 -- 9658.
https://doi.org/10.1073/pnas.90.20.9654
56
MODELING SYSTEMS RECONSOLIDATION
Maviel, T., Durkin, T. P., Menzaghi, F., & Bontempi, B. (2004). Sites of Neocortical
Reorganization Critical for Remote Spatial Memory. Science, 305(5680), 96 -- 99.
https://doi.org/10.1126/science.1098180
McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary
learning systems in the hippocampus and neocortex: Insights from the successes and
failures of connectionist models of learning and memory. Psychological Review, 102,
419 -- 457. https://doi.org/10.1037/0033-295X.102.3.419
McCormack, S. G., Stornetta, R. L., & Zhu, J. J. (2006). Synaptic AMPA receptor exchange
maintains bidirectional plasticity. Neuron, 50(1), 75 -- 88.
https://doi.org/10.1016/j.neuron.2006.02.027
Meeter, M., & Murre, J. M. J. (2005). Tracelink: A model of consolidation and amnesia.
Cognitive Neuropsychology, 22(5), 559 -- 587.
https://doi.org/10.1080/02643290442000194
Migues, P. V., Hardt, O., Wu, D. C., Gamache, K., Sacktor, T. C., Wang, Y. T., & Nader, K.
(2010). PKMzeta maintains memories by regulating GluR2-dependent AMPA receptor
trafficking. Nature Neuroscience, 13(5), 630 -- 634. https://doi.org/10.1038/nn.2531
Migues, P. V., Liu, L., Archbold, G. E. B., Einarsson, E. Ö., Wong, J., Bonasia, K., … Hardt, O.
(2016). Blocking Synaptic Removal of GluA2-Containing AMPA Receptors Prevents the
Natural Forgetting of Long-Term Memories. Journal of Neuroscience, 36(12), 3481 --
3494. https://doi.org/10.1523/JNEUROSCI.3333-15.2016
Milekic, M. H., & Alberini, C. M. (2002). Temporally graded requirement for protein synthesis
following memory reactivation. Neuron, 36(3), 521 -- 525. https://doi.org/10.1016/S0896-
6273(02)00976-5
57
MODELING SYSTEMS RECONSOLIDATION
Milner, P. (1989). A Cell Assembly Theory of Hippocampal Amnesia. Neuropsychologia, 27(1),
23 -- 30. https://doi.org/10.1016/0028-3932(89)90087-0
Milton, A. L., Merlo, E., Ratano, P., Gregory, B. L., Dumbreck, J. K., & Everitt, B. J. (2013).
Double Dissociation of the Requirement for GluN2B- and GluN2A- Containing NMDA
Receptors in the Destabilization and Restabilization of a Reconsolidating Memory.
Journal of Neuroscience, 33(3), 1109 -- 1115. https://doi.org/10.1523/JNEUROSCI.3273-
12.2013
Misanin, J. R., Miller, R. R., & Lewis, D. J. (1968). Retrograde Amnesia Produced by
Electroconvulsive Shock after Reactivation of a Consolidated Memory Trace. Science,
160(3827), 554 -- 555. https://doi.org/10.1126/science.160.3827.554
Miserendino, M. J. D., Sananes, C. B., Melia, K. R., & Davis, M. (1990). Blocking of acquisition
but not expression of conditioned fear-potentiated startle by NMDA antagonists in the
amygdala. Nature, 345(6277), 716. https://doi.org/10.1038/345716a0
Monfils, M.-H., Cowansage, K. K., Klann, E., & LeDoux, J. E. (2009). Extinction-
Reconsolidation Boundaries: Key to Persistent Attenuation of Fear Memories. Science,
324(5929), 951 -- 955. https://doi.org/10.1126/science.1167975
Moscovitch, M., Cabeza, R., Winocur, G., & Nadel, L. (2016). Episodic Memory and Beyond:
The Hippocampus and Neocortex in Transformation. Annual Review of Psychology,
67(1), 105 -- 134. https://doi.org/10.1146/annurev-psych-113011-143733
Murre, J. M. J. (1996). TraceLink: A model of amnesia and consolidation of memory.
Hippocampus, 6(6), 675 -- 684.
58
MODELING SYSTEMS RECONSOLIDATION
Nabavi, S., Fox, R., Proulx, C. D., Lin, J. Y., Tsien, R. Y., & Malinow, R. (2014). Engineering a
memory with LTD and LTP. Nature, 511(7509), 348-+.
https://doi.org/10.1038/nature13294
Nadel, L., & Hardt, O. (2010). Update on memory systems and processes.
Neuropsychopharmacology, 36(1), 251 -- 273.
Nadel, L., & Moscovitch, M. (1997). Memory consolidation, retrograde amnesia and the
hippocampal complex. Current Opinion in Neurobiology, 7(2), 217 -- 227.
https://doi.org/10.1016/S0959-4388(97)80010-4
Nadel, L., Winocur, G., Ryan, L., & Moscovitch, M. (2007). Systems consolidation and
hippocampus: Two views. Debates in Neuroscience, 1(2 -- 4), 55 -- 66.
https://doi.org/10.1007/s11559-007-9003-9
Nader, K., & Einarsson, E. Ö. (2010). Memory reconsolidation: An update. Annals of the New
York Academy of Sciences, 1191, 27 -- 41. https://doi.org/10.1111/j.1749-
6632.2010.05443.x
Nader, K., Schafe, G. E., & Le Doux, J. E. (2000). Fear memories require protein synthesis in the
amygdala for reconsolidation after retrieval. Nature, 406(6797), 722 -- 726.
https://doi.org/10.1038/35021052
Okubo-Suzuki, R., Saitoh, Y., Shehata, M., Zhao, Q., Enomoto, H., & Inokuchi, K. (2016).
Frequency-specific stimulations induce reconsolidation of long-term potentiation in
freely moving rats. Molecular Brain, 9, 36. https://doi.org/10.1186/s13041-016-0216-4
Pastalkova, E., Serrano, P., Pinkhasova, D., Wallace, E., Fenton, A. A., & Sacktor, T. C. (2006).
Storage of spatial information by the maintenance mechanism of LTP. Science,
313(5790), 1141 -- 1144. https://doi.org/10.1126/science.1128657
59
MODELING SYSTEMS RECONSOLIDATION
Penfield, W., & Milner, B. (1958). Memory Deficit Produced by Bilateral Lesions in the
Hippocampal Zone. A.M.A. Archives of Neurology & Psychiatry, 79(5), 475 -- 497.
https://doi.org/10.1001/archneurpsyc.1958.02340050003001
Plant, K., Pelkey, K. A., Bortolotto, Z. A., Morita, D., Terashima, A., McBain, C. J., … Isaac, J.
T. R. (2006). Transient incorporation of native GluR2-lacking AMPA receptors during
hippocampal long-term potentiation. Nature Neuroscience, 9(5), 602 -- 604.
https://doi.org/10.1038/nn1678
Platt, S. R. (2007). The role of glutamate in central nervous system health and disease -- A
review. The Veterinary Journal, 173(2), 278 -- 286.
https://doi.org/10.1016/j.tvjl.2005.11.007
Przybyslawski, J., & Sara, S. J. (1997). Reconsolidation of memory after its reactivation.
Behavioural Brain Research, 84(1 -- 2), 241 -- 246. https://doi.org/10.1016/S0166-
4328(96)00153-2
Quirk, G. J., & Mueller, D. (2008). Neural mechanisms of extinction learning and retrieval.
Neuropsychopharmacology, 33(1), 56 -- 72. https://doi.org/10.1038/sj.npp.1301555
Rao-Ruiz, P., Rotaru, D. C., van der Loo, R. J., Mansvelder, H. D., Stiedl, O., Smit, A. B., &
Spijker, S. (2011). Retrieval-specific endocytosis of GluA2-AMPARs underlies adaptive
reconsolidation of contextual fear. Nature Neuroscience, 14(10), 1302-U117.
https://doi.org/10.1038/nn.2907
Rossato, J. I., Bevilaqua, L. R. M., Medina, J. H., Izquierdo, I., & Cammarota, M. (2006).
Retrieval induces hippocampal-dependent reconsolidation of spatial memory. Learning &
Memory, 13(4), 431 -- 440. https://doi.org/10.1101/lm.315206
60
MODELING SYSTEMS RECONSOLIDATION
Rudy, J. W., Biedenkapp, J. C., & O'Reilly, R. C. (2005). Prefrontal cortex and the organization
of recent and remote memories: An alternative view. Learning & Memory, 12(5), 445 --
446. https://doi.org/10.1101/lm.97905
Runyan, J. D., Moore, A. N., & Dash, P. K. (2019). Coordinating what we've learned about
memory consolidation: Revisiting a unified theory. Neuroscience & Biobehavioral
Reviews, 100, 77 -- 84. https://doi.org/10.1016/j.neubiorev.2019.02.010
Sachser, R. M., Santana, F., Crestani, A. P., Lunardi, P., Pedraza, L. K., Quillfeldt, J. A., … de
Oliveira Alvares, L. (2016). Forgetting of long-term memory requires activation of
NMDA receptors, L-type voltage-dependent Ca2+ channels, and calcineurin. Scientific
Reports, 6, 22771. https://doi.org/10.1038/srep22771
Sacktor, T. C. (2011). How does PKMζ maintain long-term memory? Nature Reviews
Neuroscience, 12(1), 9 -- 15. https://doi.org/10.1038/nrn2949
Schiller, D., & Phelps, E. A. (2011). Does reconsolidation occur in humans? Frontiers in
Behavioral Neuroscience, 5. https://doi.org/10.3389/fnbeh.2011.00024
Schneider, A. M., & Sherman, W. (1968). Amnesia: A function of the temporal relation of
footshock to electroconvulsive shock. Science (New York, N.Y.), 159(3811), 219 -- 221.
https://doi.org/10.1126/science.159.3811.219
Scoville, W., & Milner, B. (1957). Loss of Recent Memory After Bilateral Hippocampal
Lesions. Journal of Neurology Neurosurgery and Psychiatry, 20(1), 11 -- 21.
https://doi.org/10.1136/jnnp.20.1.11
Sederberg, P. B., Gershman, S. J., Polyn, S. M., & Norman, K. A. (2011). Human memory
reconsolidation can be explained using the temporal context model. Psychonomic
Bulletin & Review, 18, 455 -- 468. https://doi.org/10.3758/s13423-011-0086-9
61
MODELING SYSTEMS RECONSOLIDATION
Serrano, P., Yao, Y. D., & Sacktor, T. C. (2005). Persistent phosphorylation by protein kinase M
zeta maintains late-phase long-term potentiation. Journal of Neuroscience, 25(8), 1979 --
1984. https://doi.org/10.1523/JNEUROSCI.5132-04.2005
Shema, R., Hazvi, S., Sacktor, T. C., & Dudai, Y. (2009). Boundary conditions for the
maintenance of memory by PKM zeta in neocortex. Learning & Memory, 16(2), 122 --
128. https://doi.org/10.1101/lm.1183309
Shema, R., Sacktor, T. C., & Dudai, Y. (2007). Rapid erasure of long-term memory associations
in the cortex by an inhibitor of PKM zeta. Science, 317(5840), 951 -- 953.
https://doi.org/10.1126/science.1144334
Shultz, T. R. (2003). Computational developmental psychology. MIT Press.
Sierra, R. O., Pedraza, L. K., Zanona, Q. K., Santana, F., Boos, F. Z., Crestani, A. P., …
Quillfeldt, J. A. (2017). Reconsolidation-induced Rescue of a Remote Fear Memory
Blocked by an Early Cortical Inhibition: Involvement of the Anterior Cingulate Cortex
and the Mediation by the Thalamic Nucleus Reuniens. Hippocampus, 27(5), 596 -- 607.
https://doi.org/10.1002/hipo.22715
Smolen, P., Baxter, D. A., & Byrne, J. H. (2012). Molecular Constraints on Synaptic Tagging
and Maintenance of Long-Term Potentiation: A Predictive Model. Plos Computational
Biology, 8(8), e1002620. https://doi.org/10.1371/journal.pcbi.1002620
Sossin, W. S. (2008). Molecular memory traces. In W. S. Sossin, J. C. Lacaille, V. F. Castellucci,
& S. Belleville (Eds.), Essence of Memory (Vol. 169, pp. 3 -- 25). Amsterdam: Elsevier
Science Bv.
Spear, N. E. (1973). Retrieval of memory in animals. Psychological Review, 80(3), 163 -- 194.
https://doi.org/10.1037/h0034326
62
MODELING SYSTEMS RECONSOLIDATION
Squire, L. R., & Alvarez, P. (1995). Retrograde amnesia and memory consolidation: A
neurobiological perspective. Current Opinion in Neurobiology, 5(2), 169 -- 177.
https://doi.org/10.1016/0959-4388(95)80023-9
Squire, L. R., & Cohen, N. (1979). Memory and amnesia: Resistance to disruption develops for
years after learning. Behavioral and Neural Biology, 25(1), 115 -- 125.
https://doi.org/10.1016/S0163-1047(79)90841-0
Squire, L. R., Genzel, L., Wixted, J. T., & Morris, R. G. (2015). Memory Consolidation. Cold
Spring Harbor Perspectives in Biology, 7(8), a021766.
https://doi.org/10.1101/cshperspect.a021766
Squire, L. R., & Zola-Morgan, S. (1991). The Medial Temporal-Lobe Memory System. Science,
253(5026), 1380 -- 1386. https://doi.org/10.1126/science.1896849
Stevens, C. F. (1998). A million dollar question: Does LTP = memory? Neuron, 20(1), 1 -- 2.
https://doi.org/10.1016/S0896-6273(00)80426-2
Sutherland, R. J., Sparks, F. T., & Lehmann, H. (2010). Hippocampus and retrograde amnesia in
the rat model: A modest proposal for the situation of systems consolidation.
Neuropsychologia, 48(8), 2357 -- 2369.
https://doi.org/10.1016/j.neuropsychologia.2010.04.015
Takehara, K., Kawahara, S., & Kirino, Y. (2003). Time-dependent reorganization of the brain
components underlying memory retention in trace eyeblink conditioning. Journal of
Neuroscience, 23(30), 9897 -- 9905.
Teyler, T. J., & Discenna, P. (1986). The Hippocampal Memory Indexing Theory. Behavioral
Neuroscience, 100(2), 147 -- 154. https://doi.org/10.1037/0735-7044.100.2.147
63
MODELING SYSTEMS RECONSOLIDATION
Teyler, T. J., & Rudy, J. W. (2007). The hippocampal indexing theory and episodic memory:
Updating the index. Hippocampus, 17(12), 1158 -- 1169.
https://doi.org/10.1002/hipo.20350
Thompson, R., & Dean, W. (1955). A Further Study on the Retroactive Effect of Ecs. Journal of
Comparative and Physiological Psychology, 48(6), 488 -- 491.
https://doi.org/10.1037/h0041777
Walker, D. L., & Davis, M. (2002). The role of amygdala glutamate receptors in fear learning,
fear-potentiated startle, and extinction. Pharmacology Biochemistry and Behavior, 71(3),
379 -- 392. https://doi.org/10.1016/S0091-3057(01)00698-0
Walker, M. P., Brakefield, T., Hobson, J. A., & Stickgold, R. (2003). Dissociable stages of
human memory consolidation and reconsolidation. Nature, 425(6958), 616 -- 620.
https://doi.org/10.1038/nature01930
Wang, S.-H., & Morris, R. G. M. (2010). Hippocampal-neocortical interactions in memory
formation, consolidation, and reconsolidation. Annual Review of Psychology, 61, 49 -- 79,
C1-4. https://doi.org/10.1146/annurev.psych.093008.100523
Wanisch, K., & Wotjak, C. T. (2008). Time course and efficiency of protein synthesis inhibition
following intracerebral and systemic anisomycin treatment. Neurobiology of Learning
and Memory, 90(3), 485 -- 494. https://doi.org/10.1016/j.nlm.2008.02.007
Westmark, P. R., Westmark, C. J., Wang, S., Levenson, J., O'Riordan, K. J., Burger, C., &
Malter, J. S. (2010). Pin1 and PKM zeta Sequentially Control Dendritic Protein
Synthesis. Science Signaling, 3(112), ra18. https://doi.org/10.1126/scisignal.2000451
64
MODELING SYSTEMS RECONSOLIDATION
Wiltgen, B. J., Brown, R. A. M., Talton, L. E., & Silva, A. J. (2004). New Circuits for Old
Memories: The Role of the Neocortex in Consolidation. Neuron, 44(1), 101 -- 108.
https://doi.org/10.1016/j.neuron.2004.09.015
Wiltgen, B. J., Zhou, M., Cai, Y., Balaji, J., Karlsson, M. G., Parivash, S. N., … Silvaz, A. J.
(2010). The hippocampus plays a selective role in the retrieval of detailed contextual
memories. Current Biology, 20(15), 1336 -- 1344.
https://doi.org/10.1016/j.cub.2010.06.068
Winocur, G. (1990). Anterograde and retrograde amnesia in rats with dorsal hippocampal or
dorsomedial thalamic lesions. Behavioural Brain Research, 38(2), 145 -- 154.
https://doi.org/10.1016/0166-4328(90)90012-4
Winocur, G., Frankland, P. W., Sekeres, M., Fogel, S., & Moscovitch, M. (2009). Changes in
context-specificity during memory reconsolidation: Selective effects of hippocampal
lesions. Learning & Memory, 16(11), 722 -- 729. https://doi.org/10.1101/lm.1447209
Winocur, G., Moscovitch, M., & Bontempi, B. (2010). Memory formation and long-term
retention in humans and animals: Convergence towards a transformation account of
hippocampal-neocortical interactions. Neuropsychologia, 48(8), 2339 -- 2356.
https://doi.org/10.1016/j.neuropsychologia.2010.04.016
Winocur, G., Moscovitch, M., & Sekeres, M. J. (2013). Factors affecting graded and ungraded
memory loss following hippocampal lesions. Neurobiology of Learning and Memory,
106, 351 -- 364. https://doi.org/10.1016/j.nlm.2013.10.001
Winocur, G., Sekeres, M. J., Binns, M. A., & Moscovitch, M. (2013). Hippocampal lesions
produce both nongraded and temporally graded retrograde amnesia in the same rat.
Hippocampus, 23(5), 330 -- 341. https://doi.org/10.1002/hipo.22093
65
MODELING SYSTEMS RECONSOLIDATION
Wittenberg, G., & Tsien, J. (2002). An emerging molecular and cellular framework for memory
processing by the hippocampus. Trends in Neurosciences, 25(10), 501 -- 505.
https://doi.org/10.1016/S0166-2236(02)02231-2
Zhang, Y., Smolen, P., Baxter, D. A., & Byrne, J. H. (2010). The sensitivity of memory
consolidation and reconsolidation to inhibitors of protein synthesis and kinases:
Computational analysis. Learning & Memory, 17(9), 428 -- 439.
https://doi.org/10.1101/lm.1844010
Zola-Morgan, S. M., & Squire, L. R. (1990). The primate hippocampal formation: Evidence for a
time-limited role in memory storage. Science (New York, N.Y.), 250(4978), 288 -- 290.
https://doi.org/10.1126/science.2218534
66
|
1608.04935 | 1 | 1608 | 2016-08-17T12:07:14 | Identification of interictal epileptic networks from dense-EEG | [
"q-bio.NC"
] | Epilepsy is a network disease. The epileptic network usually involves spatially distributed brain regions. In this context, noninvasive M/EEG source connectivity is an emerging technique to identify functional brain networks at cortical level from noninvasive recordings. In this paper, we analyze the effect of the two key factors involved in EEG source connectivity processing: i) the algorithm used in the solution of the EEG inverse problem and ii) the method used in the estimation of the functional connectivity. We evaluate four inverse solutions algorithms and four connectivity measures on data simulated from a combined biophysical/physiological model to generate realistic interictal epileptic spikes reflected in scalp EEG. We use a new network-based similarity index (SI) to compare between the network identified by each of the inverse/connectivity combination and the original network generated in the model. The method will be also applied on real data recorded from one epileptic patient who underwent a full presurgical evaluation for drug-resistant focal epilepsy. In simulated data, results revealed that the selection of the inverse/connectivity combination has a significant impact on the identified networks. Results suggested that nonlinear methods for measuring the connectivity are more efficient than the linear one. The wMNE inverse solution showed higher performance than dSPM, cMEM and sLORETA. In real data, the combination (wMNE/PLV) led to a very good matching between the interictal epileptic network identified from noninvasive EEG recordings and the network obtained from connectivity analysis of intracerebral EEG recordings. These results suggest that source connectivity method, when appropriately configured, is able to extract highly relevant diagnostic information about networks involved in interictal epileptic spikes from non-invasive dense-EEG data. | q-bio.NC | q-bio | Identification ofinterictal epileptic networks from dense-
EEG
Mahmoud Hassan1,2, Isabelle Merlet1,2, Ahmad Mheich1,2,4, Aya Kabbara1,2,4, Arnaud Biraben1,2,3, Anca
Nica3 and Fabrice Wendling1,2
1 INSERM, U1099, Rennes, F-35000, France
2 Université de Rennes 1, LTSI, F-35000, France
3 Neurology department, CHU, Rennes, France
4AZM center-EDST, Lebanese University, Tripoli, Lebanon
Corresponding author:
Mahmoud Hassan
[email protected]
Tel: +33 2 23 23 56 05 / 62 20
Fax: +33 2 23 23 69 17
1
Abstract
Epilepsy is a network disease. The epileptic network usually involves spatially distributed brain regions. In this context,
noninvasive M/EEG source connectivity is an emerging technique to identify functional brain networks at cortical level from
noninvasive recordings. In this paper, we analyze the effect of the two key factors involved in EEG source connectivity
processing: i) the algorithm used in the solution of the EEG inverse problem and ii) the method used in the estimation of the
functional connectivity. We evaluate four inverse solutions algorithms (dSPM, wMNE, sLORETA and cMEM) and four
connectivity measures (r2, h2, PLV, and MI) on data simulated from a combined biophysical/physiological model to generate
realistic interictal epileptic spikes reflected in scalp EEG. We use a new network-based similarity index (SI) to compare
between the network identified by each of the inverse/connectivity combination and the original network generated in the
model. The method will be also applied on real data recorded from one epileptic patient who underwent a full presurgical
evaluation for drug-resistant focal epilepsy.
In simulated data, results revealed that the selection of the inverse/connectivity combination has a significant impact on the
identified networks. Results suggested that nonlinear methods (nonlinear correlation coefficient, phase synchronization and
mutual information) for measuring the connectivity are more efficient than the linear one (the cross correlation coefficient).
The wMNE inverse solution showed higher performance than dSPM, cMEM and sLORETA. In real data, the combination
(wMNE/PLV) led to a very good matching between the interictal epileptic network identified from noninvasive EEG
recordings and the network obtained from connectivity analysis of intracerebral EEG recordings. These results suggest that
source connectivity method, when appropriately configured, is able to extract highly relevant diagnostic information about
networks involved in interictal epileptic spikes from non-invasive dense-EEG data.
2
Introduction
Epilepsy is a network disease (Engel Jr et al. 2013). Over the two past decades, the concept of "epileptic focus" has
progressively evolved toward that of "epileptic network" (Kramer and Cash 2012; Laufs 2012). Indeed, with the progress of
functional neuroimaging techniques, many studies confirmed that the epileptic zone (EZ) can rarely be reduced to a
circumscribed brain area (Bartolomei et al. 2001) as it very often involves distinct brain regions generating both interictal
(Bourien et al. 2005) and ictal activity (Bourien et al. 2004). Among the investigation techniques classically used in the
diagnostic of epilepsy, electrophysiological recordings (typically, magneto- and electro-encephalography including depth-
EEG, referred to as M/EEG) are still extensively used to localize and delineate the EZ in a patient-specific context. Regarding
the numerous methods proposed to process the recorded data; those aimed at characterizing brain connectivity are particularly
suitable to identify networks involved during epileptiform activity (both interictal and ictal).
In the context of invasive EEG signals (intracranial EEG, stereo-EEG and electrocorticoGraphy –EcoG-) recorded in patients
candidate to surgery, these "connectivity" methods have been a topic of extensive research (see (van Mierlo et al. 2014) for
recent review). For instance, the coherence function was shown to localize the seizure onset (Gotman 1987), similarity
indexes were used to distinguish a preictal state from the ongoing interictal activity (Le Van Quyen et al. 2005; Mormann et
al. 2000). Nonlinear regression analysis was applied to intracerebral signals to characterize connectivity patterns at the
seizure onset (Bourien et al. 2004). Readers may refer to previous reviews for more detailed information about brain
connectivity methods applied to non-invasive (van Mierlo et al. 2014) and invasive EEG signals (Wendling et al. 2010) in
drug-resistant focal epilepsies.
In the context of scalp M/EEG recording, connectivity methods have received less attention as compared with invasive EEG.
A number of studies performed at the level of electrodes and focused on ictal periods have been reported aiming at analyzing
seizure propagation (Gotman 1983) or to determine the seizure onset side (Caparos et al. 2006), for instance. For interictal
periods, few connectivity studies made use of dense EEG and phase synchronization (Ramon and Holmes 2013) to identify
epileptic sites. One reason for this paucity of studies may lie in the intricate interpretation of connectivity measures obtained
from scalp recordings. Indeed, this interpretation is not straightforward as signals are severely corrupted by the effects of
volume conduction (Schoffelen and Gross 2009).
Interestingly, some recent studies showed how to overcome this limitation. In line with previous cognitive studies (Astolfi et
al. 2007; Babiloni et al. 2005; Betti et al. 2013; Bola and Sabel 2015; David et al. 2003; David et al. 2002; de Pasquale et al.
2010; Hassan et al. 2015a; Hassan et al. 2014; Hassan and Wendling 2015; Hipp et al. 2011; Hoechstetter et al. 2004;
Liljeström et al. 2015; Schoffelen and Gross 2009), the basic principle is to estimate functional connectivity at the level of
3
brain sources reconstructed from scalp signals. These methods, referred to as "source connectivity" were applied to both
interictal EEG (Coito et al. 2015; Song et al. 2013; Vecchio et al. 2014) and MEG signals (Dai et al. 2012; Malinowska et al.
2014) as well as to EEG signals recorded during seizures (Ding et al. 2007; Jiruska et al. 2013; Lu et al. 2012) or resting
states (Adebimpe et al. 2016; Coito et al. 2016).
Although these approaches all include two steps (M/EEG inverse problem followed by source connectivity estimation), they
strongly differ from a methodological viewpoint. Indeed, various algorithms were used to reconstruct brain sources and both
functional and effective connectivity measures were utilized to assess statistical couplings among time series associated with
reconstructed sources. Therefore, a central issue is the impact of selected methods (EEG inverse solution and connectivity
measure) on the topological/statistical properties of identified epileptic networks activated during paroxysmal activity.
In this paper, we report a quantitative comparison of methods aimed at identifying cortical epileptic networks from scalp
EEG data. The novelty of this work is twofold. First, our comparative study includes simulated dense EEGs generated from
physiologically- and biophysically-plausible models of distributed and coupled epileptic sources. To our knowledge, no
previous study has reported results on the performance of source connectivity methods based on a "ground truth" provided by
realistic computational models of interictal EEG signals (recorded later in time than the dense EEG recordings). Second, in
line with a recent analysis performed on MEG data (Malinowska et al. 2014), networks estimated from real scalp dense EEG
are compared with those obtained from depth-EEG recordings (SEEG).
Materials and Methods
A. Inverse solution algorithms
Given the equivalent current dipole model, EEG signals X(t) recorded from M channels can be considered as linear
combinations of P time-varying current dipole sources S(t):
where G[M, P] is the lead field matrix and N(t) is the noise. As G is known, the EEG inverse problem consists of estimating
the unknown sources
from X(t). Several algorithms have been proposed to solve this problem based on different
assumptions about spatial and temporal properties of sources and regularization constraints. Here, we chose to evaluate the
four algorithms implemented in Brainstorm (Tadel et al. 2011).
1) Weighted Minimum Norm Estimate (wMNE)
Minimum norm estimates (MNE) originally proposed by (Hämäläinen and Ilmoniemi 1994) are based on a search for the
solution with minimum power using the L2 norm to regularize the problem. A common expression for MNE resolution
4
X(t)GS(t)N(t)S(t)
matrix is
where
is the regularization parameter and C represents the noise covariance matrix. The weighted MNE (wMNE)
algorithm compensates for the tendency of MNE to favor weak and surface sources (Hämäläinen 2005). This is achieved by
introducing a weighting matrix WX in
that adjusts the properties of the solution by reducing the bias inherent to the standard MNE solution. Classically, W X is a
diagonal matrix built from matrix G with non-zero terms inversely proportional to the norm of the lead field vectors.
2) dynamical Statistical Parametric Mapping (dSPM)
The dSPM is based on the MNE solution (Dale et al. 2000). For dSPM, the normalization matrix contains the minimum norm
estimates of the noise at each source (Caparos et al. 2006), derived from the noise covariance matrix, defined as:
Where
.
3) Standardized low resolution brain electromagnetic tomography (sLORETA)
sLORETA uses the source distribution estimated from MNE and standardizes it a posteriori by the variance of each estimated
dipole source:
where
. The sLORETA inverse method has been originally described using the
whole brain volume as source space (Pascual-Marqui 2002). For the present study, in order to ease the comparison with other
methods, we restricted the source space to the neocortical surface.
4) Standard Maximum Entropy on the Mean (cMEM)
The Maximum Entropy on the Mean (MEM) solver is based on a probabilistic method where inference on the current source
intensities is estimated from the data, which is the basic idea of the maximum of entropy. The first application of MEM to
5
TT1MNEGGGCGS()T1TwMNEXXGWGCGWXS()dSPMdSPMMNESWS2TdSPMMNEMNEWdiagSCS()sLORETAsLORETAMNESWS2TTsLORETAMNEMNEMNEWdiagSGdiagSGGCS()(())
electromagnetic source localization was reported in (Clarke and Janday 1989). The main feature of this method is its ability to
recover the spatial extent of the underlying sources. Its solution is assessed by finding the closest distribution of source
intensities to a reference distribution in which source intensities are organized into cortical parcels showing homogeneous
activation state (parallel cortical macro-columns with synchronized activity). In addition a constraint of local spatial
smoothness in each parcel can be introduced (Chowdhury et al. 2013).
B. Connectivity measures
We selected four methods that have been widely used to estimate functional brain connectivity from electrophysiological
signals (local field potentials, depth-EEG or EEG/MEG) (see (Wendling et al. 2009) for review). These measures were
chosen to cover the main families of connectivity methods (linear and nonlinear regression, phase synchronization and
mutual information).
Briefly, concerning the regression approaches, the linear cross-correlation coefficient is only limited to the detection of the
linear properties of the relationships between time series. However, mechanisms at the origin of EEG signals were shown to
have strong nonlinear behaviors (Pereda et al. 2005). Thus, we have selected three nonlinear connectivity measures. The
nonlinear regression where the basic idea is to evaluate the dependency of two signals from signal samples only and
independently of the type of relation between the two signals. Concerning the phase synchronization measure, the method
estimates the instantaneous phase of each signal and then computes a quantity based on co-variation of extracted phases to
determine the degree of relationship. Finally, the mutual information method is based on the probability and information
theory to measures mutual dependence between two variables. More technical details about the four methods are presented
hereafter:
1) Cross-correlation coefficient (r2)
The cross-correlation coefficient measures the linear correlation between two variables x and y as a function of their time
delay (
). Referred to as the linear correlation coefficient, it is defined as:
where
and cov denote the standard deviation and the covariance, respectively.
2) Nonlinear correlation coefficient (h2)
The nonlinear correlation coefficient (h2) is a non-parametric measure of the nonlinear relationship between two time series x
and y. In practice, the nonlinear relation between the two time series is approximated by a piecewise linear curve.
6
222()()cov((),())max()xtytxyxtytr
where
and f(x) is the linear piecewise approximation of the nonlinear
regression curve.
3) Mutual information (MI)
The mutual information (MI) between signal x and y is defined as:
where
is the joint probability of x=xi and y=yj. In the case of no relationship between x and y,
=
, so that the MI
is zero for independent processes. Otherwise,
will be positive, attaining its maximal value for identical signals.
4) Phase Locking Value (PLV)
For two signals x and y, the phase locking value is defined as:
where
and
are the unwrapped phases of the signals x and y at time t. The
denotes the average over time. The
Hilbert transform was used to extract the instantaneous phase of each signal.
The h2, PLV and r2 values range from 0 (independent signals) to 1 (fully correlated signals).
C. Data
Simulations
In order to quantitatively assess the performance of source connectivity approaches, we generated simulated EEG signals
following the procedure described in (Cosandier-Rimélé et al. 2008), see figure 1A. The distributed source space consisted in
a mesh of the cortical surface (8000 vertices, ~5 mm inter-vertex spacing) that was obtained by segmenting the grey-white
matter interface from a normal subject's structural T1-weighted 3D-MRI with Freesurfer (Fischl 2012). Dipoles were located
at each vertex of this mesh and oriented radially to the surface at the midway between the white/grey matter interface and the
7
2var(()/())max1var(())xyhytxtytlogijijijxyxyxyxypMIpppijxypijxypijxyppxyMI()()xyittxyPLVe()xt()yt.
pial surface. The time-course of each dipole of the source space was generated from a modified version of the physiologically
relevant neural mass model reported in (Bourien et al. 2005; Wendling et al. 2002; Wendling et al. 2000).
In brief, this computational model was designed to represent a neuronal population with three subsets of neurons (pyramidal
cells P and interneurons I and I') interacting via synaptic transmission (Figure 1-A). Pyramidal cells (P) receive endogenous
excitatory drive (AMPAergic collateral excitation) from other local pyramidal cells and exogenous excitatory drive from
distant pyramidal cells (via external noise input p(t)). They also receive inhibitory drive (GABAergic feedback inhibition)
from both subsets of local interneurons (I and I'). In turn, interneurons receive excitatory input (AMPA) from pyramidal
cells.
A Gaussian noise was used as external input to neuronal population. The mean (m=90) and standard deviation (sigma=30)
were adjusted to represent randomly varying density of incoming action potentials (Aps). However, for the purpose of this
study, a modification was made to this noise model. Indeed, abrupt increase/decrease of the density of Aps can occur in the
external input noise at user-defined times to mimic transient AP volleys from other brain regions involved in the generation
of interictal events. Thus, in this model, a simulated IES can be viewed as the responses of a nonlinear dynamical system
(comprising positive and negative feedback loops) to transient pulses superimposed on a Gaussian noise (classically used in
the neural mass modeling approachs).
As in the standard implementation, the shape (spike component followed by a wave component) can still be controlled by
adjusting excitation and inhibition parameters of each population (gains in feedback loops). However, the aforementioned
modification offers one major advantage: as pulses in the noise input are user-defined, the occurrence times of simulated IESs
are controlled, in contrast with the standard implementation where IESs simply result from random fluctuations of the noise.
The consequence is that this new model feature allows for simulation multi-focal IESs with well-controlled time shifts.
Indeed, as illustrated in Figure 1-A, we could generate delayed epileptiform activity in multiple distant patches just by
introducing short delays between the pulses superimposed on the respective input noises of neuronal populations at each
patch.
Finally, from appropriate setting of the input noise, as well as excitation and inhibition-related parameters at each neural mass
included in simulated epileptic sources, a set of epileptiform temporal dynamics was obtained. These dynamics were assigned
to a source made of contiguous vertices (active source) manually outlined with a mesh visualization software (Paraview,
Kitware Inc., NY, US). Uncorrelated background activities were attributed to the other vertices. Once the amplitude of each
elementary dipole was known, EEG simulations were obtained by solving the forward problem in a 3-layer realistic head
model (inner skull, outer skull and the scalp with conductivity values of 0.33, 0.0042, 0.33 S/m respectively) using the
Boundary Element Method (BEM) with the OpenMEEG (Gramfort et al. 2010) implemented in Brainstorm software.
8
We considered two different scenarios. In the first one (single network), EEG simulations were generated from a single
source located in the inferior parietal region (~1000 mm²). In the second one (two interconnected networks) an additional
source (~1000 mm²) was placed in the middle temporal gyrus. In that case, the temporal dynamics of the second source were
highly correlated with those of the first source, but with a minor delay (30ms). This delay of 30 ms was in the range of 10-50
ms delays that are often observed during interictal spikes at different intracranial recording location (Alarcon et al. 1994;
Alarcon et al. 1997; Emerson et al. 1995; Merlet and Gotman 1999) or at different surface sensors (Barth et al. 1984;
Ebersole 1994) or between the peaks of dipole source activity (Baumgartner et al. 1995; Merlet and Gotman 1999). This
delay was usually interpreted as reflecting propagation between distant regions in the brain. For each scenario, 20 epochs of
60s at 512 Hz containing 30 epileptic spikes were simulated. Each epoch was obtained for a new realization of the input
random noise leading to a new realization of epileptic spikes occurring in background activity. Simulated data were imported
in Brainstorm for further analysis.
Real data
Real data were selected from a patient who underwent presurgical evaluation for drug-resistant focal epilepsy. Seizures were
stereotyped, with a sudden start, febrile motor automatisms of the upper limb, stretching of the neck and torso and no post-
ictal motor deficit. The patient had a comprehensive evaluation including detailed history and neurological examination,
neuropsychological testing, structural MRI, standard 32-channels (Micromed, Italy) as well as Dense-EEG 256-channels
(EGI, Electrical Geodesic Inc., Eugene, USA) scalp EEG with video recordings and intracerebral EEG recordings (SEEG).
MRI showed a focal cortical dysplasia in the mesial aspect of the orbito-frontal region. Dense-EEG was recorded for one
hour, at 1000 Hz following the procedure approved by the National Ethics Committee for the Protection of Persons (CPP,
agreement number 2012-A01227-36). The patient gave his written informed consent to participate in this study. This
recording revealed sub-continuous spike activity at the most left frontopolar basal electrodes. From this interictal epileptic
activity, 85 spikes were visually selected away from the occurrence of any artefacts (muscle activity, blood pulsation, eye
blinks). Each spike was centered in a 2s window and all 85 windows were concatenated for further analysis.
As part of his presurgical evaluation, the patient also underwent intracerebral SEEG recordings with 9 implanted electrodes
(10±18 contacts; length: 2 mm, diameter: 0.8 mm; 1.5 mm apart) placed intracranially according to Talairach's stereotactic
method in the left frontal and temporal region, see Figure 1C. The positioning of the electrodes was determined from
available non-invasive information and hypotheses about the localization of his epileptic zone. From these data, subsets of 25
out of the 118 original leads were selected. This selection was made according to the following criteria: i) only contacts
showing grey matter activity were retained and ii) among them, only the contact showing the maximal activity was kept when
similar intracerebral activity was observed on several contacts.
9
D. Data analysis
Scalp-EEG based interictal epileptic networks. As illustrated in Figure 1B, source activity was estimated using four inverse
algorithms (dSPM, wMNE, sLORETA and cMEM, see materials and methods section –A). A baseline of 1s length was used
to estimate the noise covariance matrix both on simulated and real scalp EEG data. For real data, source localization was
applied on averaged spikes, taking as time reference the maximum of the negative peak, while for simulated data the source
localization was applied to non-averaged spikes. The cortical surface was anatomically parcellated into 148 regions of
interest (ROI) (Destrieux et al. 2010) and then re-subdivided into ~1500 sub-ROIs using Brainstorm. The 148 ROIs provided
initially by the Destrieux Atlas (using Freesurfer) were quasi equally subdivided to obtain the 1500 sub-ROIs with 1cm2
average sizes. Time series of the reconstructed source activities were averaged over each of the 1500 ROIs.
After the reconstruction of sources (source localization and estimation of temporal dynamics), functional connectivity was
estimated using four methods (r2, h2, PLV, and MI, see materials and methods section -A-). Each quantity was computed on
the set of 2 s single spikes. All connectivity matrices (1500 x 1500) were thresholded as follows. We computed the strength
of each node of the weighted undirected graph and we kept nodes with the highest 1% strength values. The same threshold
was applied on the adjacency matrices for all combinations (inverse/connectivity). The strength was defined as the sum of all
edge weights for each node; all weights were positive and normalized between 0 and 1.
In order to define the reference networks, all the dipoles were supposed synchronized and the reference network reflected a
fully connected undirected graph. In the case of double network scenario, a number of 37 sub-regions (nodes) were
considered. The dynamics of the dipoles associated to these nodes were similar and resulting a 37x37 fully connected
network where connections (local and remote) between the 37 nodes have the same weight value.
Quantification of network similarity. In order to compare the reference brain network simulated in the model with the
network identified from simulated scalp EEG by each of the inverse/connectivity combination (Figure 1B), we used a
network similarity algorithm recently developed in our team (Mheich et al. 2015a), see supplementary materials for more
details about the algorithm. The main advantage of this algorithm is that it takes into account the spatial location (3D
coordinates) of the nodes when comparing two networks, in contrast with other methods based on the sole statistical
properties of compared graphs. The algorithm provides a normalized Similarity Index (SI): 0 for no similarity and 1 for two
identical networks (same properties and topology). The connectivity analysis, the network measures and network
visualization were performed using EEGNET (Hassan et al. 2015a; Hassan et al. 2015b).
10
Depth-EEG based interictal epileptic networks. Functional connectivity using h2 were directly computed from SEEG signals
at the 25 selected intracerebral electrode contacts. Adjacency matrices (25 x 25) were obtained and thresholded using the
same procedure than that applied to the graphs obtained for scalp dense EEG (both simulated and real).
Scalp-EEG-based vs. depth-EEG-based epileptic network matching. In order to compare the graphs in the three-dimensional
coordinates system of the cortex mesh, the 3D coordinates of the SEEG were first estimated by the co-registering the patient'
CT scan and MRI. These points were then projected on the surface mesh. The transformation from MRIs (voxels) coordinates
to surface (SCS/MNI) coordinates was realized in brainstorm. The Scalp-EEG-based and depth-EEG-based epileptic
networks were visually compared by matching the identified regions (nodes) in both networks.
Statistical analysis: On the simulated data, a Wilcoxon rank-sum test was used to compare between the SIs obtained for each
combination at each trial, corrected for multiple comparison using Bonferroni approach.
Results
A. Simulated Data: Influence of the source reconstruction/functional connectivity combination
The results obtained in the case of the single network scenario are illustrated in Figure 2, for the 16 different combinations of
the source reconstruction and functional connectivity methods. The visual investigation of these results revealed that
networks identified using the different combinations of methods were concordant with the reference network (figure 2B).
Indeed none of the identified networks had nodes in a remote region (Figure 2A). The qualitative analysis also showed that
the number of nodes and the connections between them varied according to the combination of methods used. For a given
connectivity approach, changing the localization method did not dramatically modify the network aspect, except for cMEM.
Conversely, for a given source localization approach, changing the functional connectivity measure changed, qualitatively,
the network. Although this was difficult to assess visually, h2 combined with MNE or sLORETA was giving the network that
best matched the reference network while cMEM/MI provided a result that was different from the reference network.
Quantification of these differences is provided in figure 2C. Overall, values of network similarity were relatively high and
ranged from 70 to 82%. For a given connectivity approach, changing only the localization algorithm slightly modified SI
values by 3% (h2) to 8% (MI). For a given source localization approach, the SIs varied within 9% (wMNE) to 12% (dSPM).
Results obtained using MI were on average better than PLV, r2 and h2. The combination providing the highest similarity
values between the estimated and the actual network was dSPM/MI (82.2%) followed by wMNE/MI (82%) and wMNE-PLV
(82%). The lowest similarity value was obtained with the dSPM/h2 combination. The Wilcoxon rank-sum test did not reveal
any significant difference between the similarity values obtained in this first study.
11
Results obtained in the case of two interconnected networks for the 16 combinations of the inverse/connectivity methods are
reported in Figure 3. Results indicate that the networks identified by all the combinations are relatively close to the model
network (figure 3B) since, similarly to the previously scenario, there was no node in other distant regions or in the right
hemisphere. The networks did not qualitatively change much for a given connectivity measure except for cMEM. Rather, as
observed in the first scenario, the variability between the different combinations was more related to the choice of the
connectivity measure, given a source localization approach. The results of PLV (whatever the inverse solution algorithm)
provide the closest result to the reference network. cMEM/MI shows also a relatively close network to the reference network
while cMEM/h2 indicated, visually, the farthest result from the reference network.
Values of network similarity are reported in figure 3C. These values were lower than those in the single network scenario,
ranging from 57 to 73%. For a particular connectivity measure, changing the inverse algorithm modified the SIs by 1% (r2) to
8% (h2) while for a given source reconstruction algorithm, the SIs varied between 6% (dSPM) to 13% (wMNE). The
combination providing the result closest to the reference network was wMNE/PLV (73%). High values were also obtained
with sLORETA/PLV (68%) and cMEM/PLV (66%). The cMEM/ h2 combination shows the lowest SI value (57%).
Interestingly, for scenario 2 results obtained with wMNE/PLV were significantly closest to the actual network than the other
ones (Wilcoxon rank-sum test, p<0.01, corrected using Bonferroni).
B. EEG source localization vs. functional connectivity
An essential issue that is addressed in this paper relates to the difference between the proposed "network-based" approach
and the classical approach using source localization only. In figure 4, we show two typical examples of the difference
between the proposed network-based analysis and the classical localization approach. The first example is for cMEM
combined with MI vs. cMEM only. This Figure shows that the information extracted in both cases was noticeably different.
The source connectivity approach identified a network close to the reference one (figure 4A), with nodes both in the parietal
and in the temporal region (figure 4B). There were no spurious nodes in remote regions. In contrast, with the sole source
localization, after averaging the results over a 50 ms interval around each of the epileptic peaks, the parietal source was well
retrieved while the temporal source remained almost unobserved. The second example was wMNE/PLV vs. wMNE, the
figure shows that the network-based approach was able to identify a network close to the reference with no spurious
connections in distant regions. The source localization approach identified the two regions different energies at. Moreover,
many spurious sources were observed in remote regions. Similar results were observed for single network configuration.
C. Real data: Scalp-EEG-based vs. depth-EEG-based epileptic network
12
The results obtained from real data recorded in a patient are described on Figure 5. In this patient, the sources of scalp EEG
interictal spikes were widespread over the left frontal and temporal regions. Sources with maximum activity were found in
the left frontal pole and orbitofrontal regions but a substantial activation was also detected in the left temporal as well as right
frontal poles (figure 5A, left). When combining wMNE and PLV on the same scalp EEG data, the source connectivity
approach retrieved a 5-nodes network in the left frontal lobe, involving the mesial (rectus gyrus) and lateral orbitofrontal
region as well as the anterior cingulate gyrus (figure 5A, right). This result was concordant with that the network identified
directly from intracerebral recordings by computing the functional connectivity among SEEG signals (Figure 5B right).
Indeed, the depth-EEG based network involved six nodes in the left mesial orbito-frontal (rectus gyrus), and anterior
cingulate region. All these nodes were also identified by the visual analysis (Figure 5B, left) as regions involved in the main
interictal activity (rectus gyrus) as well as in the propagated interictal activity (cingulate gyrus).
The similarity indices between networks identified by each of the combination with the depth-EEG-based network are
presented in figure 5C. Results showed that the wMNE/PLV provides the highest SI value (70%) followed by wMNE/h2
(47%) and sLORETA/PLV (47%). The cMEM method showed the lowest SI values whatever the connectivity measure (6%,
1%, 1% and 1% for cMEM/MI, cMEM/PLV, cMEM/h2 and cMEM/r2 respectively).
Discussion
Identifying epileptic brain networks from noninvasive M/EEG data is a challenging issue. Recently, source localization
combined with functional connectivity analysis led to encouraging findings in the estimation of functional cortical brain
networks from scalp M/EEG recordings (Coito et al. 2015; Jiruska et al. 2013; Malinowska et al. 2014). Nevertheless, the
joint use of these two approaches is still novel and raises a number of methodological issues that should be controlled in order
to get appropriate and interpretable results. In this paper, we reported a comparative study -in the context of epilepsy- of the
networks obtained from all possible combinations between four algorithms to solve the EEG inverse problem and four
methods to estimate the functional connectivity. An originality of this study is related to the use of dense EEG signals
simulated data from a realistic model of multi-focal epileptic zone as a ground truth for comparing the performance of
considered methods. To our knowledge, a model-based evaluation of source connectivity methods has not been performed
yet. A second – and still novel - aspect is the use of depth-EEG signals (intracerebral recordings performed during presurgical
evaluation of drug-resistant epilepsy) to evaluate the relevance of networks identified from scalp EEG data. Overall results
obtained on simulated as well on real data indicated that the combination of the wMNE and the PLV methods leads to the
13
most relevant networks as compared with the ground-truth (simulations) or with the intracerebrally-identified networks
(patient data). Results are more specifically discussed hereafter.
Methodological considerations
The connectivity matrices were thresholded by keeping the nodes with highest strength values (strongest 1%). This procedure
was used to standardize the comparison between all the combinations. We were aware about the effect of this threshold and
we realized the comparative study using different threshold values. All threshold values were found to lead to the same
differences between the method (inverse/connectivity) combinations.
In this paper, we have averaged the reconstructed sources within the same region of interest defined by the parcellation
process based on Destrieux atlas. This approach was frequently used in the context of M/EEG source connectivity (de
Pasquale et al. 2010; Fraschini et al. 2016; Hassan et al. 2015a). However, such an averaging procedure may increase the
noise power since its computation is performed over sources that, with some probability, may not exhibit correlation
(Brookes et al. 2014) where the need of alternative approaches such as flipping the sign of the sources in each ROIs before
averaging the regional time series (Fraschini et al. 2016) or developing methods based on probabilistic maps, a widely
approach used in the fMRI-based analysis, for instance.
Although EEG source connectivity reduced the problem of volume conduction as compared with scalp EEG connectivity, it
does not yet provide a perfect solution. The volume conduction effect is a challenging issue when performing EEG/MEG
inverse solution (Schoffelen and Gross 2009). In the connectivity context, the main effect of the volume conduction is the
appearance of 'artificial' connections among close sources, a problem often referred to as 'source leakage'. The use of a high
spatial resolution (high number of ROIs) may increase this problem. A few approaches have been proposed recently to deal
with the source leakage by either normalizing the edges weights by the distance between the nodes or removing the edges
between very close sources. Although, these approaches have some advantages, it was shown that, in most cases, they also
remove 'real' connections (Schoffelen and Gross 2009). In this context, some connectivity methods such as PLV have been
shown to reduce the volume conduction (Hipp et al. 2011). This may explain the good performance of this method. Indeed, in
the double network scenario, PLV was able to detect the long-range connections between parietal and temporal networks.
Four inverse/connectivity algorithms were evaluated in this paper. It is worth mentioning that some other inverse algorithms
like MUSIC-based and beamforming as well some connectivity measures such as power envelope correlation (O'Neill et al.
2015) were not included in this study. Moreover, we were focusing in this paper on evaluating different families of
'functional' connectivity methods regardless the directionality of these connections. Nevertheless, we consider that the
analyses of the 'effective' connectivity methods that investigate the causality between brain regions may be of interest in the
14
context of epilepsy (Coito et al. 2016; Coito et al. 2015). In this perspective, methods such as the granger causality, the
transfer entropy could be added to expand this comparative study. In addition, all methods evaluated in the paper were
bivariate, multivariate methods such as those based on the MVAR model were not included in our study. Different
methodological questions appear when using MVAR-based approaches. First, the successful estimation of MVAR such as
Partial Directed Coherence (PDC) or Directed Transfer Function (DTF) depends largely on the fitted MAR model, since all
information is resulting from the estimated model parameters. In practice, this issue is difficult and directly related to the
choice of an optimal model order and an optimal epoch length. Concerning the optimal model, most of the criteria were
originally proposed for univariate AR modeling and no consensus was reported about multivariate ones. The second crucial
question is how to choose the proper window size specially that MVAR model assumes that the underlying process is
stationary, while neurophysiological activity are transient and may rapidly change their states representing high nonstationary
behaviors (Pereda et al. 2005). Nevertheless the MVAR (when carefully applied) could provide complementary information
not only about the link exists between two signals but also if one structure drive another of if there is feedback between these
structures (Kuś et al. 2004). The directionality could be also defined as 'time-delay' between two regional time series which
can be computed using linear or nonlinear correlation coefficients. As our main objective in this study was to compare
inverse algorithms and 'functional' connectivity methods using same criteria (here we used similarity between reference and
estimated undirected graphs), we didn't investigate the time-delays in the presented quantitative analysis. In addition this
feature cannot be computed for all the selected methods (the case of the phase synchronization method for instance). We
believe that the directionality, estimated from Granger causality or/and time delays, is indeed an interesting supplementary
feature in the context of epileptic seizure propagation and will be a potential element for further analysis.
The head model used in this study was computed using the Boundary Element Method (BEM) with three layers (skin, skull
and brain). This model was widely used in the context of M/EEG source estimation (Fuchs et al. 2007) as a compromise
between computational cost and accuracy. Nevertheless, other methods exist to compute the head model such as the Finite
Element Method (FEM) or adding other layers such as cerebrospinal fluid (CSF). These methods can possibly have effect of
the resultant network (Cho et al. 2015). The evaluation of the above mentioned parameters/factors may be the topic of further
investigation.
Identification of interictal epileptic networks from scalp dense-EEG data
A salient feature of epilepsy in general and epileptic networks is the increased synchronization among interconnected
neuronal populations distributed over distant areas. This "hyper"- synchronization often leads to an increase of brain
connectivity, not only during the transition to seizures but also during interictal periods, as shown in many studies based on
15
intracranial recordings (see (Wendling et al. 2010) for review). In this context, the combination of the M/EEG source imaging
with the functional connectivity measures has recently disclosed promising findings to identify pathological brain networks,
at the cortical level (Dai et al. 2012; Lu et al. 2012; Malinowska et al. 2014; Song et al. 2013).
However, two factors seem to be crucial for reliable estimation of EEG source connectivity: i) the number of scalp electrodes
and ii) the combination between the inverse solution algorithm and the functional connectivity measure. Concerning the
number of electrodes, it was reported that the increase of the spatial resolution by using dense EEG may dramatically
improve the accuracy of the source localization analysis (Michel and Murray 2012; Song et al. 2015). In addition, the use of
dense EEG, as compared to classical montages (32 or 64 electrodes), is needed to accurately identify functional brain
networks from scalp EEG (Hassan et al. 2014). To overcome this issue, dense-EEG (256 electrodes) recordings were used in
this study. The main feature of this system is the excellent coverage of the subject's head by surface electrodes allowing for
improved reconstruction of the cortical activity from non-invasive scalp measurements, as compared with more standard 32-
128 electrode systems (Song et al. 2015). Regarding the combination of inverse/connectivity methods, most of reported
studies have empirically selected a combination while this selection was shown to have a dramatic impacts on results, in term
of identified network topology (Hassan et al. 2014). The present analysis brings further confirmation of this high variability
observed when different inverse solutions and/or connectivity measures are being used in the pipeline leading to cortical
networks from EEG signals.
A major advantage of the EEG source connectivity approach as presented here is that reconstructed sources and associated
networks were identified for the whole brain. Then graph-based metrics (strength values) were computed to characterize the
networks and the similarity index was used to compare the networks obtained from various method combinations. In
addition, functional connectivity was applied directly to the reconstructed signals and not on derived components. In this
regard, this study differs from (Malinowska et al. 2014) where connectivity was estimated on signals components obtained by
ICA decomposition. Although the methodological issue of measuring connectivity between independent components still
holds, a future interesting study will compare the results obtained from the ICA-based approach to those reported here from
source connectivity.
EEG source localization vs. functional connectivity
Source localization methods have been widely applied to interictal epileptic spikes (Becker et al. 2014). The goal of these
approaches is the localization of brain generators of epileptic activity from scalp recordings. A fundamental question that is
addressed in this paper relates to the difference between the source connectivity and the source localization approach. This
study indicated that the information extracted from dense-EEG recordings in both cases can differ dramatically. First, the
16
connectivity is an additional step to the simple source reconstruction/localization. Second, the fundamental difference
between both methods is that the source localization ignores all possible communications between brain regions. When
performing source localization analysis, the sources with highest amplitude (averaged at given time period or computed at the
instant of peak amplitude of the signal) are classically kept. However, to some extent (depending on threshold), this approach
may neglect the possible contribution of "low energy" sources participating into the network activity.
Conversely, the hypothesis behind the network-based approach (typically when phase synchronization methods are used as
connectivity measure) is that sources can be synchronized regardless their amplitude. To this extent, we believe that the
network-based approach allows for revealing networks that are more specific to epileptic networks, as hyper-synchronization
phenomena constitute a typical hallmark of such networks. An illustrative example in this paper is the poor involvements of
the temporal lobe region when the sole source localization approach (in the case of cMEM) was applied while both parietal
and temporal networks (as a priori introduced in the EEG generation model) are retrieved by the connectivity-based approach
(cMEM/h2). Note that we have averaged the source localization results in a time window of 50ms to cover the time delay of
30ms set in the model between the two brain regions. Different time window were used to avoid the effect of the window
length. All tested windows (30ms, 40ms, 50ms and 60ms) showed similar observations i.e. the absence of the temporal
sources (not shown here). The fact that epilepsy is considered as a network disease can explain the low performance of some
of the inverse methods as these methods were originally developed to localize 'local' epileptic foci characterized by high-
energy sources regardless the interrelationships between brain regions. Our results show that EEG source connectivity
methods are more suited in the case of multi-focal epileptic zone. More generally, they support the recent tendency in brain
disorder research which is the necessity to move from localizing 'pathological areas' to identifying 'altered networks'
(Diessen et al. 2013; Fornito et al. 2015).
Epilepsy as a network disorder
There is increasing evidence that epileptic activity involves brain networks rather than a single well circumscribed region and
that these dysfunctional networks contribute to both ictal and interictal activity (Bourien et al. 2005; Bourien et al. 2004;
Coito et al. 2015; Engel Jr et al. 2013; Hipp et al. 2011). Functional connectivity was widely applied to depth-EEG data to
predict seizures (Mormann et al. 2000) and identify epileptic networks in partial epilepsies (Bartolomei et al. 2001). These
studies showed alterations of synchronization in brain networks during interictal to ictal transition (Wendling et al. 2003) as
well as during seizures (Diessen et al. 2013; Jiruska et al. 2013; Schindler et al. 2008). Most of these studies were performed
using invasively-recorded data in patient's candidate to surgery. Interestingly, our results show that pathological networks
involved during epileptiform activity can also be identified from scalp EEG.
17
Indeed, we have evaluated the performance of a relatively new approach aiming at identifying epileptic brain networks from
scalp EEG. The application of the method on real data showed the good performance of this method in term of network
identified from scalp EEG as compared with those identified from intracerebral EEG. Note that the comparison was done
only by computing h2 between the intracerebral signals based on a large number of studies showing that h2 is one of the most
adapted metrics to compute functional connectivity between intracerebral recordings (Bettus et al. 2008; Wendling et al.
2010). Although it is obviously difficult to conclude on a single patient analysis, results showed good matching between
scalp-EEG based networks and both the depth-EEG based networks and the expert judgment. Therefore, future work will
consist in the application of the EEG source connectivity on a big database of real dense EEG data recorded from epileptic
patients. In these patients candidate to surgery, we plan to use also intracerebral EEG signals as a reference to validate the
identified networks. In addition, due to the excellent temporal resolution of the EEG, the dynamic behaviors of the epileptic
networks will be also explored (Hassan et al. 2015a; Mheich et al. 2015b).
Finally, the capacity to describe epileptic activity not only according to the sites where epileptiform activity is generated but
also according to the abnormal functional relationships between these sites can definitively improve the surgical approach.
We speculate that in order to better understand and ultimately prevent seizures, it is essential to identify and then
remove/disconnect pathological nodes of the network (exhibiting abnormal hyper-synchronization capability). The proposed
method contributes to this aim and reported results constitute a first step toward the development of more efficient non-
invasive diagnostic methods for clinical epileptology.
Acknowledgements
This work has received a French government support granted to the CominLabs excellence laboratory and managed by the
National Research Agency in the "Investing for the Future" program under reference ANR-10-LABX-07-01. It was also
financed by the Rennes University Hospital (COREC Project named conneXion, 2012-14).
References
Adebimpe A, Aarabi A, Bourel-Ponchel E, Mahmoudzadeh M, Wallois F (2016) EEG resting state functional connectivity analysis in children with benign
epilepsy with centrotemporal spikes Frontiers in Neuroscience 10
Alarcon G, Guy C, Binnie C, Walker S, Elwes R, Polkey C (1994) Intracerebral propagation of interictal activity in partial epilepsy: implications for source
localisation Journal of Neurology, Neurosurgery & Psychiatry 57:435-449
Alarcon G et al. (1997) Origin and propagation of interictal discharges in the acute electrocorticogram. Implications for pathophysiology and surgical
treatment of temporal lobe epilepsy Brain 120:2259-2282
Astolfi L et al. (2007) Comparison of different cortical connectivity estimators for high‐resolution EEG recordings Human brain mapping 28:143-157
Babiloni F et al. (2005) Estimation of the cortical functional connectivity with the multimodal integration of high-resolution EEG and fMRI data by directed
transfer function Neuroimage 24:118-131
Barth DS, Sutherling W, Engle J, Beatty J (1984) Neuromagnetic evidence of spatially distributed sources underlying epileptiform spikes in the human brain
Science 223:293-296
18
Bartolomei F, Wendling F, Bellanger J-J, Régis J, Chauvel P (2001) Neural networks involving the medial temporal structures in temporal lobe epilepsy
Clinical neurophysiology 112:1746-1760
Baumgartner C et al. (1995) Propagation of interictal epileptic activity in temporal lobe epilepsy Neurology 45:118-122
Becker H et al. (2014) EEG extended source localization: tensor-based vs. conventional methods NeuroImage 96:143-157
Betti V, Della Penna S, de Pasquale F, Mantini D, Marzetti L, Romani GL, Corbetta M (2013) Natural scenes viewing alters the dynamics of functional
connectivity in the human brain Neuron 79:782-797
Bettus G, Wendling F, Guye M, Valton L, Regis J, Chauvel P, Bartolomei F (2008) Enhanced EEG functional connectivity in mesial temporal lobe epilepsy
Epilepsy Res 81:58-68 doi:10.1016/j.eplepsyres.2008.04.020
Bola M, Sabel BA (2015) Dynamic reorganization of brain functional networks during cognition NeuroImage 114:398-413
Bourien J, Bartolomei F, Bellanger JJ, Gavaret M, Chauvel P, Wendling F (2005) A method to identify reproducible subsets of co-activated structures during
interictal spikes. Application to intracerebral EEG in temporal lobe epilepsy Clinical neurophysiology : official journal of the International
Federation of Clinical Neurophysiology 116:443-455 doi:10.1016/j.clinph.2004.08.010
Bourien J, Bellanger JJ, Bartolomei F, Chauvel P, Wendling F (2004) Mining reproducible activation patterns in epileptic intracerebral EEG signals:
application to interictal activity IEEE transactions on bio-medical engineering 51:304-315 doi:10.1109/TBME.2003.820397
Brookes MJ et al. (2014) Measuring temporal, spectral and spatial changes in electrophysiological brain network connectivity Neuroimage 91:282-299
Caparos M, Louis-Dorr V, Wendling F, Maillard L, Wolf D (2006) Automatic lateralization of temporal lobe epilepsy based on scalp EEG Clinical
neurophysiology 117:2414-2423
Cho J-H, Vorwerk J, Wolters CH, Knösche TR (2015) Influence of the head model on EEG and MEG source connectivity analysis Neuroimage 110:60-77
Chowdhury RA, Lina JM, Kobayashi E, Grova C (2013) MEG source localization of spatially extended generators of epileptic activity: comparing entropic
and hierarchical Bayesian approaches PloS one 8:e55969
Clarke C, Janday B (1989) The solution of the biomagnetic inverse problem by maximum statistical entropy Inverse Problems 5:483
Coito A et al. (2016) Altered directed functional connectivity in temporal lobe epilepsy in the absence of interictal spikes: A high density EEG study
Epilepsia
Coito A et al. (2015) Dynamic directed interictal connectivity in left and right temporal lobe epilepsy Epilepsia 56:207-217
Cosandier-Rimélé D, Merlet I, Badier J-M, Chauvel P, Wendling F (2008) The neuronal sources of EEG: modeling of simultaneous scalp and intracerebral
recordings in epilepsy NeuroImage 42:135-146
Dai Y, Zhang W, Dickens DL, He B (2012) Source connectivity analysis from MEG and its application to epilepsy source localization Brain topography
25:157-166
Dale AM, Liu AK, Fischl BR, Buckner RL, Belliveau JW, Lewine JD, Halgren E (2000) Dynamic statistical parametric mapping: combining fMRI and
MEG for high-resolution imaging of cortical activity Neuron 26:55-67
David O, Cosmelli D, Hasboun D, Garnero L (2003) A multitrial analysis for revealing significant corticocortical networks in magnetoencephalography and
electroencephalography Neuroimage 20:186-201
David O, Garnero L, Cosmelli D, Varela FJ (2002) Estimation of neural dynamics from MEG/EEG cortical current density maps: application to the
reconstruction of large-scale cortical synchrony IEEE Transactions on Biomedical Engineering, 49:975-987
de Pasquale F et al. (2010) Temporal dynamics of spontaneous MEG activity in brain networks Proceedings of the National Academy of Sciences 107:6040-
6045
Destrieux C, Fischl B, Dale A, Halgren E (2010) Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature
Neuroimage 53:1-15
Diessen E, Diederen SJ, Braun KP, Jansen FE, Stam CJ (2013) Functional and structural brain networks in epilepsy: what have we learned? Epilepsia
54:1855-1865
Ding L, Worrell GA, Lagerlund TD, He B (2007) Ictal source analysis: localization and imaging of causal interactions in humans Neuroimage 34:575-586
Ebersole J (1994) Non‐invasive localization of the epileptogenic focus by EEG dipole modeling Acta Neurologica Scandinavica 89:20-28
Emerson RG, Turner CA, Pedley TA, Walczak TS, Forgione M (1995) Propagation patterns of temporal spikes Electroencephalography and clinical
neurophysiology 94:338-348
Engel Jr J, Thompson PM, Stern JM, Staba RJ, Bragin A, Mody I (2013) Connectomics and epilepsy Current opinion in neurology 26:186
Fischl B (2012) FreeSurfer Neuroimage 62:774-781
Fornito A, Zalesky A, Breakspear M (2015) The connectomics of brain disorders Nature Reviews Neuroscience 16:159-172
Fraschini M, Demuru M, Crobe A, Marrosu F, Stam CJ, Hillebrand A (2016) The effect of epoch length on estimated EEG functional connectivity and brain
network organisation Journal of Neural Engineering 13:036015
Fuchs M, Wagner M, Kastner J (2007) Development of volume conductor and source models to localize epileptic foci Journal of Clinical Neurophysiology
24:101-119
Gotman J (1983) Measurement of small time differences between EEG channels: method and application to epileptic seizure propagation
Electroencephalography and clinical neurophysiology 56:501-514
Gotman J (1987) Interhemispheric interactions in seizures of focal onset: data from human intracranial recordings Electroencephalography and clinical
neurophysiology 67:120-133
Gramfort A, Papadopoulo T, Olivi E, Clerc M (2010) OpenMEEG: opensource software for quasistatic bioelectromagnetics Biomed Eng Online 9:45
Hämäläinen M (2005) MNE software user's guide NMR Center, Mass General Hospital, Harvard University 58:59-75
Hämäläinen MS, Ilmoniemi R (1994) Interpreting magnetic fields of the brain: minimum norm estimates Medical & biological engineering & computing
32:35-42
Hassan M, Benquet P, Biraben A, Berrou C, Dufor O, Wendling F (2015a) Dynamic reorganization of functional brain networks during picture naming
Cortex 73:276-288
Hassan M, Dufor O, Merlet I, Berrou C, Wendling F (2014) EEG Source Connectivity Analysis: From Dense Array Recordings to Brain Networks PloS one
9:e105041
Hassan M, Shamas M, Khalil M, El Falou W, Wendling F (2015b) EEGNET: An Open Source Tool for Analyzing and Visualizing M/EEG Connectome
PloS one 10:e0138297
Hassan M, Wendling F (2015) Tracking dynamics of functional brain networks using dense EEG IRBM 36:324-328
Hipp JF, Engel AK, Siegel M (2011) Oscillatory synchronization in large-scale cortical networks predicts perception Neuron 69:387-396
Hoechstetter K, Bornfleth H, Weckesser D, Ille N, Berg P, Scherg M (2004) BESA source coherence: a new method to study cortical oscillatory coupling
Brain topography 16:233-238
Jiruska P, de Curtis M, Jefferys JG, Schevon CA, Schiff SJ, Schindler K (2013) Synchronization and desynchronization in epilepsy: controversies and
hypotheses The Journal of physiology 591:787-797 doi:10.1113/jphysiol.2012.239590
Kramer MA, Cash SS (2012) Epilepsy as a disorder of cortical network organization The Neuroscientist 18:360-372
19
Kuś R, Kamiński M, Blinowska KJ (2004) Determination of EEG activity propagation: pair-wise versus multichannel estimate IEEE Transactions on
Biomedical Engineering, 51:1501-1510
Laufs H (2012) Functional imaging of seizures and epilepsy: evolution from zones to networks Current opinion in neurology 25:194-200
Le Van Quyen M, Soss J, Navarro V, Robertson R, Chavez M, Baulac M, Martinerie J (2005) Preictal state identification by synchronization changes in
long-term intracranial EEG recordings Clinical Neurophysiology 116:559-568
Liljeström M, Kujala J, Stevenson C, Salmelin R (2015) Dynamic reconfiguration of the language network preceding onset of speech in picture naming
Human brain mapping 36:1202-1216
Lu Y, Yang L, Worrell GA, He B (2012) Seizure source imaging by means of FINE spatio-temporal dipole localization and directed transfer function in
partial epilepsy patients Clinical Neurophysiology 123:1275-1283
Malinowska U, Badier JM, Gavaret M, Bartolomei F, Chauvel P, Bénar CG (2014) Interictal networks in magnetoencephalography Human brain mapping
35:2789-2805
Merlet I, Gotman J (1999) Reliability of dipole models of epileptic spikes Clinical Neurophysiology 110:1013-1028
Mheich A, Hassan M, Gripon V, Dufor O, Khalil M, Berrou C, Wendling F A novel algorithm for measuring graph similarity: application to brain networks.
In: 7th International IEEE EMBS Neural engineering Conference, Montpellier, France, 2015a.
Mheich A, Hassan M, Khalil M, Berrou C, Wendling F (2015b) A new algorithm for spatiotemporal analysis of brain functional connectivity Journal of
neuroscience methods 242: 77–81
Michel CM, Murray MM (2012) Towards the utilization of EEG as a brain imaging tool Neuroimage 61:371-385
Mormann F, Lehnertz K, David P, Elger CE (2000) Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy
patients Physica D: Nonlinear Phenomena 144:358-369
O'Neill GC, Barratt EL, Hunt BA, Tewarie PK, Brookes MJ (2015) Measuring electrophysiological connectivity by power envelope correlation: a technical
review on MEG methods Physics in medicine and biology 60:R271
Pascual-Marqui RD (2002) Standardized low-resolution brain electromagnetic tomography (sLORETA): technical details Methods Find Exp Clin Pharmacol
24:5-12
Pereda E, Quiroga RQ, Bhattacharya J (2005) Nonlinear multivariate analysis of neurophysiological signals Progress in neurobiology 77:1-37
Ramon C, Holmes MD (2013) Noninvasive localization of epileptic sites from stable phase synchronization patterns on different days derived from short
duration interictal scalp dEEG Brain topography 26:1-8
Schindler KA, Bialonski S, Horstmann MT, Elger CE, Lehnertz K (2008) Evolving functional network properties and synchronizability during human
epileptic seizures Chaos 18:033119 doi:10.1063/1.2966112
Schoffelen JM, Gross J (2009) Source connectivity analysis with MEG and EEG Human brain mapping 30:1857-1865
Song J et al. (2015) EEG source localization: Sensor density and head surface coverage Journal of neuroscience methods 256:9-21
Song J, Tucker DM, Gilbert T, Hou J, Mattson C, Luu P, Holmes MD (2013) Methods for examining electrophysiological coherence in epileptic networks
Frontiers in neurology 4
Tadel F, Baillet S, Mosher JC, Pantazis D, Leahy RM (2011) Brainstorm: a user-friendly application for MEG/EEG analysis Computational intelligence and
neuroscience 2011:8
van Mierlo P, Papadopoulou M, Carrette E, Boon P, Vandenberghe S, Vonck K, Marinazzo D (2014) Functional brain connectivity from EEG in epilepsy:
Seizure prediction and epileptogenic focus localization Progress in neurobiology 121:19-35
Vecchio F et al. (2014) Cortical connectivity in fronto-temporal focal epilepsy from EEG analysis: A study via graph theory Clinical Neurophysiology
Wendling F, Ansari-Asl K, Bartolomei F, Senhadji L (2009) From EEG signals to brain connectivity: a model-based evaluation of interdependence measures
Journal of neuroscience methods 183:9-18
Wendling F, Bartolomei F, Bellanger J-J, Bourien J, Chauvel P (2003) Epileptic fast intracerebral EEG activity: evidence for spatial decorrelation at seizure
onset Brain 126:1449-1459
Wendling F, Bartolomei F, Bellanger JJ, Chauvel P (2002) Epileptic fast activity can be explained by a model of impaired GABAergic dendritic inhibition
The European journal of neuroscience 15:1499-1508
Wendling F, Bellanger J-J, Bartolomei F, Chauvel P (2000) Relevance of nonlinear lumped-parameter models in the analysis of depth-EEG epileptic signals
Biological cybernetics 83:367-378
Wendling F, Chauvel P, Biraben A, Bartolomei F (2010) From intracerebral EEG signals to brain connectivity: identification of epileptogenic networks in
partial epilepsy Frontiers in systems neuroscience 4
20
21
Figure 1: Structure of the investigation. A) Simulated epileptic spikes: Model used to generate epileptic spikes (see
Materials and methods/Data/Simulations section for detailed description), B) Identification of interictal epileptic network:
First, a network is generated by the model and considered as the 'ground truth'. By solving the forward problem, synthetic
dense EEG data are generated. These signals are then used to solve the inverse problem in order to reconstruct the dynamics
of sources using three different inverse solutions (wMNE, sLORETA, dSPM and cMEM). The statistical couplings are then
computed between the reconstructed sources using three different methods (r2, PLV, h2 and MI). The identified network by
each combination (inverse/connectivity) was then compared with the original network using a 'network similarity' algorithm
[13] and C) Intracerebral recordings: The positions of the intracerebral SEEG signals used in the real application. The
corrdinates of the electrode's contacts was obtianed by the CT/MRI coregistration. Abbreviations: wMNE: weighted
Minimum Norm Estimate; sLORETA: Standardized low resolution brain electromagnetic tomography; dSPM: dynamical
Statistical Parametric Mapping; cMEM: standard Maximum Entropy on the Mean; r2: linear correlation coefficient; PLV:
Phase Locking Value; h2: nonlinear correlation coefficient; MI: mutual information, P: pyramidal cells, I: Inhibitory
interneurons.
22
23
Figure 2: One network scenario. A) Brain networks obtained by using the different inverse and connectivity methods, B)
The original network (ground truth) is shown and C) Values (mean ± standard deviation) of the similarity indices computed
between the network identified by each combination and the model network.
24
25
Figure 3: Two networks scenario. A) Brain networks obtained by using the different inverse and connectivity methods. B)
The original network (ground truth) is shown and C) Values (mean ± standard deviation) of the similarity indices computed
between the network identified by each combination and the model network.
26
27
Figure 4: Source localization vs. source connectivity. A. the reference network. B. Results obtained by the network-based
approach (cMEM/MI and wMNE/PLV) and C. Results obtained by the localization based approach (cMEM and wMNE)
using same window of analysis. Results were averaged over a 50 ms interval around each of the spike peaks. Red color
represents the sources with the highest energy.
28
29
Figure 5: Application on real data. A) Scalp EEG: Results of the source localization approach using wMNE (right) and
source connectivity using wMNE/PLV (left), B. Intracerebral EEG: Regions visually identified by the epileptologist (right)
and the network obtained by computing the functional connectivity between the intracerebral EEG signals (left), C.
Similarity indices: The SI values obtained between the network identified by each of the combination and the intracerebral-
EEG-based network.
30
|
1702.08421 | 1 | 1702 | 2017-02-27T18:27:48 | The role of the observer in goal-directed behavior | [
"q-bio.NC"
] | In goal-directed behavior, a large number of possible initial states end up in the pursued goal. The accompanying information loss implies that goal-oriented behavior is in one-to-one correspondence with an open subsystem whose entropy decreases in time. Yet ultimately, the laws of physics are reversible, so systems capable of yielding goal-directed behavior must transfer the information about initial conditions to other degrees of freedom outside the boundaries of the agent. To operate steadily, they must consume ordered degrees of freedom provided as input, and be dispensed of disordered outputs that act as wastes from the point of view of the aimed objective. Here I argue that a physical system may or may not display goal-directed behavior depending on what exactly is defined as the agent. The borders of the agent must be carefully tailored so as to entail the appropriate information balance sheet. In this game, observers play the role of tailors: They design agents by setting the limits of the system of interest. Brain-guided subjects perform this creative observation task naturally, implying that the observation of goal-oriented behavior is a goal-oriented behavior in itself. Minds evolved to cut out pieces of reality and endow them with intentionality, because ascribing intentionality is an efficient way of modeling the world, and making predictions. One most remarkable agent of whom we have indisputable evidence of its goal-pursuing attitude is the self. Notably, this agent is simultaneously the subject and the object of observation. | q-bio.NC | q-bio | The role of the observer in goal-directed behavior
In´es Samengo
7
1
0
2
b
e
F
7
2
]
.
C
N
o
i
b
-
q
[
1
v
1
2
4
8
0
.
2
0
7
1
:
v
i
X
r
a
Abstract
In goal-directed behavior, a large number of possible initial states end up in the pursued goal.
The accompanying information loss implies that goal-oriented behavior is in one-to-one corre-
spondence with an open subsystem whose entropy decreases in time. Yet ultimately, the laws of
physics are reversible, so entropy variations are necessarily a consequence of the way a system
is described. In order to reconcile different levels of description, systems capable of yielding
goal-directed behavior must transfer the information about initial conditions to other degrees of
freedom outside the boundaries of the agent. To operate steadily, they must consume ordered
degrees of freedom provided as input, and be dispensed of disordered outputs that act as wastes
from the point of view of the aimed objective. Broadly speaking, hence, goal-oriented behav-
ior requires metabolism, even if conducted by non-living agents. Here I argue that a physical
system may or may not display goal-directed behavior depending on what exactly is defined as
the agent. The borders of the agent must be carefully tailored so as to entail the appropriate
information balance sheet. In this game, observers play the role of tailors: They design agents
by setting the limits of the system of interest. Their computation may be iterated to produce a
hierarchy of ever more complex agents, aiming at increasingly sophisticated goals, as observed
in darwinian evolution. Brain-guided subjects perform this creative observation task naturally,
implying that the observation of goal-oriented behavior is a goal-oriented behavior in itself.
Minds evolved to cut out pieces of reality and endow them with intentionality, because ascrib-
ing intentionality is an efficient way of modeling the world, and making predictions. One most
remarkable agent of whom we have indisputable evidence of its goal-pursuing attitude is the
self. Notably, this agent is simultaneously the subject and the object of observation.
1
A bunch of nucleic acids swim among many other organic compounds forming a cytoplas-
matic soup, and somehow, manage to arrange themselves into precisely the sequence required
for DNA replication. Carbon dioxide molecules steadily stick to one another materializing a
solid tree trunk out of a tiny seed. Owls eat the young bats with poor navigation ability, thereby
improving the eco-location proficiency of the species. The neurons in a dog's brain fire pre-
cisely in the required sequence to have the dog bury its bone, hiding it from other dogs. The
wheels, breaks, and clutch of a self-driving car coordinate their actions in order to reach the
parking area of a soccer field, no matter the initial location of the car, nor the traffic along the
way. The limbs of the Argentine soccer players display a complex pattern of movements that
carry the ball, through kicks and headers, at Messi's feet in front of the keeper... kick... goal!
This essay is about goals. In all the above examples, a collection of basic elements, follow-
ing local and apparently purpose-less laws, manage to steer the value of certain variables into
some desired regime. The initial state is rather arbitrary, and yet, the agents manage to adap-
tively select, out of many possible actions, the maneuvers that are suited to conduct the system
to the desired goal. Throughout these seemingly intelligent choices, order appears to raise from
disorder. Scattered nucleotides become DNA. Air and dust become trees. Owl hunger becomes
sophisticated eco-location organs. Neural activity becomes a buried bone. A car anywhere in
the city becomes a car at a specific location. A football anywhere in the stadium becomes a
football in the goal. How do the components of each system know what to do, and what not to
do, in order to reach the goal? This is the question that will entertain us here.
Initial
G
G
A
T
C
T
A
C
C
C
T
A
G
A
G
T
A
T
C
G
C
T
A
G
Final
G
A
A
T
C
T
G
C
Figure 1:
In DNA replication, many initial
states are mapped onto a single final state. The
entropy is therefore high at the beginning and
low at the end.
One important characteristic of goal-
directed agents is that they are flexible: They
reach the goal from multiple initial condi-
tions, and are typically able to circumvent ob-
stacles. For example, in DNA replication, the
initial state is one out of many configurations
in which nucleotides can be spatially dis-
tributed in a solution of organic compounds.
The final state, the goal, is the precise spatial
arrangement of those same nucleotides within
the newly constructed DNA strand.
In the
soccer stadium, the ball may be initially in
any location, the final state is the ball at the
goal. Multiple initial states are hence mapped
onto a single final state, as in Fig. 1. In physical terms, the non-injective nature of this mapping
implies a reduction in entropy.
2
Admittedly, the final state need not be strictly unique. In DNA replication, permutations
of equal nucleotides are still allowed in the final state, and occasionally, there might also be
a few errors in the replication process. Dogs may consider more than a single location for the
concealed bounty, and Messi may choose to shoot the ball anywhere inside the 24 ft wide by 8 ft
high of the goal. Such restricted amounts of freedom, or even the occasional failures to reach the
final state (shooting an own goal, for example), by no means compensate the abrupt reduction
in entropy that takes place throughout the process. In fact, were entropy not to decrease, the
system would not exhibit goal-oriented behavior.
We are used to associating entropy increments with information losses, and entropy re-
ductions with information gains. Here I am taking the opposite view: Entropy reductions are
associated with information losses. The two views are not incompatible, they simply refer to
different things. The first case deals with a closed system and information about macroscopic
variables. The second, with an open system and microscopic variables. When a closed macro-
scopic system evolves in the direction that maximizes the entropy of all compatible microscopic
states (the usual case in closed thermodynamical systems), the final macro-state does not allow
us to deduce the initial macro-state, since the mapping between them is non-injective. Were we
to know the detailed final micro-state, however, we would be able to deduce the initial micro-
state. When a goal-oriented system evolves in the direction of decreasing entropy, the final
micro-state does not allow us to deduce the initial micro-state, but for a different reason: goal-
oriented systems are open, and they interact with degrees of freedom we are not keeping track
of. In this essay, the distinction between micro and macro-states is not emphasized, because the
phenomena we deal with are not always divisible into separate scales.
The notion of goal-oriented behavior that is used here always brings about an entropy reduc-
tion. I now want to demonstrate the reciprocal statement: If a system reduces its entropy, a goal
can be ascribed to the process. Therefore, entropy reduction and goal-oriented behavior are in a
one-to-one correspondence. The goal in question can always be defined by the restricted set of
values that the variables acquire in the final state: the target DNA sequence, the buried bone, the
ball at the goal. Of course, the reduction in entropy must first be verified: a broad set of initial
states must evolve into a small final set. A car that in a single trial travels from one location to
another is not guaranteed to be a self-driving car. Only if the initial location has proven to be
arbitrary, and the traffic conditions variable, can goal-directed behavior be arrogated.
The notion of entropy is subtle, since it not only characterizes a physical system, but also,
the way it is described. When the universe is described at its utmost basic level (assuming there
is one such level), all we have is a collection of fundamental particles evolving from some initial
state. If the state of all particles is specified, the total entropy of the universe vanishes. Time
3
reversibility of the laws of physics dictates entropy to remain zero for all past and future times.
Therefore, there is no way to attain neither an increase nor a reduction in entropy. Energy
dissipation and goal-directed behavior, hence, are absent from the complete description. We
need to blur our point of view to give them a chance, either by restricting the description to
macroscopic variables, or to subsystems. In fact, the main conclusion of this essay is that an
observer with a very special point of view is required for agency to exist.
If the information about the initial conditions is apparently lost in goal-oriented behavior,
then such information must be somehow concealed in degrees of freedom we are not keeping
track of. They may have been moved into too microscopic states to be monitored, or into fluid
degrees of freedom that, by the time the goal is reached, have already exited the subsystem
under study. What we track, and what we ignore, hence, plays a crucial role in agency.
hole
sliding
door
Figure 2: A demon controls the sliding door, al-
lowing particles to pass from right to left, but
not the other way round. The initial state of ev-
ery molecule that the demon has already acted
on (letting it pass or not) is recorded in its mem-
ory, and depicted in blue.
To be consistent with the second law
of thermodynamics, processes where entropy
decreases are only possible in open sys-
tems that somehow interact with the external
world. Originally, they were supposed to re-
quire an energy influx. This is, however, not
a necessary condition: Sometimes, the sole
exchange of information suffices. A good ex-
ample is Maxwell's Demon [1]. Suppose we
have a gas enclosed in two adjacent chambers
communicated by a small hole in the wall be-
tween them (Fig. 2). The hole may or may
not be covered by a sliding door controlled
by a demon.
Initially, both chambers have
equal pressure and density. The demon then
opens or closes the hole selectively, depend-
ing on whether a molecule approaches from
one side, or the other. Molecules coming from the right are allowed to pass into the left cham-
ber, but not the other way round. As time goes by, molecules accumulate on the left side,
eventually leaving the right side empty. The collection of all gas molecules can be interpreted
as performing goal-directed behavior: No matter the initial state, gas is gradually compressed
into the left chamber. This final state can be conceived as a goal, and it comprises a reduction
in entropy: initially each particle can be anywhere in the two chambers, and in the final state,
they are all in the left side. Arrogating purpose, in this case, is to assume that the gas -- who
takes the role of the agent -- wants to shrink. Other verbs may be used (tends to, is inclined to,
4
etc.), but the phrasing is irrelevant. As uncanny as it may seem, arrogating purpose to the gas is
a rather accurate description of the gas' phenomenology.
The gas + demon is a toy model of a closed system, so no interaction with the outside world
is allowed. To perform the task, the demon needs to acquire information on the location of each
molecule approaching the hole, to then decide whether to let it pass or not. In a slightly modified
version of this system, Bennett [2] demonstrated that the storage of information in the demon's
memory can be done with no energy expenditure, as long as the memory is initially blank, and
there is plenty of storage capacity. The work required to move and stop the door, as well as the
energy needed to measure the position of particles and to maintain the demon alive, can also
be made as small as desired, simply diminishing mechanical friction, and moving slowly. The
demon is however not allowed to delete the acquired information, because information erasure
requires energy consumption, at a minimal cost of kBT per erased bit [3]. Therefore, as time
goes by, the information of the initial location of each gas molecule is erased from the gas, and
copied onto the demon's memory.
The gas gradually reduces its entropy only if we are careful to exclude the demon from what
we define as the system. If we include the demon (and its memory), entropy simply remains
constant, since all the details of the initial state are still stored. Depending on the observer's
choices, then, entropy may or may not decrease, meaning that arrogating agency may or may
not be possible.
A subsystem can only decrease its entropy if it somehow gets rid of initial conditions. In
DNA replication, after the addition of each new nucleotide to the developing strand, the ini-
tial location of the free nucleotide determines the final configuration of the mediating enzimes,
thereby transferring the information of the initial state to a change in the 3-dimensional configu-
ration of nearby proteins. If enzimes are not restored to their functional state, the process cannot
be iterated. So enzimes, in turn, must pass the information on somewhere else. This transfer is
actually the important point in the emergence of goal-directed behavior. Energy consumption is
only helpful if energy is degraded in the process: ordered energy sources must be transformed
into disordered products. In animal cells, order arrives as glucose and oxygen molecules. Dis-
order exits as carbon dioxide, water and faster molecular motion (heat). The input degrees of
freedom, specifically in the case of glucose, are conformed of atoms tidily organized into large
molecules. The output degrees of freedom are transported by smaller molecules, amenable to
be arranged in many more configurations.
The laws of physics are ultimately reversible, so initial conditions cannot be truly erased,
they can only be shuffled around. As an example, Edward Fredkin studied how non-dissipative
systems, such as our universe, may perform the usual logical computations (AND, OR, etc),
5
which are themselves non-invertible [4]. We know that 0 AND 1 = 0. However, knowing that
the result of the operation is 0 does not suffice to identify the two input variables. If the compu-
tation is performed by an ultimately non-dissipative system, the information of the initial input
variables must be somehow moved into some other variable, albeit perhaps not in a manner that
is easily accessible. Fredkin's solution was to prove that computing required some extra input
variables, not needed for the computation per se, but mandatory for the information balance.
When performing a single logical operation (say, for example, AND), the additional variables
are in a well defined state (no uncertainty), and throughout the computation, they acquire so-
called garbage values (garbage because they are not required to perform the computation), that
represent those input degrees of freedom that cannot be deduced from the output. Copying
part of the input into garbage variables ensures that no information is lost, and the computation
becomes feasible in a non-dissipative physical substrate.
Input
Output
Figure 3: Goal-directed systems (blue ball) eat
up ordered degrees of freedom, and produce
disordered degrees of freedom.
Ascribing agency is all about ignoring
who really did the job (the Universe, to put
it grandly), and arrogating intentionality to an
entropy-reducing subsystem. The task of the
observer is to design the borders of the sub-
system so as to allow ordered degrees of free-
dom to be progressively incorporated, and/or
disordered ones to be eliminated. If the goal
is to be achieved repeatedly, a steady flow of
order is required, as well as a regular garbage collection service. Purposeful agents, hence,
only emerge from sub systems that eat up order (Fig. 3). Broadly speaking, they can be said
to breathe, or to be endowed with metabolism, even if they need not be alive in the biological
sense.
Maxwell's demon hid the initial conditions of the gas in its memory. The dog, the self-
driving car, and the soccer players, all hide their own initial state and that of the environment
inside their memories. Memories can of course be erased, but erasures consume energy, and
they are ultimately no more than flushing initial conditions into the high-entropy products of
energy degradation. For a long time, scientists failed to include memories as part of the sys-
tems under study, so goal-directed behavior sometimes appeared paradoxical. Here I argue that
observers attribute agency by disregarding initial conditions. Of course, observers are free to
delineate the borders of the subsystem under study as they wish. They can always shape the
limits of what they define as the agent in such a way as to have it do all sorts of wonderful
things, as achieve goals, and reduce entropy. The agent must be fed with order, and the mess
must be cleaned up, but still, it can be done. The natural question is therefore: What is in-
6
teresting in goal-directed behavior if the observer is allowed to engineer the very definition of
the agent, in order to get the desired result? Plants grow because what we define as a plant
is the stuff that grows every spring, and not the dirt left on the ground every autumn. Species
improve because we restrict the definition of a species to the material that a posteriori is seen as
successful, and exclude the corpses left behind of those who failed. Cell division seems to be a
productive business because the waste products are not defined to be part of cells. Returning to
the question posed above "How do the components of the system know what to do, and what not
to do, in order to reach the goal?", we can now provide an answer. Components know nothing,
observers do. Just as photographers select an arbitrary plane in the visual world where to focus
an image and engender a sharp object (Fig. 4), so do observers choose which variables compose
the system, and which do not, so that a goal emerges.
Should we be amazed that the world we
live in allows observers to create agents?
Could we not live in a universe where as-
signing agency were downright impossible?
I would be very much surprised if it were so.
The impossibility to define goal-directed be-
havior would mean that no subsystems exist
where entropy decreases. The global entropy
growth that takes place in the whole universe
should develop uniformly and monotonously
all throughout space and time, allowing for no
local oscillations. That would be ordered in-
deed! I do not expect disorder to arise in such
an orderly manner.
Figure 4: Observers, just as photographers, se-
lectively focus on an aspect of reality, to satisfy
their cognitive appetite [6]. Potograph kindly
supplied by Luc´ıa Samengo.
Within this picture, all the interesting events seem to take place in the observer's creative act.
Any local decrement of entropy, no matter how trivial, appears to suffice for an observer to be
able to ascribe agency. We are demanding little of the world, and a lot of the observer. But does
the evolving world not have organizational merits of its own? If we look at the history of events
taking place in our planet, as time goes by, agents seem to become increasingly sophisticated.
Compare the strategy of a replicating DNA molecule in a bacterium with the one of Menelaus
of Sparta to recover Helen of Troy, and thereby, ultimately manage bisexual reproduction. All
the complexity of the bacterium's strategy is present in Menelaus', but not the other way round.
Evolution seems to be striving towards what appears to be a runaway escalation of sophistica-
tion and design [5]. Is the development of refined agents something that only depends on the
observer's creativity, or is it something actually taking place independently of observers?
7
Merits are shared, I believe. Observers produce agents. In the absence of agents, no sub-
systems are cut out of the wholeness of the cosmos, and complexity cannot be measured. Once
observers are in play, even if they might have never intended it, it turns out that the computa-
tion they perform is liable to iteration. Agency allocation implies that many equal final states
are produced from many different initial states. The final states are similar to one another, and
similarity is a form of order. Such final states can therefore be used as the ordered degrees
of freedom that a higher-level process may use as fuel. This fuel needs to enter into a noisy
system for higher-level agents to emerge (Fig. 5). Noise is typically instantiated by a changing
environment, often in combination with the occasional mistakes that may have happened in the
lower-level process, as mutations in DNA replication. By iterating the algorithm, profuse RNA
replication in free solutions can be observed to give rise to prokaryote cells, who in turn evolve
into eukaryotes, from which multi-cellular organisms appear, all the way up to the ever growing
branches of the tree of life. In the way, conscious humans, civilization, and artificial intelligence
emerge. As well as a lot of garbage, as environmentalists wisely remind us.
Final state of
1st level, and
ordered fuel
of 2nd level
Initial state
(1st level)
l
e
u
f
d
e
r
e
d
r
O
)
l
e
v
e
l
t
s
1
(
Initial state
(2nd level)
Final state
of 2nd level
and ...
The tree of life develops as a continuous
process, irrespectively of whether observers
interpret it or not. Observers are required to
dissect it into agents, and to evaluate their so-
phistication. As we climb the ladder of evo-
lutionary design, the fuel degrees of freedom
diminish in number, and increase in complex-
1st level wastes
2nd level wastes
ity, the latter defined as the number of bits re-
Figure 5: The goal achieved by low-level agents
can become the ordered fuel of higher-level
agents.
quired to describe their inner structure. If the
resources at the bottom level are finite, the
process cannot be iterated indefinitely, since
eventually, too few and too complex degrees
of freedom may not be identifiable as multiple instances of one single ordered pattern.
Brain-guided observers are continuously ascribing agency. They do so because the role
of a brain is to model the world around its carrier, so that effective survival strategies can be
implemented. They have evolved to do so. Mental models must capture the regularities of the
world, and discard the noise. Here, noise is defined as the degrees of freedom that are irrelevant
to predicting those features of the environment that affect the observer's fitness. It would be
a waste of resources, if not impossible, for us to represent in our brains all what happens in
a dog's brain. Much more efficient is to ascribe agency, and conclude that the dog wants to
bury the bone. We cannot follow the evolution of all the bats that were eaten by owls, we
therefore conclude that the predation of owls sharpens the eco-location capacity of bats. We do
8
not care for the details with which self-driving cars are programmed, we just think of them as
goal-directed. We need an economic description, so we assign agency.
Observers do not assign agency to all the entropy-reducing systems they meet. Purpose is
only arrogated to subsystems for which there is no evident source of order, or for sources that
are too costly to represent. The cost of a representation is judged in terms of its contribution
to prediction accuracy. If we only look at the gas controlled by Maxwell's demon, ignoring the
demon itself, we conclude that the gas wants to shrink. If, however, the demon takes weekends
off, the purposeful model of the gas loses accuracy. A more sophisticated representation dis-
cerning between week days and weekends is needed. Assigning agency may or may not be a
convenient strategy, depending on the trade-off between the economy of the representation and
the prediction errors it induces. Arrogating agency in excess, for example by believing that all
what happens is maneuvered by some obscure intentionality, yields a poor prediction strategy.
Observation is the result of development: Observers learn how to observe, and they do so
within the framework of learning theory [7]. They are first exposed to multiple examples of the
process, that act as the training set. Before learning, the final state can only be predicted from
the initial one if all degrees of freedom are tracked - a representation capacity that observers
typically lack. Making the best use of their resources, observers explore the power set of the
system (the set of all subsets of the system) and search for some entropy-reducing subset from
which an agent and a goal can be defined. They then discard the superfluous degrees of freedom,
thereby compressing information. Yet, if the subsystem does indeed reduce entropy, they are
still able to make predictions. With successful predictions the world begins to makes sense, so
ascribing agency is in a way equivalent to constructing knowledge. In fact, the construction of
knowledge can be argued to be the essence of a mind.
In the last paragraphs, we have been observing observers. In doing so, we have placed our-
selves one step above the hierarchical ladder of observation. We have concluded that purposeful
agents do not exist per se, they are a mental construct of observers. This view may be easily
accepted when regarding the agents (let us call them "zombies") around us, but is more prob-
lematic when it applies to ourselves. In the end, we experience our own purposes in a most
irrefutable manner. In the present context, a self with purposes is no more than a compressed
representation of an observer. Within a physicalist's point of view, the observer of the self can-
not be situated in any other place than in the same brain where the self emerges. Whether the
observer coincides with the self, whether it only partially overlaps with it, whether it contains
it, or is contained by it, I do not dare to assert. I conjecture, however, that brains create selves
following the same principles with which they ascribe agency to external factors, the difference
being that the creation of the self involves a vastly larger number of degrees of freedom. Those
9
degrees of freedom, moreover, are typically only accessible to the local subsystem. They in-
clude the mental processes of which we have conscious access, encompassing external sensory
input, and the detailed state of our body. The latter has been proposed as the base for emotion,
and the higher-level neural patterns triggered by such a state, the base of feeling [8].
Within the self, the subject and the object of observation seem to coincide, forming a strange
loop. Douglas Hofstadter [9] has suggested that the circular nature of the mind observing itself
is essential to the self. I am not sure, however, whether this recursive hypothesis constitutes an
actual explanation of the self, or simply a way to bind the two loose ends together and worry
no more. It could also be the case that what we perceive as a unitary self is in fact a whole
collection of disperse mental processes, inside which multiple observers coexist, although sep-
arately unaccessible. In the end, consciousness has been equated with complex and indivisible
information processing [10], so accessing subprocesses may not be possible. Dennet however
very strongly argues that if such mental subprocesses can be considered multiple observers [11],
there is no such thing as a hierarchy, and even less, an ultimate observer.
I am afraid I am unable to provide a finished picture of agency when going all the way up to
the self. I hope, however, to have built a sensible image of other less intimate agents. The main
conclusion of this essay is that the interesting part of agency is the observer. Physics does not
make sense, observers make sense of it. Life does not have a meaning, we give it a meaning.
Life may not even be fundamentally different from non-life, it may just be a collection of
subsystems that appear to have goals. Goal-directed behavior does not exist if we do not define
our variables in such a way as to bring goals into existence. Bringing goals into existence is a
task that brains perform naturally, because they have evolved to model and predict the future.
One fundamental agent that has emerged inside each one of us is the self. The mechanisms
behind this process remain unclear, but its evolutionary utility is undisputed. Were we not able
to produce meaning, we would not manage to distinguish ourselves as a special part of the
cosmos. We would not have a sense of identity, a sense of self-preservation, nor the ability to
think. The self is required to enunciate even the most basic statements, all the way up from
cogito ergo sum.
10
References
[1] Maxwell J C (1908). Theory of heat. London, New York: Longmans, Green.
[2] Bennett C H (1982). The thermodynamics of computation -- a review. Int.J.Theoret.Phys.
21 (12): 905940.
[3] Landauer RW (1961). Irreversibility and Heat Generation in the Computing Process IBM
J.Res. Dev. 5(3): 183 - 191.
[4] Fredkin E, Toffoli T (1982). Conservative Logic. Int.J. Theoret.Phys. 21 (3-4): 219253.
[5] Dennet DC (1995). Darwin's dangerous idea: Evolution and the meaning of life. New
York: Simon & Schuster.
[6] Norretranders T (1999). The User Illusion: Cutting Consciousness Down to Size. New
York: Viking.
[7] MacKay D (2003). Information theory, inference and learning algorithms. New York:
Cambridge University Press.
[8] Damasio AR (1999). The Feeling of What Happens: Body and Emotion in the Making of
Consciousness. San Diego: Harcourt Brace & Co.
[9] Hofstadter D (2007). I am a strange loop. New York: Basic Books.
[10] Tononi G (2008). Consciousness as Integrated Information: a Provisional Manifesto. Biol.
Bull. 215: 216 242.
[11] Dennet DC (1991) Consciousness explained. Boston: Little, Brown and Co.
11
|
1907.08801 | 2 | 1907 | 2019-12-19T05:40:07 | Learning spatiotemporal signals using a recurrent spiking network that discretizes time | [
"q-bio.NC",
"cs.NE",
"nlin.AO",
"physics.bio-ph"
] | Learning to produce spatiotemporal sequences is a common task that the brain has to solve. The same neural substrate may be used by the brain to produce different sequential behaviours. The way the brain learns and encodes such tasks remains unknown as current computational models do not typically use realistic biologically-plausible learning. Here, we propose a model where a spiking recurrent network of excitatory and inhibitory biophysical neurons drives a read-out layer: the dynamics of the driver recurrent network is trained to encode time which is then mapped through the read-out neurons to encode another dimension, such as space or a phase. Different spatiotemporal patterns can be learned and encoded through the synaptic weights to the read-out neurons that follow common Hebbian learning rules. We demonstrate that the model is able to learn spatiotemporal dynamics on time scales that are behaviourally relevant and we show that the learned sequences are robustly replayed during a regime of spontaneous activity. | q-bio.NC | q-bio | Learning spatiotemporal signals using a recurrent spiking network
that discretizes time
Amadeus Maes1, Mauricio Barahona2, and Claudia Clopath1
1Department of Bioengineering, Imperial College London, United Kingdom
2Department of Mathematics, Imperial College London, United Kingdom
9
1
0
2
c
e
D
9
1
]
.
C
N
o
i
b
-
q
[
2
v
1
0
8
8
0
.
7
0
9
1
:
v
i
X
r
a
Abstract
Learning to produce spatiotemporal sequences is a common task that the brain has to solve. The same neural
substrate may be used by the brain to produce different sequential behaviours. The way the brain learns and
encodes such tasks remains unknown as current computational models do not typically use realistic biologically-
plausible learning. Here, we propose a model where a spiking recurrent network of excitatory and inhibitory
biophysical neurons drives a read-out layer: the dynamics of the driver recurrent network is trained to encode
time which is then mapped through the read-out neurons to encode another dimension, such as space or a
phase. Different spatiotemporal patterns can be learned and encoded through the synaptic weights to the
read-out neurons that follow common Hebbian learning rules. We demonstrate that the model is able to learn
spatiotemporal dynamics on time scales that are behaviourally relevant and we show that the learned sequences
are robustly replayed during a regime of spontaneous activity.
Author summary
The brain has the ability to learn flexible behaviours on a wide range of time scales. Previous studies have
successfully built spiking network models that learn a variety of computational tasks, yet often the learning
involved is not biologically plausible. Here, we investigate a model that uses biological-plausible neurons and
learning rules to learn a specific computational task: the learning of spatiotemporal sequences (i.e., the temporal
evolution of an observable such as space, frequency or channel index). The model architecture facilitates the
learning by separating the temporal information from the other dimension. The time component is encoded
into a recurrent network that exhibits sequential dynamics on a behavioural time scale, and this network is then
used as an engine to drive the read-out neurons that encode the spatial information (i.e., the second dimension).
We demonstrate that the model can learn complex spatiotemporal spiking dynamics, such as the song of a bird,
and replay the song robustly spontaneously.
Introduction
Neuronal networks perform flexible computations on a wide range of time scales. While individual neurons
operate on the millisecond time scale, behaviour time scales typically span from a few milliseconds to hun-
dreds of milliseconds and longer. Building functional models that bridge this time gap is of increasing in-
terest [Abbott et al., 2016], especially now that the activity of many neurons can be recorded simultaneously
[Jun et al., 2017, Maass, 2016]. Many tasks and behaviours in neuroscience consist of learning and producing
flexible spatiotemporal sequences, e.g. a 2-dimensional pattern with time on the x-axis and any other observable
on the y-axis which we denote here in general terms as the "spatial information". For example, songbirds pro-
duce their songs through a specialized circuit: neurons in the HVC nucleus burst sparsely at very precise times
to drive the robust nucleus of the arcopallium which in its turn drives motor neurons [Hahnloser et al., 2002,
Leonardo and Fee, 2005]. For different motor tasks, sequential neuronal activity is recorded in various brain
regions [Pastalkova et al., 2008, Itskov et al., 2011, Harvey et al., 2012, Peters et al., 2014, Adler et al., 2019],
and while the different tasks involve different sets of muscles, the underlying computation on a more fundamental
level might be similar [Rhodes et al., 2004].
Theoretical and computational studies have shown that synaptic weights of recurrent networks can be set
appropriately so that dynamics on a wide range of time scales is produced [Litwin-Kumar and Doiron, 2014,
Zenke et al., 2015, Tully et al., 2016]. In general, these synaptic weights are engineered to generate a range of
interesting dynamics. In slow-switching dynamics, for instance, the wide range of time scales is produced by
1
having stochastic transitions between clusters of neurons [Schaub et al., 2015]. Another example is sequential
dynamics, where longer time scales are obtained by clusters of neurons that activate each other in a sequence.
This sequential dynamics can emerge by a specific connectivity in the excitatory neurons [Chenkov et al., 2017,
Setareh et al., 2018] or in the inhibitory neurons [Billeh and Schaub, 2018]. However, it is unclear how the brain
learns these dynamics, as most of the current approaches use non biologically plausible ways to set or "train"
the synaptic weights. For example, FORCE training [Sussillo and Abbott, 2009, Laje and Buonomano, 2013,
Nicola and Clopath, 2017] or backpropagation through time [Werbos, 1990] use non-local information either in
space or in time to update weights. Such information is not available to the synaptic connection, which only
has access to the presynaptic and postsynaptic variables at the current time.
Here, we propose to learn a spatiotemporal task over biologically relevant time scales using a spiking
recurrent network driving a read-out layer where the neurons and synaptic plasticity rules are biologically
plausible. Specifically, all synapses are plastic under typical spike-timing dependent Hebbian learning rules
[Clopath et al., 2010, Litwin-Kumar and Doiron, 2014]. Our model architecture decomposes the problem into
two parts. First, we train a recurrent network to generate a sequential activity which serves as a temporal
backbone so that it operates as a 'neuronal clock' driving the downstream learning. The sequential activity is
generated by clusters of neurons activated one after the other: as clusters are highly recurrently connected, each
cluster undergoes reverberating activity that lasts longer than neural time scale so that the sequential cluster
activation is long enough to be behaviourally relevant. This construction allows us to bridge the neural and the
behavioural time scales. Second, we use Hebbian learning to encode the target spatiotemporal dynamics in the
read-out neurons. In this way, the recurrent network encodes time and the read-out neurons encode 'space'. As
discussed above, we use the term 'space' to denote a temporally-dependent observable, be it spatial position,
or phase, or a time-dependent frequency, or a more abstract state-space. Similar to the liquid state-machine,
where the activity in a recurrent network is linearly read-out by a set of neurons, we can learn different dynamics
in parallel in different read-out populations [Jaeger et al., 2007]. We also show that learning in the recurrent
network is stable during spontaneous activity and that the model is robust to synaptic failure.
Results
Model architecture
The model consists of two separate modules: a recurrent network and a read-out layer (Fig 1A). Learning
happens in two stages. In the first stage, we learn the weights of the recurrent network so that the network
exhibits a sequential dynamics. The ensuing recurrent neuronal network (RNN) effectively serves as a temporal
backbone driving the learning of the downstream read-out layer.
In the second stage, a target sequence is
learned in the read-out layer.
Architecture: The recurrent network is organized in C clusters of excitatory neurons and a central clus-
ter of inhibitory neurons. All excitatory neurons follow adaptive exponential integrate-and-fire dynamics
[Brette and Gerstner, 2005] while all inhibitory neurons follow a leaky integrate-and-fire dynamics. The in-
hibitory neurons in the RNN prevent pathological dynamics. The aim of this module is to discretize time
into C sequential intervals, associated with each of the C clusters. This is achieved by learning the weights of
the recurrent network. The neurons in the excitatory clusters then drive read-out neurons through all-to-all
feedforward connections. The read-out neurons are not interconnected. The target sequence is learned via the
weights between the driver RNN and the read-out neurons.
In previous models, the learning schemes are typically not biologically plausible because the
Plasticity:
plasticity depends on non-local information. Here, however, we use the voltage-based STDP plasticity rule
in all the connections between excitatory neurons (Fig 1B). This is paired with weight normalization in the
recurrent network (Fig 1C) and weight dependent potentiation in the read-out synapses (Fig 1D). Inhibitory
plasticity [Vogels et al., 2011] finds good parameters aiding the sequential dynamics (Fig S5).
Learning scheme: During the first stage of learning, all neurons in each cluster receive the same input in
a sequential manner. As a result of this learning stage, the recurrent spiking network displays a sequential
dynamics of the C clusters of excitatory neurons. Neurons within each cluster spike over a time interval (while
all neurons from other clusters are silent), with the activity switching clusters at points t = [t0, t1, ..., tC] so that
cluster i is active during time interval [ti−1, ti]. Thus, time is effectively discretized in the RNN.
During the second stage of learning, the read-out neurons receive input from a set of excitatory supervisor
neurons. The discretization of time enables Hebbian plasticity to form strong connections from the neurons in
the relevant time bin to the read-out neurons. For instance, if we want to learn a signal which is 'on' during
[ti−1, ti] and 'off' otherwise, a supervisor neuron can activate the read-out neuron during that time interval so
2
Fig 1. Model architecture. (A) The recurrent network consists of both inhibitory (in blue) and excitatory
(in red) neurons. The connectivity is sparse in the recurrent network. The temporal backbone is established
in the recurrent network after a learning phase. Inset: zoom of recurrent network showing the macroscopic
recurrent structure after learning, here for 7 clusters. The excitatory neurons in the recurrent network project
all-to-all to the read-out neurons. The read-out neurons are not interconnected. (B) All excitatory to excitatory
connections are plastic under the voltage-based STDP rule (see Methods for details). The red lines are spikes of
neuron j (top) and neuron i (bottom). When neurons j and i are very active together, they form bidirectional
connections strengthening both Wij and Wji. Connections Wij are unidirectionally strengthened when neuron
j fires before neuron i. (C) The incoming excitatory weights are L1 normalized in the recurrent network, i.e.
the sum of all incoming excitatory weights is kept constant. (D) Potentiation of the plastic read-out synapses
is linearly dependent on the weight. This gives weights a soft upper bound.
3
Model architectureRecurrent networkRead-out neurons...ZoomAVoltage-based STDPWWminWmaxALTP0AALTP = A Wmax- WWmax- WminjiWijWjiBidirectionalUnidirectionalWijWjiWijWeight dependent potentiation in read-out synapsesW1W2W1+ W2 = CNormalization in recurrent networkW1W2WBDC that connections from cluster i to the read-out neuron are potentiated through activity (who fires together, wires
together). This means that, after learning, the read-out neuron will be activated when cluster i is activated.
In general, the read-out layer learns a multivariate signal of time, i.e., the neurons in the read-out layer encode
the D different dimensions of the target signal: t → φ(t) = [φ1(t), φ2(t), ..., φD(t)].
A recurrent network that encodes discrete time
We give here further details of the first learning stage, where a recurrent network is trained to produce a
sequential dynamics. To this end, we initialize the weight matrix so that each synaptic weight between two
neurons is non-zero with probability p. The weights that are zero remain zero at all times, i.e. the topology
is fixed. We set the initial values of the non-zero weights in the recurrent network such that the dynamics is
irregular and asynchronous (i.e., a balanced network, see Methods for details).
We stimulate the C clusters with an external input in a sequential manner (Fig 2A): neurons in cluster i
each receive external Poisson spike trains (rate of 18 kHz for 10 ms, assuming a large input population). After
this, there is a time gap where no clusters receive input (5 ms). This is followed by a stimulation of cluster i + 1.
This continues until the last cluster is reached and then it links back to the first cluster (i.e. a circular boundary
condition). During the stimulation, neurons in the same cluster fire spikes together strengthening the intra-
cluster connections bidirectionally through the voltage-based STDP rule [Clopath et al., 2010, Ko et al., 2013].
Additionally, there is a pre/post pairing between adjacent clusters. Neurons in cluster i + 1 fire after neurons in
cluster i. The weights from cluster i to cluster i+1 strengthen unidirectionally (Fig 2B). If the time gap between
sequential stimulations is increased during the training phase, so that the gap becomes too long with respect to
the STDP time window, then there is no pre/post pairing between clusters and the ensuing dynamics loses its
sequential nature and becomes a slow-switching dynamics [Litwin-Kumar and Doiron, 2014, Schaub et al., 2015]
(Fig S1). In slow-switching dynamics, clusters of neurons are active over long time scales, but both the length of
activation and the switching between clusters is random. This is because the outgoing connections from cluster
i to all other clusters are the same in a slow-switching network. To summarize, a connectivity structure emerges
through biophysically plausible potentiation by sequentially stimulating the clusters of excitatory neurons in
the recurrent network. When the gap between activation intervals is sufficiently small compared to the STDP
window, the connectivity structure is such that intra-cluster weights and the weights from successive clusters
i → i + 1 are strong.
After the synaptic weights have converged, the external sequential input is shut-down and spontaneous
dynamics is simulated so that external excitatory Poisson spike trains without spatial or temporal structure
drive the RNN. Under such random drive, the sequence of clusters reactivates spontaneously and ensures that
both the intra-cluster and the connections from cluster i to cluster i+1 remain strong. In general, the interaction
between plasticity, connectivity and spontaneous dynamics can degrade the learned connectivity and lead to
unstable dynamics [Morrison et al., 2007]. To test the stability of the learned connectivity, we track the changes
in the off-diagonal weights (i.e. the connections from cluster i to cluster i + 1). After the external sequential
input is shut-down, we copy the weight matrix and freeze the weights of this copy. We run the dynamics of the
recurrent network using the copied frozen weights and apply plastic changes to the original weight matrix. This
means that we effectively decouple the plasticity from the dynamics. Indeed, when the dynamics is sequential,
the off-diagonal structure is reinforced. When the off-diagonal structure is removed from the frozen copied
weight matrix, the dynamics is not sequential anymore. In this case, the off-diagonal structure degrades. We
conclude that the connectivity pattern is therefore stable under spontaneous dynamics (Fig S2).
We next studied how the spectrum of the recurrent weight matrix is linked to the sequential dynamics. In
linear systems, the eigenvalues of the connectivity matrix determine the dynamics of the system. In a nonlinear
spiking model, the relationship between connectivity and dynamics is less clear. The connectivity after learning
can be seen as a low-dimensional perturbation of a random matrix. Such low-dimensional perturbations create
outliers in the spectrum [Tao, 2013] and change the dynamics [Mastrogiuseppe and Ostojic, 2018]. Here, we
have carried out a similar spectral analysis to that presented in [Schaub et al., 2015] (see Figs 7 in the Methods
and S2 in the Supplementary material). The weight matrix has most of its eigenvalues in a circle in the complex
plane (Fig 2B) with eigenvalues associated both with the balanced nature of the network, but, importantly,
also with the sequential structure (Fig 2B). As the temporal backbone develops through learning (as seen in
Fig S2), it establishes a spectral structure in which the pairs of leading eigenvalues with large real parts have
almost constant imaginary parts.
A simplified analysis of a reduced weight matrix (where nodes are associated with groups of neurons) shows
that the imaginary parts of the dominant eigenvalues depend linearly on the strength of the weights from cluster
i to cluster i+1 (see Methods, Fig 7). Hence for this simplified linearised rate model, this results in an oscillatory
dynamics where the imaginary part determines the frequency by which the pattern of activation returns due to
the periodic excitation pattern. As shown in [Schaub et al., 2015], these properties of the linear system carry
over to the nonlinear spiking model, i.e., the imaginary parts of the eigenvalues with large real parts determine
the time scales of the sequential activity (Fig S2).
4
Fig 2. Learning a sequential dynamics stably under plasticity. (A) The excitatory neurons receive
sequential clustered inputs. Excitatory neurons are grouped in 30 disjoint clusters of 80 neurons each.
(7
clusters shown in the cartoon for simplicity) (B) The weight matrix after training (only the first five clusters
shown) exhibits the learned connectivity structure, e.g., neurons within cluster 1 are highly interconnected and
also project to neurons in cluster 2, same for cluster 2 to cluster 3, etc. The spectrum of the full weight matrix
after training shows most eigenvalues in a circle in the complex plane (as in a random graph) with two other
eigenvalues signifying the balancing of the network, and a series of dominant eigenvalues in pairs that encode
the feedforward embedding. (C) Raster plot of the total network consisting of 2400 excitatory (in red) and 600
inhibitory (in blue) neurons. After learning, the spontaneous dynamics exhibits a stable periodic trajectory
'going around the clock'. The excitatory clusters discretize time (see zoom) and the network has an overall
period of about 450 ms.
5
ASequential external input during learningLearning a feedforward structure through the recurrent networkTime123456710 ms5 msTime123456710 ms5 msAfter learning RNN discretizes time under spontaneous random activityCBWeight matrix of RNN after learningDetail of weight matrix(5 first clusters)FeBcnSpectrum(full matrix)Rdc Under spontaneous activity, each cluster is active for about 15 ms, due to the recurrent connectivity within
the cluster. A large adaptation current counteracts the recurrent reverberating activity to turn off the activity
reliably. Therefore, as each cluster is spontaneously active in a sequence, the sequence length reaches behavioural
time scales (Fig 2C). In summary, the network exhibits sequential dynamics, serving as a temporal backbone
where time is discretized over behavioural time scales.
Learning a non-Markovian sequence
After the sequential temporal backbone is learnt via the RNN, we can then learn a spatiotemporal sequence
via the read-out neurons. To achieve this, during the second stage of training, the read-out neurons receive
additional input from supervisor neurons and from interneurons (Fig 3A). The supervisor neurons receive an
external Poisson input with rate modulated by the target sequence to be learned (Fig 3B).
As a first example, consider a target sequence composed of states A, B, C activated in the following
deterministic order: ABCBA. This is a non-Markovian state sequence because the transition from state B to
the next state (A or C) requires knowledge about the previous state [Brea et al., 2013], a non trivial task that
requires information to be stored about previous network states, potentially over long time periods. Previous
studies have proposed various solutions for this task [Maass and Markram, 2002, Brea et al., 2013]. However,
separating the problem of sequence learning in two stages solves this in a natural way.
The recurrent network trained in the first stage (Fig 2) is used to encode time. The underlying assumption
is that a starting signal activates both the first cluster of the recurrent network and the external input to the
supervisor neurons, which activate the read-out neurons.
After the training period, the interneurons and supervisor neurons stop firing (Fig 3C) and the target
sequence is stored in the read-out weight matrix (Fig 3D). During spontaneous activity, clusters in the RNN
reactivate in a sequential manner driving the learned sequence in the read-out neurons. Hence the spike sequence
of the read-out neurons is a noisy version of the target signal (Fig 3E). Learning the same target signal several
times results in slightly different read-out spike sequences each time (Fig S3). The firing rates of neurons in the
read-out corresponds to the target sequence (Fig 3F). In summary, our results show that the model is able to
learn simple but non-trivial spatiotemporal signals that are non-Markovian.
Learning sequences in parallel
We next wondered how multiple spatiotemporal signals can be learned. We hypothesized that, once the temporal
backbone is established, multiple spatiotemporal sequences can easily be learned in parallel. As an example, we
learn two sequences: ABCBA and DEDED. Here, D and E denote two additional read-out neurons (Fig 4A).
We assume that the model observes each sequence alternately for 2 seconds at a time (Fig 4B), although in
principle it could also been shown simultaneously. After learning, the target sequences are encoded in the
read-out weight matrix (Fig 4C). In a regime of spontaneous dynamics the learned sequences can be replayed
(Fig 4D). We conclude that multiple sequences can be learned in parallel. Each separate sequence requires a
separate set of read-out neurons. As such, the number of read-out neurons required increases linearly with the
number of target sequences.
Properties of the model: scaling, robustness and temporal variability
We investigate several scaling properties of the network. We first assess how the sequential dynamics in the RNN
depends on the cluster size by increasing the number of excitatory neurons in each cluster (NC), preserving the
ratio of excitatory to inhibitory neurons (NE/NI). To preserve the magnitude of the currents in the network, the
sparseness of the connectivity (p) also varies with NC such that pNC is constant. The same training protocols
are used for each network configuration as described in the Methods.
In Fig 5A, we show that for a fixed
number of clusters C, the mean period of the sequential dynamics exhibited by the RNN is largely independent
of cluster size NC. If we fix the number of neurons in the RNN, and in this way change the number of clusters
C, the mean period of the sequential dynamics decreases with increasing cluster size NC. We conclude that the
sequential dynamics is preserved over a wide range of network configurations. The time scales in the dynamics
depend on the number of clusters and network size.
Another way to modulate the period of the sequential dynamics is to change the unstructured Poisson
input to the RNN during spontaneous dynamics (i.e., after the first stage of learning). When the rate of
the external excitatory input is increased/decreased, the mean period of the sequential dynamics in the RNN
decreases/increases (Fig 5B). These results suggest that the network could learn even if the supervisor signal
changes in length at each presentation, assuming that both the supervisor and external Poisson input are
modulated by the same mechanism.
We next looked at the robustness of the learning of our model under random perturbations and network
size. In this context, we consider the effect of cluster size and the deletion of synapses in the read-out layer
6
Fig 3. Learning a non-Markovian sequence via the read-out neurons. (A) Excitatory neurons in the
recurrent network are all-to-all connected to the read-out neurons. The read-out neurons receive additional
excitatory input from the supervisor neurons and inhibitory input from interneurons. The supervisor neurons
receive spike trains that are drawn from a Poisson process with a rate determined by the target sequence.
The read-out synapses are plastic under the voltage-based STDP rule. (B) The rate of the input signal to
the supervisor neurons A, B and C. The supervisor sequence is ABCBA where each letter represents a 75
ms external stimulation of 10 kHz of the respective supervisor neuron.
(C) After learning, the supervisor
input and plasticity are turned off. The read-out neurons are now solely driven by the recurrent network. (D)
The read-out weight matrix W RE after 12 seconds of learning. (E) Under spontaneous activity, the spikes of
recurrent network (top) and read-out (bottom) neurons. Excitatory neurons in the recurrent network reliably
drive sequence replays. (F) The target rate (above) and the rate of the read-out neurons (below) computed
using a one sequence replay and normalized to [0, 1]. The spikes of the read-out neurons are convolved with a
Gaussian kernel with a width of ∼ 12 ms.
7
Learning a sequence through the read-out neuronsRead-out neuron50150250350CBATarget sequence50150250350Time [ms]CBARead-out neuron rate InterneuronSupervisorneuronTime-discretizing sequential RNNSpontaneous random activity (Poisson)Supervisor input(target)CBATime-discretizing sequential RNNSpontaneous random activity (Poisson)Read-out neuronsReplaying a learned sequenceSupervisor signalRead-out weight matrix after learningRate of read-out neuronsSpontaneous sequence reactivationsA Fig 4. Learning sequences in parallel. (A) The recurrent network projects to two sets of neurons. (B) Two
different sequences, ABCBA and DEDED, are learned by alternating between them and presenting each for 2
seconds at a time. (C) The read-out weight matrix after 24 seconds of learning. (D) Raster plot of spontaneous
sequence reactivations, where an external inhibitory current is assumed to control which sequence is replayed.
8
0100200300400Time [ms]ABCDESupervisor input rate180016002400Recurrent network neuron indexEDCBARead-out neuron02400.511.52800160024003000Recurrent networkneuron index00.511.52Time [s]ABCDERead-out neuronLearning twosequences in parallel CBATime-discretizing sequential RNNSpontaneous random activity (Poisson)Read-out neuronsAED Fig 5. Scaling, robustness and time variability of the model. (A) Change of the mean period of the
sequential dynamics as the number of clusters grows with: (i) the total number of excitatory neurons kept
constant (red line); (ii) the total number of neurons increasing with cluster size (blue line). Error bar shows a
standard deviation. (B) Dynamics with varying level of external excitatory input for four different cluster sizes
and NE = 2400. The external input can modulate the period of the sequential dynamics by ∼ 10%. (C) Recall
performance of the learned sequence ABCBA for varying cluster sizes and NE = 30NC under synapse deletion
(computed over 20 repeats). The learning time depends on the cluster size: ∆t = 960s/NC. (D) The ABCBA
sequence is learned with a network of 120 excitatory neurons connected in one large chain and readout neurons
with maximum synaptic read-out strength increased to W AE
max = 75 pF . The network is driven by a low external
ext = 2.75 kHz). When, at t = 500 ms a single synapse is deleted, the dynamics breaks down and parts
input (rEE
of the sequence are randomly activated by the external input. Top: spike raster of the excitatory neurons of
the RNN. Bottom: spike raster of the read-out neurons. (E) (Left) Histogram of the variability of the period of
the sequential activity of the RNN over 79 trials (Right) The standard deviation of the cluster activation time,
µt (root mean squared
σt, increases as the square root of µt, the mean time of cluster activation: σt = 0.213
error = 0.223 ms).
√
9
Time [s]Deleted read-out synapses per clusterPerformance02040608010012014000.20.40.60.81NC=80NC=120NC=160NC=200Scaling the networkPerformance under synapse deletionSequential dynamics breaks down in synfire chainTemporal variability in sequential activity0200400600Mean time [ms]0246Standard deviation [ms]simulationfit44.254.54.755External input [kHz]250300350400450500550Period [ms]NC= 80NC= 120NC= 160NC= 200External input can modulate the period ABCDEBroken chain after learning. We learn the simple ABCBA sequence (Fig 3) in the read-out neurons using a RNN with a fixed
number of clusters C but varying the cluster size NC. The total learning time (∆t) is varied with the cluster
size, NC∆t, because smaller clusters learn slower, since smaller clusters need larger read-out synaptic strengths
to drive the same read-out neuron. We also eliminate an increasing number of synapses in the read-out layer.
Performance is quantified as the number of spikes elicited by the read-out neurons after deletion of read-out
synapses normalized by the number of spikes elicited before deletion. Networks with larger clusters show a more
robust performance under noise (Fig 5C and Fig S4). These results show that, not surprisingly, larger clusters
drive read-out neurons more robustly and learn faster.
We then tested the limits of time discretization in our model. To that end, we hardcoded a recurrent network
with clusters as small as one neuron. In that extreme case, our network becomes a synfire chain with a single
neuron in every layer [Abeles, 1991]. In this case, randomly removing a synapse in the network will break the
sequential dynamics (Fig 5D). Hence, although a spatiotemporal signal can be learned in the read-out neurons,
the signal is not stable under a perturbation of the synfire chain. In summary, the choice of cluster size is a
trade-off between network size on the one hand and robustness on the other hand. Large clusters: (i) require
a large network to produce sequential dynamics with the same period; (ii) are less prone to a failure of the
sequential dynamics; (iii) can learn a spatiotemporal signal faster; and (iv) drive the read-out neurons more
robustly.
We have also characterized the variability in the duration of the sequential activity, i.e., the period of
the RNN. Since the neural activity does not move through the successive clusters with the same speed in each
reactivation, we wondered how the variance in the period of our RNN network compared to Weber's law. Weber's
law predicts that the standard deviation of reactions in a timing task grows linearly with time [Gibbon, 1977,
Hardy and Buonomano, 2018]. Since our model operates on a time scale that is behaviourally relevant, it is
interesting to look at how the variability increases with increasing time. Because time in our RNN is discretized
by clusters of neurons that activate each other sequentially, the variability increases over time as in a standard
T rather than
Markov chain diffusive process. Hence the variability of the duration T is expected to grow as
linearly. This is indeed what our network displays (Fig 5E). Here, we scaled the network up and increased the
period of the recurrent network by increasing the network size (80 excitatory clusters of 80 neurons each, see
Methods for details). By doing so, we can look at how the standard deviation of the duration of the RNN
activity grows with time.
√
Learning a complex sequence
In the non-Markovian target sequence ABCBA, the states have the same duration and the same amplitude
(Fig 3B). To test whether we could learn more complex sequences, the model was trained using a spatiotemporal
signal with components of varying durations and amplitudes. As an example, we use a 'spatio'-temporal signal
consisting of a 600 ms meadowlark song (Fig 6A). The spectrogram of the sound is normalized and used as the
time-varying and amplitude-varying rate of the external Poisson input to the supervisor neurons. Each read-
out and supervisor neuron encodes a different frequency range, hence in this example our 'space' dimension is
frequency.
We first trained a RNN of 6400 excitatory neurons (Fig 5E, see Methods) in order to discretize the time
interval spanning the full duration of the song. We then trained the read-out layer. The learned read-out
weight matrix reflects the structure of the target sequence (Fig 6B). Under spontaneous activity, the supervisor
neurons and interneurons stop firing and the recurrent network drives song replays (Fig 6C), and the learned
spatiotemporal signal broadly follows the target sequence (Fig 6A). The model performs worse when the target
dynamics has time-variations that are faster than or of the same order as the time discretization in the RNN.
Thus, we conclude that the model can learn interesting spiking dynamics up to a resolution of time features
limited by the time discretization in the recurrent network.
Discussion
We have proposed here a neuronal network architecture based on biophysically plausible neurons and plasticity
rules in order to learn spatiotemporal signals. The architecture is formed by two modules. The first module pro-
vides a temporal backbone that discretizes time, implemented by a recurrent network where excitatory neurons
are trained into clusters that become sequentially active due to strong inter-cluster and cluster i to cluster i + 1
weights. All of the excitatory clusters are connected to a central cluster of inhibitory neurons. As previously
shown for randomly switching clustered dynamics [Litwin-Kumar and Doiron, 2014], the ongoing spontaneous
activity does not degrade the connectivity patterns: the set of plasticity rules and sequential dynamics reinforce
each other. This stable sequential dynamics provides a downstream linear decoder with the possibility to read
out time at behavioural time scales. The second module is a set of read-out neurons that encode another dimen-
sion of a signal, which we generically denote as 'space' but can correspond to any time-varying observable, e.g.,
10
Fig 6. Learning a complex sequence. (A) Target sequence (top). The amplitude shows the rate of the
Poisson input to the supervisor neurons and is normalized between 0 and 10 kHz. Rate of read-out neurons for
one sample reactivation after learning 6 seconds (bottom). 45 read-out neurons encode the different frequencies
in the song. Neuron i encodes a frequency interval of [684 + 171i, 855 + 171i]Hz. (B) The read-out weight
matrix after learning 6 seconds. (C) Sequence replays showing the spike trains of both the recurrent network
neurons (top, excitatory neurons in red and inhibitory neurons in blue), and the read-out neurons (bottom).
spatial coordinates, frequency, discrete states, etc. The read-out neurons learn spike sequences in a supervised
manner, and the supervisor sequence is encoded into the read-out weight matrix. Bringing together elements
from different studies [Brea et al., 2013, Nicola and Clopath, 2017, Gilra and Gerstner, 2017], our model ex-
ploits a clock-like dynamics encoded in the RNN to learn a mapping to read-out neurons so as to perform the
computational task of learning and replaying spatiotemporal sequences. We illustrated the application of our
scheme on a simple non-Markovian state transition sequence, a combination of two such simple sequences, and
a time series from bird singing with more complex dynamics.
Other studies have focused on the classification of spatiotemporal signals. The tempotron classifies a spa-
tiotemporal pattern by either producing a spike or not [Gütig and Sompolinsky, 2006]. More recent studies
on sequential working memory propose similar model architectures that enable the use of Hebbian plasticity
[Manohar et al., 2019]. For example, spatiotemporal input patterns can be encoded in a set of feedforward
synapses using STDP-type rules [Park et al., 2017, Lee et al., 2019]. Combining these approaches with our
model might be an interesting line of future research.
The dynamics of the recurrent network spans three time scales: (i) individual neurons fire at the millisecond
time scale; (ii) clusters of neurons fire at the "tick of the clock", τc, i.e., the time scale that determines the time
discretization of our temporal backbone; and (iii) the slowest time scale is at the level of the entire network, i.e.
the period of the sequential activity, τp, achieved over the cascade of sequential activations of the clusters (see
Fig 5A). The time scales τc and τp are dependent on several model parameters: the cluster and network size,
the average connection strengths within the clusters, and adaptation. Smaller cluster sizes lead to a smaller τc
when the network size is fixed and conversely τp increases with network size when the cluster size is fixed.
The recurrent network is the "engine" that, once established, drives read-out dynamics. Our model can learn
different read-out synapses in parallel (Fig 4) and is robust to synapse failure (Fig 5C). This robustness is a con-
sequence of the clustered organization of the recurrent network. Previously proposed models are also robust to
similar levels of noise [Nicola and Clopath, 2017, Nicola and Clopath, 2019]. While an exact comparison is hard
to draw, we have shown that it is possible to retain robustness while moving towards a more biological learning
rule. The development of a clustered organization in the RNN allows a large drive for the read-out neurons while
keeping the individual synaptic strengths reasonably small. If the clusters become small, larger read-out synap-
tic strengths are required, and the dynamics become less robust. Indeed, the sequential dynamics is especially
fragile in the limit where every cluster has exactly one neuron. Furthermore, we show that learning is faster
with more neurons per cluster since relatively small changes in the synapses are sufficient to learn the target.
This is consistent with the intuitive idea that some redundancy in the network can lead to an increased learning
speed [Raman et al., 2019]. In its current form, the target pattern needs to be presented repeatedly to the net-
11
B Read-out weight matrixC Spontaneous sequence reactivations12004006004525Targetsequence545255452554525451200400600Time [ms]Read-outneuron rate6400RNN neuron indexRead-outneuron index 320018000RNNneuron index00.511.52Read-outneuron indexTime [s]00.511.52A Target and learned sequence4000 work and does not support one-shot learning through a single supervisor presentation. Although increasing the
cluster size NC can reduce the number of presentations, the size of the clusters would need to be impractically
large for one-shot learning. Alternatively, increasing the learning rate of the plastic read-out weights could be a
way to reduce the number of target presentations to just one [Nicola and Clopath, 2019]. However, it is unclear
at present whether such high learning rates are supported by experiments [Sjöström et al., 2001].
Taken together, our numerical simulations suggest ways to scale our network. An optimal network con-
figuration can be chosen given the temporal length of the target sequence, and requirements on the temporal
precision and robustness. For example, we have shown that a network with NE = 6400 and NC = 200 can be
chosen for a 400 ms target sequence that can be learned fast with good temporal precision and a robust replay.
If the network configuration and size are fixed, this severely constrains the sequences that can be learned and
how they are replayed.
In this paper, we use local Hebbian learning to produce a sequential dynamics in the recurrent network. This
is in contrast with previous studies, where often a recursive least squares method is used to train the weights
of the recurrent network [Rajan et al., 2016, Hardy and Buonomano, 2018]. Hardcoding a weight structure
into the recurrent network has been shown to result in a similar sequential dynamics [Chenkov et al., 2017,
Setareh et al., 2018, Spreizer et al., 2019]. Studies that do incorporate realistic plasticity rules are mostly lim-
ited to purely feedforward synfire chains [Fiete et al., 2010, Waddington et al., 2012, Zheng and Triesch, 2014].
Those studies focus on the generation of a dynamics that is sequential. In this regard, the contribution of our
work is to use the sequential dynamics as a key element to learning spatiotemporal spiking patterns. The ubiquity
of sequential dynamics in various brain regions [Ikegaya et al., 2004, Jin et al., 2009, Mackevicius et al., 2019]
and the architecture of the songbird system [Fee and Scharff, 2010] were an inspiration for the proposed sepa-
ration of temporal and spatial information in our setup. As we have shown, this separation enables the use of a
local Hebbian plasticity rule in the read-out synapses. Our model would therefore predict that perturbing the
sequential activity should lead to learning impairments. Further perturbation experiments can test this idea
and shed light on the mechanisms of sequential learning [Hemberger et al., 2019].
Previous studies have discussed whether sequences are learned and executed serially or hierarchically [Lashley, 1951].
Our recurrent network has a serial organization. When the sequential activity breaks down halfway, the re-
maining clusters are not activated further. A hierarchical structure would avoid such complete breakdowns
at the cost of a more complicated hardware to control the system. Sequences that are chunked in sub-
sequences can be learned separately and chained together. When there are errors in early sub-sequences
this will less likely affect the later sub-sequences. A hierarchical organization might also improve the ca-
pacity of the network.
In our proposed serial organization, the number of spatiotemporal patterns that
can be stored is equal to the number of read-out neurons. A hierarchical system could be one way to
extract general patterns and reduce the number of necessary read-out neurons. Evidence for hierarchical
structures is found throughout the literature [Sakai et al., 2003, Glaze and Troyer, 2006, Okubo et al., 2015].
The basal ganglia is for example thought to play an important role in shaping and controlling action se-
quences [Tanji, 2001, Jin and Costa, 2015, Geddes et al., 2018]. Another reason why a hierarchical organization
seems beneficial is inherent to the sequential dynamics. The time-variability of the sequential activity grows by
t (see Fig 5E)). While on a time scale of a few hundreds of milliseconds, this does not pose a
approximately
problem, for longer target sequences this variability would exceed the plasticity time constants. The presented
model could thus serve as an elementary building block of a more complex hierarchy.
In summary, we have demonstrated that a clustered network organization can be a powerful substrate for
learning, moving biological learning systems closer to machine learning performance. Specifically, the model
dissociates temporal and spatial information and therefore can make use of Hebbian learning to learn spa-
tiotemporal sequences over behavioural time scales. More general, the backbone as a clustered connectivity
might encode any variable x and enable downstream read-out neurons to learn and compute any function of
this variable, φ(x).
√
Methods
Neuron and synapse models
Excitatory neurons are modelled with the adaptive exponential integrate-and-fire model [Brette and Gerstner, 2005].
A classical integrate-and-fire model is used for the inhibitory neurons. All excitatory to excitatory recurrent
synapses are plastic under the voltage-based STDP rule [Clopath et al., 2010]. This enables the creation of neu-
ronal clusters and a feedforward structure. Normalization and weight bounds are used to introduce competition
and keep the recurrent network stable. Synapses from inhibitory to excitatory neurons in the recurrent network
are also plastic under a local plasticity rule [Vogels et al., 2011]. In general, it prevents runaway dynamics and
allows for an automatic search of good parameters (Fig S5). The connections from the recurrent network to
the read-out neurons are plastic under the same voltage-based STDP rule. However, potentiation of read-out
12
Table 1: Initialization of network
0
Constant Value
N E
2400
N I
600
0.2
p
2.83 pF
wEE
1.96 pF
wIE
62.87 pF
wEI
0
20.91 pF
wII
0 pF
wRE
200 pF
wRS
200 pF
wRH
200 pF
wHR
0
Description
Number of recurrent E neurons
Number of recurrent I neurons
Recurrent network connection probability
Initial E to E synaptic strength
E to I synaptic strength
Initial I to E synaptic strength
I to I synaptic strength
Initial E to R synaptic strength
S to R synaptic strength
H to R synaptic strength
R to H synaptic strength
synapses is linearly dependent on the strength of the synapses. There is no normalization here to allow a contin-
uous weight distribution. The dynamics was chosen based on previous models, with parameters for the dynamics
and plasticity to a large extent conserved [Litwin-Kumar and Doiron, 2014]. More simple integrate-and-fire dy-
namics should lead to the same qualitative results, given that the parameters are appropriately changed (data
not shown).
Network dynamics
Recurrent network A network with N E excitatory (E) and N I inhibitory (I) neurons is homogeneously
recurrently connected with connection probability p. Our network is balanced in terms of inhibition and excita-
tion, so that it displays irregular and asynchronous spiking. This is signalled by the coefficient of variation (CV)
of the inter-spike intervals of the neurons being CV ∼ 1, thus indicating Poisson-like spiking [Brunel, 2000]. In
our construction, we initialise the weights of the network near the balanced state by scaling the weights of the
balanced RNN in Ref. [Litwin-Kumar and Doiron, 2014] by the square root of the relative network size. We
then verify that the scaled parameters indeed lead to irregular dynamics. The spiking dynamics is slightly more
regular on average, with a mean CV ∼ 0.8 for excitatory neurons and a mean CV ∼ 0.9 for inhibitory neurons.
Read-out neurons The N E excitatory neurons from the recurrent network are all-to-all connected to N R
excitatory read-out (R) neurons. This weight matrix is denoted by W RE and it is where the learned sequence is
stored. To help learning, there are two additional types of neurons in the read-out network. During learning, the
read-out neurons receive supervisory input from N R excitatory supervisor (S) neurons. The connection from
supervisor neurons to read-out neurons is one-to-one and fixed, wRS. Also during learning, N R interneurons
(H) are one-to-one and bidirectionally connected to the read-out neurons with fixed connection strengths, wRH
and wHR (see Table 1 for the recurrent network and read-out parameters). The E to E, I to E and the E to
R connections are plastic.
Membrane potential dynamics There are two different regimes, one for each part of the model. Excitatory
neurons in the recurrent network have a high adaptation current while excitatory neurons in the read-out network
have no adaptation. This is to allow for a wide range of firing rates in the read-out network, while spiking is
more restricted in the recurrent network. Differences in the refractory period are there for the same reason,
but are not crucial. The membrane potential of the excitatory neurons (V E) in the recurrent network has the
following dynamics:
(cid:18)
(cid:18) V E − V E
T
(cid:19)(cid:19)
dV E
dt
=
1
τ E
L − V E + ∆E
EE
T exp
∆E
T
+ gEE EE − V E
C
+ gEI EI − V E
C
− aE
C
(1)
where τ E is the membrane time constant, EE
T is the slope of the exponential, C is
the capacitance, gEE, gEI are synaptic input from excitatory and inhibitory neurons respectively and EE, EI
are the excitatory and inhibitory reversal potentials respectively. When the membrane potential diverges and
exceeds 20 mV, the neuron fires a spike and the membrane potential is reset to Vr. This reset potential is the
same for all neurons in the model. There is an absolute refractory period of τabs. The parameter V E
T is adaptive
for excitatory neurons and set to VT + AT after a spike, relaxing back to VT with time constant τT :
L is the reversal potential, ∆E
τT
dV E
T
dt
= VT − V E
T .
13
(2)
The adaptation current aE for recurrent excitatory neurons follows:
τa
daE
dt
= −aE.
(3)
where τa is the time constant for the adaptation current (see also Fig S6). The adaptation current is increased
with a constant β when the neuron spikes. The membrane potential of the read-out (V R) neurons has no
adaptation current:
dV R
dt
=
1
τ E
L − V R + ∆E
EE
T exp
+ gRE EE − V R
C
+ gRS EE − V R
C
+ gRH EI − V R
C
(4)
(cid:18) V R − V R
T
(cid:19)(cid:19)
∆E
T
(cid:18)
L , ∆E
where τ E, EE
T , EE, EI and C are as defined before. gRE is the excitatory input from the recurrent network.
gRS is the excitatory input from the supervisor neuron (supervisor input only non-zero during learning, when
the target sequence is repeatedly presented). gRH is the inhibitory input from the interneuron (only non-zero
during learning, to have a gradual learning in the read-out synapses). The absolute refractory period is τabsR.
The threshold V R
T , with the same parameters. The membrane potential of
the supervisor neurons (V S) has no inhibitory input and no adaptation current:
T follows the same dynamics as V E
dV S
dt
=
1
τ E
L − V S + ∆E
EE
T exp
+ gSE EE − V S
C
(5)
(cid:18) V S − V S
T
(cid:19)(cid:19)
∆E
T
(cid:18)
where the constant parameters are defined as before and gSE is the external excitatory input from the target
sequence. The absolute refractory period is τabsS. The threshold V S
T follows again the same dynamics as V E
T ,
with the same parameters. The membrane potential of the inhibitory neurons (V I) in the recurrent network
has the following dynamics:
dV I
dt
=
L − V I
EI
τ I
+ gIE EE − V I
C
+ gII EI − V I
C
.
(6)
L is the inhibitory reversal potential and EE, EI are
where τ I is the inhibitory membrane time constant, EI
the excitatory and inhibitory resting potentials respectively. gEE and gEI are synaptic input from recurrent
excitatory and inhibitory neurons respectively. Inhibitory neurons spike when the membrane potential crosses
the threshold VT , which is non-adaptive. After this, there is an absolute refractory period of τabs. There is no
adaptation current. The membrane potential of the interneurons (V H) follow the same dynamics and has the
same parameters, but there is no inhibitory input:
dV H
dt
=
L − V H
EI
τ I
+ gHE EE − V H
C
(7)
where the excitatory input gHE comes from both the read-out neuron it is attached to and external input.
After the threshold VT is crossed, the interneuron spikes and an absolute refractory period of τabsH follows.
The interneurons inhibit the read-out neurons stronger when they receive strong inputs from the read-out
neurons. This slows the potentiation of the read-out synapses down and keeps the synapses from potentiating
exponentially (see Table 2 for the parameters of the membrane dynamics).
Synaptic dynamics The synaptic conductance of a neuron i is time dependent, it is a convolution of a kernel
with the total input to the neuron i:
(t) = K Y (t) ∗
gXY
i
ext sX
i,ext +
W XY
ij
sY
j (t)
(8)
W X
(cid:88)
j
.
where X and Y denote two different neuron types in the model (E, I, R, S or H). K is the difference of
exponentials kernel:
K Y (t) =
d − e−t/τ Y
e−t/τ Y
d − τ Y
τ Y
r
r
,
with a decay time τd and a rise time τr dependent only on whether the neuron is excitatory or inhibitory. There
is no external inhibitory input to the supervisor and inter- neurons. During spontaneous activity, there is no
external inhibitory input to the recurrent network and a fixed rate. The external input to the interneurons has
a fixed rate during learning as well. The external input to the supervisor neurons is dependent on the specific
learning task. There is no external input to the read-out neurons. The externally incoming spike trains sX
ext
are generated from a Poisson process with rates rX
ext. The externally generated spike trains enter the network
through synapses W X
ext (see Table 3 for the parameters of the synaptic dynamics).
14
Table 2: Neuronal membrane dynamics parameters
Description
E membrane potential time constant
I membrane potential time constant
Refractory period of E and I neurons
R neurons refractory period
S neurons refractory period
H neurons refractory period
excitatory reversal potential
inhibitory reversal potential
excitatory resting potential
inhibitory resting potential
Constant Value
20 ms
τE
20 ms
τI
5 ms
τabs
1 ms
τabsR
1 ms
τabsS
1 ms
τabsH
0 mV
EE
−75 mV
EI
−70 mV
EE
−62 mV
L
EI
−60 mV Reset potential (for all neurons the same)
L
Vr
300 pF
C
2 mV
∆E
T
30 ms
τT
−52 mV Membrane potential threshold
VT
10 mV
AT
100 ms
τa
1000 pA Adaptation current increase constant of recurrent network neurons
β
Capacitance
Exponential slope
Adaptive threshold time constant
Adaptive threshold increase constant
Adaptation current time constant
Table 3: Synaptic dynamics parameters
Description
E decay time constant
E rise time constant
I rise time constant
I rise time constant
External input synaptic strength to E neurons
Rate of external input to E neurons
External input synaptic strength to I neurons
Constant Value
6 ms
τ E
d
1 ms
τ E
r
2 ms
τ I
d
0.5 ms
τ I
r
1.6 pF
W E
ext
4.5 kHz
rE
ext
1.52 pF
W I
2.25 kHz Rate of external input to I neurons
rI
ext
1.6 pF
W S
ext
1.6 pF
W H
ext
1.0 kHz
rH
ext
External input synaptic strength to S neurons
External input synaptic strength to H neurons
Rate of external input to H neurons
ext
15
Table 4: Excitatory plasticity parameters
Constant Value
ALT D
A
θLT D
θLT P
τu
0.0014 pA mV −2
0.0008 pA mV −1
−70 mV
−49 mV
10 ms
τv
τxEE
τxRE
W EE
min
W EE
max
W RE
min
W RE
max
7 ms
3.5 ms
5 ms
1.45 pF
32.68 pF
0 pF
25 pF
Description
LTD amplitude
LTP amplitude (in RNN: ALT P = A)
LTD threshold
LTP threshold
Time constant of low pass filtered
postsynaptic membrane potential (LTD)
Time constant of low pass filtered
postsynaptic membrane potential (LTP)
Time constant of low pass filtered
presynaptic spike train in recurrent network
Time constant of low pass filtered
presynaptic spike train for read-out synapses
Minimum E to E weight
Maximum E to E weight
Minimum E to R weight
Maximum E to R weight
Plasticity
Excitatory plasticity The voltage-based STDP rule is used [Clopath et al., 2010]. The synaptic weight from
excitatory neuron j to excitatory neuron i is changed according to the following differential equation:
= −ALT D sj(t) R(cid:0)ui(t) − θLT D
(cid:1) + ALT P xj(t) R(cid:0)Vi(t) − θLT P
(cid:1) R(cid:0)vi(t) − θLT D
(cid:1).
dWij
dt
(9)
Here, ALT D and ALT P are the amplitude of depression and potentiation respectively. θLT D and θLT P are
the voltage thresholds to recruit depression and potentiation respectively, as R(.) denotes the linear-rectifying
function (R(x) = 0 if x < 0 and else R(x) = x). Vi is the postsynaptic membrane potential, ui and vi are
low-pass filtered versions of Vi, with respectively time constants τu and τv (see also Fig S6):
τu
τv
dui
dt
dvi
dt
= Vi − ui
= Vi − vi
where sj is the presynaptic spike train and xj is the low-pass filtered version of sj with time constant τx:
τx
dxj
dt
= sj − xj.
(10)
(11)
(12)
a neuron constant: (cid:80)
j Wij = K. E to E weights have a lower and upper bound [W EE
Here the time constant τx is dependent on whether learning happens inside (E to E) or outside (E to R) the
recurrent network. sj(t) = 1 if neuron j spikes at time t and zero otherwise. Competition between synapses in
the recurrent network is enforced by a hard L1 normalization every 20 ms, keeping the sum of all weights onto
max]. The minimum
and maximum strengths are important parameters and determine the position of the dominant eigenvalues of
W . Potentiation of the read-out synapses is weight dependent. Assuming that stronger synapses are harder to
potentiate [Debanne et al., 1999], ALT P reduces linearly with W RE:
max − W RE
W RE
max − W RE
W RE
min, W EE
ALT P = A
(13)
.
min
The maximum LTP amplitude A is reached when W RE = W RE
plasticity rule).
min (see Table 4 for the parameters of the excitatory
Inhibitory plasticity Inhibitory plasticity acts as a homeostatic mechanism, previously shown to prevent
runaway dynamics [Vogels et al., 2011, Litwin-Kumar and Doiron, 2014, Zenke et al., 2015]. Here, it allows to
automatically find good parameters (see also Fig S5). Excitatory neurons that fire with a higher frequency
will receive more inhibition. The I to E weights are changed when the presynaptic inhibitory neuron or the
postsynaptic excitatory neuron fires [Vogels et al., 2011]:
i (t) − 2r0τy
j (t) + Ainh yI
(cid:1) sI
(cid:0)yE
j (t) sE
= Ainh
dWij
(14)
i (t)
dt
16
Table 5: Inhibitory plasticity parameters
Description
Constant Value
Ainh
r0
τy
W EI
min
W EI
max
10−5 AHz Amplitude of inhibitory plasticity
3 Hz
20 ms
48.7 pF
243 pF
Target firing rate
Time constant of low pass filtered spike train
Minimum I to E weight
Maximum I to E weight
Table 6: Parameters for the large recurrent network (all the other parameters are the same as the smaller
network)
0
Constant Value
N E
6400
N I
1600
1.73 pF
wEE
1.20 pF
wIE
40 pF
wEI
0
12.80 pF
wII
1.27 pF
W EE
min
30.5 pF
W EE
max
40 pF
W EI
min
200 pF
W EI
max
15 pF
W RE
max
Description
Number of recurrent E neurons
Number of recurrent I neurons
baseline E to E synaptic strength
E to I synaptic strength
Initial I to E synaptic strength
I to I synaptic strength
Minimum E to E weight
Maximum E to E weight
Minimum I to E weight
Maximum I to E weight
Maximum E to R weight
where r0 is a constant target rate for the postsynaptic excitatory neuron. sE and sI are the spike trains of the
postsynaptic E and presynaptic I neuron respectively. The spike trains are low pass filtered with time constant
τy to obtain yE and yI (as in equation 12). Table 5 shows parameter values for the inhibitory plasticity rule.
The I to E synapses have a lower and upper bound [W IE
min, W IE
max].
Learning protocol
Learning happens in two stages. First a sequential dynamics is learned in the RNN. Once this temporal backbone
is established connections to read-out neurons can be learned. Read-out neurons are not interconnected and
can learn in parallel.
Recurrent network The network is divided in 30 disjoint clusters of 80 neurons. The clusters are sequentially
stimulated for a time duration of 60 minutes by a large external current where externally incoming spikes are
drawn from a Poisson process with rate 18 kHz. This high input rate does not originate from a single external
neuron but rather assumes a large external input population. Each cluster is stimulated for 10 ms and in between
cluster stimulations there are 5 ms gaps (see also Fig S6 for different gaps). During excitatory stimulation of
a cluster, all other clusters receive an external inhibitory input with rate 4.5 kHz and external input weight
ext = 2.4 pF . There is a periodic boundary condition, i.e. after the last cluster is activated, the first cluster
W I
is activated again. After the sequential stimulation, the network is spontaneously active for 60 minutes. The
connectivity stabilizes during the spontaneous dynamics. Learning in scaled versions of this network happens
in exactly the same way (Fig 5A). The recurrent weight matrix of the large network (80 clusters of 80 neurons,
Figs 5E and 6) is learned using the same protocol. The recurrent weight matrix reaches a stable structure after
three hours of sequential stimulation followed by three hours of spontaneous dynamics. Parameters that change
for the scaled up version are summarized in Table 6. For randomly switching dynamics, a similar protocol is
followed (Fig S1). The weight matrix used to plot the spectrum of the recurrent network in Figs 2 and S2 is:
(cid:18) W EE W IE
−W EI −W II
(cid:19)
W =
.
Read-out network During learning of the read-out synapses, external input drives the supervisor and inter-
neurons. The rate of the external Poisson input to the supervisor neurons reflects the sequence that has to
be learned. The rate is normalized to be between 0 kHz and 10 kHz. During learning, W RE changes. After
learning, the external input to the supervisor and inter- neurons is turned off and both stop firing. The read-out
neurons are now solely driven by the recurrent network. Plasticity is frozen in the read-out synapses after
learning. With plasticity on during spontaneous dynamics, the read-out synapses would continue to potentiate
17
because of the coactivation of clusters in the recurrent network and read-out neurons. This would lead to
read-out synapses that are all saturated at the upper weight bound.
Simulations The code used for the training and simulation of the recurrent network is built on top of the
code from [Litwin-Kumar and Doiron, 2014] in Julia. The code used for learning spatiotemporal sequences
using read-out neurons is written in Matlab. Forward Euler discretization with a time step of 0.1 ms is used.
The code is available for the reviewers on ModelDB (http://modeldb.yale.edu/257609) with access code
'SSRSNT'. The code will be made public after publication.
Linear rate model: spectral analysis
A linear rate model can give insight into the dynamics of a large nonlinear structured spiking network [Schaub et al., 2015].
The dynamics of a simplified rate model with the same feedforward structure as in the RNN is as follows:
= −x + Ax + ξ
dx
dt
(15)
where x is a multidimensional variable consisting of the rates of all excitatory and inhibitory clusters, A is
the weight matrix, and ξ is white noise. The matrix A is a coarse-grained version of the weight matrix of the
recurrent network in Fig 2B averaged over each cluster. In order to obtain analytical expressions, we consider a
network with 3 excitatory clusters and 1 inhibitory cluster. The connectivity of this model can be parametrized
as:
where δ > 0, w = δ + + 1, > 1 guarantees sequential dynamics, and k > 1 guarantees a balanced network.
The Schur decomposition A = U T U T gives eigenvalues and Schur vectors ui:
(16)
(17)
(18)
−kw
1 −kw
δ −kw
3 −kw
A =
√
w
1
w
3
1
δ
w
3
δ
(cid:1)
3w(cid:0)k + 1
, u3 =
1
0
0
0
1
3
1−3
0
0
0
0
δ − +1
√
2 ( − 1)
3
2
3
− √
2 ( − 1)
δ − +1
2
−1
, u4 =
0
1
0
1√
2
1√
6
−1
.
2−1
0
−w(k − 1)
, u2 =
0
0
0
2
√
1
3
T =
1
1
1
1
u1 =
1
2
The first mode u1 decays fast and uniformly over the four neuronal groups. The second mode u2 decays more
slowly, and indicates the interplay between excitatory groups (x1, x2, x3) and the inhibitory group x4. The
eigenspace associated with the pair {u3, u4} has complex conjugate eigenvalues and is localized on the three
excitatory groups. An increase of activity in one excitatory group is coupled with decreased activities in the
other groups. If the real part δ− +1
2 < 1 then these modes are linearly stable but if the real part becomes closer
to one means a slower decay of this mode. Importantly, the imaginary part of the eigenvalues is ±√
3( − 1)/2;
hence it grows linearly with the strength of the feedforward structure ( − 1) (Fig 7A). This leads to oscillatory
behavior, which determines the time scale of the sequential switching.
18
Fig 7. Spectral analysis of reduced linear model. (A) Cartoon of a simplified linearised rate model with
three nodes x1, x2, x3 corresponding to three clusters of excitatory neurons with recurrent strength δ connected
to a central cluster of inhibitory neurons x4. The cyclic connections are stronger clockwise than anticlockwise
since > 1. (B) The spectrum shows a conjugate complex eigenvalue pair with large real part (2δ − − 1)/2
and an imaginary part ±√
3( − 1)/2 which grows linearly with the asymmetry of the clockwise/anticlockwise
strength ( − 1). This pair of eigenvalues dominates the dynamics as their real parts are close to 1 and leads to
the periodic behaviour corresponding to propagation around the cycle x1 → x2 → x3 → x1 . . ..
References
[Abbott et al., 2016] Abbott, L. F., DePasquale, B., and Memmesheimer, R. M. (2016). Building functional
networks of spiking model neurons. Nature Neuroscience, 19(3):350 -- 355.
[Abeles, 1991] Abeles, M. (1991). Corticonics: Neural circuits of the cerebral cortex. Cambridge University
Press.
[Adler et al., 2019] Adler, A., Zhao, R., Shin, M. E., Yasuda, R., and Gan, W.-B. (2019). Somatostatin-
Expressing Interneurons Enable and Maintain Learning-Dependent Sequential Activation of Pyramidal Neu-
rons. Neuron, 102(1):202 -- 216.e7.
[Billeh and Schaub, 2018] Billeh, Y. N. and Schaub, M. T. (2018). Feedforward architectures driven by in-
hibitory interactions. Journal of Computational Neuroscience, 44(1):63 -- 74.
[Brea et al., 2013] Brea, J., Senn, W., and Pfister, J.-P. (2013). Matching Recall and Storage in Sequence
Learning with Spiking Neural Networks. Journal of Neuroscience, 33(23):9565 -- 9575.
[Brette and Gerstner, 2005] Brette, R. and Gerstner, W. (2005). Adaptive exponential integrate-and-fire model
as an effective description of neuronal activity. Journal of Neurophysiology, 94(5):3637 -- 3642.
[Brunel, 2000] Brunel, N. (2000). Dynamics of Sparsely Connected Networks of Excitatory and Inhibitory
Spiking Neurons. Journal of Computational Neuroscience, 8(3):183 -- 208.
[Chenkov et al., 2017] Chenkov, N., Sprekeler, H., and Kempter, R. (2017). Memory replay in balanced recur-
rent networks. PLoS Computational Biology, 13(1):e1005359.
[Clopath et al., 2010] Clopath, C., Büsing, L., Vasilaki, E., and Gerstner, W. (2010). Connectivity reflects
coding: A model of voltage-based STDP with homeostasis. Nature Neuroscience, 13(3):344 -- 352.
[Debanne et al., 1999] Debanne, D., Gähwiler, B. H., and Thompson, S. M. (1999). Heterogeneity of Synaptic
Plasticity at Unitary CA3 -- CA1 and CA3 -- CA3 Connections in Rat Hippocampal Slice Cultures. The Journal
of Neuroscience, 19(24):10664 -- 10671.
[Fee and Scharff, 2010] Fee, M. S. and Scharff, C. (2010). The Songbird as a Model for the Generation of
Complex Behaviors. ILAR journal, 51(4):362 -- 377.
[Fiete et al., 2010] Fiete, I. R., Senn, W., Wang, C. Z., and Hahnloser, R. H. (2010). Spike-Time-Dependent
Plasticity and Heterosynaptic Competition Organize Networks to Produce Long Scale-Free Sequences of
Neural Activity. Neuron, 65(4):563 -- 576.
[Geddes et al., 2018] Geddes, C. E., Li, H., and Jin, X. (2018). Optogenetic Editing Reveals the Hierarchical
Organization of Learned Action Sequences. Cell, 174(1):32 -- 43.e15.
19
x4A Cartoon of linear rate modelB Eigenvalues of linear rate model-kww/31-101Real axis-0.0500.05Imaginary axisx1x2x33 [Gibbon, 1977] Gibbon, J. (1977). Scalar expectancy theory and Weber's law in animal timing. Psychological
Review, 84(3):279 -- 325.
[Gilra and Gerstner, 2017] Gilra, A. and Gerstner, W. (2017). Predicting non-linear dynamics by stable local
learning in a recurrent spiking neural network. Elife, 6(e28295):1 -- 38.
[Glaze and Troyer, 2006] Glaze, C. M. and Troyer, T. W. (2006). Temporal Structure in Zebra Finch Song:
Implications for Motor Coding. Journal of Neuroscience, 26(3):991 -- 1005.
[Gütig and Sompolinsky, 2006] Gütig, R. and Sompolinsky, H. (2006). The tempotron: A neuron that learns
spike timing-based decisions. Nature Neuroscience, 9(3):420 -- 428.
[Hahnloser et al., 2002] Hahnloser, R. H., Kozhevnikov, A. A., and Fee, M. S. (2002). An ultra-sparse code
underlies the generation of neural sequences in a songbird. Nature, 419(6902):65 -- 70.
[Hardy and Buonomano, 2018] Hardy, N. F. and Buonomano, D. (2018). Encoding Time in Feedforward Tra-
jectories of a Recurrent Neural Network Model. Neural Computation, 30(2):378 -- 396.
[Harvey et al., 2012] Harvey, C. D., Coen, P., and Tank, D. W. (2012). Choice-specific sequences in parietal
cortex during a virtual-navigation decision task. Nature, 484(7392):62 -- 68.
[Hemberger et al., 2019] Hemberger, M., Shein-Idelson, M., Pammer, L., and Laurent, G. (2019). Reliable
Sequential Activation of Neural Assemblies by Single Pyramidal Cells in a Three-Layered Cortex. Neuron,
104(2):353 -- 369.e5.
[Ikegaya et al., 2004] Ikegaya, Y., Aaron, G., Cossart, R., Aronov, D., Lampl, I., Ferster, D., and Yuste, R.
(2004). Synfire Chains and Cortical Songs: Temporal Modules of Cortical Activity. Science, 304(5670):559 --
564.
[Itskov et al., 2011] Itskov, V., Curto, C., Pastalkova, E., and Buzsaki, G. (2011). Cell Assembly Sequences
Arising from Spike Threshold Adaptation Keep Track of Time in the Hippocampus. Journal of Neuroscience,
31(8):2828 -- 2834.
[Jaeger et al., 2007] Jaeger, H., Maass, W., and Principe, J. (2007). Special issue on echo state networks and
liquid state machines. Neural Networks, 20(3):287 -- 289.
[Jin et al., 2009] Jin, D. Z., Fujii, N., and Graybiel, A. M. (2009). Neural representation of time in cortico-
basal ganglia circuits. Proceedings of the National Academy of Sciences of the United States of America,
106(45):19156 -- 19161.
[Jin and Costa, 2015] Jin, X. and Costa, R. M. (2015). Shaping action sequences in basal ganglia circuits.
Current Opinion in Neurobiology, 33:188 -- 196.
[Jun et al., 2017] Jun, J. J., Steinmetz, N. A., Siegle, J. H., Denman, D. J., Bauza, M., Barbarits, B., Lee,
A. K., Anastassiou, C. A., Andrei, A., Aydin, Ç., Barbic, M., Blanche, T. J., Bonin, V., Couto, J., Dutta, B.,
Gratiy, S. L., Gutnisky, D. A., Häusser, M., Karsh, B., Ledochowitsch, P., Lopez, C. M., Mitelut, C., Musa,
S., Okun, M., Pachitariu, M., Putzeys, J., Rich, P. D., Rossant, C., Sun, W. L., Svoboda, K., Carandini, M.,
Harris, K. D., Koch, C., O'Keefe, J., and Harris, T. D. (2017). Fully integrated silicon probes for high-density
recording of neural activity. Nature, 551(7679):232 -- 236.
[Ko et al., 2013] Ko, H., Cossell, L., Baragli, C., Antolik, J., Clopath, C., Hofer, S. B., and Mrsic-Flogel, T. D.
(2013). The emergence of functional microcircuits in visual cortex. Nature, 496(7443):96 -- 100.
[Laje and Buonomano, 2013] Laje, R. and Buonomano, D. V. (2013). Robust timing and motor patterns by
taming chaos in recurrent neural networks. Nature neuroscience, 16(7):925 -- 33.
[Lashley, 1951] Lashley, K. S. (1951). The Problem of Serial Order in Behavior. Cerebral Mechanisms in
Behavior, 21:112 -- 146.
[Lee et al., 2019] Lee, H., Choi, W., Park, Y., and Paik, S.-B. (2019). Distinct role of flexible and stable
encoding in sequential working memory. Neural Networks, 121:419 -- 429.
[Leonardo and Fee, 2005] Leonardo, A. and Fee, M. S. (2005). Ensemble Coding of Vocal Control in Birdsong.
Journal of Neuroscience, 25(3):652 -- 661.
[Litwin-Kumar and Doiron, 2014] Litwin-Kumar, A. and Doiron, B. (2014). Formation and maintenance of
neuronal assemblies through synaptic plasticity. Nature Communications, 5(5319):1 -- 12.
20
[Maass, 2016] Maass, W. (2016). Searching for principles of brain computation. Current Opinion in Behavioral
Sciences, 11:81 -- 92.
[Maass and Markram, 2002] Maass, W. and Markram, H. (2002). Synapses as dynamic memory buffers. Neural
Networks, 15(2):155 -- 161.
[Mackevicius et al., 2019] Mackevicius, E. L., Bahle, A. H., Williams, A. H., Gu, S., Denisenko, N. I., Goldman,
M. S., and Fee, M. S. (2019). Unsupervised discovery of temporal sequences in high-dimensional datasets,
with applications to neuroscience. eLife, 8(e38471):1 -- 42.
[Manohar et al., 2019] Manohar, S. G., Zokaei, N., Fallon, S. J., Vogels, T. P., and Husain, M. (2019). Neural
mechanisms of attending to items in working memory. Neuroscience and Biobehavioral Reviews, 101:1 -- 12.
[Mastrogiuseppe and Ostojic, 2018] Mastrogiuseppe, F. and Ostojic, S. (2018). Linking Connectivity, Dynam-
ics, and Computations in Low-Rank Recurrent Neural Networks. Neuron, 99(3):609 -- 623.e29.
[Morrison et al., 2007] Morrison, A., Aertsen, A., and Diesmann, M. (2007). Spike-timing dependent plasticity
in balanced random networks. Neural Computation, 19:1437 -- 1467.
[Nicola and Clopath, 2017] Nicola, W. and Clopath, C. (2017). Supervised learning in spiking neural networks
with FORCE training. Nature Communications, 8(1):1 -- 15.
[Nicola and Clopath, 2019] Nicola, W. and Clopath, C. (2019). A diversity of interneurons and Hebbian plas-
ticity facilitate rapid compressible learning in the hippocampus. Nature Neuroscience, 22(7):1168 -- 1181.
[Okubo et al., 2015] Okubo, T. S., Mackevicius, E. L., Payne, H. L., Lynch, G. F., and Fee, M. S. (2015).
Growth and splitting of neural sequences in songbird vocal development. Nature, 528(7582):352 -- 357.
[Park et al., 2017] Park, Y., Choi, W., and Paik, S. B. (2017). Symmetry of learning rate in synaptic plasticity
modulates formation of flexible and stable memories. Scientific Reports, 7(1):1 -- 12.
[Pastalkova et al., 2008] Pastalkova, E., Itskov, V., Amarasingham, A., and Buzsáki, G. (2008).
generated cell assembly sequences in the rat hippocampus. Science, 321(5894):1322 -- 1327.
Internally
[Peters et al., 2014] Peters, A. J., Chen, S. X., and Komiyama, T. (2014). Emergence of reproducible spa-
tiotemporal activity during motor learning. Nature, 510(7504):263 -- 267.
[Rajan et al., 2016] Rajan, K., Harvey, C. D., and Tank, D. W. (2016). Recurrent Network Models of Sequence
Generation and Memory. Neuron, 90(1):128 -- 142.
[Raman et al., 2019] Raman, D. V., Rotondo, A. P., and O'Leary, T. (2019). Fundamental bounds on learning
performance in neural circuits. In Proceedings of the National Academy of Sciences.
[Rhodes et al., 2004] Rhodes, B. J., Bullock, D., Verwey, W. B., Averbeck, B. B., and Page, M. P. (2004).
Learning and production of movement sequences: Behavioral, neurophysiological, and modeling perspectives.
Human Movement Science, 23(5):699 -- 746.
[Sakai et al., 2003] Sakai, K., Kitaguchi, K., and Hikosaka, O. (2003). Chunking during human visuomotor
sequence learning. Experimental Brain Research, 152(2):229 -- 242.
[Schaub et al., 2015] Schaub, M. T., Billeh, Y. N., Anastassiou, C. A., Koch, C., and Barahona, M. (2015).
Emergence of Slow-Switching Assemblies in Structured Neuronal Networks. PLoS Computational Biology,
11(7):1 -- 28.
[Setareh et al., 2018] Setareh, H., Deger, M., and Gerstner, W. (2018). Excitable neuronal assemblies with
adaptation as a building block of brain circuits for velocity-controlled signal propagation. PLoS Computational
Biology, 14(7):e1006216.
[Sjöström et al., 2001] Sjöström, P. J., Turrigiano, G. G., and Nelson, S. B. (2001). Rate, timing, and cooper-
ativity jointly determine cortical synaptic plasticity. Neuron, 32(6):1149 -- 1164.
[Spreizer et al., 2019] Spreizer, S., Aertsen, A., and Kumar, A. (2019). From space to time: Spatial inhomo-
geneities lead to the emergence of spatio-temporal activity sequences in spiking neuronal networks. PLoS
Computional Biology, 15(10):e1007432.
[Sussillo and Abbott, 2009] Sussillo, D. and Abbott, L. F. (2009). Generating Coherent Patterns of Activity
from Chaotic Neural Networks. Neuron, 63(4):544 -- 557.
21
[Tanji, 2001] Tanji, J. (2001). Sequential Organization of Multiple Movements : Involvement of Cortical Motor
Areas. Annual Review of Neuroscience, 24:631 -- 651.
[Tao, 2013] Tao, T. (2013). Outliers in the spectrum of iid matrices with bounded rank perturbations. Probability
Theory and Related Fields, 155(1-2):231 -- 263.
[Tully et al., 2016] Tully, P. J., Lindén, H., Hennig, M. H., and Lansner, A. (2016). Spike-Based Bayesian-
Hebbian Learning of Temporal Sequences. PLoS Computational Biology, 12(5).
[Vogels et al., 2011] Vogels, T. P., Sprekeler, H., Zenke, F., Clopath, C., and Gerstner, W. (2011). Inhibitory
Plasticity Balances Excitation and Inhibition in Sensory Pathways and Memory Networks. Science, 334(De-
cember 2011):1569 -- 1573.
[Waddington et al., 2012] Waddington, A., de Kamps, M., and Cohen, N. (2012). Triphasic spike-timing-
dependent plasticity organizes networks to produce robust sequences of neural activity. Frontiers in Compu-
tational Neuroscience, 6:1 -- 14.
[Werbos, 1990] Werbos, P. J. (1990). Backpropagation Through Time: What It Does and How to Do It.
Proceedings of the IEEE, 78(10):1550 -- 1560.
[Zenke et al., 2015] Zenke, F., Agnes, E. J., and Gerstner, W. (2015). Diverse synaptic plasticity mechanisms
orchestrated to form and retrieve memories in spiking neural networks. Nature Communications, 6(6922):1 --
13.
[Zheng and Triesch, 2014] Zheng, P. and Triesch, J. (2014). Robust development of synfire chains from multiple
plasticity mechanisms. Frontiers in Computational Neuroscience, 8:1 -- 10.
22
Supplementary material
Fig S1. Randomly switching dynamics. The recurrent network is stimulated with external input that
is spatially clustered, but temporally uncorrelated. Each cluster is stimulated for 50 ms, with 50 ms gaps in
between stimulations. The rate of external stimulation is 18 kHz during training. This is repeated for 20
minutes after which the network stabilizes during 20 minutes of spontaneous activity. (A) A diagonal structure
is learned in the recurrent weight matrix. Since there are no temporal correlations in the external input,
there is no off-diagonal structure. (B) The spectrum shows an eigenvalue gap. This indicates the emergence
of a slower time scale. The leading eigenvalues do not have an imaginary part, pointing at the absence of
feedforward structure and thus there is no sequential dynamics. (C) Under a regime of spontaneous dynamics
(i.e. uncorrelated Poisson inputs), the clusters are randomly reactivated.
23
ADetail of recurrent weight matrix(5 first clusters) 8016024032080160240320Postsynaptic neuron indexPresynaptic neuron indexBC-5000500Real part-1000-50005001000Imaginary partSpectrum of RNN weight matrix(full matrix)800160024003000Recurrent networkneuron index Random switching between clusters during spontaneous dynamicsTime [s]0.210.6 Fig S2. The connectivity structure is stable under spontaneous dynamics. (A) After 60 minutes
of training, the network stabilizes during spontaneous activity. During the first 30 minutes of spontaneous
dynamics, the connectivity still changes. More specifically, the imaginary parts of the leading eigenvalues
increase. This leads to a higher switching frequency and as such a smaller period in the sequential activity.
After around 30 minutes, a fixed point is reached. The first row shows spike trains at different times, for one
second of spontaneous activity. The second row shows the spectra of the weight matrix at those times. (B)
After 60 minutes of sequential stimulation, we test reinforcement and degradation of the learned connectivity
by decoupling the plasticity from the dynamics. We plot the evolution of the off-diagonal weights during
spontaneous dynamics in two separate cases: (i) we run the dynamics of the network using a frozen copy of the
learned weight matrix and apply plastic changes that result from the dynamics to the original weight matrix
(blue curve); (ii) we run the dynamics of the network using a frozen copy of the learned weight matrix where
the off-diagonal structure was removed and apply plastic changes that result from the dynamics to the original
weight matrix (red curve). We can see that in the former, the learned connectivity is reinforced and in the
latter, the learned connectivity degrades. Off-diagonal weights (the y-axis) are quantified by averaging over the
weights in the 80 by 80 blocks in the lower diagonal, for the 30 different clusters. The curves are the means
over the 30 clusters and the error bars one standard deviation.
24
Imaginary axis 10002000-2000-100000800160024003000Recurrent networkneuron index 00.51Time [s] 800160024003000800160024003000Development and stability of the temporal backbone0204060Time [min]1.522.5Off-diagonalweightsDegradation and reinforcement of connectivity during spontaneous dynamicsReal axis -5000500After 60 minutes ofspontaneous dynamics Real axis -5000500Real axis -5000500After 30 minutes ofspontaneous dynamics After 60 minutes ofsequential stimulation 00.51Time [s] 00.51Time [s] SpectrumSpectrumSpectrumEnd of trainingSpontaneous random activityABDegradationReinforcement10002000-2000-10000010002000-2000-100000 Fig S3. Noisy learning. The sequence ABCBA is relearned four times for 12 seconds each. Before relearning,
the read-out weight matrix W RE was always reset. When active, read-out neurons fire two spikes on average
+/− one spike. This variability is a consequence of the noisy learning process.
Fig S4. Details of some network properties. (A) Duration that a cluster is activated as a function of
network size (B) Raster plot of sequential dynamics for NE = 1200 and NC = 40, after training. We observe
that by reducing the cluster size, the irregularities in the sequential dynamics are increased (compare with
Fig 2). (C) Two raster plots showing two different levels of robustness (summary plot in Fig 5B). In both cases,
at t = 1s (purple arrow), 40 read-out synapses are deleted for each cluster. Left panel: NC = 120, each read-out
neuron fires two spikes before deletion and one spike after deletion resulting in ∼ 50% performance. Right panel:
NC = 200, each read-out neuron fires two spikes before deletion and one or two spikes after deletion resulting
in a higher performance (∼ 80%).
25
00.20.40.60.8Time [s]ABCRead-out neuron00.20.40.60.8Time [s]ABCRead-out neuron00.20.40.60.8Time [s]ABCRead-out neuron00.20.40.60.8Time [s]ABCRead-out neuron012200040006000Recurrent networkneuron index012Time [s]ABCRead-out neuronSyanp Duration of cluster activations in various networks Networks with small clusters show more irregular sequential dynamicsC Spike rasters showing in more detail the robustness experiment for two casesABNC = 120NC = 200Synapse deletionSynapse deletion Fig S5. The role of inhibition. (A) Inhibitory neurons are necessary to prevent pathological excitatory
activity. (B) The weights projecting from the inhibitory neurons to the excitatory neurons without inhibitory
plasticity are random (left panel). The weights projecting from the inhibitory neurons to the excitatory neurons
with inhibitory plasticity shows some structure (right panel). (C) The full spectrum of the recurrent weight
matrix after learning without inhibitory plasticity. (D) Without inhibitory plasticity, the sequential dynamics
shows irregularities. The inhibitory plasticity allows for better parameters to be found to stabilize the sequential
dynamics in the recurrent network.
Fig S6. Sensitivity to parameters. The periods of the sequential dynamics are computed after one hour of
external stimulation and one hour of spontaneous dynamics. Only one parameter at a time is changed. (A) The
adaptation time constant is varied. (B) The time gap between external sequential stimulations is varied. (C)
The time constants of the voltage-based STDP rule are varied. The lines are guides to the eye and the error
bars indicate one standard deviation.
26
04080Time [ms]180016002400Recurrent networkneuron indexA No inhibitory neurons leads to pathological activity-5000500Real axis-2000-1000010002000Imaginary axis00.511.52Time [s]800160024003000Recurrent networkneuron indexC Spectrum after learning without inhibitory plasticityD Spike raster shows irregularities15080Presynaptic Ineuron index180160240320400Postsynaptic Eneuron index0102030405060150100Presynaptic Ineuron index180160240320400Postsynaptic Eneuron index050100150200B Inhibitory to excitatory weights without and with inhibitory plasticity 357Pairing gap [ms]440460480Period [ms]-0.50+0.5-10+1-20+2Change in time constant [ms]420440460480500Period [ms]vxu50100150Adaptation timeconstant [ms]440460480Period [ms]CAB |
1301.7747 | 1 | 1301 | 2013-01-31T20:47:53 | Entrainment of the intrinsic dynamics of single isolated neurons by natural-like input | [
"q-bio.NC"
] | Neuronal dynamics is intrinsically unstable, producing activity fluctuations that are essentially scale-free. Here we show that while these scale-free fluctuations are independent of temporal input statistics, they can be entrained by input variation. Joint input output statistics and spike train reproducibility in synaptically isolated cortical neurons were measured in response to various input regimes over extended time scales (many minutes). Response entrainment was found to be maximal when the input itself possesses natural-like, scale-free statistics. We conclude that preference for natural stimuli, often observed at the system level, exists already at the elementary, single neuron level. | q-bio.NC | q-bio |
Entrainment of the intrinsic dynamics of single isolated neurons
by natural-like input
Asaf Gal1,2,∗, Shimon Marom1
1 Network Biology Research Laboratories, Lorry Lokey Interdisciplinary Center
for Life Sciences and Engineering, Technion, Haifa, Israel
2 The Interdisciplinary Center for Neural Computation (ICNC), The Hebrew
University, Jerusalem, Israel
∗ E-mail: [email protected]
Abstract
Neuronal dynamics is intrinsically unstable, producing activity fluctuations that are essentially
scale-free. Here we show that while these scale-free fluctuations are independent of temporal
input statistics, they can be entrained by input variation. Joint input output statistics and
spike train reproducibility in synaptically isolated cortical neurons were measured in response
to various input regimes over extended time scales (many minutes). Response entrainment
was found to be maximal when the input itself possesses natural-like, scale-free statistics. We
conclude that preference for natural stimuli, often observed at the system level, exists already
at the elementary, single neuron level.
Introduction
Variability is a most prominent property of neural activity and neural response: neurons and
neural networks behave in an irregular and indeterministic manner both spontaneously, and in
response to series of stimuli. At the single neuron level, variability is observable in practically all
aspects of evoked activity: irregularity of the spike train, trial-to-trial variability in spike counts,
as well as irreproducibility of train structure evoked by identical input series (Faisal et al., 2008;
Yarom and Hounsgaard, 2011). Generally, variability in responses to repeated presentation of a
1
stimulus is a significant constraint on information carrying and processing capacity. However, as
demonstrated in several cases, response variability might be quenched by a variation introduced
to the input itself (Churchland et al., 2010).
At the single neuron level, it was demonstrated that when stimulated with constant input,
the neuronal spike train differ substantially from trial to trial (Bryant and Segundo, 1976;
Mainen and Sejnowski, 1995). In contrast, when stimulated with a fluctuating (filtered white
noise) input, the reproducibility of the spike train is dramatically improved to the point of
perfect repeatability, locking itself to (i.e. entrained by) input fluctuations, reliably encoding its
structure. This key property was reproduced in a stochastic simulation of a Hodgkin-Huxley
neuron, relating it to the properties of the underlying ion channels (Schneidman et al., 1998).
Other works have demonstrated the existence of repeatable spike patterns under different types
of stimuli (Fellous et al., 2004). While these measurements and simulations were limited to
timescale of seconds, it is known that when neuronal activity is observed over extended time
scales, slower effects are recruited and excitability dynamics becomes rich (Marom, 2010). In
a recent work (Gal et al., 2010) we have shown that, indeed, when presented with long (> 1h)
sequences of pulse stimuli, single neuron response dynamics becomes intermittent and irregular,
exhibiting scale-free fluctuations (e.g. with auto-correlation that lacks a characteristic scale)
and transitions between quasi-stable response pattern modes. Given these slower modulatory
processes, it is not obvious that the statistically unstructured random input series, that are
capable of quenching response variation over limited time scales, will effectively entrain response
variability over extended durations (minutes and more). The biophysical mechanism underlying
the capacity of unstructured random input to entrain response variability relies on matching
between time scales of input variations and time scales of the stochastic processes that generate
the action potential.
In contrast, when longer stochastic processes are allowed, they are left
unmatched by the above unstructured input. It is therefore natural to hypothesize that in order
to entrain neuronal response over extended time scales, the variations of the input series must
match the scale-free temporal structure of intrinsic neuronal response dynamics. Indeed, at least
2
at the system level, neuronal response variability is reduced under natural or natural-like sensory
input (Aertsen and Johannesma, 1981; Baddeley et al., 1997; de Ruyter van Steveninck et al.,
1997; Garcia-Lazaro et al., 2006, 2011; Yu et al., 2005). These natural-like signals are often
characterized by long range temporal correlations, and a general scale-free temporal structure
(De Coensel et al., 2003; Simoncelli, 2003; Voss and Clarke, 1975).
In this study we directly measure the impacts of input temporal structure on response
variability over extended time scales in isolated cultured cortical neurons. We show that while
the response of neurons is temporally scale-free, independently of input statistics, entrainment is
maximal when the input itself has a matching, scale-free structure. We also perform analogous
analyses to those of Mainen and Sejnowski, quantifying the reproducibility of spike trains under
different types of input. Here too, natural-like input minimizes the trial-to-trial variability of
the spike train. We conclude that the rich and complex neuronal dynamics enable the neuron
to match its dynamics to that of the natural environment, and that "tuning" to natural input
statistics arises already at the atomic level of neural processing.
Materials and Methods
Culture preparation. Cortical tissues were obtained from newborn (< 24h) rats (Sprague-
Dawley) and dissociated following procedures described in earlier studies (Marom and Shahaf,
2002). The cells were plated directly onto substrate-integrated multi-electrode arrays (MEA)
and developed for a time period of 2-3 weeks before their use. A total amount of approxi-
mately 106 cells was seeded on poly-ethyene-immine (PEI) coated MEAs. The preparations
were kept in Minimal Essential Medium (MEM) supplemented with heat-inactivated horse
serum (5%), glutamine (0.5mM ), glucose (20mM ), and gentamycin (10µg/ml), and main-
tained in an atmosphere of 37◦C, 5% CO2 and 95% air in an incubator as well as during
the recording phases. An array of Ti/Au/TiN extracellular electrodes, 30µm in diameter,
and spaced either 500µm or 200µm from each other (MultiChannelSystems - MCS, Reutlin-
gen, Germany) were used. Synaptic transmission in the network was completely blocked by
3
adding 20µM APV (amino-5-phosphonovaleric acid), 10µM CNQX (6-cyano-7-nitroquinoxaline-
2,3-dione), and 5µM Bicuculline-methiodide to the bathing solution. All experiments were
performed in accordance with the regulations (and under the supervision) of the Technion -
Israel Institute of Technology animal care committee.
Measurements and stimulation. A commercial amplifier (MEA-1060-inv-BC, MCS) with
frequency limits of 150-3000Hz and a gain of x1024 was used. Rectangular 200µs biphasic 600-
800mV voltage stimulation through extracellular electrodes was performed using a dedicated
stimulus generator (STG4004, MCS). In the context of this study, no difference was observed in
the behavior of neurons under current or voltage stimulation. Data was digitized to 16bit using a
USB-ME256 system (MCS). Each recorded channel was sampled at a frequency of 20KHz. One
hour after the addition of synaptic blockers, the stimulation electrode was selected as one evoking
well-isolated spikes with high signal to noise ratio, in as many recording electrodes as possible.
From the selected recording electrodes, voltage traces of 15-20ms post stimulus were collected.
Spike detection was performed off-line by a manual threshold-based procedure. A 3ms long spike
shape was extracted for each response for further noise cleaning and analysis. Stability of spike
shape and activity dynamics criteria were applied in order to validate experimental stability, as
described in (Gal et al., 2010). Random stimulation sequences were generated by modulating
a constant stimulus interval sequence with a noise signal. This noise signal was either a white
Gaussian noise or 1/f Gaussian noise generated by weighting the frequency components of the
Gaussian white noise. In both cases a low cutoff was applied to have a minimum interval of
20ms. The SD of the noisy stimulation interval sequence was set such that its CV will match the
CV of its response constant interval. For example, if a neuron responded to a 100ms constant
interval sequence with an average inter spike interval of 200ms and SD of 40ms, the SD of the
noisy interval sequence was set to 20ms.
Data analysis. Analysis throughout this study was performed on either the spike time series,
or on a smoothened firing rate time series, produced by filtering the spike train with a sliding
4
rectangular window (Figure 1C). Spectral analysis was performed on a binned time series, with a
bin size of 1s. The Power Spectrum Density (PSD) was estimated using a modified Periodogram.
The Allan variance, which is commonly used to identify fractal point processes (Lowen and Teich,
1996, 2005) was calculated by binning the time series with different bin sizes T . For each binned
time series, the Allan variance is defined as A(T ) = <ZT (k)−ZT (k+1)>2
, where ZT is the binned
2·<ZT >
series. Another measure, widely used to characterize the temporal statistics of long memory
and non stationary time series, is Detrended Fluctuation Analysis (DFA, Peng et al., 1995), in
which the fluctuations (in terms of mean square error) around a piecewise linear fit to the time
series are quantified, for different segment durations. Since both the Periodogram and the Allan
variance are can be regarded as power law for the purpose of this work, power law functions
were fitted to the tail of these curves, and their exponents were used as descriptive statistics
(see Gal et al., 2010).
The similarity between two spike trains was assessed with two distance measures: (1) A
correlation based metric, defined as one minus the correlation between binned time series (Za,Zb):
d = 1 − corr(Za, Zb). This is a rate-based measure, which is insensitive to temporal features of
the spike train below resolution dictated by the bin size. (2) The Victor-Purpura spike train
metric (Toups et al., 2011; Victor, 2005; Victor and Purpura, 1997), which defines the distance
between the two spike trains as the minimal cost of transforming one into the other. Briefly, a
spike train is modified by a combination three possible steps: inserting or deleting a spike, with
a cost of +1, and moving a spike in time, with a cost of q·dt , where q is a temporal resolution
parameter. The value of q sets the sensitivity of the metric to fine temporal features, and is set
here to a default value of q = 1s. The dependence of the results on the value of q is shown in
Figure 4. The Victor-Purpura metric was calculated using Matlab code downloaded from the
website of Jonathan Victor.
5
Results
Unexplained response variability is minimized by scale-free input
In this paper we investigate the response properties of individual neurons, independent of synap-
tic and network effects. To that end, experiments are conducted on cultured cortical neurons,
functionally isolated from their network by means of pharmacological block of both glutamatergic
and GABAergic synapses (see Gal et al., 2010, and Materials and Methods). Individual neurons
are stimulated with sequences of short, identical extracellular electrical pulses. In response to a
single pulse, a neuron either responds by emitting a spike, or fails to do so (Figure 1A); neuronal
responses are monitored by extracellular recording electrodes. As previously reported (Gal et
al., 2010), when repeatedly stimulated over extended durations (minutes and more) with a low
(∼1Hz) stimulation rate, a neuron responds to each and every stimulation pulse (1:1 mode).
As stimulation rate increases, the excitability of the neuron declines, and at a certain, neuron-
specific critical stimulation rate, the 1:1 response mode breaks, and the neuron exhibits rich and
complex response dynamics (Figure 1B).
Here, we study the properties of response dynamics evoked by three statistically different
regimes of stimulation series: (i) constant interval regime, (ii) white noise regime in which the
interval series is modulated by a Gaussian white noise process, and (iii) scale-free regime in
which the intervals are modulated by a 1/f β process, with parameter β = 1, which is more
representative of natural sensory input (De Coensel et al., 2003; Simoncelli, 2003; Voss and
Clarke, 1975) and similar to activity properties of cortical neurons in-vivo (Bhattacharya et al.,
2005; Lowen et al., 2001; Lowen and Teich, 1993, 1996; Teich et al., 1997). The term scale-free
is used here to designate a timeseries with an autocorrelation that decays slowly, usually as a
power-law without a typical scale. White noise on the other hand, which has a delta-function
autocorrelation, is regarded as a zero scale signal. The choice of β = 1 is typical to a wide range
of natural signals, in natural environments and in biological systems. As will be shown, the
results of this work are not sensitive to the exact value of β, consistent with previous studies
(Garcia-Lazaro et al., 2006). Interval sequences in all stimulation regimes were normalized to
6
have the same mean and standard deviation (the latter is applicable only to the second and
third regime). The mean stimulation rate (the reciprocal of the mean stimulation interval)
was set to a high enough value to drive the neuron beyond its critical point, leading to response
failures and intermittency (Gal et al., 2010). The interval standard deviation (SD) was chosen to
approximately match the intrinsic SD of the response to constant interval stimulation. Figure 1D
shows examples of extracts from the three stimulation regimes, as well as their autocorrelation
functions and the Power Spectral Densities (PSD).
Our analysis starts by stimulating neurons with long sequences (> 1h) of each of the stim-
ulation regimes. Figure 2A shows an example of the response of the same neuron to the three
regimes. It is immediately obvious that the three input regimes did not cause a significant dif-
ference in the statistical properties of the responses. This is formally shown in the plots of the
PSD, Allan variance, detrended Fluctuation Analysis (DFA), rate histograms and inter-spike
interval (ISI) histograms (Figure 2B-F). Clearly, the macroscopic properties of the evoked spike
trains are the same under all stimulation regimes: the neuron exhibits scale-free dynamics, char-
acterized by power law statistics, in accordance with previously published analysis (Gal et al.,
2010). In order to confirm the insignificance of the differences between the different statistical
measures under the three regimes, a control experiment was performed, in which 4 neurons were
stimulated and recorded, and each stimulation block was repeated 10 times. For each statistical
analysis presented in Figure 2B-F, the relevant parameter was estimated independently from
the response to each repetition. The value ranges of these parameters from a single neuron
are depicted in the insets of each panel, and show that indeed the temporal statistics are the
same under the three regimes. We emphasize that the claim made here is not that the specific
model chosen for each statistics is the correct one (i.e. that the PSD is an exact power law),
rather that these fits are good enough representative shapes, useful for comparing the population
statistics. The above results imply that the various membrane and cellular processes underlying
the stochastic response fluctuations are effectively insensitive to the statistical structure of the
input regime.
7
1/f -type response statistics can be interpreted as a modulation of neuronal excitability by a
cascade of oscillating processes at various time scales. An oscillator can be entrained (i.e. phase-
locked) by a driving stimulus at a frequency that matches the oscillator's natural frequency, and
with a magnitude proportional to the oscillation amplitude. This suggests that the 1/f intrinsic
fluctuations could be entrainable by a matching 1/f stimulation.
Figure 3 demonstrates this effect: The stimulation and response of a neuron are plotted
in two timescales, for the three stimulation regimes. On the short timescale (panel A, 5s bin
size), the response in the white noise and scale-free regimes nicely follows the stimulation, while
in the constant interval regime, the input has no bearing on timing of the output fluctua-
tions. On a longer timescale (panel B, bin size of 100s), the white noise input is practically
flattened, becoming constant; as a result, it fails to entrain the response. In contrast, in the
scale-free regime entrainment is evident throughout. Figure 3C shows a scatter plot of the
input and output rates of the neuron, calculated with a 5s bin size. As expected from the
data of Figure 2, the input and output ranges of the white-noise (red) and scale-free (green)
are practically identical. However, the correlation in the scale-free data is much higher, as
expressed in the reduced variability around the linear trend line. This indicates that indeed the
scale-free fluctuations in neuronal dynamics are best entrained by scale-free input. For 13 of
the 17 recorded neurons, the input output correlation coefficient significantly increases under
scale-free input regime, compared to white noise regime. A Wilcoxon signed-rank test yields
p < 0.01 for the effect of the input regime on input/output correlation. Thus, under scale-free
input, there is no change in the amount of variability in neuronal response, but more of it is
explained by the input. Figure 3D shows a scaled correlation analysis: The correlation between
input and output is calculated for responses in white noise and scale-free regimes on different
timescales. The correlation in a given timescale T is calculated by smoothing the stimulation
and response series with a rectangular window of width T , and subtracting from it a version
of the series, smoothed with a window of size 25T . This effectively band-passes the time series
around T . While the correlation in the short timescale regime is similar for the white noise and
8
scale free regimes (the apparent increase for white noise is non typical, a Wilcoxon signed-rank
test yields insignificant difference), for longer timescales it decreases in the case of white noise
compared to scale free (p < 0.01, Wilcoxon signed-rank test).
The above results show that there is a significant increase in correlation upon change from
white noise (i.e. β = 0) to scale free (β = 1) input. However, there is nothing unique about
β = 1: other types of inputs might serve just as well. In order to asses the sensitivity of the
increased input-output correlation to the value of β, neurons (n=31) were exposed to 9 stim-
ulation blocks, each characterized by a different value of β, ranging from 0 to 2, and lasting
70 minutes. For each block, the correlation between input and response was computed, and
compared with the correlation for β = 0. The population statistics of this correlation ratio are
depicted in Figure 3E, and show a clear concave shape, with a peak in the mid-range values, and
a decrease toward lower and higher values of β. A similar peak is observed when plotting the
distribution of preferred exponents (i.e. the exponents which results in the strongest correlation
for each neuron, Figure 3F). These peaks around β = 1, however, are wide, suggesting that the
variability between neurons is considerable, and that inputs characterized by a wide range of
exponents are equally effective in entraining neuronal responses.
The decrease of the correlation for input with large exponents, which are dominated by slow
oscillations, is important for the understanding of this phenomena. Figure 3G shows extracts
from the response to 1/f stimulation (left) and 1/f 2 stimulation (right). It is easy to see how,
for 1/f 2, the faster fluctuations remain un-entrained, while slower oscillations are nicely locked
by modulations in the input.
Neuronal response repeatability is maximized by scale-free input
Another functionally relevant aspect of the entrainment described above, concerns response
repeatability. It is well established that in spite of the extensive variability of neuronal responses,
spike train structure is repeatable when the input is fluctuating, both at the single neuron and
sensory system levels (Churchland et al., 2010; de Ruyter van Steveninck et al., 1997; Mainen
9
and Sejnowski, 1995). In the following set of experiments we ask whether a scale-free stimulation
regime enhances repeatability in general, and over long time scales in particular. To this aim
we stimulated neurons with 10 repetitions of the same input sequence under each of the three
stimulation regimes: constant intervals, white noise and scale-free. Each sequence lasted 10
minutes, separated by a 10 minute break. Figure 4A shows responses of one neuron to ten
identical sequences under each stimulation regime.
It is immediately obvious that although
some reproducibility exists under the white noise regime, the reproducibility of responses to
scale-free sequences is much higher. This is a direct consequence of the previous section analysis:
if responses are more correlated to the input, they will be more correlated between themselves
under repetitions of the same input sequence. This observation is quantified using two kinds
of spike train similarity measure: a rate-based measure (the correlation distance between spike
histograms, calculated with 1s bins; see Methods section), and a time-based measure (the Victor-
Purpura distance, Toups et al., 2011; Victor, 2005; Victor and Purpura, 1997). Figure 4B
depicts the pairwise distance matrices of the responses of Figure 4A, calculated using these two
metrics. Responses to white noise stimulation are significantly more reproducible than responses
to constant input; this is in agreement with (Mainen and Sejnowski, 1995). But as expected from
the results presented above (Figure 3) the responses to the scale-free sequences are significantly
more reproducible than those of white noise input. The purple lines of Figures 4C and 4D depict
the mean and SD of the values in the distance matrices shown in Figure 4B. Of 14 recorded
neurons, 11 neurons showed significant improvement in reproducibility for scale-free sequences
compared to constant interval and white noise, as quantified by the two metrics (there were no
cases of disagreement). The grey lines depict the trends of mean distances calculated for all of
the 14 recorded neurons. A Wilcoxon signed-rank test for the mean pair-wise distances shows an
overall significant effect for the input regime on response repeatability (p < 0.01 for both white
noise vs. constant and for scale-free vs. white noise, for both metrics).
The results are insensitive to the choice of the temporal parameter q of the spike train metric
(Figure 4F), which may point to the lack of characteristic scale in this phenomenon.
10
The above enhanced reproducibility does not stem from a decreasing spike rate. Next to the
preservation of the mean stimulation rate in all sequences, the spike rate itself does not decrease
for scale-free input (see Figure 4E). For all of the 14 cells recorded, there was no significant
decrease in firing rate under the scale-free regime compared with constant and white noise.
Since the scale-free input itself is structured, it is expected that even a "Bernoulli" neuron
that has a constant probability of response to a pulse stimulation, will have some reproducibility
of its output spike train. Figure 5A shows a comparison between metric analyses on actual
responses (grey), and on a surrogate Bernoulli neuron (purple) that responds with a constant
probability (set to the mean response probability of the real neuron). As expected, the responses
of the simulated neuron are indeed more reproducible for scale-free input, but not as consistent
as the real neuron. It is also possible to construct a more detailed neuronal response curve, which
takes into account the dependency of the response probability on the last interval between stimuli
(an example for such a curve is given in Figure 5B). Interestingly, the curves of the white noise
and scale-free regimes substantially differ, pointing to a strong history dependence in response
probability. As shown in Figure 5A, the resulting metric analyses (orange) behave more like the
real neuron when compared to the Bernoulli neuron, yet cannot account for the entire effect.
It is reasonable to believe that one might construct a response model that takes into account
deeper history of the stimulation and response sequences, to produce better fitting. It should be
emphasized though, that while these neuronal response models do reproduce the repeatability
effect to a significant extent, they do not reproduce the intrinsic scale-free fluctuations, and can
not be considered as successful explanatory models.
A scale-free input is characterized by the abundance of relatively long "breaks" in stimulation,
or long periods with low stimulation rate, enabling recovery of internal processes from previous
activations. As repeatedly shown over the past 15 years, the longer a neuron is exposed to
repeated activations, longer recovery times are required (Ellerkmann et al., 2001; Fairhall et al.,
2001; Lundstrom et al., 2008; Marom, 2009; Toib et al., 1998). A scale-free input statistics is
inherently matched to such a mechanistic context: elongating a scale-free input series naturally
11
gives rise to longer breaks, hence allowing for stabilization on every scale. An illustration of this
property is provided by reanalyzing the data of Figure 4 over blocks of increasing lengths. The
results are summarized in Figures 5C and 5D, showing that the accumulation of variability (or
divergence of response) with increasing block size (1-10 minutes range) is significantly slower
than its accumulation in responses to other stimulation regimes.
Discussion
In this paper we have shown how a natural-like, scale-free input entrains fluctuations of single
neuron responses over extended timescales. We have demonstrated this property by comparing
neuronal responses in three different stimulation regimes: constant interval, white noise and
scale-free. In the case of the scale-free regime, the correlation between the input and the re-
sponse is significantly higher, and the repeatability of response is considerably enhanced. These
characteristics are stable over long, practically unlimited durations. While the results do show
a preference to mid-range values of β (around 1), there is nothing special about the exact value;
what seems to be important is that the entrainment decreases when the slow frequency compo-
nent in the input becomes either too dominant (large β) or marginal (low β).
It has long been acknowledged that responsiveness of neural systems is optimized to the
ranges of statistics found in natural inputs (Aertsen and Johannesma, 1981; Baddeley et al.,
1997; de Ruyter van Steveninck et al., 1997; Garcia-Lazaro et al., 2006, 2011; Yu et al., 2005).
Here we show that preference to natural statistics is not limited to large-scale neural systems;
rather, it goes all the way down, to the atomic level of neural organization, namely the single,
isolated neuron.
At the shorter timescales, the entrainment of neuronal fluctuations by white noise input is
explained by the fast stochastic processes underlying the generation of action potentials (Schnei-
dman et al., 1998). There, a variation in input allows for recovery of inactivation processes,
unlike the case of a constant input that drives the neuron to operate around the limit of channel
availability threshold, making it highly sensitive to stochastic events. It is reasonable to assume
12
that a similar explanation would also be appropriate over extended timescales: a long break
(or a period of low-rate stimulation) enables recovery of slow processes, in contrast to constant
input (or shortly correlated input) that drives these processes to a highly stochastic operation
point. It is natural to assume that such processes include slow inactivation properties of the ionic
channels themselves (Ellerkmann et al., 2001; Marom, 2009; Soudry and Meir, 2010; Toib et al.,
1998), or other cellular modulatory processes (e.g., protein phosphorylation, protein synthesis
and metabolic cycles).
From the more abstract, functional point of view, when the temporal structure of the input
is relatively dull, it can only entrain a narrow range of cellular processes underlying neuronal
dynamics. Under these conditions, a large fraction of the response variability is tagged "un-
explained". However, when the input is temporally rich, it matches the temporal manifold
complex structure of the intrinsic dynamics, and the former "unexplained" variability becomes
information-carrying. Thus, a neuron can be viewed as a collection of entangled information
channels, distributed over a continuum of timescales. In this picture, information transfer is
maximized when there is information to be transferred on any given scale.
Acknowledgments
The authors thank Danni Dagan, Danny Eytan and Avner Wallach for helpful discussions and
insightful input, and Eleonora Lyakhov and Vladimir Lyakhov for invaluable technical assistance
in conducting the experiments. The research leading to these results has received funding from
the European Unions Seventh Framework Program FP7 under grant agreement 269459 and was
also supported by a grant of the Ministry of Science and Technology of the State of Israel and
MATERA grant agreement 3-7878.
References
Aertsen A, Johannesma P (1981) A comparison of the spectro-temporal sensitivity of auditory
neurons to tonal and natural stimuli. Biological Cybernetics 42:145 -- 156.
13
Baddeley R, Abbott LF, Booth MC, Sengpiel F, Freeman T, Wakeman EA, Rolls ET (1997)
Responses of neurons in primary and inferior temporal visual cortices to natural scenes. Proc.
R. Soc. Lond. B 264:1775 -- 83.
Bhattacharya J, Edwards J, Mamelak aN, Schuman EM (2005) Long-range temporal correlations
in the spontaneous spiking of neurons in the hippocampal-amygdala complex of humans.
Neuroscience 131:547 -- 55.
Bryant H, Segundo J (1976) Spike initiation by transmembrane current: a white-noise analysis.
The Journal of Physiology pp. 279 -- 314.
Churchland MM, Yu BM, Cunningham JP, Sugrue LP, Cohen MR, Corrado GS, Newsome WT,
Clark AM, Hosseini P, Scott BB, Bradley DC, Smith Ma, Kohn A, Movshon JA, Armstrong
KM, Moore T, Chang SW, Snyder LH, Lisberger SG, Priebe NJ, Finn IM, Ferster D, Ryu
SI, Santhanam G, Sahani M, Shenoy KV (2010) Stimulus onset quenches neural variability:
a widespread cortical phenomenon. Nature Neuroscience 13:369 -- 78.
De Coensel B, Botteldooren D, De Muer T (2003) 1/f Noise in Rural and Urban Soundscapes.
Acta Acustica united with Acustica 89:9.
De Ruyter Van Steveninck RR, Lewen GD, Strong SP, Koberle R, Bialek W (1997) Repro-
ducibility and variability in neural spike trains. Science 275:1805 -- 1808.
Ellerkmann RK, Riazanski V, Elger CE, Urban BW, Beck H (2001) Slow recovery from inacti-
vation regulates the availability of voltage-dependent Na(+) channels in hippocampal granule
cells, hilar neurons and basket cells. The Journal of Physiology 532:385 -- 97.
Fairhall AL, Lewen GD, Bialek W, de Ruyter Van Steveninck RR (2001) Efficiency and ambi-
guity in an adaptive neural code. Nature 412:787 -- 792.
Faisal AA, Selen LPJ, Wolpert DM (2008) Noise in the nervous system. Nature Reviews
Neuroscience 9:292 -- 303.
14
Fellous J, Tiesinga P, Thomas P, Sejnowski T (2004) Discovering spike patterns in neuronal
responses. The Journal of Neuroscience 24:2989 -- 3001.
Gal A, Eytan D, Wallach A, Sandler M, Schiller J, Marom S (2010) Dynamics of excitability over
extended timescales in cultured cortical neurons. The Journal of Neuroscience 30:16332 -- 42.
Garcia-Lazaro JA, Ahmed B, Schnupp JWH (2006) Tuning to natural stimulus dynamics in
primary auditory cortex. Current Biology 16:264 -- 271.
Garcia-Lazaro Ja, Ahmed B, Schnupp JWH (2011) Emergence of Tuning to Natural Stimulus
Statistics along the Central Auditory Pathway. PLoS ONE 6:e22584.
Lowen SB, Ozaki T, Kaplan E, Saleh BE, Teich MC (2001) Fractal features of dark, maintained,
and driven neural discharges in the cat visual system. Methods 24:377 -- 394.
Lowen SB, Teich MC (1993) Fractal renewal processes generate 1/f noise. Physical Review
E 47:992 -- 1001.
Lowen SB, Teich MC (1996) The periodogram and Allan variance reveal fractal exponents
greater than unity in auditory-nerve spike trains. The Journal of the Acoustical Society of
America 99:3585 -- 3591.
Lowen SB, Teich MC (2005) Fractal-based point processes Wiley-Interscience, Hoboken, N.J.
Lundstrom BN, Higgs MH, Spain WJ, Fairhall AL (2008) Fractional differentiation by neocor-
tical pyramidal neurons. Nature Neuroscience 11:1335 -- 1342.
Mainen ZF, Sejnowski TJ (1995) Reliability of spike timing in neocortical neurons. Sci-
ence 268:1503 -- 1506.
Marom S (2009) Adaptive transition rates in excitable membranes. Frontiers in Computational
Neuroscience 3:2.
Marom S (2010) Neural timescales or lack thereof. Progress in Neurobiology 90:16 -- 28.
15
Marom S, Shahaf G (2002) Development, learning and memory in large random networks of
cortical neurons: lessons beyond anatomy. Quarterly Reviews of Biophysics 35:63 -- 87.
Peng CK, Havlin S, Stanley HE, Goldberger AL (1995) Quantification of scaling exponents and
crossover phenomena in nonstationary heartbeat time series. Chaos 5:82.
Schneidman E, Freedman B, Segev I (1998) Ion channel stochasticity may be critical in deter-
mining the reliability and precision of spike timing. Neural Computation 10:1679 -- 1703.
Simoncelli EP (2003) Vision and the statistics of the visual environment. Current Opinion in
Neurobiology 13:144 -- 149.
Soudry D, Meir R (2010) History-dependent Dynamics in a Generic Model of Ion Channels -
an Analytic Study. Frontiers in Computational Neuroscience 4:1 -- 19.
Teich MC, Heneghan C, Lowen SB, Ozaki T, Kaplan E (1997) Fractal character of the neu-
ral spike train in the visual system of the cat. Journal of the Optical Society of America
A 14:529 -- 546.
Toib A, Lyakhov V, Marom S (1998) Interaction between duration of activity and time course
of recovery from slow inactivation in mammalian brain Na+ channels. The Journal of Neu-
roscience 18:1893 -- 1903.
Toups JV, Fellous JM, Thomas PJ, Sejnowski TJ, Tiesinga PH (2011) Finding the event struc-
ture of neuronal spike trains. Neural Computation 23:2169 -- 208.
Victor JD (2005) Spike train metrics. Current opinion in neurobiology 15:585 -- 92.
Victor JD, Purpura KP (1997) Metric-space analysis of spike trains: theory, algorithms and
application. Network: Computation in Neural Systems 8:127 -- 164.
Voss RF, Clarke J (1975) 1/f noise in music and speech. Nature 258:317 -- 318.
Yarom Y, Hounsgaard J (2011) Voltage fluctuations in neurons: signal or noise? Physiological
Reviews 91:917 -- 29.
16
Yu Y, Romero R, Lee T (2005) Preference of Sensory Neural Coding for 1/f Signals. Physical
Review Letters 94:18 -- 21.
17
Figure captions
Figure 1. Data analysis and stimulation regimes. (A) Examples of voltage traces recorded
following several stimulation pulses, delivered at 15Hz. In response to such a pulse, a neuron
sometimes emits a spike (blue traces) or fail to do so (red traces). Stimulation pulses are 400µs
wide and start at t = 0. (B) Color representation of response traces. Each line represents
a single response trace. Responses to consecutive stimulation pulses, delivered at 15Hz, are
ordered top to bottom. Voltage is color coded, red for high voltage and blue for low. It can be
seen that spikes are fired in response to some of the stimulations, in a seemingly random and
complex manner. (C) Examples for stimulation sequences, each of 1h length. The signals shown
are stimulation rates, the reciprocals of the stimulation interval series. Examples are given of
the three stimulation regimes (see Results section): constant interval (upper panel, blue), white
noise (middle panel, red) and scale-free (bottom panel, green). (D) The autocorrelation function
for the white noise (red) and scale free (green) sequences. (E) Power Spectral Density (PSD)
for the white noise (red) and scale free (green) sequences.
Figure 2. Response statistics under the three stimulation regimes. Main panels show raw
data and statistical measures of a typical neuron to the three regimes (1h of stimulation for
each sequence): constant interval (blue), white noise (red) and scale-free (green). Insets show
results from a neuron subjected to a control experiment where the significance of the differences
in responses to each of the regimes are assessed by 10 repetitions of each stimulation block. (A)
Extracts from the firing rate of the neuron in response to the three stimulation types (8min
length, 1s bin width). (B) Histograms of firing rate values from one repetition, calculated with
5s bin size. The inset shows the distribution of mean firing rate value from each of the 10
repetitions in the control experiment. Filled box represents the 25%-75% range, and whiskers
extend to the extreme values. (C) Histograms of ISI values from one repetition. The inset
plots the range of values of the ISI coefficient of variation under 10 repetitions. (D) Response
PSD from one repetition, on double logarithmic axes. The inset depicts the range of exponent
18
values for a power law fit to the low frequency tail of the PSD. (E) Detrended Fluctuation
Analysis (DFA, see Methods) of the responses from one repetition. The inset depicts the range
of exponent values for a power law fit to the Fluctuation curve. (F) Allan variance of the
responses of the responses from one repetition. The inset depicts the range of exponent values
for a power law fit to the linear tail of the curves.
Figure 3. Input and output correlations under the three stimulation regimes. (A) Extracts
from the response rates of a neuron to long stimulation (1 h) under the three regimes. Responses
and stimulation are binned with a 5s bin size, 30 bins are shown. Mean is subtracted to have
the input and output aligned. The responses under white noise and scale-free regimes follow
closely the stimulation (black line), while for constant interval the variability is freely running.
(B) Extracts from the same experiment, using a 100s bin size, 30 bins are shown. The response
in the scale-free regime is still locked to stimulus fluctuations, while input fluctuations in the
white noise regime are substantially diminished, therefore unable to lock response variability.
(C) A scatter plot of the stimulation rate against response rates, calculated with a 5s bin size.
White noise input in red, scale-free input in green. Mean is removed for visual clarity, and both
axes are in standardized units. While the marginal distributions for both inputs are similar, the
correlation in the scale-free case is significantly higher. (D) Scaled correlation analysis. The
correlation between input and output rates are calculated in different timescales, for white noise
and scale-free regimes. The correlation in a given timescale T is calculated by smoothing the
series with a rectangular window of width T , and subtracting a version of the series smoothed
with a window of size 25T , effectively band-passing the time series. (E) Dependence of the
input output correlation on the exponent β of the input. Neurons (n=31) were stimulated with
blocks of 70 minutes duration with exponent ranging from 0 to 2. The input output correlation
(cβ) was calculated from the responses of each block, using a 1s time bin. The graph shows the
population statistics of the ratio cβ/c0 for each β, using box and whisker plot. The red horizon-
tal line marks the median, the box marks the lower and upper quartiles and the whiskers the
19
range of data values. Outliers (values outside the range of 3 SD units from the average value)
are marked with red points. (F) Distribution of preferred exponent values (the exponents which
results in the strongest correlation for each neuron). (G) Extracts comparing the entrainment
for β = 1 (left) and β = 2 (right), from a single neuron, demonstrating lack of entrainment for
the fast fluctuations for the latter case. Traces are smoothed with a 2s rectangular window, time
scale is identical for the two extracts.
Figure 4. Repeatability of neuronal response. (A) The firing rates of a neuron under 10
identical stimulation sequences, under the three regimes. Responses to the constant interval
stimulation show no reproducibility at all. Responses to the white noise input show repro-
ducibility on short timescales, as can be seen in the inset. At longer timescales there is again
no response repeatability. The responses to the scale-free sequence are the most reproducible,
and lock to the input on many timescales. Stimulation rates are plotted in black for the three
plots, but are normalized to have the mean and SD of the responses for clear visualization. (B)
Pairwise distance matrices between responses to repetitions of the same stimulation sequence,
calculated according to correlation metric (left) and the Victor-Purpura (VP) distance with tem-
poral parameter q = 1 (right, see Methods section for details). Distance values are color coded,
and for each metric the color scale was normalized to the maximal value. Diagonal pixels were
whitened for visual clarity. (C) The purple line designates the mean and SD of the pairwise
distances depicted in B, for the correlation metric, quantifying the change in repeatability under
the three regimes. Also shown are the means for all 14 recorded neurons (grey for those which
showed significant improvement for scale-free input regime, brown for those which didn't). (D)
Same as C, but for VP distance. (E) Mean and SD of the average firing rate per trial for the
neuron of A (purple). The firing rate significantly increases for scale-free input for this neuron.
While this is not generally true across the population of neurons recorded, the firing rate never
significantly decreases for scale-free input. (F) The effect of the temporal scale q on the results
for the spike train metric. q is varied from 0.25s (light grey) to 10s (dark grey). While the
20
typical values differ, the effect exists for any choice of q, as can be expected from the lack of
typical scale for the phenomena. The red curve is for q = 1.
Figure 5. (A) Metric analysis on surrogate data. As explained in the Results section, the
purple curve is metric analysis results, as in Figure 4, on surrogate data generated by applying
a Bernoulli response model with a mean equaling the average response of the neuron, using VP
distance with q = 1. The orange curve is results of the same analysis applied to data generated
by a Bernoulli response model conditioned on the last interval, according to the curve in B. The
grey curve is the actual experimental results for this neuron. (B) The response probability of
a neuron, conditioned on the last stimulation interval, for white noise input (red, lower curve)
and scale-free input (green, upper curve). Both curves, as expected, are mostly increasing. The
difference between the curves implies the dependence of the probability on history longer than
the last interval. Since the scale-free input contains correlations between intervals, its response
curve differs from the white noise curve. (C) The effect of block length. The metric analysis
of Figure 4 was repeated with various block sizes, i.e. with analysis performed on the first T
seconds of each repetition, using the correlation metric. Data was taken from the onset of the
stimulation block, including the transient phase. This analysis shows how the responses to the
constant interval input and white noise input are drifting apart relatively rapidly (mean distance
increases), while the responses to the scale-free input are forced together by the input dynamics,
and show only a slow and moderate gain in distance. (D) Block length analysis as in C, using
VP distance . It should be pointed that both the metrics used here are effectively normalized to
the input length, in contrast to metrics like Euclidean and others which are extensive in input
length.
21
Figure 1
22
51015−200−150−100−50050100time (ms)v (µV)interval −1 (1/s) −3003000.51time lag (s)corr0.010.11101001000frequency (Hz)psd3ms1minABCDE Figure 2
23
101001000110100bin size (s)F0.60.8051015sr (spk/s)Prob456012ISI (s)Log Prob1.61.822.2.001.01.11.1110100f (Hz)PSD−0.8−0.61101001000110100bin size (s)Allan Variance0.811.22minABCDEF Figure 3
24
−4−2024−4−2024stim ratespike rate05010000.20.40.60.8timescale (s)correlation0.250.751.251.750.60.811.21.41.61.82βcorrelation ratio0.250.751.251.750246810βcount10min30s1minABCDEFG Figure 4
25
cwsf0.51dcwsf0.1.2.3.4dcwsf468sr (spk/s)cwsf.1.3.5.7Dspt1min0100.4ABCDEF Figure 5
26
cwsf.15.20.25.30d0.1.2.3.6.9Interval (s)Spike Prob0120240360480600.2.4.6.81T (s)d0120240360480600.2.4T (s)dABCD |
1301.5148 | 1 | 1301 | 2013-01-22T11:17:38 | Limit-cycle-based control of the myogenic wingbeat rhythm in the fruit fly Drosophila | [
"q-bio.NC",
"nlin.AO"
] | In many animals, rhythmic motor activity is governed by neural limit cycle oscillations under the control of sensory feedback. In the fruit fly Drosophila melanogaster, the wingbeat rhythm is generated myogenically by stretch-activated muscles and hence independently from direct neural input. In this study, we explored if generation and cycle-by-cycle control of Drosophila's wingbeat are functionally separated, or if the steering muscles instead couple into the myogenic rhythm as a weak forcing of a limit cycle oscillator. We behaviourally tested tethered flying flies for characteristic properties of limit cycle oscillators. To this end, we mechanically stimulated the fly's gyroscopic organs, the halteres, and determined the phase relationship between the wing motion and stimulus. The flies synchronized with the stimulus for specific ranges of stimulus amplitude and frequency, revealing the characteristic Arnold tongues of a forced limit cycle oscillator. Rapid periodic modulation of the wingbeat frequency prior to locking demonstrates the involvement of the fast steering muscles in the observed control of the wingbeat frequency. We propose that the mechanical forcing of a myogenic limit cycle oscillator permits flies to avoid the comparatively slow control based on a neural central pattern generator. | q-bio.NC | q-bio | Limit-cycle-based control of the myogenic wingbeat rhythm
in the fruit fly Drosophila
Jan Bartussek1, A. Kadir Mutlu1, Martin Zapotocky2,3, Steven N. Fry4
1 Institute of Neuroinformatics, University of Zurich and ETH Zurich, Winterthurerstrasse 190, 8057 Zurich, Switzerland
2 Institute of Physiology, Academy of Sciences of the Czech Republic, Videnska 1083, 14220 Prague 4, Czech Republic
3 Institute of Biophysics and Informatics, First Faculty of Medicine, Charles University, Salmovska 1, 12000 Prague 2, Czech Republic
4 SciTrackS GmbH, Lohzelgstrasse 7, 8118 Pfaffhausen, Switzerland
Author Generated Postprint, published in J R Soc Interface 10:20121013. http://dx.doi.org/10.1098/rsif.2012.1013
Electronic Supplementary Material available under
http://rsif.royalsocietypublishing.org/content/suppl/2012/12/28/rsif.2012.1013.DC1.htm l
Author for correspondence: Jan Bartussek, e-mail: [email protected]
Abstract
In many animals, rhythmic motor activity is governed by neural limit cycle oscillations under the control of sensory
feedback. In the fruit fly Drosophila melanogaster, the wingbeat rhythm is generated myogenically by stretch -activated
muscles and hence independently from direct neural input. In this study, we explored if generation and cycle-by-cycle control
of Drosophila’s wingbeat are functionally separated, or if the steering muscles instead couple into the myogenic rhythm a s a
weak forcing of a limit cycle oscillator. We behaviourally tested tethered flying flies for characteristic properties of limi t
cycle oscillators. To this end, we mechanically stimulated the fly’s ‘gyroscopic’ organs, the halteres, and determined the
phase relationship between the wing motion and stimulus. The flies synchronized with the stimulus for specific ranges of
stimulus amplitude and frequency, revealing the characteristic Arnol´d tongues of a forced limit cycle oscillator. Rapid
periodic modulation of the wingbeat frequency prior to locking demonstrates the involvement of the fast steering muscles in
the observed control of the wingbeat frequency. We propose that the mechanical forcing of a myogenic limit cycle oscillator
permits flies to avoid the comparatively slow control based on a neural central pattern generator.
1.
Introduction
Locomotion plays a profound role in essential animal behaviours, such as foraging for food, finding mates and evading
predation. In most cases, animals rely on the rhythmic actuation of appendages, such as legs, fins and wings for locomotion.
While providing reaction forces for forward propulsion or to remain airborne, these same structures typically also serve as
control surfaces to achieve stability of motion, as well as perform impressive manoeuvres.
In many animals, the rhythmic activity underlying locomotion originates in neural central pattern generators (CPGs), whose
cyclic activity is modulated by reafferent sensory feedback. CPGs represent a ubiquitous neural mechanism by which animals
generate and control rhythmic activity for various bodily functions, such as respiration, chewing and limb actuation [1].
CPGs have been successfully modelled using oscillators with a stable limit cycle (limit cycle oscillators , LCOs) [2-4]. The
intrins ic properties of such nonlinear oscillators are illustrated in Fig. 1 by example of the Van der Pol oscillator [5]. The
oscillation of position in time (black trace in Fig. 1A) represents a limit cycle orbit in the position -velocity phase space (black
curve in Fig. 1B). Limit cycle orbits are asymptotically stable, i.e. the system converges to the limit cycle following a
transient perturbation. In the example shown, the delivery of pulse perturbations during three consecutive cycles (Fig. 1A)
temporarily affects the periodic activity (green curve), which subsequently returns to its stable limit cycle (Fig. 1B).
Depending on the cycle phase, in which such a forcing is applied, the frequency of the LCO is transiently affected [6]. In the
present case, the delivery of the pulses leads to a frequency increase (compare green and dashed lines in Fig. 1 A).
In such a way, precisely timed input can control the rhythmic pattern of LCOs efficiently and rapidly – a phenomenon that
has also been experimentally demonstrated in CPGs [7, 8]. Computational modelling has shown that phase-dependent
reafferent feedback from the periphery to the CPG can ensure robustness of the rhythm in presence of environmental
perturbations [9-11]. Such limit-cycle-based control therefore provides a conceptual framework in which to understand how
animals generate and control complex motor patterns robustly and efficiently [12].
Figure 1. A) Time course
of a van der Pol oscillator perturbed by external force
. The equation of motion is
with
. Blue trace: perturbation
. Black trace:
before and after perturbation. Green trace:
during perturbation. Dashed trace: continuation of unperturbed oscillation (perturbation not switched on,
). B) Limit cycle of
van der Pol oscillator. The black and green traces from A) are replotted in the phase space (velocity
vs. coordinate
); increasing
time
corresponds to clockwise rotation.
CPGs were originally discovered in locusts [13], in which CPGs provide the neural rhythm for the activity of the thoracic
flight muscles giving rise to the wing motion. Synaptic input from the sensory system modulates these patterns to achieve
flight control [14-16]. An entirely different actuation mechanism is found in flies, whose highly specialized flight apparatus
likewise serves as a model to explore fundamental control mechanisms. The sophisticated flight control abilities of flies are
exemplified by the tiny fruit fly Drosophila melanogaster , which performs sharp turning manoeuvres within a fraction of a
second (around 50 milliseconds, or 10 wingbeats) [17]. Unlike in locusts, the fast and power demanding wingbeat of fruit
flies arises myogenically, without direct neural control [18]: Antagonistic sets of stretch act ivated muscles bring the thorax
and with it the wings into resonant oscillation at around 200 Hz. At the time scale of single wing strokes, the wing motion i s
finely modulated by minuscule steering muscles, which insert directly at the sclerit es of the wing articulation and are under
immediate neural control [19-22]. The activity of these direct steering muscles depends on a reflexive feedback loop
involving specialized mechanosensory organs, the halteres, which sense the rotational velocity of the fly’s body [23-29].
According to the current understanding, the powerful, but “dumb” [30] stretch activated muscles supply a periodic force to
drive the roughly sinusoidal wing motion [31, 32], while the weak, but fast direct steering muscles modulate stroke position
and amplitude to stabilize flight and manoeuvre.
In this way, the highly differentiated muscle types of flies are assumed to bring about the generation and control of the
wingbeat rhythm as functionally separate processes [30, 33] - unlike in locusts, in which actuation and control are closely
integrated within the framework of a LCO forced by afferent neural input. In the fly, it remains unclear if and how the
myogenic rhythm, which cannot be controlled on a cycle-by-cycle basis through direct neuronal input, may be efficiently
regulated by sensory feedback. While cycle-by-cycle control of wing stroke amplitude, stroke deviation and mean stroke
position by the direct action of the steering muscles has been extensively studied [33-36], the fast control of the thorax/wing
oscillator frequency has not been investigated in the previous literature. To achieve such frequency control, the myogenic
rhythm would have to be coupled to the only system that is able to act on such fast time scales, i. e., the steering muscles. The
null hypothesis of a functional separation between power and steering muscles predicts that flies are unable to control their
wingbeat frequency at the time scale of single wing strokes.
While flies apparently do not rely on CPGs to provide their wingbeat rhythm, it is intriguing to consider the possibility that
the indirect muscle actuation mechanism itself represents a LCO, which is forced in a phase-dependent manner by the
)(tx)(tF)()1(2tFxxxx3.0)(tF)(tx)(tx0)(tFdtdx/xt mechanical activity of the steering muscles. This alternative hypothesis is consistent with the “limit-cycle control” scheme
for insect flight suggested in [37, 38]. If confirmed, the flight control strategy of flies, and the functional role of their steering
muscles in particular, would appear in a new light altogether.
We explored this possibility by testing fruit flies for a functional coupling of the steering muscles onto the wing stroke
pattern generated by the stretch activated muscles. We evoked steering muscle activity from a periodic mechanic stim ulation
of the halteres, which are known to provide direct input to the steering muscles [21]. We used a laser vibrometer to measure
stroke-by-stroke variations in the wingbeat frequency of tethered flying flies and attribute these changes unambiguously to
the activity of the direct steering muscles, which are the only muscles in the flight apparatus of flies that are capable of
responding at the time scale of a single wing stroke.
In the case of sustained periodic forcing, a LCO may adjust its frequency and eventually synchronize with the forcing, a
phenomenon that has been observed in various biological limit cycle oscillators (e.g. [39, 40], review [41]). Depending on the
value of the forcing frequency and amplitude, one can expect to obtain 1:1 synchronization (entrainment), or a regime in
which the forcing frequency and the LCO frequency are other rational multiples of each other, such as 2:1, 3:2, etc..
Consistent with this generic property of periodically forced LCOs, we found 4 distinct regions of s timulus parameter space
(so called Arnol’d tongues), in which the flies synchronized their wingbeat with the forcing. Furthermore, we found that the
stimulation caused a fast modulation of the wingbeat frequency just outside of these parameter regions, demonstrating control
of the wingbeat rhythm on a cycle-by-cycle basis. Our results indicate that the direct steering muscles of flies function as a
weak mechanical forcing of a limit cycle oscillator embodied in the myogenic wing actuation mechanism.
2.1 Flies
2. Materials & Methods
We obtained fruit flies (Drosophila melanogaster Meigen) from our laboratory stock (descended from a wild-caught
population of 200 mated females). A standard breeding procedure was applied (25 females and 10 males, 12:12 hour
light/dark cycle, standard nutritive medium). The experiments were performed during the first 8 hours of a subjective day
with 5-10 days old female flies.
2.2 Experiments
We cold- anaesthetized single fruit flies and glued them by their thorax to the tip a steel tether (Fig. 2A). We then attached
the tether to a piezoelectrical actuator (P830.40, Physik Instrumente, Germany). We used a function generator (Agilent
33120A, Hewlett-Packard, USA) to create a sinusoidal voltage signal, amplified it using a custom built amplifier and applied
to the piezo. The piezo-induced displacement of the tether was proportional to the voltage with a conversion factor of 0.9
µm/V. The resulting sinusoidal displacement of the tethered fly caused an inertial force Fi on the endknobs of its halteres,
given by
, with
the amplitude of the tether displacement ,
the signal’s frequency and
the
mass of the haltere endknob [24].
We oscillated the flies along the anterior-posterior axis, applying either voltage frequency sweeps (0.1-500 Hz, sweep rate 5
Hz/s, holding constant the voltage at 1, 2, 4, 6, 8 or 10 volts) or voltage amplitude sweeps (0.1-8 V, sweep rate 0.8 V/s,
holding constant the frequencies close to the baseline wingbeat frequency of the fly currently being tested). Note that as the
inertial force is proportional to the square of the stimulus frequency, the forcing amplitude slowly increases during a voltage
frequency sweep - despite the voltage amplitude at the piezo staying constant.
When the stimulus frequency
is equal to (or very near) the wingbeat frequency, and when the stimulus and the wingbeat
are in the appropriate phase relation, the inertial force Fi resembles the Coriolis forces acting on the halteres during pitch
manoeuvres. Specifically, to mimic the time course of the Coriolis forces, the phase relation must be such that the tether
acceleration is zero at the time points corresponding to the dorsa l and ventral wing reversal [28]. The amplitude range of our
stimulation (1 - 10 V) then corresponds to pitch rotations of 120 - 1200 °/s (see electronic supplementary material), which
lies in the range typical for free flight manoeuvres [42].
We recorded the stimulation and the fly’s response together by measuring the velocity of the tether vibration about 1 mm
above the fly’s body using a laser Doppler vibrometer (MSA-500, Polytec, Germany), sampling at 10240 Hz. We chose
HTimtAtFsin)(2TAHm appropriate tethers (stainless steel, 1cm length, 200µm diameter) to tune the system suitably for our measurements. In each
wing cycle, the wings generate a sharp force peak that is known to occur near the ventral reversal phase of the wingbeat [43].
This leads to an overdamped, high-frequency (>1000 Hz) oscillation of the tether within each wing cycle. In the measured
tether velocity trace, this fly-induced signal was overlaid with a low frequency signal corresponding to the piezo actuator
(0.1-500 Hz) – see Fig. 2B, top trace.
Figure 2. A) Experimental setup. Flies were tethered to a steel pin,
which was attached to a piezoelectric actuator. The piezo oscillated
the tether, and with it the fly, according to the applied voltage s ignals
created by a function generator. In addition, the fly’s wingbeat
causes the tether to vibrate; we measured the total tether oscillation
using a laser Doppler vibrometer. B) Data processing. Top panel:
Raw data of the tether vibration generated by a flying fly and a
sinusoidal piezo oscillation with a frequency close to the baseline
WBF of the fly (about 200 Hz). Middle panel: The raw data was
high-pass and low-pass filtered to separate the fly’s signal (spiky
trace, red) from the piezo signal (sinusoidal trace, blue). Bottom
panel: Instantaneous phase of the piezo oscillation (blue trace) and
the phase in which the main peak in the fly’s signal occurred (red
dots). C) Schematic Arnol’d tongues. Four selected regions of
synchronization (filled triangles) for a generic limit cycle oscillator
with natural frequency
, forced by an external stimulus of
frequency
and amplitude
. At low stimulus amplitudes,
locking of the oscillator to the stimulus occurs when
is close
to a rational multiple of
; the frequency range increases with
increasing
. Arrows indicate amplitude and frequency sweeps
(see text).
2.3 Data analysis
Phase extraction
To extract the wing stroke phase from the tether vibration, we applied a 7 th order high-pass Butterworth filter in a zero phase
lag configuration with a cut-off frequency of 1000 Hz (red trace in Fig. 2B, middle panel). We then up-sampled the data to 50
kHz and applied a peak detection algorithm to determine the times
of the prominent peak in this fly-induced s ignal. The
times
approximately coincide with the ventral reversal of the wings.
To extract the stimulus phase, we applied a 7 th order low-pass Butterworth filter with a cut-off frequency of 500 Hz (blue
(blue trace in Fig.
trace in Fig. 2B, middle panel). We then applied the Hilbert transform to obtain the forcing phase
2B, bottom panel). Zero phase (i.e.,
= 0 modulo
) was chosen to coincide with the time points at which the tether
natfstimfstimastimfnatfstimaktkt)(t2 moves with maximal velocity in the direction towards the posterior direction of the fly (Fig 2A). Note that when
= 0
and
, the acceleration of the tether becomes zero.
Synchrogram construction
We applied a method called synchrogram analysis to reveal synchronization between the stimulus and the response in our
data. This method was developed and successfully applied to detect phase-locking in nonstationary and irregular signals [44,
45]. A synchrogram can be described as a ‘phase stroboscope’: Intervals of synchronization between an ex ternal forcing and
an oscillator are revealed by plotting the phase of the oscillator at periodic instances of the forcing, or vice versa.
We constructed the synchrograms by computing the phase
of the forcing at the periodic instances of the wingbeat,
(red dots, Fig. 2C). When the fly’s wing stroke and the applied force are not synchronized,
changes rapidly (Fig. 2B,
bottom panel, 0 - 0.7s). When the fly phase-locks to the stimulus and beats its wings at the f orcing frequency (1:1
synchronization),
remains constant (single, almost horizontal line in Fig. 2B, bottom panel, 0.7 - 1s). In the case of
higher-order synchronization (i.e. n wingbeats fit to m stimulus cycles , n:m locking), the phase points form m roughly
horizontal lines in the synchrogram.
Automated synchrogram evaluation
To determine the stimulus parameter regions in which synchronization took place (i.e. the Arnol’d tongues), we adapted a
synchronization detection algorithm for synchrograms from [46] and [47]. Because higher-order lockings tend to be difficult
to detect [6], we restricted our analys is to four low-order lockings (1:2, 1:1, 3:2 and 2:1). The synchronizat ion detection
algorithm operates as follows on the synchrogram time series
. First, the phase is rewrapped according to the locking
index m, such that
modulo
in the vicinity of each potential n:m locking region. Next, for each data
point in the series, a window spanning 10 wingbeat cycles preceding it and following it is defined, and the mean and the
standard deviation of
in this window are calculated. If the standard deviation is higher than a threshold of 2.5% of
,
the data point is deleted. Following this step, the time series is defined on a single or multiple intervals in which the
wingbeat is phase locked with the stimulus. Multiple intervals are present if a transient loss of locking occurs while still
within the synchronization region (so called phase slips [6], see Sec. 3.1 for an example). To determine the width of the
synchronization region, we therefore merged multiple intervals by linear interpolation if 1) their length was at least 15 poi nts
(wingbeats) and 2) their distance was less than 10 points. The longest remaining interval was considered to be the candidate
region of synchronization.
To exclude false positives, i.e. intervals in which an increase in WBF was not caused by the stimulat ion but by random
fluctuations in the wingbeat frequency, the candidate region was classified as “synchronized” only if the total number of
points in the interval was above a threshold value
. To determine
, we constructed synchrograms from control data
measured without piezo stimulation. This control synchrogram was constructed as described above, but in this case the
measured phase
of the stimulus was replaced by the phase of a fictitious stimulus. The synchrogram analys is was then
run as described in the previous paragraph. We found that by setting
for 1:2 locking,
for 1:1 locking,
for 3:2 locking and
for 2:1 locking, our algorithm discarded at least 99.5 % of lockings (of each order)
in the control synchrogram. Consequently, we used these threshold values of interval length to declare a candidate region as
“synchronized” in our test data sets (with piezo stimulation). The first and the last point of the synchronized interval were
taken as beginning and end of the respective Arnol’d tongue.
Construction of Arnol’d tongues
To place these border points into the frequency–amplitude parameter space of the forcing, we determined the stimulus
amplitudes, i.e., the acceleration amplitude of the tether at the beginning and end of the synchronized interval. The
acceleration was calculated from the measured velocity signal according to [48]. To account for possible differences in the
baseline frequency (i.e., the wingbeat frequency in absence of piezo stimulation) in between tests and in the 6 flies tested; we
divided the stimulus frequency at the border points by the fly’s mean wingbeat frequency in the 3 seconds preceding the
synchronized interval plus the 3 seconds following the synchronized interval.
For each frequency sweep test during which synchronization of a given order occurred, we therefore placed two border points
into the parameter space, with the ratio of stimulus frequency to baseline frequency on the abscissa, and the acceleration
delivered by the piezo on the ordinate. The set of all such border points (from the 6 flies tested) defines the Arnol’d tongu e.
)(t)(t)(kt)(kt)(kt)(kt)()(kkmttm2mm2nmNnmN4512N14011N11732N21821N 3. Results
3.1 Nonlinear response to periodic mechanical stimulus
When we stimulated flies periodically with the piezo, we observed nonlinear responses typical for forced limit cycle
oscillators. We will first present three examples to describe these phenomena in detail.
Periodically forced limit cycle oscillators are generically expected to exhibit multiple Arnol’d tongues, corresponding to
synchronization at various commensurate ratios of stimulation frequency and oscillator frequency. The frequency range in
which locking of a given order occurs typically increases with the forcing amplitude and gives rise to the characteristic
tongue shape (Fig. 2C). A transition to synchronization should occur when the stimulus parameters are varied so that the
border of the tongue is crossed. This can be achieved e.g. by a stimulus amplitude sweep (vertical arrow in Fig. 2C) or
stimulus frequency sweep (horizontal arrow in Fig. 2C).
To test for such a transition at the border of the 1:1 tongue (wingbeat frequency equal to stimulus frequency), we stimulate d
flies with amplitude sweeps at fixed frequencies close to the fly’s baseline frequency. For example, we increased the
amplitude from 0 to 3.8 m/s at a fixed detuning (difference of baseline frequency and stimulus frequency) of about 5 Hz
(Fig.3A). Here, the synchrogram reveals synchronization as a line of constant relative phase
(black dots, top trace),
while the stimulus amplitude is above 2.3 m/s 2 (red dotted line, bottom trace). This phase locking lasts for about 7 seconds or
about 1650 wingbeats. It is interrupted by three intervals, during which the phase changes rapidly (arrows in Fig. 3A). The
occasional occurrence of such rapid phase changes, called phase slips, is typical for biological oscillators due to their inherent
noisiness [6]. The calculated wingbeat frequency (WBF, black line, bottom trace) shows that the fly returns to its baseline
frequency during the phase slips .
Figure 3. A) Example of a recorded amplitude sweep. We
stimulated with
Hz while linearly first increasing
and then decreasing
, thus crossing the 1:1 Arnol’d
tongue border vertically. The fly’s baseline WBF was 235 ± 3
Hz. Top: Relative phase
between stimulation and
wingbeat. Bottom: WBF,
(dashed, blue),
(dotted,
red). Arrows on top indicate phase slips. B) Example of a
recorded frequency sweep. We stimulated with mean
m/s2 while linearly increasing
, thus
crossing the 1:1 Arnol’d tongue border horizontally. The fly’s
baseline frequency was 211 ± 2 Hz. Top: Relative phase
between stimulation and wingbeat. Bottom: WBF,
(dashed,
blue). Arrows
indicate
quasiperiodic
modulation of the WBF.
.
Similarly, we tested for a transition to synchronization when decreasing the detuning. We stimulat ed flies with frequency
sweeps with starting frequencies well below their WBFs . We then increased the stimulus frequency while keeping the applied
voltage amplitude constant. In the example shown in figure 3B, the fly’s baseline WBF was 209 ± 4 Hz (mean ± std.) and we
stimulated with a frequency sweep (0.1-500 Hz, 10 V). The fly locked to the forcing when the stimulation frequency was
above 199 Hz and de-locked when it reached 220 Hz. The locking lasted for about 4 seconds or 835 wingbeats. Before and
after looking, the WBF oscillated (see arrows). This oscillation reflects a regime in which the forcing frequency is too far
away from fly’s baseline frequency to fully entrain the fly; however it is close enough to significantly influence the oscill ator.
Indeed, similar frequency oscillations are generically expected for limit cycle oscillators when the forcing parameters are just
230stimfstima)(ktstimfstima1.8stimastimf)(ktstimf)(kmt outside the Arnol’d tongues (called phase walk-through or quasiperiodicity [6, 49]). The phenomenon is likewise visible in
Fig. 3A just before the onset of locking.
Besides 1:1 entrainment of the oscillator frequency to stimulus frequency, forced limit cycle oscillators typically also show
higher-order synchronization (also called subharmonic and superharmonic entrainment in the literature). The corresponding
Arnol’d tongues touch the abscissa axis (Fig. 2C) at points obtained as rational fractions of stimulus frequency and baseline
frequency, fStim: fNat = n: m, where n, m are relatively prime integers larger than 1. The most robust tongues are generically
expected for n, m small (for example 1:2, 3:2), while the tongues with large n, m tend to be very narrow and difficult to
observe in the presence of noise.
Figure 4. Synchrogram of a full frequency sweep. Relative phase
between stimulation and wingbeat. We stimulated with a
frequency sweep from 0.1 to 500 Hz with 10V applied piezo voltage. The baseline frequency of the fly was 209 ± 4Hz. The continuous
recording is shown divided in three rows for clarity. Phase locking leads to the occurrence of horizontal or weakly tilted lines of phase
points; for identified lockings of more than 100 wingbeats the corresponding Arnol’d tongue order n:m is indicated on top. Arrow indicates
phase slip at 2:1 locking.
To test for the existence of higher-order synchronization, we applied frequency sweeps and constructed synchrograms in a
range from 40 to 455 Hz. As the baseline wingbeat frequencies were around 200 Hz, the sweeps were expected to cross the
borders of multiple Arnol'd tongues. If, for example, a fly beating with 195 Hz is stimulated at 400 Hz, we expect the fly to
increase its frequency to 200 Hz, so that the frequency ratio adjusts to the locking order 2:1. An example synchrogram of a
full recording is shown in Fig. 4. Here, the synchrogram reveals not only the extended 1:1 locking region (around 210 Hz),
but also several lockings of higher order n:m. The fly synchronized in intervals in which the phase points form a single
horizontal line (for locking orders with m=1, like 1:1 and 2:1) or m horizontal lines (for m>1, see also Sec. 2.2). In the
presented example, all prominent lockings with a minimum length of 100 wingbeats are marked on top with the
corresponding locking order. Note the phase slip at the beginning of the 2:1 locking. This measurement contains 4 prominent
lockings, with 3:2 being the highest locking order.
3.2 Locking statistics and Arnol’d tongues
To extract the Arnol’d tongue boundaries and to quantify the occurrence of phase-locking, we focused on the frequency
sweep experiments. This had the major advantage that by applying a single sweep, we crossed the borders of multiple Arnol'd
tongues and, in case of synchronization, obtained both an entry and an exit point of the respective Arnol ’d tongue. We tested
)(kt 6 flies and stimulated each with 6 frequency sweeps of varying voltage amplitude (see Sec. 2.3). We obtained results from 33
measurements (Table 1); during 3 trials, the flies stopped flying and the respective measurements were discard ed. 1:1 locking
occurred in 16 out of 33 trials (48%). 1:2, 3:2 and 2:1 lockings occurred in 56%, 52% and 33% cases, respectively. We found
1:1 entrainment and higher-order locking in every fly. The percentage of sweeps, in which locking was observed in a fly, was
47 ± 10% (mean ± std.).
Figure 5. Extracted Arnol’d tongues. Start (>) and end
points (<) of synchronization as automatically detected from
the frequency sweep synchrograms. For the 1:1 lockings we
show linear fits of the boundaries of the tongue (robust fits,
forced through [1,0]). From the control experiments with
flies whose haltere endknobs were ablated, we obtained only
two lockings (►◄) with markedly reduced width.
From the identified locking intervals, we reconstructed the Arnol’d tongue boundaries, as described in Sec. 2.3. We found
that the 1:1 tongue is roughly triangular, i.e. an increasing forcing strength leads to a roughly proportional increase in locking
width, as expected for a weakly forced limit cycle oscillator (Fig. 4, second panel). The width of the tongue shows that the
synchronization is maintained for a detuning up to ± 5% of the WBF (which corresponds to about ±1 0 Hz). The shapes of the
higher-order tongues (Fig. 5, panels 1, 3 and 4) are more inconclusive and do not appear to be triangular. Still, all tongues
cover extended areas in the frequency-amplitude space.
Note that each measured tongue is nearly symmetrical, which is a generic property of very weakly forced limit cycle
oscillators [6]. While it cannot be entirely excluded that a hysteretic effect was nevertheless present due to a frequency
scanning in one direction only, it is rather unlikely that this would have exactly compensated an existing intrinsic asymmetric
shape. Furthermore, the possible existence of a hysteresis effect does not challenge any of the drawn conclusions in this
work.
Table 1. Locking quantification. Lockings rates, i.e. percentage of frequency sweeps in which phase locking was observed, obtained from
33 frequency sweeps at piezo voltage amplitudes of 1-10 V.
Fly 1
Fly 2
Fly 3
Fly 4
Fly 5
Fly 6
Locking rate per locking order
Locking rate per
1:2
100
25
67
50
60
33
1:1
100
50
67
33
20
17
3:2
2:1
83
50
67
33
60
17
83
25
67
0
20
0
fly
92
38
67
29
40
17
Mean locking rate
per locking order
56
48
52
33
rate per fly ± std.
Mean locking
47±10
3.3 Fast and phase-dependent modulation of wingbeat frequency
We showed that sustained mechanical stimulation at a fixed frequency (or slowly increasing frequency in the case of a
frequency sweep) acts as a forcing that can entrain the wingbeat rhythm. This result, however, does not tell us the time scal e
on which the forcing acts. To deduce this time scale, we analysed the intervals of fast WBF modulation prior to 1:1 locking
(see Sec. 3.1) in more detail. Such intervals were readily identified for stimulus amplitudes of more than 2 m/s 2. In these
intervals, the WBF was found to oscillate with a frequency that increased with the difference between baseline WBF and
stimulation frequency. We observed WBF oscillations with frequencies up to 40 Hz. In the example shown in Fig. 6A, the
time courses of WBF and the phase of stimulation reveal that the WBF varies from cycle to cycle, and follows the phase
of stimulus relative to the wingbeat.
For further analys is, we selected recordings, in which the WBF oscillation lasted for at least 2 seconds (7 measurements from
4 flies). The range of modulation frequencies was 11 to 18 Hz (mean 14Hz, std. 2.5 Hz). We found that the response to
stimulus is cons istent across the recordings, and that the resulting change in WBF varies approximately sinusoidally wit h the
) can be well fitted (r2 =
0.92) with the first two components of its Fourier series (Fig. 6B). A maximal decrease in WBF is observed around relative
(Fig. 6B). The observed response funct ion (which by definition is periodic in
relative phase
phase
, while a phase offset of 0 led to a maximal increase in WBF. Stimulation with phase offsets of
and
did not elicit a response in WBF. Recall that when the stimulus is applied with the frequency close to WBF and with the
relative phase
or
, the resulting acceleration has a time course mimicking the Coriolis force acting on the
haltere knobs during a pitch manoeuvre in free flight (see Sec. 2.3).
For other values of
, the time course of the acceleration does not match the Coriolis force in any rotational manoeuvre
during free flight. The response of the wing/thorax oscillator is thus maximal when the stimulus mimics the Coriolis force, to
which the haltere mechanosensory pathway is known to be highly sensitive [25].
Figure 6. A) Example of fast modulation of wingbeat frequency during stimulus frequency sweep. Recording of 1 second of flight
prior to the onset of 1:1 locking. Upper trace: Relative phase
between stimulation and wingbeat. Lower trace: WBF. B) Phase-
coupled frequency response. Pooled results from 7 recordings of stimulus frequency sweeps in 4 flies. For each recording, we
calculated the difference
instantaneous WBF and average WBF over the interval [ -1.5s:-0.5s] prior to locking, and normalized the
time series by its variance. We then binned the data with respect to the stimulus phase. Pooling the responses from the 7 rec ordings, we
calculated the median (open circle) and the interquartile range (error bars). Trace: best fit
for the two leading Fourier components of the periodic response function.
)(ktf)1.32sin(06.0)52.1sin(55.0xxy2/2/0 Based on the observed phase-dependent response, we can make a deduction on how quickly the stimulus affects the WBF.
Suppose that a fixed time delay
elapses between the mechanical stimulus and the resulting WBF adjustment. In the
response curve in Fig. 6B, the time delay would appear as a phase delay
, where
is the period of the modulation of
WBF. The response curve was obtained, however, from data segments in which the modulation frequency varied widely (by
60%). Therefore a sizeable time delay would result in a range of phase delays, and after averaging, a clear phase-dependent
response (as in Fig. 6B) would likely not be obtained. In addition, the coincidence of the optimal stimulus (resulting in
maximal response) with the Coriolis force would be lost. This suggests that the time delay
is small compared to the typical
period of WBF modulation (about 0.07 sec, or 14 wingbeat cycles). Our analysis therefore indicates that the mechanical
stimulus affects the WBF within a few wingbeat cycles.
3.4 Control experiments
To test if the observed entrainment was indeed mediated by the halteres, we measured the responses of flies whose haltere
endknobs had been ablated (which reduce the haltere sensitivity by about 90% without damaging the sensory fields at the
haltere base [25]). We performed 6 frequency sweeps per fly. In 4 out of the 6 flies with ablated halteres, we found no
entrainment. In two flies we observed a single, reduced locking region at the highest stimulation strengths (Fig. 5, filled
triangles). Our direct observations of the wingbeat of tethered flying flies revealed no gross difference between flies with
ablated halteres and untreated flies. Likewise, no qualitative differences were found in the vibrometry traces . Comparison of
wingbeat frequencies (extracted from the tether vibration) showed, however, an increased mean baseline WBF (225 ± 14 Hz)
in flies with ablated halteres, which was about 12 % elevated, compared to untreated flies (202 ± 16 Hz). A similar
observation was made by Dickinson [25], who reported an increase in WBF of about 24% following endknob ablation.
4. Discussion & Conclusions
In this study, we explore the possibility that the steering muscles of flies affect the myogenic wingbeat rhythm as a
mechanical forcing of a limit cycle oscillator (LCO). Our synchrogram analysis on the phase data of a mechanical stimulus
and the wing stroke reveals several characteristic features of a LCO, namely phase-locking, phase-slips, higher-order locking
and quasiperiodicity (Sec. 3.1, 3.2). We interpret the measured fast, phase-dependent control of the wingbeat frequency (Sec.
3.3) to result from the activity of the direct steering muscles that function as a mechanical forcing of a LCO embodied in the
myogenic wing actuation mechanism. Our data refute the null hypothesis according to which the power-muscle generated
rhythm is functionally decoupled from the activity of the direct steering muscles.
We adopted our experimental approach from the seminal studies performed by Nalbach, who vibrated blowflies back -and-
forth to emulate the Coriolis forces acting on the halteres of flies during rotations of the body in free flight [24, 28, 50, 51].
While these studies already revealed the basic synchronization phenomenon (1:1 locking, [50, 51])), they did not
systematically explore the possible LCO properties of the flight apparatus. We extended the previous approach by sampling
at a high rate the phase relationship between the wing stroke and the mechanical stimulus, using laser vibrometry.
Furthermore, we systematically varied the frequency and amplitude of the applied mechanical stimulus to determine the
Arnol’d tongues (i.e., regions of synchronization) of the flight control system, and to induce rapid modulation of wingbeat
frequency prior to entrainment.
Our results show that the fly responds to mechanical stimulation as a nonlinear oscillator. The essentially nonlinear prope rties
we observed included transitions to / from synchronization, subharmonic / superharmonic entrainment, and phase slips. More
specifically, our measurements showed behaviour typical for a nonlinear oscillator with a stable limit cycle [4, 6]. A special
type of chaotic oscillator (with narrow strange attractor and a high degree of phase coherence) could have similar
synchronization properties (see, e.g., [6], Sec. 10.1.2). Our observations are, however, most straightforwardly understood in
the framework of a periodically forced LCO, and we have found no signs of an attractor other than a single limit cycle.
Nalbach interpreted the 1:1 entrainment observed in blowflies as arising from a pitch illusion reflex [50]. Our experimental
approach, however, revealed synchronization also at higher frequency ratios (1:2 and 3:2), at which the mechanical stimuli
can no longer mimic Coriolis forces [28]. Hence, we interpret the observed synchronization as an emergent property of a
periodically forced LCO embodied in the myogenic wing actuation mechanism of flies.
T/T
Our results are highly consistent with the notion that fast sensory input from the halteres mediates the forcing. First, the
maximal response was obtained when the stimulus mimicked the time course of the Cori olis force that would act on the
haltere knobs during a pitch manoeuvre (Sec. 3.3). This agrees with previous studies in which blowflies [24, 50] and fruit
flies [25] were mechanically stimulated. Second, the responses were strongly reduced after ablating the end-knobs of the
halteres (see control experiments described in Sec. 3.4).
In our study, we used observations of rapid phase-dependent modulation of the wingbeat frequency prior to locking to infer
that the forcing acts on a time scale comparable to one wingbeat cycle (Sec. 3.3). We attribute the periodic forcing to the
mechanical action of direct steering muscles, which have the fast response properties that can explain the measured strict
phase relationship between the mechanical stimulus and the fr equency response (Fig. 6). The basalar muscle M.b1 could
serve this function, as it is known to modulate the wing kinematics according to the phase of its single contractions during
successive wing strokes and receives fast monosynaptic input from the halt eres [21, 22, 33, 34].
It might be argued that the halteres directly affect the thoracic rhythm mechanically, rather than acting through the
mechanosensory system and steering muscles. We evaluate this possibility as follows. We first consider the order of
magnitudes of forces exerted on the thorax by the halteres. During 1:1 locking in our experiments, the acceleration reaction
force due to the piezo actuation peaks at around
N (see electronic supplementary material). We then compar e it with
the estimated force exerted by a steering muscle. In the blowfly Calliphora, the peak force of the steering muscle M.b1 is of
order 10-2 N and depends substantially on the phase in which it is activated (Fig. 8 in [52]). Given that Drosophila is 10 times
smaller in length than Calliphora and following the scaling laws for insect flight muscles [53], we estimate the forces
generated by the homolog M.b1 muscle in Drosophila to be 103 = 1000 times smaller than in Calliphora. Hence, the forces
generated by Drosophila‘s steering muscles (10 µN) exceed the reaction acceleration forces from the halteres (60 nN) by
more than 2 orders of magnitude. We therefore conclude that the mechanical forcing of the thoracic rhythm is predominantly
effected by the steering muscles.
While the limit cycle properties of central pattern generators (CPGs) are well understood, these are now also revealed in the
myogenic wing actuation mechanism of an insect. It is per se not very surprising that a myogenic oscillation can be described
as a limit cycle. Our study, however, reveals a coupling of the fly’s haltere mechanosensory pathway to the thorax/wing LCO
that is functionally significant for flight control. This is evidenced by substantial changes in wingbeat frequency (abou t ± 10
Hz) measured when the halteres were stimulated by forces in the range of those occurring during typical flight manoeuvres.
We concluded that this coupling is effected by the mechanical activity of the steering muscles. In this way , the flight
apparatus of the fly is able to avoid the computationally expensive (and comparatively sluggish) neural mechanism of CPGs,
and instead replace it with a direct realization of a mechanical limit cycle oscillator. A mechanical system of this type has
been recently implemented in a swimming robot by Seo, Chung and Slotine [54]. Myogenic limit cycle – based control
provides an elegant conceptual framework in which to understand how flies can achieve extremely fast and precise flight
control with minimal neural resources. The fruit fly offers itself as a model system to explore limit cycle – based control
mechanisms that minimize the required computational power and operate at high frequencies. Such knowledge could serve
the design of biomimetic micro air vehicles (MAVs), which like flies are under severe constraints in terms of mass and power
consumption, and therefore depend on highly efficient control strategies [55, 56].
Acknowledgements
This work was supported by the Volkswagen foundation (I/80 984 - 986), Swiss National Science Foundation (CR23I2-
130111 / 1) and the Czech Republic (AV0Z50110509 and P304/12/G069). We thank Jonathon Howard, Henri Saleh, Vasco
Medici and Hannah Haberkern for valuable discussions and support and an anonymous reviewer for insightful critical
comments..
References
1
2
3
4
Delcomyn, F. 1980 Neural basis of rhythmic behavior in animals. Science 210, 492-498. (doi:10.1126/science.7423199)
Cohen, A. H., Holmes, P. J., Rand, R. H. 1982 The nature of the coupling between segmental oscillators of the lamprey spinal
generator for locomotion: A mathematical model. J. Math. Biol. 13, 345-369. (doi:10.1007/bf00276069)
Collins, J., Richmond, S. 1994 Hard-wired central pattern generators for quadrupedal locomotion. Biol. Cybern. 71, 375-385.
(doi:10.1007/bf00198915)
Glass, L., Mackey, M. C. From clocks to chaos: The rhythms of life. 1988.
8106 5
6
7
8
9
10
11
van der Pol, B. 1927 On relaxation-oscillations. The London, Edinburgh and Dublin Phil. Mag. & J. of Sci. 2, 978-992.
Pikovsky, A., Rosenblum, M. & Kurths, J. 2001 Synchronization - A universal concept in nonlinear sciences . Cambridge:
Cambridge University Press.
Vogelstein, R. J., Etienne-Cummings, R., Thakor, N. V., Cohen, A. H. 2006 Phase-dependent effects of spinal cord stimulation on
locomotor activity. Neural Systems and Rehabilitation Engineering, IEEE Trans. 14, 257-265. (doi:10.1109/tnsre.2006.881586)
Vogelstein, R., Tenore, F., Etienne-Cummings, R., Lewis, M., Cohen, A. 2006 Dynamic control of the central pattern generator
for locomotion. Biol. Cybern. 95, 555-566. (doi:10.1007/s00422-006-0119-z)
Taga, G., Yamaguchi, Y., Shimizu, H. 1991 Self-organized control of bipedal locomotion by neural oscillators in unpredictable
environment. Biol. Cybern. 65, 147-159. (doi:10.1007/bf00198086)
Kuo, A. D. 2002 The relative roles of feedforward and feedback in the control of rhythmic movements. Motor Control 6, 129-145.
Laszlo, J., van de Panne, M., Fiume, E. Limit cycle control and its application to the animation of balancing and walking. Proc .
23rd annual conf. on Computer graphics and interactive techniques. 237231: ACM; 1996. p. 155-162.
12
Ijspeert, A. J., Crespi, A., Ryczko, D., Cabelguen, J.-M. 2007 From swimming to walking with a salamander robot driven by a
spinal cord model. Science 315, 1416-1420. (doi:10.1126/science.1138353)
Wilson, D. M. 1961 The central nervous control of flight in a locust. J. Exp. Biol. 38, 471-490.
Wendler, G. 1974 The influence of proprioceptive feedback on locust flight co-ordination. J. Comp. Physiol. A 88, 173-200.
13
14
(doi:10.1007/bf00695406)
15
Ausborn, J., Stein, W., Wolf, H. 2007 Frequency control of motor patterning by negative sensory feedback. J. Neurosci. 27, 9319-
9328. (doi:10.1523/jneurosci.0907 -07.2007)
16
Schmeling, F., Stange, G., Homberg, U. 2010 Synchronization of wing beat cycle of the desert locust, Schistocerca gregaria, by
periodic light flashes. J. Comp. Physiol. A 196, 199-211. (doi:10.1007/s00359-010-0505-9)
17
Fry, S. N., Sayaman, R., Dickinson, M. H. 2003 The aerodynamics of free-flight maneuvers in Drosophila. Science 300, 495-498-
-. (doi:10.1126/science.1081944)
18
19
Pringle, J. W. S. 1949 The excitation and contraction of the flight muscles of insects. J. Physiol. (Lond.) 108, 226-232.
Heide, G. 1974 The influence of wingbeat synchronous feedback on the motor output systems in flies. Z. Naturforschung C C 29,
739-744.
20
Heide, G. 1983 Neural mechanisms of flight control in Diptera. In Biona- report 2 (ed. W. Nachtigall), pp. 35 -52, Stuttgart, New
York: G. Fischer Verlag.
21
Fayyazuddin, A., Dickinson, M. H. 1996 Haltere afferents provide direct, electrotonic input to a steering motor neuron in the
blowfly, Calliphora. J. Neurosci. 16, 5225-5232.
22
Fayyazuddin, A., Dickinson, M. H. 1999 Convergent mechanosensory input structures the firing phase of a steering motor neuron
in the blowfly, Calliphora. J. Neurophysiol. 82, 1916-1926.
23
Fox, J. & Daniel, T. 2008 A neural basis for gyroscopic force measurement in the halteres of Holorusia. J. Comp. Physiol. A 194,
887-897. (doi:10.1007/s00359-008-0361-z)
24
Nalbach, G., Hengstenberg, R. 1994 The halteres of the blowfly calliphora: I I. Three-dimensional organization of compensatory
reactions to real and simulated rotations. J. Comp. Physiol. A 175, 695-708. (doi:10.1007/BF00191842)
25
Dickinson, M. H. 1999 Haltere-mediated equilibrium reflexes of the fruit fly, Drosophila melanogaster. Phil. Trans. R. Soc. B
354, 903-916.
26
Fraenkel, G. & Pringle, J. W. S. 1938 Halteres of flies as gyroscopic organs of equilibrium. Nature 141, 919-920.
(doi:10.1038/141919a0)
27
28
Pringle, J. W. S. 1948 The gyroscopic mechanism of the halteres of Diptera. Phil. Trans. R. Soc. B 233, 347-384.
Nalbach, G. 1993 The halteres of the blowfly Calliphora: I. Kinematics and dynamics. J. Comp. Physiol. A 173, 293-300.
(doi:10.1007/BF00212693)
29
Nalbach, G. 1994 Extremely non-orthogonal axes in a sense organ for rotation: Behavioural analysis of the Dipteran haltere
system. Neurosc. 61, 149-163. (doi:10.1016/0306-4522(94)90068-X)
30
31
32
Dickinson, M. H., Tu, M. S. 1997 The function of Dipteran flight muscle. Comp. Biochem. Phys. A 116, 223-238.
Greenewalt, C. H. 1960 The wings of insects and birds as mechanical oscillators. Proc. Am. Phil. Soc. 104, 605-611.
Ellington, C. P. 1984 The aerodynamics of hovering insect flight. III. Kinematics. Phil. Trans. R. Soc. B 305, 41. (doi:citeulike-
article-id:8698273)
33
Heide, G., Götz, K. G. 1996 Optomotor control of course and altitude in Drosophila melanogaster is correlated with distinct
activities of at least three pairs of flight steering muscles. J. Exp. Biol. 199, 1711-1726.
34
Tu, M. S., Dickinson, M. H. 1996 The control of wing kinematics by two steering muscles of the blowfly (Calliphora vicina). J.
Comp. Physiol. A 178, 813-830. (doi:10.1007/BF00225830)
35
Balint, C. N., Dickinson, M. H. 2001 The correlation between wing kinematics and steering muscle activity in the blowfly
Calliphora vicina. J. Exp. Biol. 204, 4213-4226.
36
Balint, C. N., Dickinson, M. H. 2004 Neuromuscular control of aerodynamic forces and moments in the blowfly, Calliphora
vicina. J Exp Biol. 207, 3813-3838. (doi:10.1242/jeb.01229)
37
Taylor, G. K., Zbikowski, R. 2005 Nonlinear time-periodic models of the longitudinal flight dynamics of desert locusts
Schistocerca gregaria. J. R. Soc. Interface 2, 197-221.
38
Żbikowski, R., et al. 2006 On mathematical modelling of insect flight dynamics in the context of micro air vehicles. Bioinspir.
Biomim. 1, R26. (doi:10.1088/1748-3182/1/2/R02)
39
Stoop, R., Schindler, K., Bunimovich, L. A. 2000 When pyramidal neurons lock, when they respond chaotically, and when they
like to synchronize. Neurosc. Res. 36, 81-91. (doi:10.1016/s0168-0102(99)00108-x)
40
Buck, J., Buck, E., Case, J. F., Hanson, F. E. 1981 Control of flashing in fireflies. J. Comp. Physiol. A 144, 287-298.
(doi:10.1007/BF00612560)
41
42
Glass, L. 2001 Synchronization and rhythmic processes in phys iology. Nature 410, 277-284. (doi::10.1038/35065745)
Dickson, W. B., Straw, A. D., Dickinson, M. H. 2008 Integrative model of Drosophila flight. AIAA Journal 46, 14.
(doi:10.2514/1.29862)
43
Dickinson, M. H., Götz, K. G. 1996 The wake dynamics and flight forces of the fruit fly Drosophila melanogaster. J. Exp. Biol.
199, 2085-2104.
44
Schäfer, C., Rosenblum, M. G., Kurths, J., Abel, H.-H. 1998 Heartbeat synchronized with ventilation. Nature 392, 239-240.
(doi:10.1038/32567)
45
Schäfer, C., Rosenblum, M. G., Abel, H.-H., Kurths, J. 1999 Synchronization in the human cardiorespiratory system. Phys. Rev. E
60, 857-870.
46
Bartsch, R., Kantelhardt, J. W., Penzel, T., Havlin, S. 2007 Experimental evidence for phase synchronization transitions in the
human cardiorespiratory system. Phys. Rev. Lett. 98, 054102. (doi:10.1103/PhysRevLett.98.054102)
47
Hamann, C., Bartsch, R. P., Schumann, A. Y., Penzel, T., Havlin, S., Kantelhardt, J. W. 2009 Automated synchrogram analysis
applied to heartbeat and reconstructed respiration. Chaos 19, 015106. (doi:10.1063/1.3096415)
48
Woltring, H. J. 1985 On optimal smoothing and derivative estimation from noisy displacement data in biomechanics. Hum.
Movement. Sci. 4, 229-245. (doi:10.1016/0167-9457(85)90004-1)
49
Ermentrout, G. B., Rinzel, J. 1984 Beyond a pacemaker's entrainment limit: Phase walk-through. Am. J. Physiol. Regul. Integr.
Comp. Physiol. 246, R102-106.
50
Nalbach, G. Verhaltensuntersuchungen zur Funktion der Halteren bei der Schmeissfliege Calliphora erythrocephala mit echten
und simulierten Drehreizen. PhD- Thesis, Eberhard-Karls-Universität Tübingen; 1991.
51
Nalbach, G 1988. Linear oscillations elicit haltere mediated turning illusions and entrainment in the blowfly Calliphora. In
Proc.16th Neurobiol. Conf. 1988; Göttingen.
52
Tu, M. S., Dickinson, M. H. 1994 Modulation of negative work output from a steering muscle of the blowfly Calliphora vicina. J.
Exp. Biol.192, 207-224.
53
Schilder, R. J., Marden, J. H. 2004 A hierarchical analysis of the scaling of force and power production by dragonfly flight
motors. J. Exp. Biol.207, 767-776.(doi:10.1242/jeb.00817)
54
Seo, K., Chung, S.-J., Slotine, J.-J. 2010 CPG-based control of a turtle-like underwater vehicle. Auton. Robot. 28, 247-269.
(doi:10.1007/s10514-009-9169-0)
55
Xinyan, D., Schenato, L. & Sastry, S. S. 2003 Model identification and attitude control for a micromechanical flying insect
including thorax and sensor models. In Proc. IEEE Intern. Conf. on Robotics and Automation, ICRA '03. (doi:
10.1109/ROBOT.2003.1241748)
56
Chung S.-J.; Dorothy M. 2010 Neurobiologically inspired control of engineered flapping flight. J. Guid. Control Dynam. 33, 440-
453. (doi:10.2514/1.45311)
|
1611.00388 | 2 | 1611 | 2017-01-22T01:48:48 | Automated scalable segmentation of neurons from multispectral images | [
"q-bio.NC",
"q-bio.SC"
] | Reconstruction of neuroanatomy is a fundamental problem in neuroscience. Stochastic expression of colors in individual cells is a promising tool, although its use in the nervous system has been limited due to various sources of variability in expression. Moreover, the intermingled anatomy of neuronal trees is challenging for existing segmentation algorithms. Here, we propose a method to automate the segmentation of neurons in such (potentially pseudo-colored) images. The method uses spatio-color relations between the voxels, generates supervoxels to reduce the problem size by four orders of magnitude before the final segmentation, and is parallelizable over the supervoxels. To quantify performance and gain insight, we generate simulated images, where the noise level and characteristics, the density of expression, and the number of fluorophore types are variable. We also present segmentations of real Brainbow images of the mouse hippocampus, which reveal many of the dendritic segments. | q-bio.NC | q-bio | Automated scalable segmentation of neurons from
multispectral images
Uygar Sümbül
Grossman Center for the Statistics of Mind
and Dept. of Statistics, Columbia University
Douglas Roossien Jr.
University of Michigan Medical School
Fei Chen
MIT Media Lab and McGovern Institute
Nicholas Barry
MIT Media Lab and McGovern Institute
Edward S. Boyden
MIT Media Lab and McGovern Institute
Dawen Cai
University of Michigan Medical School
John P. Cunningham
Grossman Center for the Statistics of Mind
and Dept. of Statistics, Columbia University
Liam Paninski
Grossman Center for the Statistics of Mind
and Dept. of Statistics, Columbia University
Abstract
Reconstruction of neuroanatomy is a fundamental problem in neuroscience.
Stochastic expression of colors in individual cells is a promising tool, although its
use in the nervous system has been limited due to various sources of variability in
expression. Moreover, the intermingled anatomy of neuronal trees is challenging
for existing segmentation algorithms. Here, we propose a method to automate the
segmentation of neurons in such (potentially pseudo-colored) images. The method
uses spatio-color relations between the voxels, generates supervoxels to reduce
the problem size by four orders of magnitude before the final segmentation, and is
parallelizable over the supervoxels. To quantify performance and gain insight, we
generate simulated images, where the noise level and characteristics, the density
of expression, and the number of fluorophore types are variable. We also present
segmentations of real Brainbow images of the mouse hippocampus, which reveal
many of the dendritic segments.
1 Introduction
Studying the anatomy of individual neurons and the circuits they form is a classical approach
to understanding how nervous systems function since Ramón y Cajal's founding work. Despite
a century of research, the problem remains open due to a lack of technological tools: mapping
neuronal structures requires a large field of view, a high resolution, a robust labeling technique, and
computational methods to sort the data. Stochastic labeling methods have been developed to endow
individual neurons with color tags [1, 2]. This approach to neural circuit mapping can utilize the
light microscope, provides a high-throughput and the potential to monitor the circuits over time, and
complements the dense, small scale connectomic studies using electron microscopy [3] with its large
field-of-view. However, its use has been limited due to its reliance on manual segmentation.
The initial stochastic, spectral labeling (Brainbow) method had a number of limitations for neuro-
science applications including incomplete filling of neuronal arbors, disproportionate expression
of the nonrecombined fluorescent proteins in the transgene, suboptimal fluorescence intensity, and
color shift during imaging. Many of these limitations have since improved [4] and developments in
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
arXiv:1611.00388v2 [q-bio.NC] 22 Jan 2017
various aspects of light microscopy provide further opportunities [5–8]. Moreover, recent approaches
promise a dramatic increase in the number of (pseudo) color sources [9–11]. Taken together, these ad-
vances have made light microscopy a much more powerful tool for neuroanatomy and connectomics.
However, existing automated segmentation methods are inadequate due to the spatio-color nature
of the problem, the size of the images, and the complicated anatomy of neuronal arbors. Scalable
methods that take into account the high-dimensional nature of the problem are needed.
Here, we propose a series of operations to segment 3-D images of stochastically tagged nervous
tissues. Fundamentally, the computational problem arises due to insufficient color consistency within
individual cells, and the voxels occupied by more than one neuron. We denoise the image stack
through collaborative filtering [12], and obtain a supervoxel representation that reduces the problem
size by four orders of magnitude. We consider the segmentation of neurons as a graph segmentation
problem [13], where the nodes are the supervoxels. Spatial discontinuities and color inhomogeneities
within segmented neurons are penalized using this graph representation. While we concentrate on
neuron segmentation in this paper, our method should be equally applicable to the segmentation of
other cell classes such as glia.
To study various aspects of stochastic multispectral labeling, we present a basic simulation algorithm
that starts from actual single neuron reconstructions. We apply our method on such simulated images
of retinal ganglion cells, and on two different real Brainbow images of hippocampal neurons, where
one dataset is obtained by expansion microscopy [5].
2 Methods
Successful segmentations of color-coded neural images should consider both the connected nature
of neuronal anatomy and the color consistency of the Brainbow construct. However, the size and
the noise level of the problem prohibit a voxel-level approach (Fig. 1). Methods that are popular in
hyperspectral imaging applications, such as nonnegative matrix factorization [14], are not immediately
suitable either because the number of color channels are too few and it is not easy to model neuronal
anatomy within these frameworks. Therefore, we develop (i) a supervoxelization strategy, (ii)
explicitly define graph representations on the set of supervoxels, and (iii) design the edge weights to
capture the spatio-color relations (Fig. 2a).
2.1 Denoising the image stack
Voxel colors within a neurite can drift along the neurite, exhibit high frequency variations, and differ
between the membrane and the cytoplasm when the expressed fluorescent protein is membrane-
binding (Fig. 1). Collaborative filtering generates an extra dimension consisting of similar patches
within the stack, and applies filtering in this extra dimension rather than the physical dimensions.
We use the BM4D denoiser [12] on individual channels of the datasets, assuming that the noise is
Gaussian. Figure 2 demonstrates that the boundaries are preserved in the denoised image.
2.2 Dimensionality reduction
We make two basic observations to reduce the size of the dataset: (i) Voxels expressing fluorescent
proteins form the foreground, and the dark voxels form the much larger background in typical
Brainbow settings. (ii) The basic promise of Brainbow suggests that nearby voxels within a neurite
have very similar colors. Hence, after denoising, there must be many topologically connected voxel
sets that also have consistent colors.
The watershed transform [15] considers its input as a topographic map and identifies regions associated
with local minima ("catchment basins" in a flooding interpretation of the topographic map). It can
be considered as a minimum spanning forest algorithm, and obtained in linear time with respect
to the input size [16, 17]. For an image volume V = V (x, y, z, c), we propose to calculate the
topographical map T (disaffinity map) as
T (x, y, z) = max
t∈{x,y,z}
max
c
Gt(x, y, z, c),
(1)
where x, y, z denote the spatial coordinates, c denotes the color coordinate, and Gx, Gy, Gz denote
the spatial gradients of V (nearest neighbor differencing). That is, any edge with significant deviation
in any color channel will correspond to a "mountain" in the topographic map. A flooding parameter,
f, assigns the local minima of T to catchment basins, which partition V together with the boundary
voxels. We assign the boundaries to neighboring basins based on color proximity. The background is
the largest and darkest basin. We call the remaining objects supervoxels [18, 19]. Let F denote the
binary image identifying all of the foreground voxels.
2
Figure 1: Multiple noise sources affect the color consistency in Brainbow images. a, An 85×121
Brainbow image patch from a single slice (physical size: 8.5µ × 12.1µ). Expression level differs
significantly between the membrane and the cytoplasm along a neurite (arrows). b, A maximum
intensity projection view of the 3-d image stack. Color shifts along a single neurite, which travels
to the top edge and into the page (arrows). c, A 300 × 300 image patch from a single slice of a
different Brainbow image (physical size: 30µ × 30µ). d, The intensity variations of the different
color channels along the horizontal line in c. e, Same as d for the vertical line in c. f, The image
patch in c after denoising. g–h, Same as d and e after denoising. For the plots, the range of individual
color channels is [0, 1].
Objects without interior voxels (e.g., single-voxel thick dendritic segments) may not be detected by
Eq. 1 (Supp. Fig. 1). We recover such "bridges" using a topology-preserving warping (in this case,
only shrinking is used.) of the thresholded image stack into F [20, 21]:
B = W(Iθ, F ),
(2)
where Iθ is binary and obtained by thresholding the intensity image at θ. W returns a binary image
B such that B has the same topology as Iθ and agrees with F as much as possible. Each connected
component of B ∧ ¯F (foreground of B and background of F ) is added to a neighboring supervoxel
based on color proximity, and discarded if no spatial neighbors exist (Supp. Text).
We ensure the color homogeneity within supervoxels by dividing non-homogeneous supervoxels (e.g.,
large color variation across voxels) into connected subcomponents based on color until the desired
homogeneity is achieved (Supp. Text). We summarize each supervoxel's color by its mean color.
We apply local heuristics and spatio-color constraints iteratively to further reduce the data size and
demix overlapping neurons in voxel space (Fig. 2f,g and Supp. Text). Supp. Text provides details on
the parallelization and complexity of these steps and the method in general.
3
A
C
F
B
E
H
D
G
5
x−position (µm)
10
15
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
normalized intensity
0
0
2
4
6
8
y−position (µm)
10
12
14
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
normalized intensity
0
0
2
4
6
8
y−position (µm)
10
12
14
0.7
0.6
0.5
0.4
0.3
0.2
0.1
normalized intensity
5
x−position (µm)
10
15
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
normalized intensity
of
Figure 2: Best
viewed digitally.
a, A schematic
of the processing
steps b, Max.
intensity
pro-
jection of a raw
Brainbow image
c, Max. intensity
projection of the
denoised
image
d, A zoomed-in
version
the
patch indicated by
the dashed square
in b.
e, The
corresponding
denoised image. f,
One-third of the
supervoxels in the
top-left quadrant
(randomly
cho-
sen). g, Same as f
after the merging
h1-h4,
step.
Same as b,c,f,g
for
simulated
data. Scale bars,
20µm.
2.3 Clustering the supervoxel set
We consider the supervoxels as the nodes of a graph and express their spatio-color similarities
through the existence (and the strength) of the edges connecting them, summarized by a highly
sparse adjacency matrix. Removing edges between supervoxels that aren't spatio-color neighbors
avoids spurious links. However, this procedure also removes many genuine links due to high color
variability (Fig. 1). Moreover, it cannot identify disconnected segments of the same neuron (e.g., due
to limited field-of-view). Instead, we adjust the spatio-color neighborhoods based on the "reliability"
of the colors of the supervoxels. Let S denote the set of supervoxels in the dataset. We define
the sets of reliable and unreliable supervoxels as Sr = {s ∈ S : n(s) > ts, h(s) < td} and
Su = S \ Sr, respectively, where n(s) denotes the number of voxels in s, h(s) is a measure of the
color heterogeneity (e.g., the maximum difference between intensities across all color channels), ts
and td are the corresponding thresholds.
We describe a graph G = (V, E), where V denotes the vertex set (supervoxels) and E = Es∪Ec∪E¯s
denotes the edges between them:
Es = {(ij) : δij < s, i 6= j}
Ec = {(ij) : si, sj ∈ Sr, dij < c, i 6= j}
E¯s = {(ij), (ji) : si ∈ Su, (ij) /∈ Es, Oi(j) < kmin − Ki, i 6= j},
(3)
where δij, dij are the spatial and color distances between si and sj, respectively. s and c are
the corresponding maximum distances. An unreliable supervoxel with too few spatial neighbors is
allowed to have up to kmin edges via proximity in color space. Here, Oi(j) is the order of supervoxel
sj in terms of the color distance from supervoxel si, and Ki is the number of s-spatial neighbors of
si. (Note the symmetric formulation in E¯s.) Then, we construct the adjacency matrix as
A(i, j) =(cid:26) e−αd2
0,
4
ij ,
(ij) ∈ E
otherwise
(4)
where α controls the decay in affinity with respect to distance in color. We use k-d tree structures
to efficiently retrieve the color neighborhoods [22]. Here, the distance between two supervoxels is
minv∈V,u∈U D(v, u), where V and U are the voxel sets of the two supervoxels and D(v, u) is the
Euclidean distance between voxels v and u.
A classical way of partitioning graph nodes that are nonlinearly separable is by minimizing a function
(e.g., the sum or the maximum) of the edge weights that are severed during the partitioning [23].
Here, we use the normalized cuts algorithm [24, 13] with two simple modifications: the k-means step
is weighted by the sizes of the supervoxels and initialized by a few iterations of k-means clustering
of the supervoxel colors only (Supp. Text). The resulting clusters partition the image stack (together
with the background), and represent a segmentation of the individual neurons within the image stack.
An estimate of the number of neurons can be obtained from a Dirichlet process mixture model [25].
While this estimate is often rough [26], the segmentation accuracy appears resilient to imperfect
estimates (Fig. 4c).
2.4 Simulating Brainbow tissues
We create basic simulated Brainbow image stacks from volumetric reconstructions of single neurons
(Algorithm 1). For simplicity, we model the neuron color shifts by a Brownian noise component on
the tree, and the background intensity by a white Gaussian noise component (Supp. Text).
We quantify the segmentation quality of the voxels using the adjusted Rand index (ARI), whose
maximum value is 1 (perfect agreement), and expected value is 0 for random clusters [27]. (Supp.
Text)
background noise variability σ1, neural color variability σ2, r, saturation level M
Shift and rotate neuron ni to minimize overlap with existing neurons in the stack
Generate a uniformly random color vector vi of length C
Identify the connected components of cij of ni within the stack
for cij ∈ {cij}j do
Algorithm 1 Brainbow image stack simulation
Require: number of color channels C, set of neural shapes S = {ni}i, stack (empty, 3d space + color),
1: for ni ∈ S do
2:
3:
4:
5:
6:
7:
8:
9:
10: end for
11: Add white noise to each voxel generated by N (0, σ2
2I)
12: if brightness exceeds M then
13:
14: end if
15: return stack
Saturate at M
Pre-assign vi to r% of the voxels of cij
C-dimensional random walk on cij with steps N (0, σ2
end for
Add neuron ni to the stack (with additive colors for shared voxels)
1I) (Supp. Text)
3 Datasets
To simulate Brainbow image stacks, we used volumetric single neuron reconstructions of mouse
retinal ganglion cells in Algorithm 1. The dataset is obtained from previously published studies [28,
29]. Briefly, the voxel size of the images is 0.4µ × 0.4µ × 0.5µ, and the field of view of individual
stacks is 320µ × 320µ × 70µ or larger. We evaluate the effects of different conditions on a central
portion of the simulated image stack.
Both real datasets are images of the mouse hippocampal tissue. The first dataset has 1020×1020×225
voxels (voxel size: 0.1× 0.1× 0.3µ3), and the tissue was imaged at 4 different frequencies (channels).
The second dataset has 1080 × 1280 × 134 voxels with an effective voxel size of 70 × 70 × 40nm,
where the tissue was 4× linearly expanded [5], and imaged at 3 different channels. The Brainbow
constructs were delivered virally, and approximately 5% of the neurons express a fluorescence gene.
4 Results
Parameters used in the experiments are reported in Supp. Text.
Fig. 1b, d, and e depict the variability of color within individual neurites in a single slice and through
the imaging plane. Together, they demonstrate that the voxel colors of even a small segment of a
5
Figure 3: Segmentation
of a simulated Brainbow
image stack. Adjusted
Rand index of the fore-
ground is 0.80. Pseudo-
color representation of 4-
channel data. Top: max-
imum intensity projection
of the ground truth. Only
the supervoxels that are oc-
cupied by a single neuron
are shown. Bottom: max-
imum intensity projection
of the reconstruction. The
top-left corners show the
whole image stack. All
other panels show the max-
imum intensity projections
of the supervoxels assigned
to a single cluster (inferred
neuron).
Figure 4: Segmentation accuracy of simulated data a, Expression density (ratio of voxels occupied
by at least one neuron) vs. ARI. b, σ1 (Algorithm 1) vs. ARI. c, Channel count vs. ARI for a 9-neuron
simulation, where K ∈ [6, 12]. ARI is calculated for the foreground voxels. See Supp. Fig. 7 for
ARI values for all voxels.
neuron's arbor can occupy a significant portion of the dynamic range in color with the state-of-the-
art Brainbow data. Fig. 1c-e show that collaborative denoising removes much of this noise while
preserving the edges, which is crucial for segmentation. Fig. 2b-e and h demonstrate a similar effect
on a larger scale with real and simulated Brainbow images.
Fig. 2 shows the raw and denoised versions of the 1020 × 1020 × 225 image, and a randomly chosen
subset of its supervoxels (one-third). The original set had 6.2 × 104 supervoxels, and the merging
routine decreased this number to 3.9 × 104. The individual supervoxels grew in size while avoiding
mergers with supervoxels of different neurons. This set of supervoxels, together with a (sparse)
spatial connectivity matrix, characterizes the image stack. Similar reductions are obtained for all the
real and simulated datasets.
Fig. 3 shows the segmentation of a simulated 200×200×100 (physical size: 80µ×80µ×50µ) image
patch. (Supp. Fig. 2 shows all three projections, and Supp. Fig. 3 shows the density plot through
the z-axis.) In this particular example, the number of neurons within the image is 9, σ1 = 0.04,
σ2 = 0.1, and the simulated tissue is imaged using 4 independent channels. Supp. Fig. 4 shows a
patch from a single slice to visualize the amount of noise. The segmentation has an adjusted Rand
index of 0.80 when calculated for the detected foreground voxels, and 0.73 when calculated for all
voxels. (In some cases, the value based on all voxels is higher.) The ground truth image displays only
those supervoxels all of whose voxels belong to a single neuron. The bottom part of Fig. 3 shows
6
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
3
4
5
true (9)
channel count
6
7
8
10
11
12
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0
0.02
0.04
0.06
0.08
0.1
3 ch.
step size (σ1) −− range per channel: [0, 1]
4 ch.
5 ch.
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.06
0.08
0.1
0.12
0.14
0.16
0.18
expression density (ratio of occupied voxels)
3 ch.
adjusted Rand index
4 ch.
5 ch.
Figure 5: Segmentation of a Brainbow stack – best viewed digitally. Pseudo-color represen-
tation of 4-channel data. The physical size of the stack is 102µ × 102µ × 68µ. The top-left
corner shows the maximum intensity projection of the whole image stack, all other panels show the
maximum intensity projections of the supervoxels assigned to a single cluster (inferred neuron).
7
that many of these supervoxels are correctly clustered to preserve the connectivity of neuronal arbors.
There are two important mistakes in clusters 4 (merger) and 9 (spurious cluster). These are caused by
aggressive merging of supervoxels (Supp. Fig. 5), and the segmentation quality improves with the
inclusion of an extra imaging channel and more conservative merging (Supp. Fig. 6). We plot the
performance of our method under different conditions in Fig. 4 (and Supp. Fig. 7). We set the noise
standard deviation to σ1 in the denoiser, and ignored the contribution of σ2. Increasing the number
of observation channels improves the segmentation performance. The clustering accuracy degrades
gradually with increasing neuron-color noise (σ1) in the reported range (Fig. 4b). The accuracy does
not seem to degrade when the cluster count is mildly overestimated, while it decays quickly when the
count is underestimated (Fig. 4c).
Fig. 5 displays the segmentation of the 1020 × 1020 × 225 image. While some mistakes can be
spotted by eye, most of the neurites can be identified and simple tracing tools can be used to obtain
final skeletons/segmentations [30, 31]. In particular, the identified clusters exhibit homogeneous
colors and dendritic pieces that either form connected components or miss small pieces that do not
preclude the use of those tracing tools. Some clusters appear empty while a few others seem to
comprise segments from more than one neuron, in line with the simulation image (Fig. 2.4).
Supp. Fig. 8 displays the segmentation of the 4× expanded, 1080× 1280× 134 image. While the two
real datasets have different characteristics and voxel sizes, we used essentially the same parameters
for both of them throughout denoising, supervoxelization, merging, and clustering (Supp. Text).
Similar to Fig. 5, many of the processes can be identified easily. On the other hand, Supp. Fig. 8
appears more fragmented, which can be explained by the smaller number of color channels (Fig. 4).
5 Discussion
Tagging individual cells with (pseudo)colors stochastically is an important tool in biological sciences.
The versatility of genetic tools for tagging synapses or cell types and the large field-of-view of light
microscopy positions multispectral labeling as a complementary approach to electron microscopy
based, small-scale, dense reconstructions [3]. However, its use in neuroscience has been limited due
to various sources of variability in expression. Here, we demonstrate that automated segmentation of
neurons in such image stacks is possible. Our approach considers both accuracy and scalability as
design goals.
The basic simulation proposed here (Algo. 1) captures the key aspects of the problem and may
guide the relevant genetics research. Yet, more detailed biophysical simulations represent a valuable
direction for future work. Our simulations suggest that the segmentation accuracy increases signifi-
cantly with the inclusion of additional color channels, which coincides with ongoing experimental
efforts [9–11]. We also note that color constancy of individual neurons plays an important role both
in the accuracy of the segmentation (Fig. 4) and the supervoxelized problem size.
While we did not focus on post-processing in this paper, basic algorithms (e.g., reassignment of small,
isolated supervoxels) may improve both the visualization and the segmentation quality. Similarly,
more elaborate formulations of the adjacency relationship between supervoxels can increase the
accuracy. Finally, supervised learning of this relationship (when labeled data is present) is a promising
direction, and our methods can significantly accelerate the generation of training sets.
6 Acknowledgments
The authors thank Suraj Keshri and Min-hwan Oh (Columbia University) for useful conversations.
Funding for this research was provided by ARO MURI W911NF-12-1-0594, DARPA N66001-
15-C-4032 (SIMPLEX), and a Google Faculty Research award; in addition, this work was supported
by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior
Business Center (DoI/IBC) contract number D16PC00008. The U.S. Government is authorized to
reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation
thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should
not be interpreted as necessarily representing the official policies or endorsements, either expressed
or implied, of IARPA, DoI/IBC, or the U.S. Government.
8
References
[1] J Livet et al. Transgenic strategies for combinatorial expression of fluorescent proteins in the nervous
system. Nature, 450(7166):56–62, 2007.
[2] A H Marblestone et al. Rosetta brains: A strategy for molecularly-annotated connectomics. arXiv preprint
arXiv:1404.5103, 2014.
[3] Shin-ya Takemura, Arjun Bharioke, Zhiyuan Lu, Aljoscha Nern, Shiv Vitaladevuni, Patricia K Rivlin,
William T Katz, Donald J Olbris, Stephen M Plaza, Philip Winston, et al. A visual motion detection circuit
suggested by drosophila connectomics. Nature, 500(7461):175–181, 2013.
[4] D Cai et al. Improved tools for the brainbow toolbox. Nature methods, 10(6):540–547, 2013.
[5] F Chen et al. Expansion microscopy. Science, 347(6221):543–548, 2015.
[6] K Chung and K Deisseroth. Clarity for mapping the nervous system. Nat. methods, 10(6):508–513, 2013.
[7] E Betzig et al. Imaging intracellular fluorescent proteins at nanometer resolution. Science, 313(5793):1642–
1645, 2006.
[8] M J Rust et al. Sub-diffraction-limit imaging by stochastic optical reconstruction microscopy (storm).
Nature methods, 3(10):793–796, 2006.
[9] A M Zador et al. Sequencing the connectome. PLoS Biol, 10(10):e1001411, 2012.
[10] J H Lee et al. Highly multiplexed subcellular rna sequencing in situ. Science, 343(6177):1360–1363, 2014.
[11] K H Chen et al.
Science,
Spatially resolved, highly multiplexed rna profiling in single cells.
348(6233):aaa6090, 2015.
[12] M Maggioni et al. Nonlocal transform-domain filter for volumetric data denoising and reconstruction.
Image Processing, IEEE Transactions on, 22(1):119–133, 2013.
[13] U Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007.
[14] D Lee and H S Seung. Algorithms for non-negative matrix factorization. In Advances in neural information
processing systems, pages 556–562, 2001.
[15] F Meyer. Topographic distance and watershed lines. Signal processing, 38(1):113–125, 1994.
[16] F Meyer. Minimum spanning forests for morphological segmentation. In Mathematical morphology and
its applications to image processing, pages 77–84. Springer, 1994.
[17] J Cousty et al. Watershed cuts: Minimum spanning forests and the drop of water principle. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 31(8):1362–1374, 2009.
[18] Xiaofeng Ren and Jitendra Malik. Learning a classification model for segmentation. In Computer Vision,
2003. Proceedings. Ninth IEEE International Conference on, pages 10–17. IEEE, 2003.
[19] J S Kim et al. Space-time wiring specificity supports direction selectivity in the retina. Nature,
509(7500):331–336, 2014.
[20] Gilles Bertrand and Grégoire Malandain. A new characterization of three-dimensional simple points.
Pattern Recognition Letters, 15(2):169–175, 1994.
[21] Viren Jain, Benjamin Bollmann, Mark Richardson, Daniel R Berger, Moritz N Helmstaedter, Kevin L
Briggman, Winfried Denk, Jared B Bowden, John M Mendenhall, Wickliffe C Abraham, et al. Boundary
learning by optimization with topological constraints. In Computer Vision and Pattern Recognition (CVPR),
2010 IEEE Conference on, pages 2488–2495. IEEE, 2010.
[22] J L Bentley. Multidimensional binary search trees used for associative searching. Communications of the
ACM, 18(9):509–517, 1975.
[23] Z Wu and R Leahy. An optimal graph theoretic approach to data clustering: Theory and its application to
image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 15(11):1101–1113, 1993.
[24] A Y Ng et al. On spectral clustering: Analysis and an algorithm. Advances in neural information processing
systems, 2:849–856, 2002.
[25] K Kurihara et al. Collapsed variational dirichlet process mixture models. In IJCAI, volume 7, pages
2796–2801, 2007.
[26] Jeffrey W Miller and Matthew T Harrison. A simple example of dirichlet process mixture inconsistency for
the number of components. In Advances in neural information processing systems, pages 199–206, 2013.
[27] Lawrence Hubert and Phipps Arabie. Comparing partitions. Journal of classification, 2(1):193–218, 1985.
[28] U Sümbül et al. A genetic and computational approach to structurally classify neuronal types. Nature
communications, 5, 2014.
[29] U Sümbül et al. Automated computation of arbor densities: a step toward identifying neuronal cell types.
Frontiers in neuroscience, 2014.
[30] M H Longair et al. Simple neurite tracer: open source software for reconstruction, visualization and
analysis of neuronal processes. Bioinformatics, 27(17):2453–2454, 2011.
[31] H Peng et al. V3d enables real-time 3d visualization and quantitative analysis of large-scale biological
image data sets. Nature biotechnology, 28(4):348–353, 2010.
9
Supplementary information: Automated scalable
segmentation of neurons from multispectral images
Supplementary Text
Parameter choices and feature representations
Similar results are obtained around a neighborhood of these suggested values. Unless otherwise
noted, the same parameter values were used for all three reported experiments: two real hippocampal
datasets acquired by different labs and under different conditions, and a set of simulated retinal
datasets with different parameters. The subsection of the main text that refers to these parameters are
indicated in square brackets.
Color features: For every color triplet, we obtain the L-u-v representation (Schanda, 2007), and
calculate the top C principal components of the concatenated L-u-v representations, where C is the
number of color channels. If the data has more than 3 channels, for affinity calculations between
neighboring supervoxels, we normalize the colors before the L-u-v transformation. [Dimensionality
reduction]
Supervoxel reliability: ts = 50, td = 0.5 (before L-u-v transformation) [Clustering the supervoxel
set]
Edge set parameters: s = √3 (26-neighborhood for isotropic data), c = 20 ×pC/4 (by inspect-
ing typical color radius within individual neurons and manual adjustment), kmin = 5. [Clustering the
supervoxel set]
Edge strength decay: α = 2 × 10−3 (by inspecting typical color radius and manual adjustment)
[Clustering the supervoxel set]
Flooding parameter for watershed: f = 0.01 with 26-neighborhood. (This affects computation
time more than quality because subdividing via the maximum color perimeter can catch inhomoge-
neous supervoxels.) [Dimensionality reduction]
Maximum color perimeter for supervoxel homogeneity: (Supp. Algo. 1) p = 0.5 for each channel
when the intensity is in [0, 1] (by inspecting data – see Fig. 1). [Dimensionality reduction]
Image thresholding for warping: θ = 0.1 ×pC/4 before L-u-v transformation (for the expansion
microscopy data, θ = 0.2) [Dimensionality reduction]
Noise standard deviation for denoising: σ = 1/8 when the intensity is in [0, 1] for individual
channels. [Denoising the image stack]
Cluster (neuron) counts: For the dataset in Fig. 5, the mixture model (Kurihara et al., 2007)
suggested 52 clusters based on the colors of the supervoxels. The same routine returned 29 clusters
when run on 1
5 of the supervoxels. We chose K = 34 for a compact presentation. For the dataset in
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
arXiv:1611.00388v2 [q-bio.NC] 22 Jan 2017
Supp. Fig. 8, we used K = 19, which is what the mixture model suggested based on the colors of the
supervoxels. [Clustering the supervoxel set]
Spatial distance calculation
The spatial distance between two supervoxels is calculated as minv∈V1,u∈V2 D(v, u), where V1 and
V2 are the voxel sets of the two supervoxels and D(v, u) is the Euclidean distance between voxels v
and u. Only the boundary voxels need to be considered, and extremal values in each coordinate are
used to identify many supervoxel pairs farther than s without exact calculation over the voxels. Only
the spatial distances between nearby supervoxels need to be computed.
Color-based subdivision of supervoxels
Let the n × C matrix Vi denote the colors of all n voxels of the supervoxel si. Supp. Algo. 1 divides
the supervoxels into smaller supervoxels until the desired homogeneity is achieved.
Supplementary Algorithm 1 Subdivide supervoxels
Require: S = {si}i (set of supervoxels), {Vi}i, pmax (threshold)
1: Snew = {}
2: for si ∈ S do
3:
4:
5:
6:
7:
p = maxc∈C max(Vi(:, c)) − min(Vi(:, c))
if p < pmax then
Add si to Snew
else
Divide the voxels into 2 sets T1 and T2 based on color (e.g., using k-means, hierarchical clustering,
etc.)
Add the connected components of T1 and T2 to S
end if
Remove si from S
8:
9:
10:
11: end for
12: return Snew
Simulation data
RGC arbors stratify in the retina, distributing their dendritic length within a slab. To achieve denser
simulations, we did not shift the neurons much in z (The density numbers calculated in Fig. 4 are
obtained by considering the [35µm, 50µm] region in Supp. Fig. 3.) We obtain simulated stacks by
varying the expression density (S ∈ {5, 9, 13}), the channel count (C ∈ {3, 4, 5}), the neuron color
consistency (σ1 ∈ {0.01, 0.02, 0.03, 0.04, 0.1}), and the background noise (σ2 ∈ {0.05, 0.1}).
The random walk on a connected component assigns a color to a voxel by (i) calculating the mean
color of the neighboring voxels that were previously visited, and (ii) adding the random noise step to
this mean value.
Adjusted Rand index
We quantify the segmentation quality of the voxels of the simulated dataset using the adjusted Rand
index. The Rand index is a measure of the element pairs on which two partitions P and P of the same
set with N elements agree: R(P, P ) = 1 −(cid:0)N
2(cid:1)Pi<j δ(li, lj) − δ(li, lj), where li (li) denotes
the label of element i according to P ( P ), and δ is the indicator function. (δ(li, lj) = 1 if li = lj,
δ(li, lj) = 0 otherwise.) The adjusted Rand index corrects for chance, has a more sensitive dynamic
2
range, and is defined as A = (R − E)/(M − E), where E is the expected value of the index and
M is the maximum value of the index, based on the number of elements in individual segments. Its
maximum value is 1 (perfect agreement), and the expected value of the index is 0 for random clusters.
For the foreground based calculation, only the voxels that are assigned to the foreground after the
watershed transform and warping are considered. For the image based calculation, all voxels are
considered and the background is treated as a separate object.
Merging supervoxels
We apply local heuristics and spatio-color constraints iteratively to further reduce the data size and
demix overlapping neurons in voxel space (Fig. 2): (i) supervoxels occupied by more than one neuron
are detected and demixed by monitoring the improvement in non-negative least squares fit quality.
(Supp. Algo 2.) (ii) neighboring supervoxels with similar colors and orientations, supervoxels with
single spatial neighbors, and supervoxels all of whose neighbors have similar colors are merged. (iii)
supervoxels that are spatial neighbors and that are assigned to the same cluster by an overclustering
color k-means routine are merged. We implement (iii) to run in parallel over subgraphs of the full
graph for scalability. A rough estimate of the number of neurons required by the oversegmentation
routine is obtained by a Dirichlet process mixture model (Kurihara et al., 2007). The k-means
algorithm uses a multiple of this rough estimate. Note that only a rough estimate (Miller and Harrison,
2013) is needed because of oversegmentation (Supp Algo. 3). This algorithm can be implemented to
run in parallel over subgraphs of the full dataset.
if si has less than M voxels then
2 subject to x ≥ 0
size), ∆ (maximum color distance), f (improvement factor)
Initialize r = V (i, :)2
for each neighbor pair (i1, i2) do
2, P = (0, 0)
t = minx V ([i1, i2], :)x − V (i, :)2
if t < r then
if the minimum color distance between si and its neighbors is larger than ∆ then
Retain the neighbors that have neighbors with color distance less than ∆
if si has more than one spatial neighbors then
Supplementary Algorithm 2 Demixing of supervoxels
Require: S = {si}i, V (matrix of normalized supervoxel colors), A (spatial affinity matrix), M (maximum
1: for si ∈ S do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21: end for
22: return S, V , A
Assign the voxels of si to both of sP (1) and sP (2)
Update the spatial affinities of sP (1) and sP (2) in A accordingly
Remove si from S, V , and A
r = t, P = (i1, i2)
end if
end for
if r < (∆/f )2 then
end if
end if
end if
end if
3
factor)
Supplementary Algorithm 3 Spatio-color merging of supervoxels
Require: S = {si}i, (set of supervoxels) K (rough estimate of the number of clusters), k (oversegmentation
1: Snew = {}
2: Divide S into kK clusters based on the colors of the supervoxels, using k-means
3: for κ1 ∈ {1, . . . , kK} do
4:
5:
6: end for
7: return Snew
Find the connected components within the cluster κ1
Merge the supervoxels within the connected components of that cluster, and add to Snew
Supplementary Figure 1: Top: Maximum intensity projection of a raw Brainbow image. Bottom left: Fore-
ground after watershed transform. Arrows point to six different thin dendritic pieces ("bridges") that were
missed. Bottom right: Foreground after warping correction. Scale bar, 30µm
4
Supplementary Figure 2: The z (top left), x (top right), and y (bottom left) maximum intensity projections
of the raw simulation image shown in Fig. 3. Adjusted Rand index of the segmentation is 0.80.
5
Supplementary Figure 3: The z-profile of the ground truth simulation image with 13 neurons, showing
that a region is preferentially occupied. The range [0µm, 50µm] corresponds to slices 1 to 100 so that
most of the neuronal arbors are between slices 70 and 100.
6
10
40
distance from the top of the stack (µ m)
20
30
50
0.35
0.3
0.25
0.2
0.15
0.1
0.05
ratio of occupied voxels
0
0
Supplementary Figure 4: A 60 × 60 patch from a single slice (slice 90) of the simulation image shown in
Fig. 3. Top: raw. Bottom: denoised. Physical size: 15µm × 15µm
7
A
B
Supplementary Figure 5: Aggressive merging generates supervoxels with inconsistent colors. Top: Cluster 4
in the bottom part of Fig. 3 of the main text. Bottom: Cluster 9 in the bottom part of Fig. 3 of the main text.
Close inspection reveals that some of the multi-colored regions comprise single supervoxels.
8
Supplementary Figure 6: Segmentation of a simulated Brainbow image stack. Adjusted Rand index of
the foreground is 0.87. Pseudo-color representation of 5-channel data with more conservative supervoxel
merging compared to Fig. 3 of the main text. Maximum intensity projection of the segmentation. The top-left
corner shows the whole image stack. All other panels show the maximum intensity projections of the supervoxels
assigned to a single cluster (inferred neuron).
Supplementary Figure 7: Segmentation accuracy of simulated data a, Expression density (ratio of voxels
occupied by at least one neuron) vs. ARI. b, σ1 vs. ARI. c, Channel count vs. ARI for a 9-neuron simulation,
where K ∈ [6, 12]. ARI is calculated for all voxels.
9
0.74
0.735
0.73
0.725
0.72
0.715
0.71
0.705
0.7
0.695
3
4
true (9)
channel count
6
7
8
10
11
12
0.75
0.74
0.73
0.72
0.71
0.7
0.69
0.68
0.67
0.66
0.06
0.08
0.1
0.12
0.14
0.16
0.18
3 ch.
expression density (ratio of occupied voxels)
adjusted Rand index
4 ch.
5 ch.
0.8
0.78
0.76
0.74
0.72
0.7
0.68
0.66
0.64
0.62
0
0.02
0.04
0.06
0.08
0.1
3 ch.
step size (σ1) −− range per channel: [0, 1]
4 ch.
5 ch.
Supplementary Figure 8: Segmentation of a 4× linearly expanded Brainbow stack – best viewed digitally.
The physical size of the stack is 90µ × 76µ × 5µ. The top-left corner shows the maximum intensity projection
of the whole image stack, all other panels show the maximum intensity projections of the supervoxels assigned
to a single cluster (inferred neuron).
Parallelization and complexity
The denoising step (collaborative filtering) parallelizes over substacks (voxels) because the extra
dimension is formed by local patches.
The watershed algorithm can also be parallelized over substacks. Moreover, it has a linear run-time
with respect to the input size (Main Text).
Similarly, warping (shrinking) can be applied on those substacks. Querying the boundary voxels for
flipping at each stage, and ordering them by brightness result in a fast implementation. It has a linear
run-time with respect to the input size. (Simple voxel query can be performed over 3 × 3 × 3 patches
– See references in main text.)
10
Supervoxel merging can be performed in parallel on individual substacks. Merging subroutines use
local rules to make local merges except for the spatio-color merging step, which uses the k-means
algorithm.
Similarity calculations require extracting spatial and color neighborhoods of the supervoxels. Spatial
distance calculation is discussed above. These value are precalculated. As mentioned in the main
text, k-d tree structures are used to retrieve the color neighborhoods efficiently.
Finally, clustering is performed by the normalized cuts algorithm. We use an implicitly restarted
block Lanczos method for computing the first few eigenvectors (Baglama et al., 2003).
References
Baglama, J., Calvetti, D., and Reichel, L. (2003). Algorithm 827: irbleigs: A matlab program for computing a
few eigenpairs of a large sparse hermitian matrix. ACM Transactions on Mathematical Software (TOMS),
29(3):337–348.
Kurihara, K., Welling, M., and Teh, Y. W. (2007). Collapsed variational dirichlet process mixture models. In
IJCAI, volume 7, pages 2796–2801.
Miller, J. W. and Harrison, M. T. (2013). A simple example of dirichlet process mixture inconsistency for the
number of components. In Advances in neural information processing systems, pages 199–206.
Schanda, J. (2007). Colorimetry: understanding the CIE system. John Wiley & Sons.
11
|
1210.3555 | 2 | 1210 | 2013-10-30T15:30:56 | Task-Based Core-Periphery Organisation of Human Brain Dynamics | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.AO",
"stat.AP"
] | As a person learns a new skill, distinct synapses, brain regions, and circuits are engaged and change over time. In this paper, we develop methods to examine patterns of correlated activity across a large set of brain regions. Our goal is to identify properties that enable robust learning of a motor skill. We measure brain activity during motor sequencing and characterize network properties based on coherent activity between brain regions. Using recently developed algorithms to detect time-evolving communities, we find that the complex reconfiguration patterns of the brain's putative functional modules that control learning can be described parsimoniously by the combined presence of a relatively stiff temporal core that is composed primarily of sensorimotor and visual regions whose connectivity changes little in time and a flexible temporal periphery that is composed primarily of multimodal association regions whose connectivity changes frequently. The separation between temporal core and periphery changes over the course of training and, importantly, is a good predictor of individual differences in learning success. The core of dynamically stiff regions exhibits dense connectivity, which is consistent with notions of core-periphery organization established previously in social networks. Our results demonstrate that core-periphery organization provides an insightful way to understand how putative functional modules are linked. This, in turn, enables the prediction of fundamental human capacities, including the production of complex goal-directed behavior. | q-bio.NC | q-bio | Task-Based Core-Periphery Organization of Human Brain
Dynamics
Danielle S. Bassett1,2,∗, Nicholas F. Wymbs5, M. Puck Rombach3,4, Mason A. Porter3,4, Peter J.
Mucha6,7, Scott T. Grafton5
1 Department of Physics, University of California, Santa Barbara, CA 93106, USA
2 Sage Center for the Study of the Mind, University of California, Santa Barbara, CA
93106, USA
3 Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute,
University of Oxford, Oxford OX1 3LB, UK
4 CABDyN Complexity Centre, University of Oxford, Oxford, OX1 1HP, UK
5 Department of Psychological and Brain Sciences and UCSB Brain Imaging Center,
University of California, Santa Barbara, CA 93106, USA
6 Carolina Center for Interdisciplinary Applied Mathematics, Department of
Mathematics, University of North Carolina, Chapel Hill, NC 27599, USA
7 Institute for Advanced Materials, Nanoscience & Technology, University of North
Carolina, Chapel Hill, NC 27599, USA
∗ E-mail: Corresponding [email protected]
Abstract
As a person learns a new skill, distinct synapses, brain regions, and circuits are engaged and change over
time. In this paper, we develop methods to examine patterns of correlated activity across a large set of
brain regions. Our goal is to identify properties that enable robust learning of a motor skill. We measure
brain activity during motor sequencing and characterize network properties based on coherent activity
between brain regions. Using recently developed algorithms to detect time-evolving communities, we
find that the complex reconfiguration patterns of the brain's putative functional modules that control
learning can be described parsimoniously by the combined presence of a relatively stiff temporal core
that is composed primarily of sensorimotor and visual regions whose connectivity changes little in time
and a flexible temporal periphery that is composed primarily of multimodal association regions whose
connectivity changes frequently. The separation between temporal core and periphery changes over the
course of training and, importantly, is a good predictor of individual differences in learning success.
The core of dynamically stiff regions exhibits dense connectivity, which is consistent with notions of
core-periphery organization established previously in social networks. Our results demonstrate that core-
periphery organization provides an insightful way to understand how putative functional modules are
linked. This, in turn, enables the prediction of fundamental human capacities, including the production
of complex goal-directed behavior.
Author Summary
When someone learns a new skill, his/her brain dynamically alters individual synapses, regional activity,
and larger-scale circuits. In this paper, we capture some of these dynamics by measuring and character-
izing patterns of coherent brain activity during the learning of a motor skill. We extract time-evolving
communities from these patterns and find that a temporal core that is composed primarily of primary sen-
sorimotor and visual regions reconfigures little over time, whereas a periphery that is composed primarily
of multimodal association regions reconfigures frequently. The core consists of densely connected nodes,
and the periphery consists of sparsely connected nodes. Individual participants with a larger separation
between core and periphery learn better in subsequent training sessions than individuals with a smaller
separation. Conceptually, core-periphery organization provides a framework in which to understand how
arXiv:1210.3555v2 [q-bio.NC] 30 Oct 2013
2
putative functional modules are linked. This, in turn, enables the prediction of fundamental human
capacities, including the production of complex goal-directed behavior.
Introduction
Cohesive structures have long been thought to play an important role in information processing in the
human brain [1]. At the small scale of individual neurons, temporally coherent activity supports in-
formation transfer between cells [2]. At a much larger scale, simultaneously active cortical areas form
functional systems that enable behavior [1]. However, the question of precisely what type of cohesive
organization is present between the constituents of brain systems -- especially at larger scales -- has been
steeped in controversy [3, 4]. Although low-frequency interactions between pairs of brain areas are easy
to measure, the simultaneous characterization of dynamic interactions across the entire human brain
remained challenging until recent applications of network theory to neuroimaging data [5]. These efforts
have led to enormous insights, including the establishment of relationships between stationary functional
brain network configuration and intelligence [6] as well as relationships between altered brain network
organization and disease [7]. In this paper, we extend this approach to a non-stationary situation: the
change of network activity across the brain as a new skill is acquired.
Acquisition of new motor skills alters brain activity across spatial scales. At the level of individual
neurons, this induces changes in firing behavior in the motor cortex [8]. At the level of large-scale
areas, this induces changes in the interactions between primary motor cortex and premotor areas, and
these changes can influence the amount of learning [9]. Previous studies have demonstrated that pairwise
interactions between some of these premotor regions, as measured by the magnitude of coherence between
low-frequency blood-oxygen-level-dependent (BOLD) signals, strengthen with practice [10]. Furthermore,
complex contributions by non-motor systems such as prefrontal cortex are involved in the strategic control
of behavior during learning [11]. These findings reveal some of the changes in local circuits that occur
with learning. However, there remains no global assessment of changes in brain networks as a result of
learning. In this paper, we seek to find cohesive structures in global brain networks that capture dynamics
that are particularly relevant for characterizing skill learning that takes place over the relatively long time
scales of minutes to hours of practice.
To address these issues, we extract a set of functional networks from task-based functional Magnetic
Resonance Imaging (fMRI) time series that describe functional connectivity between brain regions. We
probe the dynamics of these putative interactions by subdividing time series into discrete time intervals
(of approximately two minutes in duration; see Fig. 1A) during the acquisition of a simple motor skill.
Subjects learned a set of 10-element motor sequences similar to piano arpeggios by practicing for at
least 30 days during a 6-week period. The depth of training was manipulated so that 2 sequences
were extensively practiced (EXT), 2 sequences were moderately practiced (MOD), and 2 sequences were
minimally practiced (MIN) on each day. In addition, subjects performed blocks of all of the sequences
during fMRI scanning on approximately days 1, 14, 28, and 42 of practice. Using the fMRI time series,
we extract functional networks representing the coherence between 112 cortical and subcortical areas for
each sequence block.
To characterize brain dynamics, we represent sets of functional networks as multilayer brain networks
and we identify putative functional modules -- i.e., groups of brain regions that exhibit similar BOLD time
courses -- in each 2 -- 3 minute time window. Such cohesive groups of nodes are called "communities" in the
network-science literature [12, 13], and they suggest that different sets of brain regions might be related
to one another functionally either through direct anatomical connections or through indirect activation
by an external stimulus. A community of brain regions might code for a different function (e.g., visual
processing, motor performance, or cognitive control), or it might engage in the same function using a
distinct processing stream. Characterizing changes in community structure thus makes it possible to map
meaningful dynamic patterns of functional connectivity that relate to changes in cognitive function (e.g.,
3
learning).
We employ computational tools for dynamic community detection [14, 15] for multilayer represen-
tations of temporal networks [16] and summarize our findings using diagnostics that quantify three
properties of community structure. (See Materials and Methods for their definitions and Ref. [17] for
evidence supporting the utility of these diagnostics in capturing changes in brain dynamics over 3 days of
learning.) To measure the strength of functional modularization in the brain and quantify the extent of
compartmentalization of putative functional modules, we maximize a quality function called multilayer
modularity Q to obtain a partition of the brain into communities. (The associated maximum value of
Q is known as the maximum modularity.) A high value of Q indicates that the pattern of functional
connectivity in the brain can be clustered sensibly into distinct communities of brain regions that exhibit
similar time courses. We also compute the number n of communities (i.e., putative functional modules)
in partitions of the multilayer networks. A large value of n indicates that there are a large number
of distinct temporal profiles in BOLD activations in the brain. To measure the temporal variability of
community structure, we compute the flexibility fi of each region i, as this quantifies the frequency that
a brain region i changes its allegiances to network communities over time. A high value of flexibility
indicates that a region often changes community affiliation.
Our results demonstrate that the temporal evolution of community structure is modulated strongly
by the depth of training (as reflected in the total number of practiced trials). We also show that the
temporal variability of module allegiance varies across brain regions. Sensorimotor and visual cortices
form the bulk of a relatively stiff temporal core in which module affiliations change little over a scanning
session, whereas multimodal association areas form the bulk of a relatively flexible temporal periphery
in which module affiliations change frequently. The separation between the temporal core and temporal
periphery predicts individual differences in extended learning. We combine these methods for identifying
a temporal core and periphery with a notion of core-periphery organization that originated in the social
sciences [18] to show that the organizational structure of functional networks in 2 -- 3 minute time windows
correlates with the organizational structure of the brain's temporal evolution: densely connected regions
in individual time windows tend to exhibit little change in module allegiance over time, whereas weakly
connected regions tend to exhibit significant changes. Taken together, our results suggest that core-
periphery organization is a critical property that is as important as modularity for understanding and
predicting cognition and behavior (see Fig. 1B).
Results
Dynamic Community Structure Changes with Learning
Community structure changes with the number of trials practiced, independent of when the practice
occurred in the 6 weeks. In Fig. 2, we show multilayer modularity (Q, a measure of the quality of a
partition into communities), the number of communities, and mean flexibility (F = 1
i=1 fi, a measure
of the temporal variability in module allegiance) as a function of the number of training trials completed
after a scanning session. See Materials and Methods for the definitions. After an initial increase from 50
to 200 trials practiced, multilayer modularity decreases with an increase in the number of trials practiced,
suggesting that community structure in functional brain networks becomes less pronounced with learning.
Both the number of communities and the flexibility of community structure increase with the number of
trials practiced, which is consistent with an increased specificity of functional connectivity patterns with
extended learning.
NPN
4
Temporal Core-Periphery Organization
Regional Variation in Flexibility. The mean flexibility over participants varied over brain regions.
It ranged from approximately 0.04 to approximately 0.14, which implies that brain regions changed their
modular affiliation between 4% and 14% of trial blocks on average (see Fig. 3A). The distribution of
flexibility across brain regions is decidedly non-Gaussian: the majority of brain regions have relatively
high flexibilities, but there is a left-heavy tail of regions (including a small peak) with low values of
flexibility. We characterized the distribution of flexibility over brain regions by calculating the third
(skewness) and fourth (kurtosis) central moments. The skewness was 0.50 ± 0.26, and the kurtosis was
3.04 ± 0.57. To interpret these findings, we note that a distribution's skewness is a measure of its
asymmetry, and the positive values that we observe indicate that the distributions from all participants
are skewed to the right. The kurtosis of a distribution is a combined measure of its peakedness [19] and
its bimodality [20], and it is sometimes construed as a measure of the extent that a distribution is prone
to outliers. The kurtosis values that we observe vary between 2.5 and 5, which includes the value of 3
that occurs for a Gaussian distribution.
Defining the Temporal Core and Temporal Periphery. To determine the significance of a brain
region's variation in flexibility, we compared the flexibility of brain regions in the empirical multilayer
network to that expected in a nodal null model. We can define a temporal core, bulk, and periphery
(see Fig. 3). The core is the set of regions whose flexibility is significantly less than expected in the null
model; the periphery is the set of regions whose flexibility is significantly greater than expected in the
null model; and the bulk consists of all remaining regions. As discussed in the Text S1, the delineation
of the brain into these three groups is robust both to the intensity of training (MIN, MOD, or EXT) and
to the duration of training (sessions 1 -- 4). Furthermore, the temporal core, bulk, and periphery tend to
form their own communities, although the relationship between core-periphery organization and modular
organization appears to be altered by learning (see the Text S1).
We show the anatomical locations of the temporal core, temporal bulk, and temporal periphery in
Fig. 3. The relatively stiff core is composed of 19 regions located predominantly in primary sensorimotor
areas in both left and right hemispheres. Most of the motor-related regions in the core were left-lateralized,
which is consistent with the participants' use of their right hand to perform the motor sequence. The
more flexible periphery is composed of 25 regions located predominantly in multimodal areas -- including
inferior parietal, intraparietal sulcus, temporal parietal junction, inferotemporal, fusiform gyrus, and
visual association areas. The bulk contains the remaining 68 cortical and subcortical regions -- including
large swaths of frontal, temporal, and parietal cortices. See Table S3 for a complete list of the affiliation
of each brain region to the temporal core, bulk, and periphery. The separation of a temporally stiff core
of predominantly unimodal regions that process information from single sensory modality (e.g., vision,
audition, etc.) and a flexible periphery of predominantly multimodal cortices that process information
from multiple modalities is consistent with existing understanding of the association of multimodal cortex
with the binding of different types of information and the performance of a broad range of cognitive
functions [21].
Temporal Core-Periphery Organization and Learning. One can interpret the anatomical location
of the temporally stiff core that consists primarily of unimodal regions and a flexible periphery that
consists primarily of multimodal cortices in the context of the known roles of these cortices in similar
tasks. The ability to retrieve and rapidly execute complex motor sequences requires extensive practice.
These well-learned sequences are known to be generated by "core" areas [8, 22 -- 24]. However, when first
learning a sequence, people can use a variety of cognitive strategies that are supported by other brain
systems (some of which are located in the periphery to augment performance) [25,26]. In some cases, these
strategies are detrimental to skill retention [27]. Consequently, we hypothesized that individuals whose
core and periphery are distinct -- indicating a strong separability of visuomotor and cognitive regions --
would learn better than those whose core and periphery were less distinguishable from one another.
5
.=
To test this hypothesis, we calculated the Spearman rank coefficient ρ between the skewness and
kurtosis of the flexibility distribution estimated from the fMRI data of the first scanning session and the
learning parameter κ estimated over the next 10 days of home training (see Materials and Methods). The
kurtosis (in essence) measures the separation between the temporal core and the temporal periphery and
is negatively correlated with κ (the correlation is ρ
0.027), indicating
that individuals with a narrower separation between temporal core and temporal periphery learn better
in the subsequent 10 home training sessions than individuals with a greater separation between temporal
core and temporal periphery. Skewness (in essence) measures the presence of -- rather than the separation
between -- the temporal core and the temporal periphery and is also negatively correlated with learning
(ρ
0.034), indicating that individuals whose flexibility was more skewed over brain
regions learned better than those whose flexibility over brain regions was less skewed. This finding implies
that individuals with "stronger" temporal core-periphery structure (i.e., larger values of skewness) learn
better than those with "weaker" temporal core-periphery structure (i.e., smaller values of skewness).
−0.498, and the p-value is p
.=
.=
−0.480 and p
.=
Importantly, the temporal separation of the data from the scanning session (which we used to esti-
mate brain flexibility) and the home training (which we used to estimate learning) ensures that these
correlations are predictive. Together, these results indicate that individuals with a stronger temporal core
and temporal periphery but a smooth transition between them seem to learn better than individuals with
a weaker temporal core and temporal periphery but a sharper transition between them. These results
suggest that successful brain function might depend on a delicate balance between a set of core regions
whose allegiance to putative functional modules changes little over time and a set of peripheral regions
whose allegiance to putative functional modules is flexible through time (and also on the smoothness of
the transition between these two types of regions).
Geometrical Core-Periphery Organization
Given our demonstration that there exists a temporal core in dynamic brain networks, it is important to
ask what role such core regions might play in individual network layers of the multilayer network [17].
While the roles of nodes in a static network can be studied in multiple ways [28,29], we focus on describing
the geometrical core-periphery organization -- which can be used to help characterize the organization
of edge strengths throughout a network -- to compare it with the temporal core-periphery organization
discussed above. The geometrical core of a network is composed of a set of regions that are strongly
and mutually interconnected. Measures of network centrality can be useful for identifying nodes in a
geometrical core because such measures help capture a node's relative importance within a network in
terms of its immediate connections, its distance to other nodes in the graph, or its influence on other
nodes in the graph [30, 31].
Drawing on studies of social networks [18], we examine geometrical core-periphery organization in
networks extracted from individual time windows by testing whether core nodes are densely connected to
one another and whether peripheral nodes are sparsely connected to one another. Rather than proposing
a strict separation between a single core and single periphery, we assess the role of a node along a
core-periphery spectrum using a centrality measure known as the (geometrical) core score C, which was
introduced in Ref. [30]. Network nodes with high C values are densely connected to one another, whereas
nodes with low C values are sparsely connected to one another. The method in Ref. [30] uses a two-
parameter function to interpolate between core nodes and peripheral nodes. One parameter (which is
denoted by α) sets the sharpness of the boundary between the geometrical core and the geometrical
periphery. Small values of α indicate a fuzzy boundary, whereas large values indicate a sharp transition.
The second parameter (which is denoted by β) sets the size of the geometrical core. Smaller values of β
correspond to smaller cores. We can quantify the fit of the transition function that defines the set of core
scores to the data using a summary diagnostic that is called the R-score (see Materials and Methods for
definitions). Large values of R indicate a good fit and therefore provide confidence that one has uncovered
a good estimate of a network's core-periphery organization.
6
.=
0.94) to assign a core score to each node.
In Fig. 4A, we show a typical R-score landscape in the (α, β) parameter plane. This landscape favors
a relatively small core and a medium value of the transition-sharpness parameter. To choose sensible
values of α and β for studying core-periphery organization, we examine the distributions of the relative
frequencies of α and β values that maximize the R-score for each network layer, participant, scanning
session, and sequence type (see Fig. 4B). We use the mean values of these distributions (α
0.40 and
β
In Fig. 4C, we show the shape of the "mean core"
that we obtain using these parameter values. This figure demonstrates that the typical (geometrical)
core-periphery organization in the networks under study is a mixture between a discrete core-periphery
organization, in which every node is either in the core or in the periphery, and a continuous core-periphery
organization, in which there is a continuous spectrum to describe how strongly nodes belong to a core.
In these networks (which usually possess a single core), the majority of nodes do not belong to the core,
but those nodes that do (roughly 10% of the nodes) have a continuum level of association strengths with
the core.
.=
.=
In some cases, we identified multiple competing cores, which we found by using simulated annealing
to explore local maxima of the R-score rather than only identifying a global maximum. Because of this
stochasticity in the methodology for examining core-periphery organization, we performed computations
with the chosen parameter values (α
0.94) 10 times and used the solution with the
highest R-score out of these 10 iterations for each network layer, subject, scanning session, and sequence
type.
0.40 and β
.=
An interesting question is whether geometrical core-periphery organization remains relatively constant
throughout time or whether the organization changes with learning. We observed that regions that
have a high geometrical core score in the first scanning session and in EXT blocks were likely to have
high geometrical core scores in later scanning sessions and in MOD and MIN blocks.
(See the Text
S1 for supporting results on the reliability of geometrical core-periphery organization.) In light of this
consistency, we calculate a mean geometrical core score for each node by taking the mean over all blocks
in a given scanning session (1, 2, 3, and 4) and sequence type (EXT, MOD, and MIN). The variance
of the mean geometrical core score over nodes in a network then gives an indication of the separation
between the mean core and periphery. As we show in Fig. 5, we find that the variance of the mean
geometrical core score over trials decreases as a function of learning. A high variance of the mean core
scores over nodes indicates a greater separation between the mean core and periphery as well as a high
consistency of the core score of each node over trial blocks. If a node's core score is inconsistent over
trial blocks, then the mean core score for each node in the network is expected to be similar and thus
one would expect the variance of the core scores over nodes to be small. A low variance in the mean
geometrical core score over trial blocks therefore suggests either little separation between the core and
periphery or an increased variability in core scores of a given node over trial blocks.
Relationship Between Temporal and Geometrical Core-Periphery Organiza-
tion
Given the geometrical core-periphery organization in the individual layers of the multilayer networks
and the temporal core-periphery organization in the full multilayer networks, it is important to ask
whether brain regions in the temporal core (i.e., regions with low flexibility) are also likely to be in the
geometrical core (i.e., whether they exhibit strong connectivity with other core nodes, as represented by a
high value of the geometrical core score). In Fig. 6, we show scatter plots of the flexibility and core score
for the 3 training levels (EXT, MOD, and MIN) and the 4 scanning sessions over the 6-week training
period. We find that the temporal core-periphery organization (which is a dynamic measurement) is
strongly correlated with the geometrical core score (which is a measure of network geometry and hence of
7
network structure). This indicates that regions with low temporal flexibility tend to be strongly-connected
core nodes in (static) network layers. In Fig. 6, we show that the relationship between temporal and
geometrical core-periphery organization occurs reliability across training depth, duration, and intensity.
In Fig. 7, we show that this relationship can also be identified robustly in data extracted from individual
subjects.
Discussion
We have shown how the mesoscale organization of functional brain networks changes over the course of
learning. Our results suggest that core-periphery organization is an important and predictive component
of cognitive processes that support sequential, goal-directed behavior. We summarize our findings in
Fig. 7, which demonstrates that poor learners tend to have poorer separation between core and periphery
(as indicated by straighter, shorter spirals in the figure) and that good learners tend to have greater
separation between core and periphery (curvier, longer spirals). Our findings also demonstrate that during
the generation of motor sequences, the brain consists of a temporally stable and densely connected set of
core regions complemented by a temporally flexible and sparsely connected set of peripheral regions. This
functional tradeoff between a core and periphery might provide a balance between the rigidity necessary
to maintain motor function by the core and the adaptivity of the periphery necessary to enable behavioral
change as a function of context or strategy.
In the Text S1, we provide supporting results that indicate (i) that our findings are not merely a
function of variation in region size and (ii) that they cannot be derived from the underlying block design
of the experimental task. We also show in this supplement that (arguably) simpler properties of brain
function -- such as the regional signal power of brain activity, mean connectivity strength, and parameter
estimates from a general linear model -- provide less predictive power than core-periphery organization.
Core-Periphery Organization of Human Brain Structure and Function. The notion of a core-
periphery organization is based on the structure (rather than the temporal dynamics) of a network [30].
Intuitively, a core consists of a set of highly and mutually interconnected set of regions. In this paper, we
have described what is traditionally called "core-periphery structure" using the terminology geometrical
core-periphery organization. (It is geometrical rather than topological because the networks are weighted.)
This intuitive notion was formalized in social networks by Borgatti and Everett in 1999 [18]. Available
methods to identify and quantify geometrical core-periphery organization in networks include ones based
on block models [18], k-core organization [32], and aggregation of information about connectivity and
short paths through a network [33]. Unfortunately, many methods that have been used to study cores
and peripheries in networks have binarized networks that are inherently weighted, which requires one
to throw away a lot of important information. Even the recently developed weighted extensions of k-
core decomposition [34] require a discretization of k-shells, which have been defined for both binary and
weighted networks [35].
Importantly, k-core decomposition is based on a very stringent and specific
type of core connectivity, so this measure misses important core-like structures [30, 36]. A well-known
measure called the "rich-club coefficient" (RCC) [37] considers a different but somewhat related question
of whether nodes of high degree (defined as k ≥ kt some threshold value kt) tend to connect to other
nodes of high degree. (The RCC is therefore a form of assortativity.) The RCC has also been extended
to weighted networks [38], but it still requires one to specify a threshold value of richness to enable one
to ask whether "rich" nodes tend to connect to other "rich" nodes.
The aforementioned limitations notwithstanding, several of the measures discussed above have recently
been used successfully to identify a structural core of the human brain white-matter tract network, which
is characterized not only by a k-core with a high value of the degree k (in particular, k ≥ 20) [34] and rich
club [39,40] but also by a knotty center of nodes that have a high geodesic betweenness centrality but not
8
necessarily a high degree [36]. A k-core decomposition has also been applied to functional brain imaging
data to demonstrate a relationship between network reconfiguration and errors in task performance [41].
A novel approach that is able to overcome many of these conceptual limitations is the geometrical
core-score [30], which is an inherently continuous measure, is defined for weighted networks, and can be
used to identify regions of a network core without relying solely on their degree or strength (i.e., weighted
degree). Moreover, by using this measure, one can produce (i) continuous results, which make it possible
to measure whether a brain region is more core-like or periphery-like; (ii) a discrete classification of
core versus periphery; or (iii) a finer discrete division (e.g., into 3 or more groups). In addition, this
method can identify multiple geometrical cores in a network and rank nodes in terms of how strongly
they participate in different possible cores. This sensitivity is particularly helpful for the examination
of brain networks for which multiple cores are hypothesized to mediate multimodal integration [42]. In
this paper, we have demonstrated that functional brain networks derived from task-based data acquired
during goal-directed brain activity exhibit geometrical core-periphery organization. Moreover, they are
specifically characterized by a straightforward core-periphery landscape that includes a relatively small
core composed of roughly 10% or so of the nodes in the network.
In this paper, we have introduced a method and associated definitions to identify a temporal core-
periphery organization based on changes in a node's module allegiance over time. We have defined the
notion of a temporal core as a set of regions that exhibit fewer changes in module allegiance over time
than expected in a dynamic-network null model. Neurobiologically, the temporal core contains brain areas
that show consistent task-based mesoscale functional connectivity over the course of an experiment, and
it is therefore perhaps unsurprising that their anatomical locations differ from nodes in the (k ≥ 20)-
core [34] and RCC [39, 40] of the human white matter tract network. Our approach is inspired by the
following idea: although the brain uses the function of a small subset of regions to perform a given
task (i.e., some sort of core), a set of additional regions that are associated more peripherally with the
task might also be activated in a transient manner. Indeed, several recent studies have highlighted the
possibility of a separation between groups of regions that are consistently versus transiently activated
during task-related function [43, 44], and they have demonstrated that correlations between such regions
can be altered depending on their activity [43, 45].
Given the very different definitions of the geometrical and temporal cores, it is interesting that nodes
in the temporal core are also likely to be present in the geometrical core. Importantly, the notions of
temporal and geometrical core are complementary, and they are both intuitive in the context of brain
function. A set of regions that is coherently active to perform a task (i.e., is in the geometrical core)
must remain online consistently throughout an experiment (i.e., be in the temporal core), whereas a set
of regions that might be activated less coherently (i.e., is in the geometrical periphery) can be utilized
by separate putative functional modules over time (i.e., it can be in the temporal periphery). This
interpretation is consistent with the notion that the anatomical locations of the core and periphery are
task-specific. Should brain activity during other tasks also exhibit core-periphery organization, then the
core and periphery of these other task networks could consist of a different set of anatomical regions
than those observed here. A comparison of dynamic community structure and associated mesoscale
organizational properties across brain states elicited by other tasks is outside of the scope of the present
study. However, such a study in a controlled sample with similar time-series length and experimental
task structure (e.g., trial lengths, block lengths, and rest periods) would likely yield important insights.
Modular Versus Core-Periphery Organization. Community structure and core-periphery orga-
nization are two types of mesoscale structures, and they can both be present simultaneously in a net-
work [30,36]. Moreover, both modular and core-periphery organization can in principle pertain to different
characteristics of or constraints on underlying brain function. In particular, the presence of community
structure supports the idea of the brain containing putative functional modules, whereas the presence of a
core-periphery organization underscores the fact that different brain regions likely play inherently differ-
9
ent roles in information processing. A symbiosis between these two types of organization is highlighted by
the findings that we report in this manuscript: the dynamic reconfiguration of putative functional mod-
ules can be described parsimoniously by temporal core-periphery organization, demonstrating that one
type of mesoscale structure can help to characterize another. Furthermore, the notion that the brain can
simultaneously contain functional modules (e.g., the executive network or the default-mode network) and
regions that transiently mediate interactions between modules is consistent with recent characterizations
of attention and cognitive control processes [46].
Dynamic Brain Networks.
It is increasingly apparent that functional connectivity in the brain
changes over time and that these changes are biologically meaningful. Several recent studies have high-
lighted the temporal variability [47 -- 50] and non-stationarity [51] of functional brain network organization,
and both of these features are apparent over short time intervals (less than 5 minutes in fMRI; less than
100 s in EEG) [47 -- 49]. Although temporal variability in functional connectivity was seen initially as a
signature of measurement noise [51], recent evidence suggests instead that it might provide an indirect
measurement of changing cognitive processes. Thus, it might serve as a diagnostic biomarker of dis-
ease [51, 52]. Moreover, such temporal variability appears to be modulated by exogenous inputs. For
example, Barnes et al. [53] demonstrated using a continuous acquisition "rest-task-rest" design that en-
dogenous brain dynamics do not return to their pre-task state until approximately 18 minutes following
task completion. Similar results that consider other tasks have also been reported [54]. More generally,
the dynamic nature of brain connectivity is likely linked to spontaneous cortical processing, reflecting a
combination of both stable and transient communication pathways [48, 49, 55].
Network Predictions of Future Learning.
In this study, we observed that properties of the temporal
organization of functional brain networks (e.g., on day 1 of this experiment) can be used to predict
extended motor learning (e.g., on the following 10 days of home training on a discrete sequence-production
task). Our findings are consistent with two previous studies that demonstrated a predictive connection
between both dynamic [17] and topological [56] network organization and subsequent learning. (Note
that we use the term topological because Ref. [56] considered only unweighted networks.) Reference [17]
focused on early -- rather than extended -- learning of a cued sequence-production motor task (rather than
a discrete one) and found that network flexibility on the first day of experiments predicted learning on
the second day and that flexibility on the second day predicted learning on the third day. Reference [56]
investigated participants' success in learning words of an artificial spoken language and found that network
properties from individual time windows could be used to predict such success [56]. Together with
the present study, these results highlight the potential breadth of the relationship between network
organization and learning. The presence of such a relationship has now been identified across multiple
tasks, over multiple time scales, and using both dynamic and topological network properties.
Methodological Considerations. Our study has focused on large-scale changes in dynamic com-
munity structure that are correlated with learning. Finer-scale investigations that employ alternative
parcellation schemes [57 -- 62] with greater spatial resolution or alternative neuroimaging techniques such
as EEG or MEG [55] with greater temporal resolution might uncover additional features that would
enhance understanding of functional network-based predictors of learning phenomena.
Throughout this paper, we have referred to feature similarities (which we estimated using the mag-
nitude squared coherence) between pairs of regional BOLD time series as functional connectivity [63].
As appreciated in prior literature [64 -- 66], the interpretation of functional connectivity must be made
with caution. Coherence in the activity recorded at different brain sites does not necessitate that those
sites share information with one another to enable cognitive processing, as they could instead indicate
that those two sites are activated by the same third party (either another brain region or an external
stimulus). In this paper, we do not distinguish between these two possible drivers of strong inter-regional
coherence. Future studies could employ multiple estimates of statistical associations in the form of di-
agnostics [67 -- 69] and/or models [70, 71] that might uncover other sets of interactions that could predict
the observed coherence structure and hence the observed behavior.
10
Materials and Methods
Experiment and Data Acquisition
Ethics Statement
Twenty-two right-handed participants (13 females and 9 males; the mean age was about 24) volunteered
with informed consent in accordance with the Institutional Review Board/Human Subjects Committee,
University of California, Santa Barbara.
Experiment Setup and Procedure
We excluded two participants from the investigation: one participant failed to complete the experiment,
and the other had excessive head motion. Our investigation therefore includes twenty participants, who
all had normal/corrected vision and no history of neurological disease or psychiatric disorders. Each
of these participants completed a minimum of 30 behavioral training sessions as well as 3 fMRI test
sessions and a pre-training fMRI session. Training began immediately following the initial pre-training
scan session. Test sessions occurred after every 2-week period of behavioral training, during which at
least 10 training sessions were required. The training was done on personal laptop computers using a
training module that was installed by the experimenter (N.F.W.). Participants were given instructions
for how to run the module, which they were required to do for a minimum of 10 out of 14 days in a
2-week period. Participants were scanned on the first day of the experiment (scan 1), and then a second
time approximately 14 days later (scan 2), once again approximately 14 days later (scan 3), and finally
14 days after that (scan 4). Not all participants were scanned exactly every two weeks; see Table S1 for
details of the number of days that elapsed between scanning sessions.
We asked participants to practice a set of 10-element sequences that were presented visually using a
discrete sequence-production (DSP) task by generating responses to sequentially presented stimuli (see
Fig. 8) using a laptop keyboard with their right hand. Sequences were presented using a horizontal array
of 5 square stimuli; the responses were mapped from left to right, such that the thumb corresponded to the
leftmost stimulus and the smallest finger corresponded to the rightmost stimulus. A square highlighted in
red served as the imperative to respond, and the next square in the sequence was highlighted immediately
following each correct key press. If an incorrect key was pressed, the sequence was paused at the error
and was restarted upon the generation of the appropriate key press.
Participants had an unlimited amount of time to respond and to complete each trial. All participants
trained on the same set of 6 different 10-element sequences, which were presented with 3 different levels
of exposure. We organized sequences so that each stimulus location was presented twice and included
neither stimulus repetition (e.g., "11" could not occur) nor regularities such as trills (e.g., "121") or runs
(e.g., "123"). Each training session (see Fig. 9) included 2 extensively trained sequences ("EXT") that
were each practiced for 64 trials, 2 moderately trained sequences ("MOD") that were each practiced for
10 trials, and 2 minimally trained sequences ("MIN") that were each practiced for 1 trial. (See Table S1
for details of the number of trials composed of extensively, moderately, and minimally trained sequences
during home training sessions.) Each trial began with the presentation of a sequence-identity cue. The
purpose of the identity cue was to inform the participant what sequence they were going to have to type.
For example, the EXT sequences were preceded by either a cyan (sequence A) or magenta (sequence B)
circle. Participants saw additional identity cues for the MOD sequences (red or green triangles) and for
the MIN sequences (orange or white stars, each of which was outlined in black). No participant reported
11
any difficulty viewing the different identity cues. Feedback was presented after every block of 10 trials;
this feedback detailed the number of error-free sequences that the participant produced and the mean
time it took to complete an error-free sequence.
Each fMRI test session was completed after approximately 10 home training sessions (see Table S1
for details of the number of home practice sessions between scanning sessions), and each participant
participated in 3 test sessions. In addition, each participant had a pre-training scan session that was
identical to the other test scan sessions immediately prior to the start of training (see Fig. 9). To
familiarize participants with the task, we gave a brief introduction prior to the onset of the pre-training
session. We showed the participants the mapping between the fingers and the DSP stimuli, and we
explained the significance of the sequence-identity cues.
To help ease the transition between each participant's training environment and that of the scanner,
padding was placed under his/her knees to maximize comfort. Participants made responses using a
fiber-optic response box that was designed with a similar configuration of buttons as those found on the
typical laptop used during training. See the lower left of Fig. 8 for a sketch of the button box used in the
experiments. For instance, the center-to-center spacing between the buttons on the top row was 20 mm
(compared to 20 mm from "G" to "H" on a recent MacBook Pro), and the spacing between the top row
and lower left "thumb" button was 32 mm (compared to 37 mm from "G" to the spacebar on a MacBook
Pro). The response box was supported using a board whose position could be adjusted to accommodate
a participant's reach and hand size. Additional padding was placed under the right forearm to minimize
muscle strain when a participant performed the task. Head motion was minimized by inserting padded
wedges between the participant and the head coil of the MRI scanner. The number of sequence trials
performed during each scanning session was the same for all participants, except for two abbreviated
sessions that resulted from technical problems. In each case that scanning was cut short, participants
completed 4 out of the 5 scan runs for a given session. We included data from these abbreviated sessions
in this study.
Participants were tested inside of the scanner with the same DSP task and the same 6 sequences that
they performed during training. Participants were given an unlimited time to complete trials, though
they were instructed to respond quickly but also to maintain accuracy. Trial completion was signified by
the visual presentation of a fixation mark "+", which remained on the screen until the onset of the next
sequence-identity cue. To acquire a sufficient number of events for each exposure type, all sequences were
presented with the same frequency. Identical to training, trials were organized into blocks of 10 followed
by performance feedback. Each block contained trials belonging to a single exposure type and included 5
trials for each sequence. Trials were separated by an inter-trial interval (ITI) that lasted between 0 and 6
seconds (not including any time remaining from the previous trial). Scan epochs contained 60 trials (i.e.,
6 blocks) and consisted of 20 trials for each exposure type. Each test session contained 5 scan epochs,
yielding a total of 300 trials and a variable number of brain scans depending on how quickly the task was
performed. See Table S2 for details of the number of scans in each experimental block.
Behavioral Apparatus
Stimulus presentation was controlled during training using a participant's laptop computer, which was
running Octave 3.2.4 (an open-source program that is very similar to Matlab) in conjunction with
PsychtoolBox Version 3. We controlled test sessions using a laptop computer running Matlab version
7.1 (Mathworks, Natick, MA). We collected key-press responses and response times using a custom fiber-
optic button box and transducer connected via a serial port (button box: HHSC-1 × 4-L; transducer:
fORP932; Current Designs, Philadelphia, PA).
12
Behavioral Estimates of Learning
Our goal was to study the relationship between brain organization and learning. To ensure independence
of these two variables, we extracted brain network structure during the 4 scanning sessions, and we
extracted behavioral estimates of learning in home training sessions 1 -- 10 (approximately between days
1 and 14; see Table S1), which took place before scanning session 2.
For each sequence, we defined the movement time (MT) as the difference between the time of the first
button press and the time of the last button press during a single sequence. For the set of sequences of
a single type (i.e., sequence 1, 2, 3, 4, 5, or 6), we estimated the learning rate by fitting an exponential
function (plus a constant) to the MT data [72, 73] using a robust outlier correction in Matlab (using
the function fit.m in the Curve Fitting Toolbox with option "Robust" and type "Lar"):
M T = D1et/κ + D2 ,
(1)
where t is time, κ is the exponential dropoff parameter (which we call the "learning parameter") used to
describe the early (and fast) rate of improvement, and D1 and D2 are real and positive constants. The
sum D1+D2 is an estimate of the starting speed of a given participant prior to training, and the parameter
D2 is an estimate of the fastest speed attainable by that participant after extended training. A negative
value of κ indicates a decrease in MT, which is thought to indicate that learning is occurring [74, 75].
This decrease in MT has been used to quantify learning for several decades [76, 77]. Several functional
forms have been suggested for the fit of MT [78, 79], and the exponential (plus constant) is viewed as the
most statistically robust choice [79]. Additionally, the fitting approach that we used has the advantage
of estimating the rate of learning independent of initial performance or performance ceiling.
Functional MRI (fMRI) Imaging
Imaging Procedures
We acquired fMRI signals using a 3.0 T Siemens Trio with a 12-channel phased-array head coil. For each
scan epoch, we used a single-shot echo planar imaging sequence that is sensitive to BOLD contrast to
acquire 37 slices per repetition time (TR of 2000 ms, 3 mm thickness, 0.5 mm gap) with an echo time
(TE) of 30 ms, a flip angle of 90 degrees, a field of view (FOV) of 192 mm, and a 64 × 64 acquisition
matrix. Before the collection of the first functional epoch, we acquired a high-resolution T1-weighted
sagittal sequence image of the whole brain (TR of 15.0 ms, TE of 4.2 ms, flip angle of 9 degrees, 3D
acquisition, FOV of 256 mm, slice thickness of 0.89 mm, and 256 × 256 acquisition matrix).
fMRI Data Preprocessing
We processed and analyzed functional imaging data using Statistical Parametric Mapping (SPM8, Well-
come Trust Center for Neuroimaging and University College London, UK). We first realigned raw func-
tional data, then coregistered it to the native T1 (normalized to the MNI-152 template with a re-sliced
resolution of 3× 3× 3 mm), and finally smoothed it using an isotropic Gaussian kernel of 8 mm full-width
at half-maximum. To control for potential fluctuations in signal intensity across the scanning sessions,
we normalized global intensity across all functional volumes.
Network Construction
Partitioning the Brain into Regions of Interest
Brain function is characterized by spatial specificity: different portions of the cortex emit different, task-
dependent activity patterns. To study regional specificity of the functional time series and putative
interactions between brain areas, it is common to apply a standardized atlas to raw fMRI data [7, 80, 81].
13
The choice of atlas or parcellation scheme is the topic of several recent studies in structural [57,60], resting-
state [58], and task-based [59] network architecture. The question of the most appropriate delineation of
the brain into nodes of a network is an open one and is guided by the particular scientific question at
hand [5, 61].
Consistent with previous studies of task-based functional connectivity during learning [15, 17, 82], we
parcellated the brain into 112 identifiable cortical and subcortical regions using the structural Harvard-
Oxford (HO) atlas (see Table S3) installed with the FMRIB (Oxford Centre for Functional Magnetic
Resonance Imaging of the Brain) Software Library (FSL; Version 4.1.1) [83, 84]. For each individual
participant and each of the 112 regions, we determined the regional mean BOLD time series by separately
averaging across all of the voxels in that region.
Within each HO-atlas region, we constrained voxel selection to voxels that are located within an
individual participant's gray matter. To do this, we first segmented each individual participant's T1 into
white and gray matter volumes using the DARTEL toolbox supplied with SPM8. We then restricted the
gray-matter voxels to those with an intensity of 0.3 or more (the maximum intensity was 1.0). Note that
units are based on an arbitrary scale. We then spatially normalized the participant T1 and corresponding
gray matter volume to the MNI-152 template -- using the standard SPM 12-parameter affine registration
from the native images to the MNI-152 template image -- and resampled to 3 mm isotropic voxels. We
then restricted the voxels for each HO region by using the program fslmaths [83,84] to include only voxels
that are in the individual participant's gray-matter template.
Wavelet Decomposition
Brain function is also characterized by frequency specificity. Different cognitive and physiological func-
tions are associated with different frequency bands, and this can be investigated using wavelets. Wavelet
decompositions of fMRI time series have been applied extensively in both resting-state and task-based
conditions [85, 86]. In both cases, they provide sensitivity for the detection of small signal changes in
non-stationary time series with noisy backgrounds [87].
In particular, the maximum-overlap discrete
wavelet transform (MODWT) has been used extensively in connectivity investigations of fMRI [88 -- 93].
Accordingly, we used MODWT to decompose each regional time series into wavelet scales corresponding
to specific frequency bands [94].
We were interested in quantifying high-frequency components of an fMRI signal, correlations between
which might be indicative of cooperative temporal dynamics of brain activity during a task. Because our
sampling frequency was 2 seconds (1 TR = 2 sec), wavelet scale one provides information on the frequency
band 0.125 -- 0.25 Hz and wavelet scale two provides information on the frequency band 0.06 -- 0.125 Hz.
Previous work has indicated that functional associations between low-frequency components of the fMRI
signal (0 -- 0.15 Hz) can be attributed to task-related functional connectivity, whereas associations between
high-frequency components (0.2 -- 0.4 Hz) cannot [95]. This frequency specificity of task-relevant functional
connectivity is likely due at least in part to the hemodynamic response function, which might act as a
noninvertible band-pass filter on underlying neural activity [95]. Consistent with our previous work [17],
we examined wavelet scale two, which is thought to be particularly sensitive to dynamic changes in
task-related functional brain architecture.
Construction of Dynamic Networks
For each of the 112 brain regions, we extracted the wavelet coefficients of the mean time series in temporal
windows given by trial blocks (of approximately 60 TRs; see Table S2). The leftmost temporal boundary
of each window was equal to the first TR of an experimental trial block, and the rightmost boundary was
equal to the last TR in the same block. We thereby extracted block-specific data sets from the EXT,
MOD, and MIN sequences (with 6 -- 10 blocks of each sequence type; see Table S2 for details of the number
14
of blocks of each sequence type) for each of the 20 participants participating in the experiment and for
each of the 4 scanning sessions.
For each block-specific data set, we constructed an N × N adjacency matrix W representing the
complete set of pairwise functional connections present in the brain during that window in a given
participant and for a given scan. Note that N = 112 is the number of brain regions in the full brain
atlas (see the earlier section on "Partitioning the Brain into Regions of Interest" for further details).
To quantify the weight Wij of functional connectivity between regions labeled i and j, we used the
magnitude squared spectral coherence as a measure of nonlinear functional association between any two
wavelet coefficient time series (consistent with our previous study [17]). In using the coherence, which has
been demonstrated to be useful in the context of fMRI neuroimaging data [95], we were able to measure
frequency-specific linear relationships between time series.
To examine changes in functional brain network architecture during learning, we constructed multi-
layer networks by considering the set of L adjacency matrices constructed from consecutive blocks of a
given sequence type (EXT, MOD, or MIN) in a given participant and scanning session. We combined
the matrices in each set separately to form a rank-3 adjacency tensor A per sequence type, participant,
and scan. Such a tensor can be used to represent a time-dependent network [14, 17]. In the following
sections, we describe a variety of diagnostics that can be used to characterize such multilayer structures.
Network Examination
Dynamic Community Detection
Community detection [12, 13] can be used to identify putative functional modules (i.e., sets of brain
regions that exhibit similar trajectories through time). One such technique is based on the optimization
of the modularity quality function [96 -- 98]. This allows one to identify groups that consist of nodes that
have stronger connections among themselves than they do to nodes in other groups [12]. Recently, the
modularity quality function has been generalized so that one can consider time-dependent or multiplex
networks using multilayer modularity [14]
Q =
1
2µXijlr
{(Aijl − γlPijl) δlr + δijωjlr} δ(gil, gjr) ,
(2)
where the adjacency matrix of layer l has components Aijl, the element Pijl gives the components of the
corresponding matrix for a null model, γl is the structural resolution parameter of layer l, the quantity gil
gives the community (i.e., "module") assignment of node i in layer l, the quantity gjr gives the community
assignment of node j in layer r, the parameter ωjlr is the connection strength -- i.e., "interlayer coupling
parameter", which gives an element of a tensor ω that constitutes a set of temporal resolution parameters
if one is using the adjacency tensor A to represent a time-dependent network -- between node j in layer
r and node j in layer l, the total edge weight in the network is µ = 1
layer l is κjl = kjl + cjl, the intra-layer strength of node j in layer l is kjl, and the inter-layer strength of
2Pjr κjr, the strength of node j in
node j in layer l is cjl =Pr ωjlr. We employ the Newman-Girvan null model within each layer by using
2Pij Aijl is the total edge weight in layer l. We let ωjlr ≡ ω = constant for neighboring
where ml = 1
layers (i.e., when l − r = 1) and ωjlr = 0 otherwise. We also let γl = γ = constant. In the main text,
we report results for ω = 1 and γ = 1, and we evaluate the dependence of our results on γ and ω in the
Text S1.
(3)
Pijl =
kilkjl
2ml
,
Optimization of multilayer modularity (2) yields a partition of the brain regions into communities
for each time window. To measure changes in the composition of communities across time (i.e., across
15
experimental blocks), we defined the flexibility fi of a node i to be the number of times that a node changed
community assignment throughout the set of time windows represented by the multilayer network [17]
normalized by the total number of changes that were possible (i.e., by the number of contiguous pairs
of layers in the multilayer framework, which in this study ranged from 4 to 10; see Table S2). We
then defined the flexibility of the entire network as the mean flexibility over all nodes in the network:
F = 1
i=1 fi. To examine the relationship between brain network flexibility and learning, we confined
ourselves to the two EXT (i.e., extensively trained) sequences, in which learning occurs more rapidly than
in MOD and MIN sequences. We therefore estimated flexibility from the multilayer networks constructed
from blocks of the two EXT sequences in the first scanning session.
NPN
Identification of Temporal Core, Bulk, and Periphery
We find that different brain regions have different flexibilities. To determine whether a particular brain
region is more or less flexible than expected, we constructed a nodal null model, which can be used to
probe the individual roles of nodes in a network [15, 17]. (Note that alternative null models can be used
to probe other aspects of the temporal or geometrical structure in a multilayer network [15, 17].) We
rewired the ends of the multilayer network's inter-layer edges (which connect nodes in one layer to nodes
in another) uniformly at random. After applying the associated permutation, an inter-layer edge can, for
example, connect node i in layer t with node j 6= i in layer t + 1 rather than being constrained to connect
each node i in layer t with itself in layer t + 1.
We considered 100 different rewirings to construct an ensemble of 100 nodal null-model multilayer
networks for each single multilayer network constructed from the brain data. We then estimated the
flexibility of each node in each nodal null-model network. We created a distribution of expected mean
nodal flexibility values by averaging flexibility over 100 rewirings and the 20 participants. We similarly
estimated the mean nodal flexibility of the brain data by averaging flexibility over the 20 participants
and 100 optimizations. (We optimized multilayer modularity using a Louvain-like locally greedy method
[99, 100]. This procedure is not deterministic, so different runs of the optimization procedure can yield
slightly different partitions of a network.) We considered a region to be a part of the temporal "core" if
its mean nodal flexibility was below the 2.5% confidence bound of the null-model distribution, and we
considered a region to be a part of the temporal "periphery" if its mean nodal flexibility was above the
97.5% confidence bound of the null-model distribution. Finally, we considered a region to be a part of
the temporal "bulk" if its mean nodal flexibility was between the 2.5% and 97.5% confidence bounds of
the null-model distribution.
Geometrical Core-Periphery Organization
To estimate the geometrical core-periphery organization of the (static) networks defined by each ex-
perimental block (i.e., for each layer of a multilayer network), we used the method that was recently
proposed in Ref. [30]. This method results in a "core score" (which constitutes a centrality measure) for
each node that indicates where it lies on a continuous spectrum of roles between core and periphery. This
method has numerous advantages over previous formulations used to study core-periphery organization.
In particular, it can identify multiple geometrical cores in a network, which makes it possible to take
multiple cores into account and in turn enables one to construct a detailed description of geometrical
core-periphery organization by ranking the nodes in terms of how strongly they participate in different
possible cores. Importantly, the continuous nature of the measure removes the need to use an artificial
dichotomy of being strictly a core node versus strictly a peripheral node.
In applying method, we consider a vector C with non-negative values, and we let Cij = Ci × Cj,
where i and j are two nodes in an N -node network. We then seek a core vector C that satisfies the
normalization condition
CiCj = 1
X
i,j
and is a permutation of the vector C∗ whose components specify the local (geometrical) core values
16
(4)
(5)
C∗m =
1
1 + exp{−(m − N β) × tan(πα/2)}
, m ∈ {1, . . . , N} .
We seek a permutation that maximizes the core quality
R =Xi,j
AijCiCj .
This method to compute core-periphery organization has two parameters: α ∈ [0, 1] and β ∈ [0, 1].
The parameter α sets the sharpness of the boundary between the geometrical core and the geometrical
periphery. The value α = 0 yields the fuzziest boundary, and α = 1 gives the sharpest transition (i.e.,
a binary transition): as α varies from 0 to 1, the maximum slope of C∗ varies from 0 to +∞. The
parameter β sets the size of the geometrical core: as β varies from 0 to 1, the number of nodes included
in the core varies from N to 0. One now has the choice of either taking into account the local core scores
of a node for a set of (α, β) coordinates sampled from [0, 1] × [0, 1] (where one weighs each choice by its
corresponding value of R) or one can take into account only the score for particular choices of (α, β).
Statistics and Software
We performed all data analysis and statistical tests in Matlab. We performed the dynamic community
detection procedure using freely available Matlab code [99] that optimizes multilayer modularity using
a Louvain-like locally greedy algorithm [100].
Acknowledgements
We thank Jean M. Carlson, Matthew Cieslak, John Doyle, Daniel Greenfield, and Megan T. Valentine for
helpful discussions; John Bushnell for technical support; and James Fowler, Sang Hoon Lee, and Melissa
Lever for comments on earlier versions of the manuscript.
References
1. Saalmann YB, Pinsk MA, Wang L, Li X, Kastner S (2012) The pulvinar regulates information
transmission between cortical areas based on attention demands. Science 337: 753 -- 756.
2. Josi´c K, Rubin J, Matias M, Romo R, editors (2009) Coherent Behavior in Neuronal Networks.
Springer-Verlag.
3. Op de Beeck HP, Haushofer J, Kanwisher NG (2008) Interpreting fMRI data: maps, modules and
dimensions. Nat Rev Neurosci 9: 123 -- 135.
4. Plaut DC (1995) Double dissociation without modularity: evidence from connectionist neuropsy-
chology. J Clin Exp Neuropsych 17: 291 -- 321.
5. Bullmore ET, Bassett DS (2011) Brain graphs: graphical models of the human brain connectome.
Ann Rev Clin Psych 7: 113 -- 140.
17
6. Langer N, Pedroni A, Gianotti LR, Hanggi J, Knoch D, et al. (2012) Functional brain network
efficiency predicts intelligence. Hum Brain Mapp 33: 1393 -- 1406.
7. Bassett DS, Bullmore ET (2009) Human brain networks in health and disease. Curr Opin Neurol
22: 340 -- 347.
8. Matsuzaka Y, Picard N, Strick PL (2007) Skill representation in the primary motor cortex after
long-term practice. J Neurophysiol 97: 1819 -- 1832.
9. Wymbs NF, Grafton ST (2009) Neural substrates of practice structure that support future off-line
learning. J Neurophysiol 102: 2462 -- 2476.
10. Sun FT, Miller LM, Rao AA, D'Esposito M (2007) Functional connectivity of cortical networks
involved in bimanual motor sequence learning. Cereb Cortex 17: 1227 -- 1234.
11. Mushiake H, Saito N, Sakamoto K, Itoyama Y, Tanji J (2006) Activity in the lateral prefrontal
cortex reflects multiple steps of future events in action plans. Neuron 50: 631 -- 641.
12. Porter MA, Onnela JP, Mucha PJ (2009) Communities in networks. Not Amer Math Soc 56:
1082 -- 1097, 1164 -- 1166.
13. Fortunato S (2010) Community detection in graphs. Phys Rep 486: 75 -- 174.
14. Mucha PJ, Richardson T, Macon K, Porter MA, Onnela JP (2010) Community structure in
time-dependent, multiscale, and multiplex networks. Science 328: 876 -- 878.
15. Bassett DS, Porter MA, Wymbs NF, Grafton ST, Carlson JM, et al. (2012) Robust detection of
dynamic community structure in networks. Chaos 23: 013142.
16. Holme P, Saramaki J (2012) Temporal networks. Phys Rep 519: 97 -- 125.
17. Bassett DS, Wymbs NF, Porter MA, Mucha PJ, Carlson JM, et al. (2011) Dynamic reconfiguration
of human brain networks during learning. Proc Natl Acad Sci USA 108: 7641 -- 7646.
18. Borgatti SP, Everett MG (1999) Models of core/periphery structures. Social Networks 21: 375 --
395.
19. Dodge Y (2003) The Oxford Dictionary of Statistical Terms. Oxford University Press.
20. Darlington RB (1970) Is kurtosis really 'peakedness' ? The American Statistician 24: 19 -- 22.
21. Mesulam MM (1998) From sensation to cognition. Brain 121: 1013 -- 1052.
22. Bischoff-Grethe A, Goedert KM, Willingham DT, Grafton ST (2004) Neural substrates of
response-based sequence learning using fMRI. J Cogn Neurosci 16: 127 -- 138.
23. Picard N, Strick PL (1997) Activation on the medial wall during remembered sequences of reaching
movements in monkeys. J Neurophysiol 77: 2197 -- 2201.
24. Kennerley SW, Sakai K, Rushworth MF (2004) Organization of action sequences and the role of
the pre-SMA. J Neurophysiol 91: 978 -- 993.
25. Willingham DB, Salidis J, Gabrieli JD (2002) Direct comparison of neural systems mediating
conscious and unconscious skill learning. J Neurophysiol 88: 1451 -- 1460.
18
26. Destrebecqz A, Peigneux P, Laureys S, Degueldre C, Del Fiore G, et al. (2005) The neural cor-
relates of implicit and explicit sequence learning: Interacting networks revealed by the process
dissociation procedure. Learn Mem 12: 480 -- 490.
27. Brown R, Robertson E (2007) Inducing motor skill improvements with a declarative task. Nat
Neurosci 10: 148 -- 149.
28. Guimer`a R, Amaral LAN (2005) Functional cartography of complex metabolic networks. Nature
433: 895 -- 900.
29. Guimer`a R, Sales-Pardo M, Amaral LAN (2007) Classes of complex networks defined by role-to-
role connectivity profiles. Nat Phys 3: 63 -- 69.
30. Rombach MP, Porter MA, Fowler JH, Mucha PJ (2012) Core-periphery structure in networks.
SIAM Journal of Applied Mathematics In Press.
31. Opsahl T, Agneessens F, Skvoretz J (2010) Node centrality in weighted networks: Generalizing
degree and shortest paths. Social Networks 32: 245 -- 251.
32. Holme P (2005) Core-periphery organization of complex networks. Phys Rev E 72: 046111.
33. da Silva MR, Ma H, Zeng AP (2008) Centrality, network capacity, and modularity as parameters
to analyze the core-periphery structure in metabolic networks. Proceedings of the IEEE 96:
1411-1420.
34. Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, et al. (2008) Mapping the structural
core of human cerebral cortex. PLoS Biol 6: e159.
35. Garas A, Schweitzer F, Havlin S (2012) A k-shell decomposition method for weighted networks.
New J Phys 14.
36. Shanahan M, Wildie M (2012) Knotty-centrality: finding the connective core of a complex net-
work. PLoS One 7: e36579.
37. Colizza V, Flammini A, Serrano MA, Vespignani A (2006) Detecting rich-club ordering in complex
networks. Nat Phys 2: 110 -- 116.
38. Opsahl T, Colizza V, Panzarasa P, Ramasco JJ (2008) Prominence and control: the weighted
rich-club effect. Phys Rev Lett 101: 168702.
39. van den Heuvel MP, Sporns O (2011) Rich-club organization of the human connectome. J Neurosci
31: 15775 -- 15786.
40. van den Heuvel MP, Kahn RS, Goni J, Sporns O (2012) High-cost, high-capacity backbone for
global brain communication. Proc Natl Acad Sci USA 109: 11372 -- 11377.
41. Ekman M, Derrfuss J, Tittgemeyer M, Fiebach CJ (2012) Predicting errors from reconfiguration
patterns in human brain networks. Proc Natl Acad Sci U S A 109: 16714 -- 16719.
42. Sepulcre J, Sabuncu MR, Yeo TB, Liu H, Johnson KA (2012) Stepwise connectivity of the modal
cortex reveals the multimodal organization of the human brain. J Neurosci 32: 10649-10661.
43. Fornito A, Harrison BJ, Zalesky A, Simons JS (2012) Competitive and cooperative dynamics
of large-scale brain functional networks supporting recollection. Proc Natl Acad Sci USA 109:
12788 -- 12793.
19
44. Dosenbach NU, Visscher KM, Palmer ED, Miezin KK F M andd Wenger, Kang HC, et al. (2006)
A core system for the implementation of task sets. Neuron 50: 799 -- 812.
45. de Pasquale F, Della Penna S, Snyder AZ, Marzetti L, Pizzella V, et al. (2012) A cortical core for
dynamic integration of functional networks in the resting human brain. Neuron 74: 753 -- 764.
46. Uddin LQ, Supekar KS, Ryali S, Menon V (2011) Dynamic reconfiguration of structural and
functional connectivity across core neurocognitive brain networks with development. J Neurosci
31: 18578 -- 18589.
47. Whitlow CT, Casanova R, Maldjian JA (2011) Effect of resting-state functional mr imaging
duration on stability of graph theory metrics of brain network connectivity. Radiology 259: 516 --
524.
48. Kramer MA, Eden UT, Lepage KQ, Kolaczyk ED, Bianchi MT, et al. (2011) Emergence of
persistent networks in long-term intracranial eeg recordings. J Neurosci 31: 15757 -- 15767.
49. Chu CJ, Kramer MA, Pathmanathan J, Bianchi MT, Westover MB, et al. (2012) Emergence of
stable functional networks in long-term human electroencephalography. J Neurosci 32: 2703 --
2713.
50. Allen EA, Damaraju E, Plis SM, Erhardt EB, Eichele T, et al. (2012) Tracking whole-brain
connectivity dynamics in the resting state. Cereb Cortex Epub ahead of print.
51. Jones DT, Vemuri P, Murphy MC, Gunter JL, Senjem ML, et al. (2012) Non-stationarity in the
"resting brain's" modular architecture. PLoS One 7: e39731.
52. Siebenhuhner F, Weiss SA, Coppola R, Weinberger DR, Bassett DS (2012) Intra- and inter-
frequency brain network structure in health and schizophrenia. arXiv 1209.0729.
53. Barnes A, Bullmore ET, Suckling J (2009) Endogenous human brain dynamics recover slowly
following cognitive effort. PLoS One 4: e6626.
54. Tambini A, Ketz N, Davachi L (2010) Enhanced brain correlations during rest are related to
memory for recent experiences. Neuron 65: 280 -- 290.
55. Doron KW, Bassett DS, Gazzaniga MS (2012) Dynamic network structure of interhemispheric
coordination. Proc Natl Acad Sci U S A 109: 18661 -- 18668.
56. Sheppard JP, Wang JP, Wong PC (2012) Large-scale cortical network properties predict future
sound-to-word learning success. J Cogn Neurosci 24: 1087 -- 1103.
57. Zalesky A, Fornito A, Harding IH, Cocchi L, Yucel M, et al. (2010) Whole-brain anatomical
networks: Does the choice of nodes matter? NeuroImage 50: 970 -- 983.
58. Wang J, Wang L, Zang Y, Yang H, Tang H, et al. (2009) Parcellation-dependent small-world
brain functional networks: A resting-state fMRI study. Hum Brain Mapp 30: 1511 -- 1523.
59. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, et al. (2011) Functional network organi-
zation of the human brain. Neuron 72: 665 -- 678.
60. Bassett DS, Brown JA, Deshpande V, Carlson JM, Grafton ST (2011) Conserved and variable
architecture of human white matter connectivity. NeuroImage 54: 1262 -- 1279.
61. Wig GS, Schlaggar BL, Petersen SE (2011) Concepts and principles in the analysis of brain
networks. Ann NY Acad Sci 1224: 126 -- 146.
20
62. de Reus MA, van den Heuvel MP (2013) The parcellation-based connectome: Limitations and
extensions. NeuroImage Epub ahead of print.
63. Friston KJ (2011) Functional and effective connectivity: a review. Brain Connect 1: 13 -- 36.
64. Horwitz B (2003) The elusive concept of brain connectivity. NeuroImage 19: 466 -- 470.
65. Gavrilescu M, Stuart GW, Rossell S, Henshall K, McKay C, et al. (2008) Functional connectivity
estimation in fMRI data: influence of preprocessing and time course selection. Hum Brain Mapp
29: 1040 -- 1052.
66. Jones TB, Bandettini PA, Kenworthy L, Case LK, Milleville SC, et al. (2010) Sources of group
differences in functional connectivity: an investigation applied to autism spectrum disorder. Neu-
roImage 49: 401 -- 414.
67. David O, Cosmelli D, Friston KJ (2004) Evaluation of different measures of functional connectivity
using a neural mass model. NeuroImage 21: 659 -- 673.
68. Smith SM, Miller KL, Salimi-Khorshidi G, Webster M, Beckmann CF, et al. (2011) Network
modelling methods for FMRI. NeuroImage 54: 875 -- 91.
69. Ryali S, Chen T, Supekar K, Menon V (2012) Estimation of functional connectivity in fMRI data
using stability selection-based sparse partial correlation with elastic net penalty. NeuroImage 59:
3852 -- 3861.
70. Gates KM, Molenaar PC (2012) Group search algorithm recovers effective connectivity maps for
individuals in homogeneous and heterogeneous samples. NeuroImage 63: 310 -- 319.
71. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, et al. (2013) A pairwise maximum entropy
model accurately describes resting-state human brain networks. Nat Comm 4: 1370.
72. Schmidt RA, Lee TD (2005) Motor Control and Learning: A Behavioral Emphasis. Human
Kinetics, fourth edition.
73. Rosenbaum DA (2010) Human Motor Control. Elsevier.
74. Yarrow K, Brown P, Krakauer JW (2009) Inside the brain of an elite athlete: the neural processes
that support high achievement in sports. Nat Rev Neurosci 10: 585 -- 596.
75. Dayan E, Cohen LG (2011) Neuroplasticity subserving motor skill learning. Neuron 72: 443 -- 454.
76. Snoddy GS (1926) Learning and stability: A psychophysical analysis of a case of motor learning
with clinical applications. J App Psych 10: 1 -- 36.
77. Crossman ERFW (1959) A theory of the acquisition of speed-skill. Ergonomics 2: 153 -- 166.
78. Newell KM, Rosenbloom PS (1981) Mechanisms of skill acquisition and the law of practice. In:
Anderson JR, editor, Cognitive skills and their acquisition, Lawrence Erlbaum Associates. pp.
1 -- 55.
79. Heathcote A, Brown S, Mewhort DJ (2000) The power law repealed: the case for an exponential
law of practice. Psychon Bull Rev 7: 185 -- 207.
80. Bassett DS, Bullmore ET (2006) Small-world brain networks. Neuroscientist 12: 512 -- 523.
81. Bullmore E, Sporns O (2009) Complex brain networks: Graph theoretical analysis of structural
and functional systems. Nat Rev Neurosci 10: 186 -- 198.
21
82. Mantzaris AV, Bassett DS, Wymbs NF, Estrada E, Porter MA, et al. (2013) Dynamic network
centrality summarizes learning in the human brain. Journal of Complex Networks Epub Ahead
of Print.
83. Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TEJ, et al. (2004) Advances
in functional and structural MR image analysis and implementation as FSL. NeuroImage 23:
208 -- 219.
84. Woolrich MW, Jbabdi W, Patenaude B, Chappell M, Makni S, et al. (2009) Bayesian analysis of
neuroimaging data in FSL. NeuroImage 45: S173 -- S186.
85. Bullmore E, Fadili J, Breakspear M, Salvador R, Suckling J, et al. (2003) Wavelets and statistical
analysis of functional magnetic resonance images of the human brain. Stat Methods Med Res 12:
375 -- 399.
86. Bullmore E, Fadili J, Maxim V, Sendur L, Whitcher B, et al. (2004) Wavelets and functional
magnetic resonance imaging of the human brain. NeuroImage 23: S234 -- S249.
87. Brammer MJ (1998) Multidimensional wavelet analysis of functional magnetic resonance images.
Hum Brain Mapp 6: 378 -- 382.
88. 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.
89. Bassett DS, Meyer-Lindenberg A, Achard S, Duke T, Bullmore E (2006) Adaptive reconfiguration
of fractal small-world human brain functional networks. Proc Natl Acad Sci USA 103: 19518 --
19523.
90. Achard S, Bullmore E (2007) Efficiency and cost of economical brain functional networks. PLoS
Comp Biol 3: e17.
91. Achard S, Bassett DS, Meyer-Lindenberg A, Bullmore E (2008) Fractal connectivity of long-
memory networks. Phys Rev E 77: 036104.
92. Bassett DS, Meyer-Lindenberg A, Weinberger DR, Coppola R, Bullmore E (2009) Cognitive fitness
of cost-efficient brain functional networks. Proc Natl Acad Sci USA 106: 11747 -- 11752.
93. Lynall ME, Bassett DS, Kerwin R, McKenna P, Muller U, et al. (2010) Functional connectivity
and brain networks in schizophrenia. J Neurosci 30: 9477 -- 9487.
94. Percival DB, Walden AT (2000) Wavelet Methods for Time Series Analysis. Cambridge University
Press.
95. Sun FT, Miller LM, D'Esposito M (2004) Measuring interregional functional connectivity using
coherence and partial coherence analyses of fMRI data. NeuroImage 21: 647 -- 658.
96. Newman MEJ, Girvan M (2004) Finding and evaluating community structure in networks. Phys
Rev E 69: 026113.
97. Newman MEJ (2006) Modularity and community structure in networks. Proc Natl Acad Sci USA
103: 8577 -- 8582.
98. Newman MEJ (2006) Finding community structure in networks using the eigenvectors of matrices.
Phys Rev E 74: 036104.
22
99. Jutla IS, Jeub LGS, Mucha PJ (2011 -- 2012). A generalized Louvain method for community
detection implemented in matlab. URL http://netwiki.amath.unc.edu/GenLouvain.
100. Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E (2008) Fast unfolding of community hierar-
chies in large networks. J Stat Mech : P10008.
101. Matsumoto M, Nishimura T (1998) Mersenne twister: a 623-dimensionally equidistributed uni-
form pseudo-random number generator. ACM Transactions on Modeling and Computer Simula-
tion 8: 3 -- 30.
Figure and Table Legends
Figure 1. Network Organization of Human Brain Dynamics. (A) Temporal Networks of the
Human Brain. We parcellate the brain into anatomical regions that can be represented as nodes in a
network, and we use the coherence between functional Magnetic Resonance Imaging (fMRI) time series
of each pair of nodes over a time window to determine the weight of the network edge connecting those
nodes. We determine these weights separately using approximately 10 non-overlapping time windows of
2 -- 3 min duration and thereby construct temporal networks that represent the dynamical functional
connectivity in the brain. (B) Cohesive Mesoscale Structures. (top) An example of a network with a
modular organization in which high-degree nodes (brown) are often found in the center of modules or
bridging distinct modules that are composed mostly of low-degree nodes (blue). (bottom) A network
with a core-periphery organization in which nodes in the core (purple) are more densely connected with
one another than nodes in the periphery are with one another (green).
A
brain network
nodes
20
0
−20
40
0
20
0
−20
−40
0
nodal
time series
B modular organization
20
40
60
80
100
20
40
60
80
100
time
windows
.
r=0.20 r=0.29 r=0.25 r=0.60
.
.
.
core-periphery organization
pairwise
coherence
temporal
networks
23
Figure 2. Dynamic Network Diagnostics Change with Learning. (A) Multilayer modularity,
(B) number of communities, and (C) mean flexibility calculated as a function of the number of trials
completed after a scanning session (see Table 1 for the relationship between the number of trials
practiced and training duration and intensity). We average the values for each diagnostic over the 100
multilayer modularity optimizations, and we average flexibility over the 112 brain regions (in addition
to averaging over the 100 optimizations per subject). Error bars indicate the standard error of the mean
over participants.
Figure 3. Temporal Core-Periphery Organization of the Brain determined using fMRI signals
during the performance of a simple motor learning task. (A) The core (cyan), bulk (gold), and
periphery (maroon) nodes consist, respectively, of brain regions whose mean flexibility over individuals
is less than, equal to, and greater than that expected in a null model (gray shaded region). We measure
flexibility based on the allegiance of nodes to putative functional modules. Error bars indicate the
standard error of the mean over individuals. (B) The anatomical distribution of regions in the core,
bulk, and periphery appears to be spatially contiguous. The core primarily contains sensorimotor and
visual processing areas, the periphery primarily contains multimodal association areas, and the bulk
contains the remainder of the brain (and is therefore composed predominantly of frontal and temporal
cortex).
102
number of trials
103
C
0.15
0.14
0.13
0.12
0.11
0.1
0.09
(cid:31)exibility
102
number of trials
103
4
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
number of communities
B
102
number of trials
103
0.094
0.092
0.09
0.088
0.086
0.084
0.082
0.08
0.078
multilayer modularity
A
periphery
bulk
core
B
null
model
0.1
flexibility
0.15
A
100
80
60
40
20
brain region
0
0.05
24
Figure 4. Geometrical Core-Periphery Organization in Brain Networks. (A) Core quality R
(5) in the (α, β) parameter plane for a typical participant (3), scanning session (1), sequence type
(EXT), and experimental block (1). (B) Distribution of the α and β values that maximize the R-score.
We compute this distribution over all network layers, participants, scanning sessions, and sequence
types. The β parameter is much more localized (its standard deviation is 0.05) than the α parameter
(its standard deviation is 0.26). (C) Mean core shape. We plot the ordered vector of C values. We have
set the values of α and β to the mean values of those that maximize the R-score for all network layers,
participants, scanning sessions, and sequence types.
Figure 5. Geometrical Core Scores Change with Learning. Variance of the distribution of
mean geometrical core scores over brain regions as a function of the number of trials completed after a
scanning session. (See Table 1 for the relationship between the number of trials practiced and training
duration and intensity.) Error bars indicate the standard error of the mean over participants (where the
data point from each participant is the mean geometrical core score over brain regions, scanning
sessions, sequence types, and network layers).
40
brain region
60
80
100
1
0.8
0.6
0.4
0.2
core scoreC
0.2
0.4
α
0.6
0.8
1
0.2
0.4
β
0.6
0.8
1
0
0
20
0.2
0.1
0
0
0.6
0.4
0.2
0
0
relative frequency
B
core quality
0.6
0.5
0.4
0.3
0.2
.1
00
0.2 0.4 0.6 0.8
β
1
A
α
0.2
0.4
0.6
0.8
1
102
number of trials
103
0.015
0.01
0.005
variance of geometrical core score
25
.=
Figure 6. Relationship Between Temporal and Geometrical Core-Periphery
Organizations. A strong negative correlation exists between flexibility and the geometrical core score
for networks constructed from blocks of (A) extensively, (B) moderately, and (C) minimally trained
sequences on scanning session 1 (day 1; circles), session 2 (after approximately 2 weeks of training;
squares), session 3 (after approximately 4 weeks of training; diamonds), and session 4 (after
approximately 6 weeks of training; stars). This negative correlation indicates that the temporal
core-periphery organization is mimicked in the geometrical core-periphery organization and therefore
that the core of dynamically stiff regions also exhibits dense connectivity. We show temporal core nodes
in cyan, temporal bulk nodes in gold, and temporal periphery nodes in maroon. The darkness of data
points indicates scanning session; darker colors indicate earlier scans, so the darkest colors indicate scan
1 and the lightest ones indicate scan 4. The grayscale lines indicate the best linear fits; again, darker
colors indicate earlier scans, so session 1 is in gray and session 4 is in light gray. The Pearson
correlation between the flexibility (averaged over 100 multilayer modularity optimizations, 20
participants, and 4 scanning sessions) and the geometrical core score (averaged over 20 participants and
4 scanning sessions) is significant for the EXT (r
p
3.4 × 10−45), MOD (r
−0.93,
.=
.=
.=
−0.92, p
4.8 × 10−50) data.
.=
.=
2.2 × 10−49), and MIN (r
−0.93, p
minimal training
0.4
0.3
0.2
0.1
0
−0.1
0.05
0.1
flexibility
0.15
C
geometrical core score
moderate training
0.05
0.1
flexibility
0.15
0.4
0.3
0.2
0.1
0
−0.1
geometrical core score
B
scan1
scan 2
scan 3
scan 4
Core Bulk Periphery
0.4
0.3
0.2
0.1
0
geometrical core score
−0.1
0.05
0.1
flexibility
0.15
A
extensive training
26
Figure 7. Core-Periphery Organization of Brain Dynamics During Learning. The
relationship between temporal and geometrical core-periphery organization and their associations with
learning are present in individual subjects. We represent this relationship using spirals in a plane; data
points in this plane represent brain regions located at the polar coordinates (f s,−f κ), where f is the
flexibility of the region, s is the skewness of flexibility over all regions, and κ is the learning parameter
(see the Materials and Methods) that describes each individual's relative improvement between sessions.
The skewness predicts individual differences in learning; the Spearman rank correlation is ρ
and p
good learners (curvier spirals) tend to have high skewness (long spirals). Color indicates flexibility: blue
nodes have lower flexibility, and brown nodes have higher flexibility.
−0.480
0.034. Poor learners (straighter spirals) tend to have a low skewness (short spirals), whereas
.=
.=
f
s
-fκ
0
1
2
−1
−2
−3
−4
−2
−1
0
1
2
3
4
0.14
0.13
0.12
0.11
(cid:30)exibility
0.1
0.09
0.08
0.07
0.06
0.05
geometrical core score
27
Figure 8. Trial Structure and Stimulus-Response (S-R) Mapping. (A) Each trial began with
the presentation of a sequence-identity cue that remained on screen for 2 seconds. Each of the 6 trained
sequences was paired with a unique identity cue. A discrete sequence-production (DSP) event structure
was used to guide sequence production. The onset of the initial DSP stimulus (thick square, colored red
in the task) served as the imperative to produce the sequence. A correct key press led to the immediate
presentation of the next DSP stimulus (and so on) until the 10-element sequence was correctly
executed. Participants received a feedback "+" to signal that a sequence was completed and to wait
(approximately 0 -- 6 seconds) for the start of the next trial. This waiting period is called the "inter-trial
interval" (ITI). At any point, if an incorrect key was hit, a participant would receive an error signal
(not shown in the figure) and the DSP sequence would pause until the correct response was received.
(B) There was a direct S-R mapping between a conventional keyboard or an MRI-compatible button
box (see the lower left of the figure) and a participant's right hand, so the leftmost DSP stimulus cued
the thumb and the rightmost stimulus cued the pinky finger. Note that the button location for the
thumb was positioned to the lower left to achieve maximum comfort and ease of motion.
a)
A
sequence identity cue
B
1
1
thmb
2
0
0
DSP task production
0
m
s
2
S-R mapping
indx mdl rng pnky
2 3 4 5
3
2
4
5
Feedback
+
0
0
0
m
s
+
v
a
r
i
a
b
l
e
I
T
I
28
Figure 9. Experiment Timeline. Training sessions in the MRI scanner during the collection of
blood-oxygen-level-dependent (BOLD) signals were interleaved with training sessions at home.
Participants first practiced the sequences in the MRI scanner during a baseline training session (top).
Following every approximately 10 training sessions (see Table S1), participants returned for another
scanning session. During each scanning session, a participant practiced each sequence for 50 trials.
Participants trained at home between the scanning sessions (bottom). During each home training
session, participants practiced the sequences in a random order. (We determined a random order using
the Mersenne Twister algorithm of Nishimura and Matsumoto [101] as implemented in the random
number generator rand.m of Matlab version 7.1). Each EXT sequence was practiced for 64 trials, each
MOD sequence was practiced for 10 trials, and each MIN sequence was practiced for 1 trial.
Figure 10. (Supplementary Figure S1) Reliability of Temporal Core-Periphery Structure.
Temporal core (cyan), bulk (gold), and periphery (maroon) of dynamic networks determined based on
the flexibility of trial blocks in which participants practiced sequences that would eventually be
extensively trained. (A) Flexibility of the temporal core, bulk, and periphery averaged over the 100
multilayer modularity optimizations and 20 participants for blocks composed of extensively trained
(EXT; light circles), moderately trained (MOD; squares), and minimally trained (MIN; dark diamonds)
sequences. The darkness of data points indicates scanning session; darker colors indicate earlier scans,
so the darkest colors indicate scan 1 and the lightest ones indicate scan 4. (B) The coefficient of
variation of flexibility calculated over the 100 optimizations and 3 sequence types for all brain regions.
Error bars indicate the standard error of the mean CV over participants. Both panels use data from
scanning session 1 on day 1 of the experiment (which is prior to home training).
experiment timeline
training sessions 1-10
training sessions 11-20
training sessions 21-30
SCAN
SCAN
SCAN
SCAN
"baseline training"
(50 trials/sequence)
(50 trials/sequence)
(50 trials/sequence)
home training example
EXT (64 trials /sequence)
MOD (10 trials /sequence)
MIN (1 trial /sequence)
29
Figure 11. (Supplementary Figure S2) Temporal Core-Periphery Organization Over 42
Days. Temporal core (cyan), bulk (gold), and periphery (maroon) of dynamic networks defined by trial
blocks in which participants practiced sequences that would eventually be (A) extensively trained, (B)
moderately trained, and (C) minimally trained for data from scanning sessions 2 (after approximately 2
weeks of training; circles), 3 (after approximately 4 weeks of training; squares), and 4 (after
approximately 6 weeks of training; diamonds). The darkness of data points indicates scanning session;
darker colors indicate earlier scans, so the darkest colors indicate scan 1 and the lightest ones indicate
scan 4.
Figure 12. (Supplementary Figure S3) Geometrical Core-Periphery Organization Over 42
Days. Geometrical core scores for each brain region defined by the trial blocks in which participants
practiced sequences that would eventually be (A) extensively trained, (B) moderately trained, and (C)
minimally trained for data from scanning sessions 1 (day 1; black circles), 2 (after approximately 2
weeks of training; dark gray squares), 3 (after approximately 4 weeks of training; gray diamonds), and 4
(after approximately 6 weeks of training; light gray stars). We have averaged the geometrical core
scores over blocks and over 20 participants. The order of brain regions is identical for all 3 panels
(A-C ), and we chose this order by ranking regions from high to low geometrical core scores from the
EXT blocks on scanning session 1 (on day 1 of the experiment).
Figure 13. (Supplementary Figure S4) Relationship Between Temporal Core-Periphery
Organization and Community Structure. (A) Mean-coherence matrix over all EXT blocks from
all participants on scanning day 1. The colored bars above the matrix indicate the 3 communities that
we identified from the representative partition. Mean partition similarity z-score zi over all participants
for blocks of (B) extensively, (C) moderately, and (D) minimally trained sequences for all 4 scanning
sessions over the approximately 6 weeks of training. The horizontal gray lines in panels (B-D) indicate
the zi value that corresponds to a right-tailed p-value of 0.05.
0.92. (B) Box plot over
.=
Figure 14. (Supplementary Figure S5) Region Size is Uncorrelated with Flexibility. (A)
Scatter plot of the size of the brain region in voxels (averaged over participants) versus the flexibility of
the EXT multilayer networks, which we averaged over the 100 multilayer modularity optimizations and
the 20 participants. Data points indicate brain regions. The line indicates the best linear fit. Its
Pearson correlation coefficient is r
the 20 participants of the squared Pearson correlation coefficient r2 between the participant-specific
region size in voxels and the participant-specific flexibility averaged over the 100 multilayer modularity
optimizations.
−0.009, and the associated p-value is p
.=
30
Figure 15. (Supplementary Figure S6) Temporal Core-Periphery Organization and
Task-Related Activations. Mean GLM parameter estimates for the temporal core (cyan; circles),
bulk (gold; squares), and periphery (maroon; diamonds) of dynamic networks defined by the trial blocks
in which participants practiced sequences that would eventually be (A) extensively trained, (B)
moderately trained, and (C) minimally trained for data from scanning sessions 1 (first day of training),
2 (after approximately 2 weeks of training), 3 (after approximately 4 weeks of training), and 4 (after
approximately 6 weeks of training).
Figure 16. (Supplementary Figure S7) Effect of Structural Resolution Parameter. (A,B)
Number of communities and (C,D) number of regions in the temporal core (cyan; circles), temporal
bulk (gold; squares), and temporal periphery (maroon; diamonds) as a function of the structural
resolution parameter γ, where we considered (A,C) γ ∈ [0.2, 5] in increments of ∆γ = 0.2 and (B,D)
γ ∈ [0.8, 1.8] in increments of ∆γ = 0.01. We averaged the values in panels (A) and (B) over 100
multilayer modularity optimizations and over the 20 participants.
Figure 17. (Supplementary Figure S8) Effect of Temporal Resolution Parameter. (A)
Number of communities averaged over 100 multilayer modularity optimizations and over 20 participants
as a function of the temporal resolution parameter ω. (B) Number of regions that we identified as part
of the temporal core (cyan; circles), temporal bulk (gold; squares), and temporal periphery (maroon;
diamonds) as we vary ω from 0.1 to 2 in increments of ∆ω = 0.1.
31
(Main Text) Table 1: Relationship Between Training Duration, Intensity, and Depth. We
report the number of trials (i.e., "depth") of each sequence type (i.e., "intensity") completed after each
scanning session (i.e., "duration") averaged over the 20 participants.
(Supplementary Table S1): Experimental Details for Behavioral Data Acquired Between
Scanning Sessions. We give the minimum, mean, maximum, and standard error of the mean over
participants for the following variables: the number of days between scanning sessions; the number of
practice sessions performed at home between scanning sessions; and the number of trials composed of
extensively, moderately, and minimally trained sequences during home practice between scanning sessions.
(Supplementary Table S2): Experimental Details for Brain Imaging Data Acquired Dur-
ing Scanning Sessions. In the top three rows, we give the mean, minimum, maximum, and standard
error over participants for the number of blocks composed of extensively, moderately, and minimally
trained sequences during scanning sessions. In the bottom three rows, we give (in TRs) the mean, mini-
mum, maximum, and standard error of the length over blocks composed of extensively, moderately, and
minimally trained sequences during scanning sessions.
(Supplementary Table S3): Brain regions in the Harvard-Oxford (HO) Cortical and
Subcortical Parcellation Scheme provided by FSL [83, 84] and their affiliation to the temporal
core (C; cyan), bulk (B; gold), and periphery (P; maroon) for both left (L) and right (R) hemispheres.
Session 1
Session 2
Session 3
Session 4
MIN Sequences
MOD Sequences
EXT Sequences
50
50
50
110
200
740
170
350
1430
230
500
2120
Table 1. Relationship Between Training Duration, Intensity, and Depth. We report the
number of trials (i.e., "depth") of each sequence type (i.e., "intensity") completed after each scanning
session (i.e., "duration") averaged over the 20 participants.
Supplementary Material for "Core-Periphery Organization of
Human Brain Dynamics"
Danielle S. Bassett1,2,∗, Nicholas F. Wymbs5, M. Puck Rombach3,4,
Mason A. Porter3,4, Peter J. Mucha6,7, Scott T. Grafton5
1Department of Physics, University of California, Santa Barbara, CA 93106, USA;
2 Sage Center for the Study of the Mind,
University of California, Santa Barbara, CA 93106;
3 Oxford Centre for Industrial and Applied Mathematics,
Mathematical Institute, University of Oxford, Oxford OX1 3LB, UK;
4CABDyN Complexity Centre, University of Oxford, Oxford, OX1 1HP, UK;
5Department of Psychological and Brain Sciences and UCSB Brain Imaging Center,
University of California, Santa Barbara, CA 93106, USA;
6Carolina Center for Interdisciplinary Applied Mathematics,
Department of Mathematics, University of North Carolina,
Chapel Hill, NC 27599, USA;
7Institute for Advanced Materials, Nanoscience & Technology,
University of North Carolina, Chapel Hill, NC 27599, USA;
∗Corresponding author. Email address: [email protected]
October 30, 2018
arXiv:1210.3555v2 [q-bio.NC] 30 Oct 2013
1
Figure 1: Reliability of Temporal Core-Periphery Structure. Temporal core (cyan), bulk (gold),
and periphery (maroon) of dynamic networks determined based on the flexibility of trial blocks in which
participants practiced sequences that would eventually be extensively trained. (A) Flexibility of the temporal
core, bulk, and periphery averaged over the 100 multilayer modularity optimizations and 20 participants
for blocks composed of extensively trained (EXT; light circles), moderately trained (MOD; squares), and
minimally trained (MIN; dark diamonds) sequences. The darkness of data points indicates scanning session;
darker colors indicate earlier scans, so the darkest colors indicate scan 1 and the lightest ones indicate scan
4. (B) The coefficient of variation of flexibility calculated over the 100 optimizations and 3 sequence types
for all brain regions. Error bars indicate the standard error of the mean CV over participants. Both panels
use data from scanning session 1 on day 1 of the experiment (which is prior to home training).
Reliability of Temporal Core-Periphery Organization
A brain region's role in the temporal core, bulk, and periphery is robust across levels of training. Regions
identified as part of the core, bulk, or periphery in multilayer networks constructed from the EXT blocks in
scanning session 1 have similar flexibilities in the other two levels of training (MOD and MIN; see Fig. 1A)
for the same scanning session. To quantify the variability of a brain region's flexibility, we calculated the
coefficient of variation (CV) of flexibility over the 100 optimizations and the 3 levels of training (see Fig. 1B).
The CV is defined as CV = σ/µ, where σ is the standard deviation of a given sample and µ is its mean. We
observe that the variabilities over optimizations and scans (i.e., CV) and over participants (i.e., error bars)
are largest in regions designated as part of the temporal core and smallest in regions designated as part of
the temporal periphery.
In addition, regional flexibility is also conserved across both intensity of training (MIN, MOD, and EXT)
and duration of training (sessions 1 -- 4). Observe in Fig. 2 that regions identified as part of the temporal core
in multilayer networks constructed from the EXT blocks in scanning session 1 exhibit small flexibility for all
other scanning sessions and for all 3 training levels (EXT, MOD, and MIN). Regions in the temporal bulk
and temporal periphery exhibit a similar amount of flexibility to one another.
Reliability of Geometrical Core-Periphery Organization
As we illustrate in Fig. 3, the geometrical core-periphery organization of the brain was consistent over the 42
days of practice, across sequence types, and throughout variations in the intensity of training (MIN, MOD,
and EXT) and in the duration of training (sessions 1 -- 4).
Relationship Between Temporal Core-Periphery Organization and Community Structure
The division of the brain networks into temporal core, bulk, and peripheral nodes has interesting similarities
to their partitioning into communities based on optimizing multilayer modularity. We first noted this sim-
ilarity when we examined community structure in an object that we call the mean-coherence matrix. The
mean-coherence matrix ¯A contain elements ¯Aij that are equal to the mean coherence between nodes i and
2
scan 1
core bulk periphery
40
80
brain region
120
2.5
2
1.5
1
0.5
0
0
coefficient of variation
scan 1
B
core bulk periphery
40
80
brain region
120
0.16
0.12
0.08
0.04
0
A
flexibility
extensive
moderate
minimal
Figure 2: Temporal Core-Periphery Organization Over 42 Days. Temporal core (cyan), bulk (gold),
and periphery (maroon) of dynamic networks defined by trial blocks in which participants practiced sequences
that would eventually be (A) extensively trained, (B) moderately trained, and (C) minimally trained for
data from scanning sessions 2 (after approximately 2 weeks of training; circles), 3 (after approximately 4
weeks of training; squares), and 4 (after approximately 6 weeks of training; diamonds). The darkness of data
points indicates scanning session; darker colors indicate earlier scans, so the darkest colors indicate scan 1
and the lightest ones indicate scan 4.
Figure 3: Geometrical Core-Periphery Organization Over 42 Days. Geometrical core scores for each
brain region defined by the trial blocks in which participants practiced sequences that would eventually be
(A) extensively trained, (B) moderately trained, and (C) minimally trained for data from scanning sessions
1 (day 1; black circles), 2 (after approximately 2 weeks of training; dark gray squares), 3 (after approximately
4 weeks of training; gray diamonds), and 4 (after approximately 6 weeks of training; light gray stars). We
have averaged the geometrical core scores over blocks and over 20 participants. The order of brain regions
is identical for all 3 panels (A-C ), and we chose this order by ranking regions from high to low geometrical
core scores from the EXT blocks on scanning session 1 (on day 1 of the experiment).
3
extensive training
B
moderate training
C
minimal training
core bulk periphery
40
80
brain region
120
0
40
80
brain region
120
0
40
80
brain region
120
scan 2
scan 3
scan 4
0.18
0.14
0.1
0.06
0
A
flexibility
minimal training
50
brain region
100
0.5
0.4
0.3
0.2
0.1
C
geometrical core score
0
0
moderate training
50
brain region
100
0.5
0.4
0.3
0.2
0.1
B
geometrical core score
0
0
scan 1
scan 2
scan 3
scan 4
extensive training
50
brain region
100
0.5
0.4
0.3
0.2
0.1
A
geometrical core score
0
0
j over participants and EXT blocks on day 1 of the experiment. We determined the community structure
of this mean-coherence matrix by optimizing the single-layer modularity quality function [1, 2, 3, 4, 5]:
kikj
2m(cid:21) δ(gi, gj) ,
Qsingle−layer =Xij (cid:20) ¯Aij −
(1)
where node i is assigned to community gi, node j is assigned to community gj, the Kronecker delta δ(gi, gj) =
1 if gi = gj and it equals 0 otherwise, ki is the strength of node i, and m is the mean strength of all nodes in
the network. After optimizing this single-layer quality function 100 times, we constructed a representative
partition [6] from the set of 100 partitions. (Each partition arises from a single optimization.) One community
in this representative partition, which we show in Fig. 4A, appears to have high connectivity to the other
two communities: nodes in this first community have edges with strong weights to nodes in the other two
communities. This indicates a high coherence in the BOLD time series, and this behavior is consistent with
the behavior expected from a network "core". A second community in this representative partition appears
to have low connectivity to the other two communities: nodes in this community have edges with small
weights that connect to nodes in the other two communities. This indicates a low coherence in the BOLD
time series, and this behavior is consistent with the behavior expected from a "periphery".
It is important to note that we observed this relationship between temporal core-periphery organization
and community structure in networks encoded by mean matrices. However, networks encoded by mean ma-
trices constructed by averaging correlation-based matrices often do not adequately represent the topological
or geometrical structure of the ensemble of individual networks from which they are derived [7]. We therefore
test for a relationship between the temporal core-periphery organization and community structure in the
ensemble of networks extracted from individual participants.
A division of the brain into temporal core, bulk, and peripheral regions gives a partition of the functional
brain network. We label this partition using the Greek letter ν, and we use the z-score of the Rand coefficient
[8] to test for similarities between this partition and algorithmic partitions, which we label using η, into
communities (based on optimization of multilayer modularity) for each participant, block, and optimization.
For each pair of partitions ν and η, we calculate the Rand z-score in terms of the total number of node pairs
M in the network, the number of pairs Mν that are in the same community in partition ν but not in the
partition η, the number of pairs Mη that are in the same community in partition η but not in ν, and the
number of node pairs wνη that are assigned to the same community in both partition ν and partition η. The
z-score of the Rand coefficient allows one to compare partitions η and ν, and it is given by the formula
zνη =
1
σwνη
wνη −
MνMη
M
,
(2)
where σwνη is the standard deviation of wνη. Let the mean partition similarity zi denote the mean value of
zνη over all partitions η (i.e., for all blocks and all optimizations) for participant i.
As we show in Fig. 4B-D, we find that communities identified by the optimization of the multilayer mod-
ularity quality function (see the "Materials and Methods" section in the main manuscript) have significant
overlap with the division into temporal core, bulk, and periphery during early learning. The mean values
of zi over participants indicate that there is a significant similarity between the partitions into modules and
the partitions into core, bulk, and periphery for networks representing functional connectivity during blocks
of extensively, moderately, and minimally trained sequences on scanning day 1. This similarity between
community structure and temporal core-periphery organization is also evident for blocks of moderately and
minimally trained sequences practiced during later scanning sessions. These results underscore the fact that
core-periphery organization can be consistent with community structure. Note, however, that there is no
statistical similarity between partitions into core, bulk, and periphery and partitions into communities for
later learning. (As shown in Fig. 4B-D, the z-scores for networks that represent the functional connectivity
during extensive training in scans 2 -- 4, moderate training in scans 3 -- 4, and minimal training in scan 4 are not
significantly greater than expected (i.e., under the null hypothesis of no difference between the partitions).)
Together, this set of results suggests that the relationship between these two types of mesoscale organization
can be altered by learning.
4
Figure 4: Relationship Between Temporal Core-Periphery Organization and Community Struc-
ture. (A) Mean-coherence matrix over all EXT blocks from all participants on scanning day 1. The colored
bars above the matrix indicate the 3 communities that we identified from the representative partition. Mean
partition similarity z-score zi over all participants for blocks of (B) extensively, (C) moderately, and (D)
minimally trained sequences for all 4 scanning sessions over the approximately 6 weeks of training. The
horizontal gray lines in panels (B-D) indicate the zi value that corresponds to a right-tailed p-value of 0.05.
5
extensive training
2
3
scanning session
minimal training
2
3
scanning session
4
4
1
1
012345678
012345678
B
z−score
mean squared coherence
0.8
0.7
0.6
0.5
0.4
0.3
0.2
.1
00
D
z−score
4
20 40 60 80 100
brain regions
moderate training
1
2
3
scanning session
20
40
60
80
brain regions
A
100
C
012345678
z−score
Mean Minimum Maximum Standard Error
Days
Between Scans 1 and 2
Between Scans 2 and 3
Between Scans 3 and 4
Practice Sessions
Between Scans 1 and 2
Between Scans 2 and 3
Between Scans 3 and 4
Extensively Trained Trials
Between Scans 1 and 2
Between Scans 2 and 3
Between Scans 3 and 4
Moderately Trained Trials
Between Scans 1 and 2
Between Scans 2 and 3
Between Scans 3 and 4
Minimally Trained Trials
Between Scans 1 and 2
Between Scans 2 and 3
Between Scans 3 and 4
12.00
12.45
12.10
9.70
9.75
10.05
620.80
624.00
643.20
97.00
97.50
100.50
9.70
9.75
10.05
9
10
9
8
4
7
512
256
448
80
40
70
8
4
7
14
14
22
10
14
13
640
896
832
100
140
130
10
14
13
0.34
0.29
0.63
0.14
0.44
0.32
9.40
28.57
20.48
1.46
4.46
3.20
0.14
0.44
0.32
Table 1: Experimental Details for Behavioral Data Acquired Between Scanning Sessions. We
give the minimum, mean, maximum, and standard error of the mean over participants for the following
variables: the number of days between scanning sessions; the number of practice sessions performed at home
between scanning sessions; and the number of trials composed of extensively, moderately, and minimally
trained sequences during home practice between scanning sessions.
Methodological Considerations
Experimental Factors
Effect of Region Size
Recent studies have noted that brain-region size can affect estimates of hard-wired connectivity strength
used in constructing structural connectomes [11, 12]. Although the present work is concerned with functional
connectomes, it is nevertheless relevant to consider whether or not region size could be a driving effect of
the observed core-periphery organization. Importantly, we observe no significant correlation between region
size and flexibility (see Fig. 5), which suggests that region size is not driving the reported results.
Effect of Block Design
Another important factor is the underlying experimental block design and its effect on the correlation
structure between brain regions in a single time window (i.e., in a single layer in the multilayer formalism).
Two brain regions, such as motor cortex (M1) and supplementary motor area (SMA), might be active during
the trial but quiet during the inter-trial interval (ITI). This would lead to a characteristic on-off activity
pattern that is highly correlated with all other regions that also turn on with the task and off during the
ITI. The frequency of this task-related activity (one on-off cycle per trial, where each trial is of length 4 -- 6
TRs) is included in our frequency band of interest (wavelet scale two, whose frequency range is 0.06 -- 0.12
Hz), and it therefore likely plays a role in the observed correlation patterns between brain regions in a single
time window.
Note, however, that our investigations of dynamic network structure -- namely, our computations of flex-
ibility of community allegiance -- probe functional connectivity dynamics at much larger time scales, and the
associated frequencies are an order of magnitude smaller. They lie in the range 0.0083 -- 0.012 Hz, as there
6
.=
Figure 5: Region Size is Uncorrelated with Flexibility. (A) Scatter plot of the size of the brain
region in voxels (averaged over participants) versus the flexibility of the EXT multilayer networks, which
we averaged over the 100 multilayer modularity optimizations and the 20 participants. Data points indicate
−0.009, and
brain regions. The line indicates the best linear fit. Its Pearson correlation coefficient is r
the associated p-value is p
0.92. (B) Box plot over the 20 participants of the squared Pearson correlation
coefficient r2 between the participant-specific region size in voxels and the participant-specific flexibility
averaged over the 100 multilayer modularity optimizations.
.=
is one time window every 40 -- 60 TRs. At these longer time scales, we can probe the effects of both early
learning and extended learning independently of block-design effects.
Specificity of Dynamic Network Organization as a Predictor of Learning
An important consideration is whether there exist (arguably) simpler properties of brain function than
flexibility that could be used to predict learning. We find that the power of activity, the mean connectivity
strength, and parameter estimates from a general linear model (GLM) provide less predictive power than
flexibility.
Measures of Activity and Connectivity.
It is far beyond the scope of this study to perform exhaustive
computations using all possible measures of brain-region activity, so we focus on two common diagnostics.
One is based on functional connectivity, and the other is based on brain activity. To estimate the strength of
functional connectivity, we calculated the mean pairwise coherence between regional wavelet scale-two time
series constructed from the BOLD signal, where we took the mean over all possible pairs of regions and all
EXT experimental blocks extracted from scans on day 1 for a given subject. To estimate the strength of
activity, we calculated the mean signal power of the regional wavelet scale-two time series constructed from
the BOLD signal, where we took the mean over all regions and all EXT experimental blocks extracted from
scans on day 1 for a given subject. We estimate the power Pw2 of the wavelet scale-two time series as the
square of the time series normalized by its length:
Pw2 =Xt
w2(t)2
T
,
(3)
where T is the length of the time series [13, 14].
.=
.=
.=
.=
We found that neither mean pairwise coherence nor mean power of regional activity measured during the
first scanning session could be used to predict learning during the subsequent 10 home training sessions. For
the mean pairwise coherence, we obtained a Pearson correlation of r
0.987.
For the mean power of brain-region activity, we obtained r
0.354. These results indicate
that a prediction similar to that made using the flexibility is not possible using the (arguably) simpler
properties of the mean pairwise coherence or the mean power of regional brain activity. They also suggest
that the dynamic pattern of coherent functional brain activity is more predictive than means of such activity
patterns.
−0.003 and a p-value of p
−0.218 and p
7
0.04
0.03
0.02
0.01
0
B
Pearson r2
500
0
region size (number of voxels)
1000
1500
2000
0.16
0.14
0.12
0.1
0.08
0.06
0.04
A
flexibility
Parameter Estimates for a General Linear Model. We determined relative differences in the BOLD
signal by using a GLM approach for event-related functional data [15, 16]. For each participant, we con-
structed a single design matrix for event-related fMRI by specifying the onset time and duration of all
stimulus events from each scanning session (i.e., the pre-training session and the 3 test sessions). We found
estimations of changes in the BOLD signal related to experimental conditions by using the design matrix
with the GLM. We modeled the duration of each sequence trial as the time elapsed to produce the entire
sequence; in other words, we calculated the movement time (MT), which is a direct measure of the time
spent on a task and leads to accurate modeling of BOLD signals using the GLM [17]. Separate stimulus
vectors indicate each sequence exposure type (EXT, MOD, and MIN) for each scanning session. We took
potential differences in brain activity due to rate of movement into account by using the MT for each trial as
the modeled duration for the corresponding event. We convolved events using the canonical hemodynamic
response function and temporal derivative. Using the canonical hemodynamic response function (HRF) and
its temporal derivative -- we use the implementation in the Statistical Parametric Mapping Toobox (SPM8)
[18] -- we then modeled the events that were specified in the stimulus vectors. From this procedure, we
obtained a pair of beta images for each event type. These images correspond to estimates of the HRF and
its temporal derivative. Using freely available software [18], we then combined the corresponding beta image
pairs for each event type (HRF and its temporal derivative) at the voxel level to form a magnitude image
[19]
H = sign( B1) +q( B1 + B2) ,
(4)
where H is called the "combined amplitude" of the estimation of the BOLD signal using the HRF ( B1) and
its temporal derivative ( B2). 1 This yielded separate magnitude images for each sequence exposure type
(EXT, MOD, and MIN) and session. We then calculated the mean region-based magnitude for each exposure
type and session using regions derived from each subject's grey matter-constrained Harvard-Oxford (HO)
atlas.
.=
−0.10 and the p-value is p
.=
We did not find a significant correlation between the mean parameter estimates averaged over brain
regions for the EXT trials in scanning session 1 and learning of the EXT sequences over the subsequent
approximately 10 home training sessions. The Pearson correlation is r
0.65.
Subject State-Dependence of Dynamic Network Organization
Our finding that temporal core-periphery organization predicts the rate of learning across individuals is
compelling evidence that the relationship between geometrical and temporal core-periphery organization is
related to learning. Nevertheless, it is important to ask whether changes in dynamic community structure
and associated mesoscale network organization are related to tasks or to changes in subjects' physiological
state over the course of longitudinal imaging [20]. It is clear from studies of behavior, peripheral physiology,
and fMRI that subjects can have high levels of anxiety or stress (particularly during their first exposure to
MRI) [21]. To address this issue, we describe additional evidence that supports our conclusions that the
reported changes in dynamic community structure with learning are indeed related to motor tasks.
First, we note that we observed temporal and geometrical core-periphery organization consistently over
all 4 scanning sessions. In Fig. 2 of the present document, we show that the anatomical identity of nodes
in the temporal core, bulk, and periphery are consistent over scanning sessions. In Fig. 3 of this document,
we show that the anatomical identity of nodes in the geometrical core and periphery are also consistent over
scanning sessions. Moreover, Fig. 6 in the main manuscript shows that we observe the relationship between
temporal and geometrical core-periphery organization consistently across scanning sessions.
Second, we assume that the effects of a subject's mental and physiological state (e.g., anxiety) are greatest
during the first imaging session [22]. If this is indeed the case, then there could be significant changes of
network organization between scans 1 (higher anxiety) and 2 (lower anxiety) that might lead to a spurious
interpretation of changes in core-periphery organization. To examine this possibility, we test whether the
changes in dynamic community structure and core-periphery organization with learning are robust to the
removal of scan 1. Importantly, the trends in Figs. 2 and 5 in the main manuscript remain present if we only
examine scans 2 -- 4. We use data from scan 1 for the three box plots located at the point in the horizontal
1In this equation, we use the hat notation to indicate that these values are estimated (rather than directly measured) from
a general linear model for a response variable (such as regional cerebral blood) at each voxel in a given participant [16].
8
axis at which the number of trials is equal to 50. (This is the leftmost point of each panel.) See Table 1 in
the main manuscript. The 9 box plots located at points on the horizontal axis at which the number of trials
is greater than 50 use data from scans 2 -- 4. Therefore, when we examine only scans 2 -- 4, we still observe a
decrease in maximum modularity, an increase in the number of communities, an increase in flexibility, and
a decrease in the variance of the geometrical core score with learning.
.=
.=
.=
.=
−0.52 (so p
.=
0.615, p
.=
0.00137) and a non-significant effect of scanning session (F (3, 57)
Finally, task-related fMRI BOLD activation magnitude in core, bulk, and peripheral brain regions are
not altered significantly across scanning sessions. We employed a repeated-measures analysis of variance
(ANOVA) on the training-depth-averaged GLM parameter estimates [23]. We treated core, bulk, and periph-
ery designations as categorical factors, and we treated scanning session as a repeated measure. We found a sig-
nificant main effect (i.e., single-factor effect) of core, bulk, and periphery (an F-statistic [23] of F (2, 38)
7.88
and a p-value of p
0.584).
These results suggest that a systematic change in the hemodynamic response function across scanning ses-
sions is unlikely to be responsible for the observed learning-related changes in dynamic community structure.
Furthermore, we observe that mean GLM parameter estimates in core, bulk, and peripheral brain regions
are not correlated significantly with the reported changes in core-periphery structure that accompany learn-
ing. The Pearson correlation coefficient between parameter estimates and the variance of the geometrical
core score for nodes in the temporal core is r
0.52), for nodes in the
temporal bulk is r
0.08).
These results provide further evidence that BOLD activation magnitude and dynamic community structure
provide distinct insights.
0.86), and for nodes in the temporal periphery is r
0.20 (which gives a p-value of p
.=
−0.05 (so p
.=
.=
.=
Temporal Core-Periphery Organization and Task-Related Activations
One of the strengths of our approach is that we examine the organization of whole-brain functional connec-
tivity and thereby remain sensitive to a wide variety of learning-related changes in the brain that could not
be identified using a traditional GLM analysis. Nevertheless, it is useful to explore the relationship between
dynamic community structure and task-related activations. In Fig. 6, we show that regions in the temporal
core tend to be regions with strong task-related activations, as evinced by high (and positive) values of
mean GLM parameter estimates. Conversely, regions in the temporal bulk and periphery tend to lack strong
task-related activations, as evinced by low (and negative) values of mean GLM parameter estimates. These
results are consistent with our interpretation that the temporal core consists of a small set of regions that
are required to perform a given task and that the temporal periphery consists of a set of regions that are
associated more peripherally with the task and which are activated in a transient manner.
Dynamic Community Detection
In the multilayer modularity quality function (see the "Materials and Methods" section of the main manuscript),
we need to choose values for two parameters [6]: a structural resolution parameter γ and a temporal resolution
parameter ω. We now examine the effects of these choices on our results.
Effect of Structural Resolution Parameter
In the main manuscript, we used a structural resolution parameter value of γ = 1, which is the most common
choice when optimizing the single-layer and multilayer modularity quality functions [4, 5, 24]. In this case,
A − γP = A − P, and one is simply subtracting the optimization null model P from the adjacency tensor
A. One can decrease γ to access community structure at smaller spatial scales (i.e., to examine smaller
communities) or increase it to access community structure at larger spatial scales (i.e., to examine larger
communities). By examining network diagnostics over a range of γ values, we explore the spatial specificity
of our results.
The mean number of communities in the partitions that we obtained by optimizing multilayer modularity
Q varies from the minimum (1) to the maximum (112) possible value for γ approximately in the interval
[0.8, 2.5] (see Fig. 7A). We investigate this transition in greater detail in Figs. 7C,D. Near the value γ = 1,
the number of regions in the bulk dips to about 65, whereas the number of regions in the core and periphery
rise to about 20 and 25, respectively. Observe the dip of the bulk curve and bumps of the core and periphery
curves in Fig. 7D. These features occur for γ approximately in the interval [0.88, 1.22], which corresponds
9
Figure 6: Temporal Core-Periphery Organization and Task-Related Activations. Mean GLM
parameter estimates for the temporal core (cyan; circles), bulk (gold; squares), and periphery (maroon;
diamonds) of dynamic networks defined by the trial blocks in which participants practiced sequences that
would eventually be (A) extensively trained, (B) moderately trained, and (C) minimally trained for data from
scanning sessions 1 (first day of training), 2 (after approximately 2 weeks of training), 3 (after approximately
4 weeks of training), and 4 (after approximately 6 weeks of training).
Figure 7:
(Supplementary Material) Effect of Structural Resolution Parameter. (A,B) Number
of communities and (C,D) number of regions in the temporal core (cyan; circles), temporal bulk (gold;
squares), and temporal periphery (maroon; diamonds) as a function of the structural resolution parameter
γ, where we considered (A,C) γ ∈ [0.2, 5] in increments of ∆γ = 0.2 and (B,D) γ ∈ [0.8, 1.8] in increments
of ∆γ = 0.01. We averaged the values in panels (A) and (B) over 100 multilayer modularity optimizations
and over the 20 participants.
10
minimal training
2
3
scanning session
4
1
0.5
0
−0.5
C
mean parameter estimates
4
−1
1
moderate training
1
0.5
0
−0.5
B
mean parameter estimates
extensive training
core
bulk
periphery
2
3
scanning session
4
−1
1
2
3
scanning session
1
0.5
0
−0.5
A
mean parameter estimates
−1
1
100
80
60
40
20
D
number of regions
core
bulk
periphery
80
80
60
60
40
40
20
20
0
0
1.8
0.8
0
5
structural resolution parameter
structural resolution parameter
1.2
2
1.4
3
1.6
4
1
1
100
100
B
number of communities
number of regions
100
80
60
40
20
0
0
5
structural resolution parameter
1
2
3
4
number of communitiesA
C
100
100
80
80
60
60
40
40
20
20
number of communities
number of regions
0
0
0
5
1.8
0.8
structural resolution parameter
structural resolution parameter
2
1.2
3
1.4
4
1.6
1
1
0
0.8
1.8
structural resolution parameter
1.6
1.4
1.2
1
to partitions that are composed of between approximately 3 and approximately 20 communities (with an
associated mean community size of between approximately 6 and approximately 37 brain regions; see Fig.7B).
This supports our claim that the temporal core-periphery structure that we examine in this study is a genuine
mesoscale feature of coherent brain dynamics.
Effect of Temporal Resolution Parameter
In the main manuscript, we used a temporal resolution parameter value of ω = 1. The value ω = 1 ensures
that the inter -layer coupling is equal to the maximum possible value of the intra-layer coupling, which
we compute from the magnitude-squared coherence (which is constrained to lie in the interval [0, 1]).
It
is important to examine the robustness of results for different values of this parameter, and investigating
dynamic network structure at other values of ω can also provide additional insights [6]. For example, one can
decrease ω to encourage greater variability in community assignments of nodes across individual layers (i.e.,
across time in temporal networks) or increase it to encourage such community assignments to be more similar
across layers. Recall that each node in the temporal multilayer network represents a single brain region at a
specified time, and different nodes that represent the same brain region at different times become more likely
to be assigned to the same multilayer community as ω is increased. By examining network diagnostics over
a range of ω values, we can quantify the robustness of our results to differing amounts of temporal variation
in community structure.
We varied ω from 0.1 to 2 in increments of ∆ω = 0.1. As expected, we find that the number of communities
identified in the optimization of the multilayer modularity quality function decreases as ω is increased (see
Fig. 8A). This is consistent with the fact that greater variation of community assignments across time is
possible for smaller values of ω. Variation between community assignments of nodes in individual layers
can occur in two ways: (1) a small number of regions change community membership from one layer to the
next, but the majority of regions retain their community membership; or (2) entire communities lose their
identities (via fragmentation, extinction, union, and/or recombination), such that the algorithm identifies
either the "death" of a community that was present in the previous layer but is not present in the current
layer or the "birth" of a community that was not present in the previous layer but is present in the current
layer.
For each value of ω, we examined the robustness of our division of brain regions into a temporal core, a
temporal bulk, and a temporal periphery using the same procedure that we employed for ω = 1. Namely,
we defined a temporal core and temporal periphery as those brain regions that were composed, respectively,
of the brain regions below and above the 95% confidence interval of the nodal null model. In Fig. 8B, we
report the number of regions in each group as a function of ω. Interestingly, the number of brain regions that
we identified as part of the temporal core varied little over the examined range of ω values; it remained at
approximately 17.0±1.1. In fact, 15 of the 17 regions that we identified as part of the temporal core at ω = 1
were also identified as part of the temporal core at all other values of ω that we examined. The number of
regions in the temporal bulk and temporal periphery varied more (with values of approximately 75.6 ± 7.4
for the bulk and approximately 19.4± 6.8 for the periphery), which suggests that the separation between the
temporal bulk and temporal periphery is less drastic than that between temporal core and temporal bulk.
Indeed, the mean flexibility of the core is less similar to the mean flexibility of the bulk than is the latter to
the mean flexibility of the periphery. See Fig. 3 of the main manuscript and Figs. 1 and 2 of this supplement.
References
[1] Newman MEJ, Girvan M (2004) Finding and evaluating community structure in networks. Phys Rev
E 69: 026113.
[2] Newman MEJ (2004) Fast algorithm for detecting community structure in networks. Phys Rev E 69:
066133.
[3] Newman MEJ (2006) Modularity and community structure in networks. Proc Natl Acad Sci USA 103:
8577 -- 8582.
11
Figure 8: Effect of Temporal Resolution Parameter. (A) Number of communities averaged over
100 multilayer modularity optimizations and over 20 participants as a function of the temporal resolution
parameter ω. (B) Number of regions that we identified as part of the temporal core (cyan; circles), temporal
bulk (gold; squares), and temporal periphery (maroon; diamonds) as we vary ω from 0.1 to 2 in increments
of ∆ω = 0.1.
[4] Porter MA, Onnela JP, Mucha PJ (2009) Communities in networks. Not Amer Math Soc 56: 1082 -- 1097,
1164 -- 1166.
[5] Fortunato S (2010) Community detection in graphs. Phys Rep 486: 75 -- 174.
[6] Bassett DS, Porter MA, Wymbs NF, Grafton ST, Carlson JM, et al. (2012) Robust detection of dynamic
community structure in networks. Chaos 23: 013142.
[7] Simpson SL, Moussa MN, Laurienti PJ (2012) An exponential random graph modeling approach to
creating group-based representative whole-brain connectivity networks. NeuroImage 60: 1117 -- 1126.
[8] Traud AL, Kelsic ED, Mucha PJ, Porter MA (2011) Comparing community structure to characteristics
in online collegiate social networks. SIAM Rev 53: 526 -- 543.
[9] Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TEJ, et al. (2004) Advances in func-
tional and structural MR image analysis and implementation as FSL. NeuroImage 23: 208 -- 219.
[10] Woolrich MW, Jbabdi W, Patenaude B, Chappell M, Makni S, et al. (2009) Bayesian analysis of
neuroimaging data in FSL. NeuroImage 45: S173 -- S186.
[11] Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, et al. (2008) Mapping the structural core
of human cerebral cortex. PLoS Biol 6: e159.
[12] Bassett DS, Brown JA, Deshpande V, Carlson JM, Grafton ST (2011) Conserved and variable archi-
tecture of human white matter connectivity. NeuroImage 54: 1262 -- 1279.
[13] MATLAB (2012) Measuring signal power. MATLAB's Help Menu .
[14] Smith SW (1997) Digital Signal Processing: A Guide for Engineers and Scientists. California Technical
Pub.
[15] Lazar N (2010) The Statistical Analysis of Functional MRI Data. Springer-Verlag.
[16] Friston KJ, Holmes AP, Worsley KJ, Poline JP, Frith CD, et al. (1994) Statistical parametric maps in
functional imaging: A general linear approach. Human Brain Mapping 2: 189 -- 210.
12
core
bulk periphery
B
100
80
60
40
20
number of regions
0
0
2
temporal resolution parameter
0.5
1.5
1
0
2
temporal resolution parameter
1.5
0.5
1
234567
number of communitiesA
[17] Grinband J, Wager TD, Lindquist M, Ferrera VP, Hirsch J (2008) Detection of time-varying signals in
event-related fMRI designs. NeuroImage 43: 509 -- 520.
[18] Steffener J, Tabert M, Reuben A, Stern Y (2010) Investigating hemodynamic response variability at
the group level using basis functions. NeuroImage 49: 2113 -- 2122.
[19] Calhoun VD, Adali T, Pekar JJ (2004) A method for comparing group fMRI data using independent
component analysis: application to visual, motor and visuomotor tasks. Magn Reson Imaging 22:
1181 -- 1191.
[20] Johnstone T, Somerville LH, Alexander AL, Oakes TR, Davidson RJ, et al. (2005) Stability of amygdala
BOLD response to fearful faces over multiple scan sessions. NeuroImage 25: 1112 -- 1123.
[21] Lueken U, Muehlhan M, Evens R, Wittchen HU, Kirschbaum C (2012) Within and between session
changes in subjective and neuroendocrine stress parameters during magnetic resonance imaging: A
controlled scanner training study. Psychoneuroendocrinology 37: 1299 -- 1308.
[22] Chapman HA, Bernier D, Rusak B (2010) MRI-related anxiety levels change within and between re-
peated scanning sessions. Psychiatry Res 182: 160 -- 164.
[23] Agresti A, Franklin CA (2007) Statistics: The Art and Science of Learning From Data. Prentice Hall.
[24] Reichardt J, Bornholdt S (2006) Statistical mechanics of community detection. Phys Rev E 74: 016110.
13
Minimum Mean Maximum Standard Error
Extensively Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
Moderately Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
Minimally Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
Length of Extensively Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
Length of Moderately Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
Length of Minimally Trained Blocks
During Scan 1
During Scan 2
During Scan 3
During Scan 4
6.00
10.00
10.00
8.00
5.00
10.00
10.00
8.00
7.00
10.00
10.00
8.00
52.50
35.50
35.40
34.60
50.80
39.70
37.60
37.60
52.10
44.10
42.50
39.70
9.70
10.00
10.00
9.90
9.70
10.00
10.00
9.90
9.80
10.00
10.00
9.90
61.94
42.36
40.79
40.30
61.67
47.56
45.07
43.83
61.19
50.02
47.37
45.79
10.00
10.00
10.00
10.00
11.00
10.00
10.00
10.00
11.00
10.00
10.00
10.00
72.20
45.90
45.50
45.70
72.60
57.20
52.80
50.60
70.60
57.70
54.50
54.10
0.21
0.00
0.00
0.10
0.27
0.00
0.00
0.10
0.18
0.00
0.00
0.10
1.34
0.72
0.77
0.87
1.26
0.80
0.67
0.79
1.29
0.73
0.71
0.70
Table 2: Experimental Details for Brain Imaging Data Acquired During Scanning Sessions. In
the top three rows, we give the mean, minimum, maximum, and standard error over participants for the
number of blocks composed of extensively, moderately, and minimally trained sequences during scanning
sessions. In the bottom three rows, we give (in TRs) the mean, minimum, maximum, and standard error of
the length over blocks composed of extensively, moderately, and minimally trained sequences during scanning
sessions.
14
Region Name
Frontal pole
Insular cortex
Superior frontal gyrus
Middle frontal gyrus
Affiliation
Region Name
B(R) B(L)
Cingulate gyrus, anterior
B(R) B(L)
Cingulate gyrus, posterior
B(R) B(L)
Precuneus cortex
B(R) B(L)
Cuneus cortex
Inferior frontal gyrus, pars triangularis
B(R) P(L)
Orbital frontal cortex
Affiliation
B(R) B(L)
P(R) B(L)
B(R) B(L)
C(R) C(L)
B(R) B(L)
Inferior frontal gyrus, pars opercularis
B(R) B(L)
Parahippocampal gyrus, anterior
B(R) B(L)
Precentral gyrus
Temporal pole
C(R) C(L)
Parahippocampal gyrus, posterior
B(R) P(L)
B(R) B(L)
Lingual gyrus
C(R) C(L)
Superior temporal gyrus, anterior
B(R) B(L)
Temporal fusiform cortex, anterior
P(R) B(L)
Superior temporal gyrus, posterior
B(R) B(L)
Temporal fusiform cortex, posterior
P(R) P(L)
Middle temporal gyrus, anterior
B(R) B(L)
Temporal occipital fusiform cortex
P(R) P(L)
Middle temporal gyrus, posterior
B(R) B(L)
Occipital fusiform gyrus
Middle temporal gyrus, temporooccipital
P(R) P(L)
Frontal operculum cortex
Inferior temporal gyrus, anterior
B(R) B(L)
Central opercular cortex
Inferior temporal gyrus, posterior
B(R) B(L)
Parietal operculum cortex
Inferior temporal gyrus, temporooccipital
B(R) B(L)
Planum polare
Postcentral gyrus
P(R) C(L)
Heschl's gyrus
Superior parietal lobule
B(R) C(L)
Planum temporale
Supramarginal gyrus, anterior
B(R) C(L)
Supercalcarine cortex
Supramarginal gyrus, posterior
B(R) P(L)
Occipital pole
Angular gyrus
P(R) B(L)
Caudate
Lateral occipital cortex, superior
P(R) P(L)
Putamen
Lateral occipital cortex, inferior
P(R) B(L)
Globus pallidus
Intracalcarine cortex
Frontal medial cortex
C(R) C(L)
Thalamus
B(R) B(L)
Nucleus Accumbens
Supplemental motor area
C(R) C(L)
Parahippocampal gyrus
Subcallosal cortex
Paracingulate gyrus
B(R) B(L)
Hippocampus
B(R) B(L)
Brainstem
P(R) P(L)
B(R) P(L)
B(R) B(L)
P(R) B(L)
C(R) B(L)
C(R) C(L)
B(R) B(L)
C(R) C(L)
C(R) B(L)
P(R) B(L)
P(R) B(L)
P(R) B(L)
P(R) P(L)
B(R) B(L)
B(R) B(L)
B(R) B(L)
B(R) B(L)
Table 3: Brain regions in the Harvard-Oxford (HO) Cortical and Subcortical Parcellation Scheme provided
by FSL [9, 10] and their affiliation to the temporal core (C; cyan), bulk (B; gold), and periphery (P; maroon) for both left
(L) and right (R) hemispheres.
15
|
1712.00462 | 1 | 1712 | 2017-12-01T19:09:32 | Functional brain networks reveal the existence of cognitive reserve and the interplay between network topology and dynamics | [
"q-bio.NC"
] | We investigated how the organization of functional brain networks was related to cognitive reserve (CR) during a memory task in healthy aging. We obtained the magnetoencephalographic functional networks of 20 elders with a high or low CR level to analyse the differences at network features. We reported a negative correlation between synchronization of the whole network and CR, and observed differences both at the node and at the network level in: the average shortest path and the network outreach. Individuals with high CR required functional networks with lower links to successfully carry out the memory task. These results may indicate that those individuals with low CR level exhibited a dual pattern of compensation and network impairment, since their functioning was more energetically costly to perform the task as the high CR group. Additionally, we evaluated how the dynamical properties of the different brain regions were correlated to the network parameters obtaining that entropy was positively correlated with the strength and clustering coefficient, while complexity behaved conversely. Consequently, highly connected nodes of the functional networks showed a more stochastic and less complex signal. We consider that network approach may be a relevant tool to better understand brain functioning in aging. | q-bio.NC | q-bio | Functional brain networks reveal the existence of cognitive reserve and the interplay
between network topology and dynamics
Johann H. Martínez *, María E. López *, Pedro Ariza, Mario Chavez, José A. Pineda- Pardo,
David López-Sanz, Pedro Gil, Fernando Maestú and Javier M. Buldú.
1 INSERM, Institut du Cerveau et de la Moelle Epinière (ICM). Hôpital Pitié Salpétrière, Paris, France
2 Laboratory of Biological Networks, Center for Biomedical Technology (CTB), Madrid, Spain
3 Department of Basic Psychology II, Complutense University of Madrid, Madrid, Spain
4 7CNRS-UMR 7225, Hôpital Pitié-Salpetrière, Paris, France.
5 Centro Integral de Neurociencias AC (CINAC), HM Puerta del Sur, Madrid, Spain
6 CEU San Pablo University, Madrid, Spain.
7Laboratory of Cognitive and Computational Neuroscience (UCM-UPM), Center for Biomedical
Technology (CTB), Madrid, Spain
8 Institute of Sanitary Investigation [IdISSC], San Carlos University Hospital, Madrid, Spain
9 Geriatrics Department, San Carlos University Hospital, Madrid, Spain.
10 Complex Systems Group, Universidad Rey Juan Carlos, Madrid, Spain
* These authors have contributed equally and must be considered as the first authors
ABSTRACT
We investigated how the organization of functional brain networks was related to cognitive
reserve (CR) during a memory task in healthy aging. We obtained the magnetoencephalographic
functional networks of 20 elders with a high or low CR level to analyse the differences at network
features. We reported a negative correlation between synchronization of the whole network and
CR, and observed differences both at the node and at the network level in: the average shortest
path and the network outreach. Individuals with high CR required functional networks with lower
links to successfully carry out the memory task. These results may indicate that those individuals
with low CR level exhibited a dual pattern of compensation and network impairment, since their
functioning was more energetically costly to perform the task as the high CR group. Additionally,
we evaluated how the dynamical properties of the different brain regions were correlated to the
network parameters obtaining that entropy was positively correlated with the strength and
clustering coefficient, while complexity behaved conversely. Consequently, highly connected
nodes of the functional networks showed a more stochastic and less complex signal. We consider
that network approach may be a relevant tool to better understand brain functioning in aging.
Keywords: Magnetoencephalography, Cognitive Reserve, Healthy aging, Brain networks,
Efficiency.
INTRODUCTION
In recent years, there is an increased interest in studying lifestyle factors that may help to achieve
a successful aging. These factors, such as educational level, leisure activities or intelligence
quotient (IQ) have been related to the concept of reserve, which emerges from the absence of a
direct relationship between the severity of brain pathology and its clinical manifestations 1.
Two non-mutually exclusive models have been proposed to address these counterintuitive
observations 1: a passive or brain reserve (BR) model, which considers that the amount of neural
substrate (e.g. the brain size or neuron quantity) determines the threshold beyond which clinical
and functional deficits emerge 2,3; and an active or cognitive reserve (CR) model, that refers to
the ability to use the existent resources as flexible and efficiently as possible during the execution
of cognitive tasks (neural reserve) or when coping with brain pathology (neural compensation) 4.
Healthy aging and dementia, especially Alzheimer's disease (AD), have been considered optimal
models to understand the role of reserve 4–6.
Among lifestyle variables, the educational level (alone or in combination with occupational
attainment) is one of the most studied and important CR proxies. In fact, some studies have
considered the educational level as a "protective" factor by reducing the incidence of dementia 5,
and its pathological consequences in AD patients 7,8. However, recent studies did not find this
effect 9,10, suggesting that CR actually may have a "masking" effect.
One of the new approaches to explore how brain functioning may be modulated by CR is
functional network organization (based on functional connectivity; FC), although there is still
scarcity of studies focused on this field. Solé-Padullés et al. 11 described that healthy elders with
higher CR showed reduced functional magnetic resonance imaging (fMRI) activation during
cognitive processing. In the same line, a previous study from our group 12 observed that healthy
elders with a high CR score in comparison with those with low CR score, exhibited less
magnetoencephalographic (MEG) functional connectivity during the performance of a memory
task. These results may suggest that higher CR is related to more effective use of cerebral
resources although this hypothesis has not been tested yet.
In addition, FC may be used to explore brain's organization in terms of structural and functional
networks, by the application of different methodologies coming from Network Science 13. This
kind of analysis can introduce a different perspective to better understand how CR modules
functional resilience in healthy aging. Recently, Yoo et al. 14 and Marques et al. 15 have proposed
the analysis of the topology of the associated functional networks as an alternative way of
quantifying CR. Interestingly, functional brain networks obtained during resting state show a
positive correlation between CR and network efficiency 16,17. Network efficiency, first introduced
by Latora et al. 18 is a metric evaluating the inverse of the topological distance between nodes of
a network. In the context of functional brain networks, the topological distance between two brain
regions is normally defined as the inverse of their level of coordination. In this way, nodes that
are not functionally connected (i.e., with zero coordination) have infinite distance between them.
However, information between not connected regions (at an infinite topological distance) can
travel indirectly through intermediate nodes. Network efficiency measures how close or distant
the nodes of a network are in terms of their topological distance using the functional connections
through intermediate nodes 18.Under this framework, the positive correlation between CR and
network efficiency during resting state relies on fMRI recordings, which report higher levels of
synchronization between brain regions in individuals with higher CR15,16. However, during a
cognitive process, the correlation between CR and synchronization turns to be negative, as shown
in 11. With the aim of clarifying whether CR is related to the topology of the functional networks
during a memory task, we are concerned about the biomagnetic network profiles of two groups
of healthy elders with different CR levels. Specifically, we obtained the MEG recordings of both
groups during the performance of a modified Stenberg's test 17 and we explored the topological
properties of the corresponding functional networks. Among the diversity of network metrics that
one may extract from the functional networks we focused on the most extended ones: the
network strength, which is a measure of the average level of synchronization along the
functional network 19; the weighted clustering coefficient, measuring the existence of
triangles in the network and related to the local resilience 20; the eigenvector centrality a
measure of node importance obtained from the eigenvector associated to the largest
eigenvalue of the adjacency matrix 21; the average shortest path d(i), measuring the
average number of steps to travel from one node to another through the shortest
topological distance 22; the strength of the nearest neighbors snn(i), indicating what is the
average strength (i.e., sum of the weights of a node) of the neighbors (i.e., nodes directly
connected) of a given node 19; and the outreach o(i), a network metric obtained by
multiplying the weight of the connections by their Euclidean distance 23 (see Supp. Info.
for a detailed definition of all network metrics).
We also quantified the dynamical properties of the brain regions in terms of entropy and
complexity to compare both high and low CR groups. To the best of our knowledge, the present
MEG network's approach, where the study of network metrics is combined with the dynamical
properties of the nodes, has never been used within this field of research. Combining both
approaches we have been able to detect differences between low and high CR individuals at the
network level, showing that the network outreach and, consequently, the energetic expenses, is
higher in individuals with low CR. At the same time, we show how the topological properties of
the nodes (i.e., brain regions) are related to the complexity and entropy of their corresponding
time series.
MATERIALS AND METHODS
Participants
The sample consisted of 20 healthy elders recruited from the "Geriatric Unit of the University
Hospital San Carlos" (Madrid, Spain). All of them were right-handed 24, native Spanish speakers,
and between 65 and 85 years old. Regarding the neuropsychological assessment, all of them met
the following criteria: 1) a score above 28 in MMSE 25; 2) no memory impairment as evidenced
by delayed recall from Logical Memory II subtest of the Wechsler Memory Scale Revised (WMS-
III-R; 26); and 3) normal daily living activities measured by the Spanish version of the Functional
Assessment Scale 27. Exclusion criteria included: 1) a history of psychiatric or neurological
disease; 2) psychoactive drugs consumption; and 3) severe sensory or comprehension deficits.
The 20 participants did not differ in physical, cognitive and social activities during late-life, and
were further subdivided into two subgroups according to their CR level. The CR index (CRI) was
estimated for each individual by adding the following two categories: 1) formal educational level
(from 1 –illiterate/ functional illiterate- to 5- superior studies-) and 2) occupational attainment
(from 1 –housewife- to 5 -business management/research-). According to their CRI score, 12
participants were classified into the low CR group if their CRI score was between 1 and 5; or into
the high CR group (8 subjects) if their CRI score was between 6 and 10. In addition, both groups
did not differ in age and in their performance during the cognitive task (see table 1).
Ethics statement: Methods were carried out in "accordance" with the approved guidelines. The
investigation was approved by the local Ethics Committee (Madrid) and all participants signed a
written informed consent before the MEG recordings.
Age
High- CR (n=8)
Low- CR (n=12)
P values
67.3±7.4
69.7±6.6
Educational Level
4.5±0.7
Occupation attainment
4.5±0.5
CRI score
Cognitive activity
Physical activity
Social activity
9±0.6
4.2±2.3
2.4±1.5
1.6±1.2
2.4±0.7
1.6±0.5
3.7±1.1
2.8±1.4
2.7±1
1.7±1
Acc. Smaqe
114±10.7
106±14.1
p>0.05
p<0.05*
p<0.05*
p<0.05*
p>0.05
p>0.05
p>0.05
p>0.05
Table 1. Mean values ± standard deviation for high and low CR groups. Significant p values are
p<0.05 and marked with an asterisk (*). Both groups differ in their educational level, occupation
attainment and CRI scores, but not in the other variables. Acc. Smaqe, accuracy in the modified
Stenberg's memory task.
MEG task
A modified version of the Sternberg's letter probe task was employed 17 (see Figure 1). Each
subject was asked to remember a single set of 5 letters (i.e. 'SMAQE'). After that, there were
presented a series of single letters, one at a time (500ms in duration with a random ISI between 2
and 3 seconds). They were asked to make a match/non-match decision by pressing a button with
their right hand if the presented letter was one of the five initially memorized. The experiment
consisted of 250 letters, in which half were initially presented and the rest were distracters,
meaning they were not in the initial 5 letter set. All of the participants completed a training session
until demonstrating they had correctly memorized the initial letter set. The task was projected
through a LCD video projector (SONY VPLX600E) located outside the shielded MEG room
through a series of mirrors. The screen was suspended approximately 1 meter above the
participant's face. The letters subtended 1, 8 and 3 degrees of horizontal and vertical visual angle
respectively.
Figure 1. Representation of the modified version of the Sternberg's letter probe task. In the
encoding phase, participants are instructed to memorize 5 letters (i.e.: "SMAQE"). In the
recognition phase, participants are instructed to make a match/non-match button-press to indicate
that the presented letter matched any of the encoded ones.
MEG recordings
MEG data was recorded during the execution of the modified Sternberg's task with a 254 Hz
sampling rate and a band pass of 0.5 to 50 Hz, using a 148 channel whole-head magnetometer
(MAGNES 2500 WH, 4D Neuroimaging) placed in a magnetically shielded room. In order to
reduce external noise we employed an algorithm using reference channels at a distance from the
MEG sensors. After trial segmentation, artifacts were visually inspected by an expert and epochs
containing muscular artifacts or eye movements were discarded for further analysis. Only hits
were considered, removing false alarms, correct rejections and omissions. We randomly selected
a set of 35 trials for each individual, since it was the lowest number of successful epochs for all
subjects in the study. Each trial consisted of each 230 time points for each of the 148 channels.
Trials are not necessary consecutive, however, this issue is not crucial for the analysis of the
coordination between sensors since each trial is analyzed individually and averaged later.
Functional connectivity analysis
To calculate the connectivity networks for each subject we used Synchronization Likelihood
algorithm (SL) 28 which is a nonlinear measure of the synchronized activity that has been proven
to be a suitable quantifier for datasets obtained from magnetoencephalographic recordings 29.
Specifically, the SL algorithm detects windows of repeated patterns within the time series of a
channel A and, next, checks whether a channel B also shows a repeated pattern at the same time
windows, no matter if it is the same or different to that observed in channel A. The range of values
of SL is 0 ≤ SL ≤ 1, being 0 when the time series of both channels are uncorrelated, and 1 for
maximal synchronization 28. SL was calculated for each pair of sensors since, in our study, sensors
were the nodes of the functional networks. All pairs of SL are then included into a correlation
matrix 𝑊{𝑤𝑖𝑗} where 𝑤𝑖𝑗 have values comprised between ~ 0.05 and ~ 0.5. We apply a linear
normalization that leads to a probability matrix 𝑃{𝑝𝑖𝑗}, where the values of 𝑝𝑖𝑗 are obtained as
𝑝𝑖𝑗 =
𝑤𝑖𝑗 − min [𝑤𝑖𝑗]
max[𝑤𝑖𝑗] − min [𝑤𝑖𝑗]
. (1)
In this way, the probability matrix 𝑃{𝑝𝑖𝑗}, whose values are within the interval [0,1], reflects the
probability of the presence of a link between nodes (brain regions) i and j.
Graph metrics
We calculated a series of metrics related with the role of the nodes within the network: the strength
s(i), the weighted clustering cw(i), the eigenvector centrality ec(i), the average shortest path d(i),
the strength of the nearest neighbours snn(i) and the outreach o(i). We used the inverse of the
values of the probability matrix 𝑃{𝑝𝑖𝑗} for obtaining the "topological distances" between nodes
to quantify the shortest paths. Another parameters such as the within-module degree z(i) (also
kwon as the z-score) and the participation coefficient p(i) 23,30 were also computed using the
community affiliation vector Ccom(i) extracted from the classical partition of the brain into six
lobes: left-frontal, right-frontal, central, left-temporal, right-temporal and occipital. A series of
global network features such as the global efficiency Eg and the average shortest path (d) were
also calculated. The node average (average along the same node for all subjects within the same
group) of several metrics was computed to obtain the following mean values: 𝑠(𝑖), 𝑐𝑤(𝑖), 𝑒𝑣(𝑖),
𝑧(𝑖), 𝑑(𝑖), ), 𝑠𝑛𝑛(𝑖) and 𝑜(𝑖). Network averages of the preceding features were also appraised in
order to compare both groups. Finally, we obtained the global efficiency 𝐸𝑔 of the networks from
the harmonic mean of the inverse of the shortest paths.
We constructed a set of randomized versions of the former functional networks, in order to
evaluate to what extent the deviation of the network parameters between groups was a
consequence of a topological reorganization or, on the contrary, just a matter of the number of
links and their weights. With this purpose, we generated a group of 100 networks for each of the
functional network. The randomization maintains the value of the links' weights by reshuffling
the components of the weighted probability matrix 𝑃{𝑝𝑖𝑗} (see 23 for a similar normalization). In
this way, we guaranteed that the average strength of the network was maintained. Next, we
calculated the network parameters for each of the randomized versions and obtained an average
value of each metric. Finally, we normalized all network parameters with the average of the set
of surrogate matrices, i.e., for a parameter X its normalized value would be:
𝑋 =
𝑋
𝑋𝑟𝑎𝑛𝑑
(2)
Note that this normalization allows focusing on changes related to the structure of the functional
networks, since they exclude variations of the average weights of the networks (and their
influence on the network metrics).
Evaluation of the dynamical properties of the nodes
In order to calculate the entropy and complexity brain signal, we used for each of the nodes in the
network their corresponding time series comprising the 35 trials of the whole experiment. Since
each trial has 230 time steps, we obtained times series of M = 8050 points for each node of each
functional network.
Next, we applied the methodology introduced by 31 to quantify the dynamical properties of
nonlinear dynamical systems. We calculated the normalized permutation entropy of the signal (S)
as described in 32 to capture the level of uncertainty of a signal.
We calculated the complexity (H) of the signal for each subject following the procedure described
in 33. This measure is effective to quantify the complexity of different dynamical systems 34.
Statistical analysis
We performed permutation tests as well as non-parametric ones for all statistical hypotheses of
this work. The permutation test is based on a t-statistical distance between two samples. For
robustness, we constructed a distribution of 5000 different p-values associated to the number of
permutations in each hypothesis. In each distribution, we found the percentile associated to the
original p-value, thus dividing this percentile by the number of permutations as a correction of
the resampling. We considered the rejection of the null hypothesis when the aforementioned ratio
is bellow or equal to α = 0.05 level of significance. Results of permutation tests are presented
throughout the paper. Complimentary, we performed a nonparametric Mann-Whitney U-test
(rank-sum test) to show how the results do not change. The latter results are shown in the
Supplementary Information.
RESULTS
Micro-scale: Differences at the node level
Firstly, we analysed those nodes with the highest eigenvector centrality in both groups
(𝑒𝑣ℎ𝑖𝑔ℎ,𝑙𝑜𝑤). Figure 2 shows that both groups exhibited the most influencing nodes placed at the
same cortical regions (i.e., at the occipital lobe). Nevertheless, we observed signs of a
displacement of the node centrality toward the central and left-temporal lobes.
Figure 2. Back view for eigenvector centrality (𝒆𝒗(𝒊)) for high (A) and low (B) CR groups.
Circles highlight the ten nodes with the highest 𝒆𝒗(𝒊). Note that they are localized close to the
occipital lobe in the low CR group (blue circles) and slightly shift toward the central lobe for the
high CR group (red circles).
In order to quantify the differences between groups, we obtained the average difference of the
node centrality (∆𝒆𝒗(𝒊)), within-module degree ∆𝒛(𝒊) and participation coefficient ∆𝒑(𝒊) of each
node for both groups.
Figure 3 shows the significant differences obtained in node centrality (A), within-module degree
(B) and participation coefficient (C) between groups (𝒑𝒗𝒂𝒍 ≤ 0.05). The high CR group showed
higher values of the eigenvector centrality in nodes over the central lobe, near the parieto-occipital
sulcus, as well as some tiny regions in the frontal lobe (i.e., 𝒆𝒗(𝒊)𝒉𝒊𝒈𝒉 > 𝒆𝒗(𝒊)𝒍𝒐𝒘 and
subsequently, ∆𝒆𝒗(𝒊) > 𝟎). Conversely, cortical tissue placed at the left- temporal and occipital
lobes have higher centrality in the low CR population (i.e., ∆𝒆𝒗(𝒊) < 0). High CR group presented
higher within-module degree over both temporal brain areas and in the occipital lobe, while low
CR group showed higher within-module degree in some regions located at the central lobe.
Finally, there were few statistical differences between groups in the participation coefficient,
showing the high CR group higher values than the low CR one, especially at regions placed at the
frontal lobe. Statistical analysis based on a Man-Whitney U test shows similar results (see Supp.
Info. for details).
It is worth mentioning that both the eigenvector centrality and the within-module degree are
quantifying the percentage of importance a node has with regard to the other nodes of the network.
Therefore, they should be interpreted as a way of evaluating the increase/decrease of importance
with respect to all nodes, in the case of eigenvector centrality, or the nodes belonging to the same
lobe, in the case of the within-module degree. Note that, if nodes belonging to the temporal and
occipital lobes have higher centrality, it should be at expenses of other nodes of the functional
network (in this case, those located at the central lobe), since the sum of the centralities of all
nodes is constant (due to the fact that the modulus of the eigenvector is one). In other words,
when comparing both groups, it would not be possible to observe higher centrality for all nodes
of a given group (High CR or Low CR).
Figure 3. Differences between high and low CR groups at the node level. Black dots indicate the
Euclidean position of the 148 magnetometers (nodes). Circles filled with lilac colour show nodes
with significant statistical differences in: Eigenvector centrality ∆ev(i) (A), within-module degree
z-score ∆z(i) (B) and participation coefficient ∆p(i) (C). Green borders correspond to those nodes
that have greater values in the high CR group, and red borders represent those nodes that exhibit
greater values in the low CR group. Circle sizes are proportional to the absolute value of the
differences between groups. Here we present the results obtained with the permutation test.
Results obtained with Mann-Whitney U-test are similar (see Fig. S3 of Supp. Info. for details).
Macro-scale: Analysing the topology of the whole network
We analysed how the average network parameters were modified according to the level of CR of
the participants. First, we calculated the mean and standard deviation of six network parameters
for each subject (network strength 𝑆, outreach 𝑂, weighted clustering coefficient 𝐶𝑤, average
neighbour strength 𝑆𝑛𝑛, global efficiency 𝐸𝑔 and average shortest path 𝑑). Second, we carried out
a permutation test and a nonparametric Mann-Whitney U-test (see Section Statistical analysis for
details) for each network metric. Finally, those metrics with p-values<0.05 were accepted to have
statistically significant differences.
In our results, the low CR group exhibited higher averages values than the high CR group in all
parameters calculated, except for 𝑑 (see Fig. 4 for details). However, only the network outreach
𝑂 showed statistical significant differences both in the permutation (p= 0.0232) and the rank-sum
(p=0.039) tests. Interestingly, the low CR group showed a higher 𝑂 (𝑂 low= 13.57 and 𝑂 high=12.29).
It is also remarkable that the average shortest path length 𝑑 exhibit significant differences
(p=0.0491) at the rank sum test, despite not passing the permutation test. In this case the average
values of shortest path in low CR are lower than in the high CR group (𝑑low= 12.67 and
𝑑high=14.13) (see also Tab. S1 of the Supp. Info. for the values the averages).
Low CR
High CR
Low CR
High CR
Low CR
High CR
Figure 4. Violin plots showing the distribution of subject's features and means for the six network
parameters analysed: Network strength S, outreach O, weighted clustering coefficient Cw, average
neighbour strength Snn, global efficiency Eg and average shortest path d. Red and blue plots
correspond to the low and high CR groups, respectively. Red crosses indicate the mean value.
Outreach O (B) showed significant differences between groups both with the permutation test as
well as with A Mann-Whitney U-test (indicated by two stars). Average shortest path d (F) showed
statistical significant differences between groups only with the latter test (one star). Both
statistical analyses were performed with pval < 0.05.
Dynamical Analysis of MEG Time Series
We found a group of nodes showing statistical differences between the high and low CR groups,
both for H(i) and C(i) (𝒑𝒗𝒂𝒍 < 0.05). In accordance with previous sections, the occipital lobe is
the brain region that concentrates the highest amount of nodes with statistical differences (see
Figures 5 and 6). We obtained a group of 23 nodes whose entropy was higher in the low CR group
(except for one node) (see Figure 5). On the contrary, we obtained lower complexity of the signals
in the low CR group (except for two nodes), but in a very similar brain localization that in the
case of entropy (only two nodes did not overlap) (see Figure 6). These results indicate that a
higher level of CR could be associated to a brain dynamics with lower entropy and higher
complexity.
Figure 5. Differences of entropy H(i) between low and high CR groups. In A, we show the mean
and standard deviations of the entropy of nodes that have statistical significant differences
between groups. H(i)lowCR (red squares) is higher than H(i)highCR (blue circles) for nearly all
nodes. In B, we plot the position of nodes with statistical differences. Node sizes are proportional
to ∆H(i) = H(i)highCR−H(i)lowCR. Red borders represent ∆H(i) < 0, otherwise node borders
are green. In this case, we used a permutation test but no relevant changes appear when a rank-
sum test is carried out (see Supp. Info.).
Figure 6. Differences of complexity C(i) between the Low and High CR groups. In A, we show
the mean and standard deviations of the entropy of nodes that have statistical significant
differences between groups. C(i)lowCR (red squares) is lower than C(i)highCR (blue circles) for
nearly all nodes. In B, we plot the position of nodes with statistical differences. Node sizes are
proportional to ∆C(i) = C(i)highCR−C(i)lowCR. Green borders represent ∆C(i) > 0, otherwise
node borders are red. See Supp. Info. for similar results obtained with the rank-sum test.
Correlation between Entropy and Complexity
Then, we explored the relationship between these two dynamical properties. Figure 7 shows a
complexity-entropy scatter plot for all nodes, with the inset plotting only those nodes with
statistical differences. In the figure, each point corresponds to the average of each node over the
subjects of a given group. We can observe a significant negative correlation between H(i) and
C(i) for low (𝒓 = 0.9874; 𝒑𝒗𝒂𝒍 = 0.0010) and high CR groups (𝒓 = 0.8288; 𝒑𝒗𝒂𝒍r = 0.8288; pval
= 0.0010), revealing that those nodes with less entropy are, at the same time, those nodes with
higher complexity.
Figure 7. Complexity-Entropy Diagram. Diagram of subjects' average complexity vs. entropy
for all 148 nodes (i.e., each point represents a node). Red squares correspond to low CR group
and blue circles to the high CR group. The inset shows only nodes with statistical significant
differences using the permutation test. Specifically, these nodes are the channels that have
statistical differences when comparing both the mean entropies and complexities of high vs. low
CR groups at node level. See Supp. Info. for similar results obtained with the rank-sum test.
Correlations between Topology and Dynamics
We explored the relationship between the dynamic properties of the nodes (i.e., 𝐻(𝑖), 𝐶(𝑖)) and
two of their topological parameters, namely the node strength 𝑆(𝑖) and the weighted clustering
coefficient 𝐶𝑤(𝑖) for each CR group. Figure 8 shows all possible combinations.
For the low CR group, we found a positive correlation between 𝐻(𝑖) and 𝑆(𝑖) (𝒓𝟐 =
0.5833; 𝒑𝒗𝒂𝒍 < 0.001), which indicates that nodes with higher strength, i.e. the hubs of the
network, are in turn those nodes with higher entropy at their dynamics (see Figure 8A). On the
contrary, the correlation between the node complexity and its strength was negative (𝒓𝟐 =
0.7124; 𝒑𝒗𝒂𝒍 < 0.001) (see Figure 8B). The same pattern was found between 𝐻(𝑖) and 𝐶𝑤(𝑖)
(positive correlation) (𝒓𝟐 = 0.5713; 𝒑𝒗𝒂𝒍 < 0.001), and between 𝐶(𝑖) and 𝐶𝑤(𝑖) (negative
correlation) (𝒓𝟐 = 0.7055; 𝒑𝒗𝒂𝒍 < 0.001) (see Figure 8C-D).
Correlations found for the high CR group are qualitatively similar to the low CR but the values
0.740.760.780.80.820.840.860.880.90.920.940.10.150.20.250.30.35Low CRHigh CR0.850.90.950.150.20.250.3 of the correlation parameter are systematically lower (See Tab. S2 of Supp. Info. for details).
Figure 8. Correlation between the topological and dynamical metrics of the nodes. Correlation
plots for the global dynamical properties (entropy H(i) and complexity C(i)) of the MEG time
series and their corresponding topological features (strength S(i) and clustering C(i)).
Specifically: A) H(i) vs. S(i); B) C(i) vs. S(i); C) H(i) vs. Cw(i) and D) C(i) vs. Cw(i). Red
squares correspond to low CR group and blue circles to the high CR group. All statistical tests
rejected the null hypothesis when using permutation test as well as with rank-sum test.
DISCUSSION
In the present study, we explored if CR might play a role in both the topology and the dynamics
of the functional brain networks in healthy elders while performing a memory task. To this end,
we classified the participants according to a CR index that combined educational level and
occupational attainment into two groups: high or low CR level.
We firstly focused on the topological characteristics of the two groups, finding differences at
microscale (i.e. nodes) and macroscale (i.e. average over all nodes) levels. At the microscale level
(i.e., when nodes are analyzed), we observed that the most influential nodes for both groups were
mainly placed over the occipital region, although there was a slight displacement towards the
central lobe in the group with high CR. Specifically, we found that the low CR group exhibited
higher eigenvector centrality over both temporal and occipital areas, while the high CR group
presented higher values in central areas. These observations indicate that the importance of
regions placed at the occipital and temporal regions is higher at the low CR group. At the same
time, since the sum of the eigenvector centrality of all nodes of the network is constant, other
nodes unavoidably have lower centrality values to compensate the higher values reported at the
occipital and temporal regions. These nodes are placed at the central lobe, where the high CR
group has higher centrality when compared with the low CR one. Interestingly, the within-module
degree behaves in the opposite way. Comparing Fig. 3 A and B, we can observe that, in general,
regions that have higher eigenvector centrality in the high CR group reduce their within-module
degree. The explanation of this, in principle, counter-intuitive observation is that nodes have
higher eigenvector centrality as a consequence of having higher weights at their long-range
connections (i.e., in weight of the functional connections with nodes laying outside of their
module).
Besides, high CR group showed higher participation coefficient values than the low CR group,
especially in regions located at the frontal lobe. These results pointed out that the importance of
a node and its connections with their neighbours inside and outside their community (lobe, in our
case) may be differently located during the execution of a cognitive task, according to the level
of CR.
At macroscale level, the low CR group exhibited higher values in different networks parameters
compared to the high CR group (network strength, weighted clustering coefficient, average
neighbour strength, global efficiency), being this increment statistically significant only in the
outreach parameter both for the permutation and the rank-sum tests. Although strength results did
not reach statistical significance, the low CR group presented a higher average in this network
parameter suggesting that these participants may require higher brain synchronization to perform
the memory task as well as the participants with high CR. As a result of this rise of functional
connections, the average shortest path showed lower values, however, only the rank-sum test
leaded to statistically significant differences between low and high CR groups. Along with the
average shortest path, the outreach was the parameter more altered in the low CR group (in this
case passing the two statistical tests). The higher outreach observed in the low CR group indicated
the existence of a higher amount of long-range connections, implying a greater energetic cost to
accomplish the task. Interestingly, this pattern was also described in a study carried out by our
group 23, in which MCI and healthy subjects were compared performing the same Sternberg's
letter-probe task. Network's results showed that MCI subjects exhibited an increment of strength
and outreach values, and a decrement in the average shortest path, as formerly mentioned. Taken
together all these findings, and especially the fact that the network outreach (and as a
consequence, the energy consumption) is higher in the low CR group we may consider that brain
functioning and network organization of those subjects with low CR have some similarities with
MCI subjects during the execution of a memory task, and therefore we suggest that CR should be
investigated as a possible factor with a remarkable impact on the cognitive status within the
healthy aging. However, it is worth mentioning that, in our case, none of the normalized
parameters showed significant statistical differences between low and high CR groups, while in
MCI subjects showed differences in the normalized parameters indicating that their topology was
more random than networks of the control group 23. In our case, since the normalization is made
by equivalent random networks, the lack of statistical differences between groups indicates that
we can not claim that functional networks of the low CR group have a higher/lower random
organization.
It is worth mentioning how the synchronized activity along the whole functional networks was
negatively correlated with CR during the cognitive task 11. In this sense, individuals with higher
CR required less synchronization between brain regions to successfully carry out the memory
test. On the contrary, fMRI recordings obtained during resting state reveal negative correlations,
i.e., individuals with higher CR maintain higher synchronization in resting state 15,16. This
discrepancy suggests different mechanisms behind the creation and organization of the functional
networks during resting state and cognitive tasks. Future studies should investigate this
phenomenon with resting and task data with MEG.
The interpretation of the brain network profile described in the low CR group may be explained
as the result of a neural compensatory effect 4. In this case, as both groups did not differ in their
memory task accuracy, we may consider it as a successful neural compensation since the low CR
group had to engage additional resources to maintain or improve their performance 12. This effect
may be related to the Compensation-Related Utilization of Neural Circuits Hypothesis
(CRUNCH) 35, which proposes that there is a loss of efficiency in brain networks of aging, when
compared with young adults, since additional brain areas are needed to perform cognitive tasks.
In the present study, we used a low demanding memory task, but if we would had modified its
difficulty, we would had probably found an unsuccessful compensation attempt in the low CR
group, as previously described in MCI subjects 36. The cause of this compensatory effect may be
the beginning of a network malfunctioning since individuals with low CR exhibited a lower
energetic efficiency to perform the memory task, as previously described in MCI subjects 23. In
this line, these individuals may exhibit clinical symptoms of pathological aging (i.e. MCI) earlier
or more severe than those subjects with a higher CR.
Another noticeable novelty explored in the present work was the dynamical properties of the
cortical brain regions while performing the memory task. In line with the microscale results, the
main differences between both groups were located over the parieto-occipital lobe. In fact, a lower
CR level was related with lower complexity and higher entropy. This means that the brain signals
of the low CR group presented a pattern more disorganized than the high CR group within the
occipital lobe. In addition, the low CR group exhibited the opposite pattern in the right temporal
lobe (i.e. less entropy and more complexity). We then may hypothesize that the recruitment of
posterior brain areas was not providing an efficient contribution to adequately perform the task,
and therefore the individuals with low CR needed to involve right temporal regions to achieved
it. We then verified that these two dynamical properties were inversely correlated, independently
of the CR level. Finally, one of the most original approaches performed in this study was the
exploration of the relationship between the topology (i.e. node strength and weighted clustering
coefficient) and the dynamic characteristics of the nodes (i.e. entropy and complexity). We found
that in both CR groups, greater values in both topological measures were related with a higher
entropy and a lower complexity. These results demonstrated that the most important nodes or
hubs within the functional brain networks were those that exhibited more random dynamics, no
matter what the level of CR of the individuals.
The results obtained in the current study are limited due to the lack of follow-up of the participants
and the small sample size, especially in the case of the high CR group. Notwithstanding, it still
demonstrated that CR plays a relevant role in the topological and dynamical network's properties
in healthy aging. It should be pointed that one of the advantages of conducting the analysis in the
sensor space was that MEG signal was not manipulated, affecting neither the complexity nor the
entropy. In addition, these results corroborated that network analysis may be a suitable approach
for studying CR, and that MEG may be an appropriate tool to detect functional brain changes in
healthy aging. Thus, the profiles described in this work for the low CR group indicate a dual
pattern of compensation and network impairment. As a consequence, if one of these subjects
suffers brain damage or starts a neurodegenerative disease process, their closeness to random
network organization would probably be a risk factor for developing a more severe cognitive
impairment, although some compensation mechanisms are still present. Thus, this study
highlights the importance of preventive interventions, including cognitive training, which could
induce a more protective brain network organization in the case of brain damage in elderly
subjects.
ACKNOWLEDGMENTS
This study was supported by the Spanish Ministry of Economy and Competitiveness under Project
FIS2013-41057-P, a postdoctoral fellowship from the Complutense University of Madrid to
María Eugenia López and a predoctoral fellowship to David López Sanz (BES-2013-063772),
both from the Spanish Ministry of Economy and Competitiveness.
AUTHOR CONTRIBUTIONS
JHM contributed writing paper, in-silico experimental setup, wrote code for results, analysed the
data; MEL with MEG recordings and writing paper; PA analysed the data; MC contributed with
results and statistical analysis; JAP-P worked on the statistical analysis; DL-S contributed writing
the paper; PG worked on the clinical assessment of the participants, FM helped in the design the
paradigm and JMB conceived the analysis, analyzed the data and helped writing the manuscript.
ADDITIONAL INFORMATION
Competing financial interests: The authors declare no competing financial interests.
REFERENCES
1.
Stern, Y. Cognitive reserve in ageing and Alzheimer's disease. Lancet Neurol. 11, 1006–
12 (2012).
2.
Katzman, R. Education and the prevalence of dementia and Alzheimer's disease.
Neurology 43, 13–20 (1993).
3.
Satz, P. Brain reserve capacity on symptom onset after brain injury: A formulation and
review of evidence for threshold theory. Neuropsychology (1993). doi:10.1037/0894-
4105.7.3.273
4.
5.
Stern, Y. Cognitive reserve. Neuropsychologia 47, 2015–2028 (2009).
Valenzuela, M. J. & Sachdev, P. Brain reserve and dementia: A systematic review.
Psychol. Med. 36, 441–454 (2006).
6.
Bartrés-Faz, D. & Arenaza-Urquijo, E. M. Structural and functional imaging correlates of
cognitive and brain reserve hypotheses in healthy and pathological aging. Brain Topogr.
24, 340–57 (2011).
7.
Serra, L. et al. Neuroanatomical correlates of cognitive reserve in Alzheimer disease.
Rejuvenation Res. 14, 143–51 (2011).
8.
Perneczky, R. et al. Schooling mediates brain reserve in Alzheimer's disease: findings of
fluoro-deoxy-glucose-positron emission tomography. J. Neurol. Neurosurg. Psychiatry
77, 1060–1063 (2006).
9.
Serra, L. et al. Cognitive reserve and the risk for Alzheimer's disease: a longitudinal study.
Neurobiol. Aging 36, 592–600 (2015).
10.
López, M. E. et al. Searching for Primary Predictors of Conversion from Mild Cognitive
Impairment to Alzheimer's Disease: A Multivariate Follow-Up Study. J. Alzheimers. Dis.
52, 133–143 (2016).
11.
Solé-Padullés, C. et al. Brain structure and function related to cognitive reserve variables
in normal aging, mild cognitive impairment and Alzheimer's disease. Neurobiol. Aging
30, 1114–24 (2009).
12.
López, M. E. et al. Cognitive reserve is associated with the functional organization of the
brain in healthy aging: A MEG study. Front. Aging Neurosci. 6, 1–9 (2014).
13.
Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of
structural and functional systems. Nat. Rev. Neurosci. 10, 186–98 (2009).
14. Wook Yoo, S. et al. A Network Flow-based Analysis of Cognitive Reserve in Normal
Ageing and Alzheimer's Disease. Sci. Rep. 5, 10057 (2015).
15. Marques, P. et al. The functional connectome of cognitive reserve. Hum. Brain Mapp. 37,
3310–3322 (2016).
16.
van den Heuvel, M. P., Mandl, R. C. W., Kahn, R. S. & Hulshoff Pol, H. E. Functionally
linked resting-state networks reflect the underlying structural connectivity architecture of
the human brain. Hum. Brain Mapp. 30, 3127–41 (2009).
17.
deToledo-Morrell, L. et al. A 'stress' test for memory dysfunction. Electrophysiologic
manifestations of early Alzheimer's disease. Arch. Neurol. 48, 605–9 (1991).
18.
Latora, V. & Marchiori, M. Efficient behavior of small-world networks. Phys. Rev. Lett.
87, 198701 (2001).
19. Newman, M. E. J. (Mark E. J. . Networks : an introduction. (Oxford University Press,
2010).
20. Ahnert, S. E., Garlaschelli, D., Fink, T. M. A. & Caldarelli, G. Ensemble approach to the
analysis of weighted networks. Phys. Rev. E 76, 16101 (2007).
21. Navas, A. et al. Functional Hubs in Mild Cognitive Impairment. Int. J. Bifurc. Chaos 25,
1550034 (2015).
22. Watts, D. J. & Strogatz, S. H. Collective dynamics of 'small-world' networks. Nature 393,
440–2 (1998).
23.
Buldú, J. M. et al. Reorganization of functional networks in mild cognitive impairment.
PLoS One 6, e19584 (2011).
24. Oldfield, R. C. The assessment and analysis of handedness: the Edinburgh inventory.
Neuropsychologia 9, 97–113 (1971).
25.
Lobo, A., Ezquerra, J., Gómez Burgada, F., Sala, J. M. & Seva Díaz, A. [Cognocitive
mini-test (a simple practical test to detect intellectual changes in medical patients)]. Actas
Luso. Esp. Neurol. Psiquiatr. Cienc. Afines 7, 189–202 (1979).
26. Wechsler, D. Wechsler Memory Scale - Revised. (The Psychological Corporation, 1987).
27.
Pfeffer, R. I., Kurosaki, T. T., Harrah, C. H., Chance, J. M. & Filos, S. Measurement of
functional activities in older adults in the community. J. Gerontol. 37, 323–9 (1982).
28.
Stam, C. J. & van Dijk, B. W. Synchronization likelihood: an unbiased measure of
generalized synchronization in multivariate data sets. Phys. D Nonlinear Phenom. 163,
236–251 (2002).
29.
Stam, C. J. et al. Magnetoencephalographic evaluation of resting-state functional
connectivity in Alzheimer's disease. Neuroimage 32, 1335–44 (2006).
30. Guimerà, R. & Amaral, L. A. N. Cartography of complex networks: modules and universal
roles. J. Stat. Mech. Theory Exp. 2005, P02001 (2005).
31.
Bandt, C. & Pompe, B. Permutation entropy: a natural complexity measure for time series.
Phys. Rev. Lett. 88, 174102 (2002).
32.
Rosso, O. A., Larrondo, H. A., Martin, M. T., Plastino, A. & Fuentes, M. A. Distinguishing
noise from chaos. Phys. Rev. Lett. 99, 154102 (2007).
33.
López-Ruiz, R., Mancini, H. L. & Calbet, X. A statistical measure of complexity. Phys.
Lett. A 209, 321–326 (1995).
34.
Zanin, M., Zunino, L., Rosso, O. A. & Papo, D. Permutation Entropy and Its Main
Biomedical and Econophysics Applications: A Review. Entropy 14, 1553–1577 (2012).
35.
Reuter-Lorenz, P. a. & Cappell, K. a. Neurocognitive aging and the compensation
hypothesis. Curr. Dir. Psychol. Sci. 17, 177–182 (2008).
36.
López, M. E. et al. Synchronization during an internally directed cognitive state\nin
healthy aging and mild cognitive impairment: a MEG study. Age (Omaha). 36, 9643
(2014).
SUPLEMENTARY INFORMATION
S1.- DEFINITION OF NETWORK METRICS
The strength s(i) of a node i is the sum of the weights 𝑤𝑖𝑗 of all its links:
𝒔(𝒊) = ∑ 𝒘𝒊𝒋
𝒋∈𝑵
The strength S of a functional network is the average of the strength of all its nodes.
The strength of the nearest neighbours snn(i) as the average strength of all neighbours of node i.
The outreach o(i) of a node i is the sum of the links' weights 𝑤𝑖𝑗 multiplied by the link
Euclidean lengths 𝑙𝑖𝑗 between node i and node j:
𝒐(𝒊) = ∑ 𝒍𝒊𝒋𝒘𝒊𝒋
𝒋∈𝑵
The outreach O of a functional network is the average of the outreach of all its nodes.
The weighted clustering coefficient cw(i) of a node i quantifies the percentage of neighbours of a
certain node that, in turn, are neighbours between them, taking into account the weight of the
connections:
𝒄𝒘(𝒊) =
∑ 𝒘𝒊𝒋𝒘𝒋𝒌𝒘𝒊𝒌
𝒋𝒌
∑ 𝒘𝒊𝒋𝒘𝒊𝒌
𝒋𝒌
The weighted clustering coefficient Cw of a functional network is the average of the weighted
clustering coefficient of all its nodes.
The eigenvector centrality ev(i) of a node i measures the importance of a node according to the
importance of its neighbours. It is obtained from the eigenvector associated with the largest
eigenvalue of the connectivity matrix.
The within-module degree z(i) of a node i measures the importance of a node inside its
community. In our case, the community of a node is the lobe the node belongs to:
𝒛(𝒊) =
𝒌𝒊(𝒎𝒊) − 〈𝒌𝒊(𝒎𝒊)〉
𝝈𝒌(𝒎𝒊)
where 𝒌𝒊(𝒎𝒊) is the degree of node i inside its community (lobe), and 〈𝒌𝒊(𝒎𝒊)〉 and 𝝈𝒌(𝒎𝒊) are
the average and the standard deviation of the degree inside the community, respectively.
The participation coefficient p(i) of node i quantifies the percentage of links of a node that reach
other communities:
𝒑(𝒊) = 𝟏 − ∑(
)𝟐
𝒌𝒊(𝒎)
𝒌𝒊
𝒎
where 𝒌𝒊(𝒎) is the degree of node i inside community m.
The average shortest path d of node i is the average of the minimum number of nodes to be visited
when going from node i to j. To obtain d, we weight the distances between nodes Dij, as the inverse
of the elements of the connectivity matrix 𝒘𝒊𝒋, i.e. Dij = 1/𝒘𝒊𝒋. Next, we calculate the shortest-
path distance between every pair of nodes using the Dijkstra's algorithm (Dijkstra, 1959). This
way, we obtain the shortest-path matrix 𝒅𝒊𝒔𝒊𝒋 (Newman, 2010), and finally the average shortest
path d is obtained as the average of the distance of each node to the rest of the network:
𝒅 =
𝟏
𝑵(𝑵 − 𝟏)
∑ 𝑑𝑖𝑠𝑖𝑗
𝑖≠𝑗
The global Efficiency Eg of a network, first introduced by Latora & Marchiori (2001), overcomes
the fact that certain nodes of a network could be isolated from the others, thus leading to infinite
distance between them. Mathematically, Eg is defined as the harmonic mean of the inverse of the
shortest paths between all nodes of the network, with 𝑑𝑖𝑠𝑖𝑗, being the shortest path between nodes
i and j:
𝑬𝒈 =
𝟏
𝑵(𝑵 − 𝟏)
1
𝑑𝑖𝑠𝑖𝑗
∑
𝑖≠𝑗
S2.- ENTROPY AND COMPLEXITY
In the matter of dynamical complexity, information plays an important role as a feature that
describes the outermost bounds of periodicity, chaos and complexity. In this sense, the Bandt &
Pompe (2002) BP method obtains the intrinsic temporal symbol sequences {St} from the
neighbouring steps of a time series (see Fig. S1 for a qualitative explanation). This symbol
sequences depend on an embedding dimension D = 3, 4, 5, …, which represents the amount of
past information, being D the number of neighbouring samples. In this way, D characterizes each
{Xt} time series along t = 1, 2, …, M samples. To do that, {Xt} is partitioned into (M - D)
overlapping vectors of dimension D.
Figure S1: Overlapping vectors for the case D = 3. D! is the number of patterns. This way 3! different types
𝜋 of accessible states are presented. The probability of appearance of the ordinal patterns is contained in P.
If an ordinal pattern never appears, it is called a forbidden pattern.
The greater the dimension D, the more information about the past state of our system, and the
longer the vectors are (containing the ordering of a set of D samples). Each of the vectors is
assigned to a time t, sliding the vector at every time step, to get a total of (M – D + 1) overlapping
vectors. Hence, for each (M - D) vector, the position of the lowest value will be assigned the
ordinal value zero. The position of the highest value will correspond to (D - 1) ordinal value (the
highest in the ranking). Thus, the following positions in-between the assigned zero and (D - 1)
will be assigned by rating the positions of the remaining samples in the respective ordinal values.
When all (M – D + 1) different order types in {St} are calculated, it is possible to obtain the
probability distribution function (PDF) P(𝜋), quantifying the probability of finding a certain order
pattern associated to {Xt}:
𝑷(𝜋) =
#{𝒕𝒕 ≤ 𝑴 − 𝑫, (𝒙𝒕+𝟏, … , 𝒙𝒕+𝑫)𝒉𝒂𝒔 𝒂 𝒕𝒚𝒑𝒆 𝜋}
𝑴 − 𝑫 + 𝟏
In previous equation, 𝜋 is a possible ordinal pattern presented in the sequence {St} and # is its
number of appearances. Note that each ordinal pattern is a permutation of 𝜋 = (0,1,2, …, D - 1).
In other words, D! represents all possible permutations 𝜋 of order D of the number of accessible
states (M - D). As an example, consider the case of D = 3 in Fig. 1.8. The number of patterns or
accessible states will be D! = 3! = 6 and the possible patterns 𝜋 will be: {(012), (021), (102),
(201), (120), (210)}. From the former ones, vectors that appear in {St} are called ordinal patterns
of {Xt}, those that do not appear in {St}, but belong to the possible accessible states are called
forbidden patterns. Finally, the discrete PDF of the ordinal patterns P =𝑝𝑗, ∀ j = 1,2, …, N ∧ N
= D! is calculated. This PDF, obtained from BP method, carries the temporal information of {Xt}
by comparing consecutive samples. In other words, this symbolic technique incorporates the
causality effects of a short-memory (of D steps) a time series has.
Next, we use the PDF of the ordinal patterns to define the Normalized Permutation Entropy H
(Band & Pompe, 2002):
Definition S1 - Normalized Permutation Entropy H[P]. It is given by the ratio between the
entropy S[P] of the ordinal patterns and Smax = S[Pe], being Pe the uniform probability
distribution:
H[P] = S[P]/Smax
Note that, the normalized permutation entropy H[P] is bounded between [0, 1].
Regarding the finite size effects, the normalized permutation entropy H allows to include a
uniform distribution Pe = {1/N, 1/N, …, 1/N} making H to be an intensive property. This uniform
distribution Pe, also maximizes the associate-system information entropy S[P], i.e., Smax
=log(N)= log(D!). This way, the amount of disorder H[P] based on the information measure S[P]
associated to P is defined as the Permutation Entropy because it runs over all D! permutations 𝜋
of order D.
On the other hand, the insertion of "a priori" equilibrium distribution {Pe} as a correction for the
associated entropy, leads to a discrimination between two populations. In other words, we need
to evaluate the distance between both distributions P and Pe. This fact makes S[P] not being
enough to effectively characterize {Xt} because there could be some ordinal patterns that belong
to P as well as to Pe. This distance accounts for the "order" of the system when one of the few
ordinal patterns emerges as the preferred one. The disequilibrium between statistical populations
will be the measure to distinguish this non-Euclidean distance. We quantify the disequilibrium Q
by adopting some statistical distance D between the possible and accessible states of the systems
in P and the equilibrium distribution Pe:
Definition S2 - Disequilibrium Q[P]. It evaluates the distance between P and Pe as:
Q[P] = Q0 D [P, Pe]
where Q0 is a normalization constant leading to 0 ≤ Q[P] ≤ 1. A Q ≠ 0 indicates the existence
of preferred states among the accessible ones.
In this way, the disequilibrium Q[P], discriminates ordinal patterns in P from the uniform
distribution Pe. The zero limit or the minimum disequilibrium, implies that the lowest separation
of both populations does not distinguish between ordinal patterns coming from both populations.
Meanwhile the upper limit, with a high disequilibrium, is related to the fact of the existence of
some privileged ordinal patterns in P.
Hitherto, both H and Q give some sense of the understanding of what the dynamical properties of
the system are. Nevertheless, we are concerned on evaluating the interplay between the order and
disorder of a system. Therefore, it is also desirable to complement these measures with some
metric quantifying the complexity of the system.
In this way, Bandt and Pompe define the statistical complexity of a system as:
Definition S3 - Statistical Complexity C[P]. This complexity mesures is defined as the product
between the permutation entropy H and the disequilibrium Q:
C[P] = H[P] *Q[P]
With this definition, the statistical complexity C accomplishes the first requirement since H and
Q are intensive quantities. The second requirement is also achieved since, by means of H and Q,
we are measuring the disorder of a system and its distance from the equilibrium. Note that the
statistical complexity vanishes either if the system is at equilibrium (maximum disorder) or if it
is completely ordered (maximal distance from the equilibrium). Figure 1.9 shows a qualitative
plot indicating the interplay between the three measures.
Figure S2: Statistical measures based on ordinal patterns. Schematic representation of the interplay
between the normalized permutation entropy H[P], disequilibrium Q[P] and the statistical complexity C[P],
in terms of a system that range from complete order to complete disorder.
The previous definitions of H, Q and C are usually known as Generalized Statistical Complexity
Measures (SCM). SCM capture either, the essential details of the dynamics that allow discerning
among different degrees of periodicity and randomness, as well as all possible degrees of
stochasticity when the information of {Xt} is extracted via the BP method. SCM, not only compute
randomness, but a wide range of correlation structures, not already offered by a simple entropy
analysis.
S3.- COMPLEMENTARY RESULTS
Figure S3. Differences between high and low CR groups at the node level. Black dots indicate the
Euclidean position of the 148 magnetometers (nodes). Circles filled with lilac colour show nodes with
significant statistical differences obtained with Mann-Whitney U-test in the: Eigenvector centrality ∆ev(i)
(A), within-module degree z-score ∆z(i) (B) and participation coefficient ∆p(i) (C). Green borders
correspond to those nodes that have greater values in the high CR group, and red borders represent those
nodes that exhibit greater values in the low CR group. Circle sizes are proportional to the absolute value of
the differences between groups. Here we present the results obtained with the permutation test.
Group
𝑆
𝑂 (**)
𝐶𝑤
𝑆𝑛𝑛
𝐸𝑔
𝑑 (*)
Low
13.78
13.57
0.102
13.96
0.096
12.69
High
12.75
12.29
0.093
12.76
0.089
14.12
Table S1. Average network parameters of the low and high CR groups. Specifically, the network strength
𝑆, the outreach 𝑂, the weighted clustering coefficient 𝐶𝑤, the average neighbour strength 𝑆𝑛𝑛, the global
efficiency 𝐸𝑔 and the average shortest path 𝑑. Asterisks indicate those metrics with statistically significant
differences between groups: one asterisk for the parameters passing the rank-sum test and two asterisks for
those parameters that also passed the permutation test.
Figure S4. Differences of entropy H(i) between low and high CR groups. In A, we show the mean and
standard deviations of the entropy of nodes that have statistical significant differences between groups.
H(i)lowCR (red squares) is higher than H(i)highCR (blue circles) for nearly all nodes. In B, we plot the
∆H(i) =
position of nodes with statistical differences. Node sizes are proportional
H(i)highCR−H(i)lowCR. Red borders represent ∆H(i) < 0, otherwise node borders are green. In this case,
significant statistical differences were obtained with Mann-Whitney U-test.
to
Figure S5. Differences of complexity C(i) between the Low and High CR groups. In A, we show the mean
and standard deviations of the entropy of nodes that have statistical significant differences between groups.
C(i)lowCR (red squares) is lower than C(i)highCR (blue circles) for nearly all nodes. In B, we plot the position
of nodes with statistical differences. Node sizes are proportional to ∆C(i) = C(i)highCR−C(i)lowCR. Green
borders represent ∆C(i) > 0, otherwise node borders are red. Statistical significant differences were
obtained with Mann-Whitney U-test.
Figure S6. Complexity-Entropy Diagram. Diagram of subjects' average complexity vs. entropy for all 148
nodes (i.e., each point represents a node). Red squares correspond to low CR group and blue circles to the
high CR group. The inset shows only nodes with statistical significant differences using the permutation
test. Specifically, these nodes are the channels that have statistical differences when comparing both the
mean entropies and complexities of high vs. low CR groups at node level. Statistical significant differences
were obtained with Mann-Whitney U-test.
𝑟2
(𝑆, 𝐻)𝐿𝑜𝑤 (𝑆, 𝐻)𝐻𝑖𝑔ℎ (𝑆, 𝐶)𝐿𝑜𝑤 (𝑆, 𝐶)𝐻𝑖𝑔ℎ (𝐶𝑤, 𝐻)𝐿𝑜𝑤 (𝐶𝑤, 𝐻)𝐻𝑖𝑔ℎ (𝐶𝑤, 𝐶)𝐿𝑜𝑤 (𝐶𝑤, 𝐶)𝐻𝑖𝑔ℎ
Lin. 0.5833
0.4105
0.6486
0.4902
0.5719
0.3423
0.6417
0.4246
Pol. 0.5833
0.4075
0.7124
0.5159
0.5713
0.3332
0.7055
0.4733
Table S2. Coefficient 𝑟2 for the correlations of Fig. 8 of the main text. First (second) row represent the
linear (second order polynomic) fit. Columns represent the 𝑟2 of both averaged variables (structural and
dynamical) for the low and high CR groups. Columns 1 and 2 are for Fig. 8A, columns 3 and 4 are for Fig.
8B, columns 5 and 6 are for Fig. 8C. The last two columns are for Fig. 8D. Permutation test and rank-sum
test rejected Ho with p<0.001.
REFERENCES
0.740.760.780.80.820.840.860.880.90.920.940.10.150.20.250.30.35Mann-Whitney U-testLow CRHigh CR0.850.90.950.150.20.250.3 Bandt, C. & Pompe, B. Permutation Entropy: A Natural Complexity Measure for Time Series.
Phys. Rev. Lett. 88, 174102 (2002).
Dijkstra, E.W. A note on two problems in connexion with graphs. Numerische Mathematik 1,
269–271 (1959).
Latora, V. & Marchiori, M. Efficient behavior of small-world networks. Physical Review Letters
87, 198701 (2001).
Newman, M. E. J. Networks: An introduction. (Oxford University Press, 2010).
Garibotto,V., Borroni,B., Sorbi,S., Cappa,S.F., Padovani,A., and Perani,D.(2011). Education and
occupation provide reserve in both ApoE ε4 carrier and non carrier patients with probable
Alzheimer's disease. Neurol.Sci. 33, 1037–1042. doi:10.1007/s10072-011-0889-5
|
1209.0426 | 2 | 1209 | 2012-11-21T14:30:49 | A biophysical observation model for field potentials of networks of leaky integrate-and-fire neurons | [
"q-bio.NC"
] | We present a biophysical approach for the coupling of neural network activity as resulting from proper dipole currents of cortical pyramidal neurons to the electric field in extracellular fluid. Starting from a reduced threecompartment model of a single pyramidal neuron, we derive an observation model for dendritic dipole currents in extracellular space and thereby for the dendritic field potential that contributes to the local field potential of a neural population. This work aligns and satisfies the widespread dipole assumption that is motivated by the "open-field" configuration of the dendritic field potential around cortical pyramidal cells. Our reduced three-compartment scheme allows to derive networks of leaky integrate-and-fire models, which facilitates comparison with existing neural network and observation models. In particular, by means of numerical simulations we compare our approach with an ad hoc model by Mazzoni et al. [Mazzoni, A., S. Panzeri, N. K. Logothetis, and N. Brunel (2008). Encoding of naturalistic stimuli by local field potential spectra in networks of excitatory and inhibitory neurons. PLoS Computational Biology 4 (12), e1000239], and conclude that our biophysically motivated approach yields substantial improvement. | q-bio.NC | q-bio |
A biophysical observation model for field
potentials of networks of leaky
integrate-and-fire neurons
Peter beim Graben∗
Bernstein Center for Computational Neuroscience Berlin, Germany
Department of German Language and Linguistics, Humboldt-Universitat zu Berlin
Serafim Rodrigues
Centre for Robotics and Neural Systems, School of Computing and Mathematics,
University of Plymouth, United Kingdom
Abstract
We present a biophysical approach for the coupling of neural network ac-
tivity as resulting from proper dipole currents of cortical pyramidal neurons
to the electric field in extracellular fluid. Starting from a reduced three-
compartment model of a single pyramidal neuron, we derive an observation
model for dendritic dipole currents in extracellular space and thereby for the
dendritic field potential that contributes to the local field potential of a neu-
ral population. This work aligns and satisfies the widespread dipole assump-
tion that is motivated by the "open-field" configuration of the dendritic field
potential around cortical pyramidal cells. Our reduced three-compartment
scheme allows to derive networks of leaky integrate-and-fire models, which
facilitates comparison with existing neural network and observation models.
In particular, by means of numerical simulations we compare our approach
∗Department of German Language and Linguistics
Humboldt-Universitat zu Berlin
Unter den Linden 6
D -- 10099 Berlin
Phone: +49-30-2093-9632
Fax: +49-30-2093-9729
Email address: [email protected] (Peter beim Graben)
URL: www.beimgraben.info (Peter beim Graben)
Preprint submitted to Frontiers in Computational Neuroscience
October 15, 2018
with an ad hoc model by Mazzoni et al.
[Mazzoni, A., S. Panzeri, N. K.
Logothetis, and N. Brunel (2008). Encoding of naturalistic stimuli by local
field potential spectra in networks of excitatory and inhibitory neurons. PLoS
Computational Biology 4 (12), e1000239], and conclude that our biophysically
motivated approach yields substantial improvement.
Keywords: biophysics, neural networks, leaky integrate-and-fire neuron,
current dipoles, extracellular medium, field potentials
1. Introduction
Since Hans Berger's 1924 discovery of the human electroencephalogram
(EEG) (Berger 1929), neuroscientists achieved much progress in clarifying
its neural generators (Creutzfeldt et al. 1966a,b, Nunez and Srinivasan 2006,
Schomer and Lopes da Silva 2011). These are the cortical pyramidal neu-
rons, as sketched in Fig. 1, that possess a long dendritic trunk separat-
ing mainly excitatory synapses at the apical dendritic tree from mainly in-
hibitory synapses at the soma and at the perisomatic basal dendritic tree
(Creutzfeldt et al. 1966a, Spruston 2008). In addition, they exhibit an ax-
ial symmetry and are aligned in parallel to each other, perpendicular to the
cortex' surface, thus forming a palisade of cell bodies and dendritic trunks.
When both kinds of synapses are simultaneously active, inhibitory synapses
generate current sources and excitatory synapses current sinks in extracellu-
lar space, hence causing the pyramidal cell to behave as a microscopic dipole
surrounded by its characteristic electrical field, the dendritic field potential
(DFP). The densely packed pyramidal cells form then a dipole layer whose
superimposed currents give rise to the local field potential (LFP) of neural
masses and eventually to the EEG (Lind´en et al. 2010, Lind´en et al. 2011,
Nunez and Srinivasan 2006, Schomer and Lopes da Silva 2011).
2
Figure 1: Sketch of a cortical pyramidal neuron with extracellular current dipole between
spatially separated excitatory (open bullet) and inhibitory synapses (filled bullet). Neural
in- and outputs are indicated by the jagged arrows. Dendritic current I D causes dendritic
field potential (DFP).
Despite of the progress from experimental neuroscience, theoretically un-
derstanding the coupling of complex neural network dynamics to the electro-
magnetic field in the extracellular space poses challenging problems; some of
them have been addressed to some extent by B´edard et al. (2004), B´edard and Destexhe
(2009), and B´edard and Destexhe (2012).
In computer simulation studies, neural mass potentials, such as LFP
and EEG are most realistically simulated by means of multicompartmen-
tal models (Lind´en et al. 2010, Lind´en et al. 2011, Protopapas et al. 1998,
Sargsyan et al. 2001). Lind´en et al. (2010) calculated the current dipole mo-
mentum of the DFP for single pyramidal and stellate cells, based on several
hundreds compartments of the dendritic trees. Their results were in com-
pliance with the standard dipole approximation of the electrostatic multi-
pole expansion in the far-field (more than 1 mm remote from the dendritic
trunk), but they found rather poor agreement with that approximation in
the vicinity of the cell body. For comparison they also computed a "two-
monopole" model of one synaptic current and its counterpart, the somatic
return current, estimated from the current dipole momentum of the whole
3
dendritic tree. This "two-monopole" model, which corresponds to an electri-
cally equivalent single dipole model, obtained from the decomposition of the
dendrite into two compartments, better approximates the true current dipole
momentum in the vicinity of the pyramidal neuron. By superimposing the
DFPs of pyramidal cells to the ensemble LFP, Lind´en et al. (2011) found that
LFP properties cannot be attributed to the far-field dipole approximation.
However, realistic multicompartmental models are computationally too
expensive for large-scale neural network simulations. Therefore, various
techniques have been proposed and employed to overcome computational
complexity. These include networks of point models (i.e. devoid from any
spatial representation), based on conductance models (Hodgkin and Huxley
1952, Mazzoni et al. 2008), population density models (Omurtag et al. 2000),
or firing rate models (Wilson and Cowan 1972), which can be seen as a
sub class of population density models, with uniform density distribution
(Chizhov et al. 2007). In these kinds of models, mass potentials such as LFP
or EEG are conventionally described as averaged membrane potential. A
different class of models are neural mass models (David and Friston 2003,
Jansen and Rit 1995, Wendling et al. 2000, Rodrigues et al. 2010), where
mass potentials are estimated either through sums (or actually differences) of
postsynaptic potentials (David and Friston 2003) or of postsynaptic currents
(Mazzoni et al. 2008).
In particular, the model of Mazzoni et al. (2008) which is based on Brunel and Wang
(2003), recently led to a series of follow-up studies (Mazzoni et al. 2010, 2011)
addressing the correlations between numerically simulated and experimen-
tally measured LFP/EEG with spike rates by means of statistical modeling
and information theoretic measures. In all of the above point models and
their extension to population models, it is assumed that the extracellular
space is iso-potential and the majority of studies thereby neglect the effect
of extracellular resistance. That is, the extracellular space constitutes a dif-
ferent and isolated domain with no effect on neuronal dynamics.
In this article we extend the ad hoc model of Mazzoni et al. (2008) to-
wards a biophysically better justified approach, taking the dipole character
of extracellular currents and fields into account. Basically, our model corre-
sponds to the "two-monopole", or, equivalent dipole model of Lind´en et al.
(2010) which gave a good fit of the DFP close to the cell body of a cor-
tical pyramidal neuron. However, we aim to keep the simplicity of the
Mazzoni et al. (2008) model in terms of computational complexity, by en-
dowing the extracellular space with resistance and by keeping point-like neu-
4
ronal circuits. That is, in our case we do not quite consider point neurons,
nor spatially extended models with detailed compartmental morphology, yet
an intermediate level of description is achieved. To this end we propose a
reduced three-compartmental model of a single pyramidal neuron (Destexhe
2001, beim Graben 2008, Wang et al. 2004), and derive an observation model
for the dendritic dipole currents in the extracellular space and thereby for the
DFP that contributes to the LFP of a neural population. Interestingly, our
reduced three-compartmental model enables us to derive a leaky integrate-
and-fire mechanism (as for a point model (Mazzoni et al. 2008)), with addi-
tional observation equations for the DFP, which all together allows to study
the relationship between spike rates and LFP. Our derivations also nicely map
realistic electrotonic parameters to phenomenological parameters considered
in Mazzoni et al. (2008).
2. Material and Methods
Mazzoni et al. (2008) consider three populations of neurons, namely exci-
tatory cortical pyramidal cells (population 1), inhibitory cortical interneurons
(population 2) and excitatory thalamic relay neurons (population 3), pass-
ing sensory input to the cortex that is simulated by a random (Erdos-R´enyi)
graph of K = 4000 pyramidal and L = 1000 interneurons with connection
probability P = 0.2.
2.1. Theory
We describe the ith cortical pyramidal neuron [Fig. 1] from population
1 via the electronic equivalent (reduced) three-compartment model Fig. 2
(Destexhe 2001, beim Graben 2008, Wang et al. 2004), which is parsimonious
to derive our observation model: one compartment for the apical dendritic
tree, another one for soma and perisomatic basal dendritic tree (Lind´en et al.
2010), and the third -- actually a leaky integrate-and-fire (LIF) unit -- for
the axon hillock where membrane potential is converted into spike trains by
means of an integrate-and-fire mechanism.
5
apical dendritic tree
perisomatic dendritic tree
axon hillock
exterior
Vi
ø(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
RD
i
RC
i
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
E E
i1
E E
i2
E E
i3
E I
i1
E I
i2
E I
i3
ñ
ñ
ñ
ñ
ñ
ñ
EM
ó
ó
ó
ó
ó
ó
(cid:19)
RE
i1
RE
i2
RE
i3
RI
i1
RI
i2
RI
i3
Ci
(cid:129)
RM
û
Ui
ú
(cid:13)
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
I B
i
û
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
ú
ú
€(cid:129)(cid:129)(cid:129)(cid:129)(cid:129)(cid:129) ‚
interior
RA
i
I D
i
I HH
i
RB
i
Figure 2: Proposed electronic equivalent circuit for a pyramidal neuron (reduced three
compartmental model). Note that the apical and basal dendrites are not true compart-
ments since capacitors are not explicitly represented, rather, these are implicitly taken
into account via EPSP and IPSP static functions, thus keeping computational complexity
low.
Excitatory synapses are represented by the left-most branch, where exci-
tatory postsynaptic potentials (EPSP) at a synapse between a neuron j from
population 1 or 3 and neuron i act as electromotoric forces EE
ij. These poten-
tials drive excitatory postsynaptic currents (EPSC) I E
ij, essentially consisting
of sodium ions, through the cell plasma with resistance RE
ij from the synapse
towards the axon hillock.
The middle branch describes the inhibitory synapses between a neuron
k from population 2 and neuron i. Here, inhibitory postsynaptic potentials
(IPSP) EI
ik provide a shortcut between the excitatory branch and the trigger
zone, where inhibitory postsynaptic currents (IPSC) I I
ik (essentially chloride
ions) close the loop between the apical and perisomatic dendritic trees. The
resistivity of the current paths along the cell plasma is given by RI
ik.
The cell membrane at the axon hillock itself is represented by the branch
at the right hand side. Here, a capacitor Ci reflects the temporary storage
capacity of the membrane. The serial circuit consisting of a battery EM
6
„
„
„
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
„
„
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
„
„
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
„
„
„
ƒ
…
„
„
„
„
„
ƒ
…
„
„
„
„
ƒ
…
„
„
„
„
ƒ
…
„
„
„
ƒ
…
„
„
„
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
„
„
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
ƒ
…
ƒ
…
„
ƒ
…
„
„
„
„
ƒ
…
„
„
„
„
„
„
„
„
„
and a resistor RM denotes the Nernst resting potential and the leakage con-
ductance of the membrane, respectively (Johnston and Wu 1997). Finally, a
spike generator (Hodgkin and Huxley 1952, Mazzoni et al. 2008) (indicated
by a "black box") is regarded of having infinite input impedance. Both,
EPSP and IPSP result from the interaction of postsynaptic receptor kinetics
with dendritic low-pass filtering in compartments one and two, respectively
(Destexhe et al. 1998, Lind´en et al. 2010). Hence the required capacitances,
omitted in Fig. 2, are already taken into account by EE
ik. Therefore, we
refer to our model as to a "reduced compartment model" here.
ij, EI
The three compartments are coupled through longitudinal resistors, RA
where RA
extracellular space (Holt and Koch 1999).
i denote the resistivity of the cell plasma and RC
i , RB
i , RD
i , RB
i that of
i , RC
i , RD
i
Finally, the membrane voltage at the axon hillock Ui (the dynamical state
variable) and the DFP Vi, which measures the drop in electrical potential
along the extracellular resistor RD
i are indicated. For the aim of calculation,
the mesh currents I D
i (the basal current) and I IF
i
(the integrate-and-fire current) are indicated.
i (the dendritic current), I B
The circuit in Fig. 2 obeys the following equations:
p
I D
i =
I B
i =
q
I E
ij
I I
ik
Xj=1
Xk=1
i − I B
dUi
+
dt
ijI E
ij + (RA
ikI I
ik + (RB
i I D
i
.
Ui − EM
RM
i + RD
i + RC
i = I D
I IF
i
I IF
i = Ci
EE
ij = RE
EI
ik = RI
Vi = RD
i )I D
i )I IF
i + RC
i + (RB
i )I IF
i + Ui , 1 ≤ k ≤ q
i + Ui , 1 ≤ j ≤ p
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Here, p is the number of excitatory and q is the number of inhibitory synapses
connected to neuron i.
The circuit described by Eqs. (1 -- 7) shows that the neuron i is likely to
fire when the excitatory synapses are activated. Then, the integrate-and-fire
current I IF
i . If, by contrast, also the inhibitory
synapses are active, the dendritic current I D
is shunted between the apical
i
i equals the dendritic current I D
7
and perisomatic basal dendritic trees and only a portion could evoke spikes
at the trigger zone [Eq. (4)]. On the other hand, the large dendritic current
I D
i flowing through the extracellular space of resistance RD
i , gives rise to a
large DFP Vi.
In order to simplify the following derivations, we gauge the resting po-
tential [Eq. (4)] to EM = 0, yielding
Ui
RM .
From (5) we obtain the individual EPSC's as
I IF
i = Ci
dUi
dt
+
I E
ij =
1
RE
ij (cid:2)EE
ij − (RA
i + RD
i )I D
i − (RB
i + RC
i )I IF
i − Ui(cid:3) .
And accordingly, the individual IPSC's from (6)
I I
ik =
1
RI
ik (cid:2)EI
ik − (RB
i + RC
i )I IF
i − Ui(cid:3) .
Inserting (9) into (1) yields the excitatory dendritic current
I D
i =
p
Xj=1
1
RE
ij
EE
ij − gE
i [(RA
i + RD
i )I D
i + (RB
i + RC
i )I IF
i + Ui] ,
where we have introduced the excitatory dendritic conductivity
gE
i =
p
Xj=1
1
RE
ij
.
(8)
(9)
(10)
(11)
(12)
Likewise we obtain the inhibitory dendritic currents from (2) and (10) as
I B
i =
q
Xk=1
1
RI
ik
ik − gI
EI
i[(RB
i + RC
i )I IF
i + Ui] ,
with the inhibitory dendritic conductivity
gI
i =
1
RI
ik
.
q
Xk=1
8
(13)
(14)
With these results, we obtain an interface equation for an observation
model as follows. Rearranging (11) yields
I D
i [1 + gE
i (RA
i + RD
i )] =
p
Xj=1
1
RE
ij
ij − gE
EE
i [(RB
i + RC
i )I IF
i + Ui]
(15)
Next, we eliminate I IF
i
through (8):
I D
i (cid:2)1 + gE
i (cid:0)RA
i + RD
Division by 1 + gE
i + RD
cellular dendritic dipole current:
RB
i + RC
i
RM (cid:19)(cid:21) .
i (cid:1)(cid:3) =
i (cid:0)RA
p
1
RE
ij
i + RC
ij − gE
EE
i (cid:20)Ci(cid:0)RB
+ Ui(cid:18)1 +
Xj=1
i (cid:1) gives the desired expression for the extra-
dUi
dt
i (cid:1)
I D
i =
p
Xj=1
αijEE
ij − βi
dUi
dt − γiUi ,
with the following electrotonic parameters
αij =
βi =
γi =
i )]
1
i (RA
i + RD
i + RC
i )
i + RD
i )
ij[1 + gE
RE
i (RB
CigE
1 + gE
i (RA
i (RM + RB
gE
RM[1 + gE
i (RA
i + RC
i )
i + RD
i )]
.
(16)
(17)
(18)
(19)
In order to derive the evolution equation we consider the integrate-and-
fire current I IF
that is given through (3). The individual EPSCs and IPSCs
i
have already been obtained in (9) and (10), respectively. Inserting (13) into
(3) yields
I IF
i
[1 − gI
i(RB
i + RC
i )] − gI
iUi = I D
i −
q
Xk=1
1
RI
ik
EI
ik .
Next we insert our interface equation (16) and also (8):
dUi
dt
(cid:20)Ci
+
Ui
RM(cid:21) [1 − gI
i(RB
i + RC
i )] − gI
iUi =
αijEE
ij − βi
dUi
dt − γiUi −
p
Xj=1
q
Xk=1
1
RI
ik
EI
ik
9
and obtain after some rearrangements
{Ci(cid:2)1 − gI
i(RB
i + RC
i )(cid:3) + βi}
dUi
dt
+
p
1 − gI
Xj=1
i(RB
i + RM) + RMγi
i + RC
RM
q
αijEE
ij −
Xk=1
1
RI
ik
EI
ik
and after multiplication with
ri =
RM
i + RC
i + RM) + RMγi
1 − gI
i(RB
the dynamical law for the membrane potential at axon hillock:
τi
dUi
dt
+ Ui =
p
Xj=1
wE
ij EE
ij −
q
Xk=1
wI
ik EI
ik ,
where we have introduced the following parameters:
• time constants
τi = ri{Ci(cid:2)1 − gI
i(RB
i + RC
i )(cid:3) + βi}
• excitatory synaptic weights
• inhibitory synaptic weights
wE
ij = riαij
wI
ik =
ri
RI
ik
.
Ui =
(20)
(21)
(22)
(23)
Using the result (20), we can also eliminate the temporal derivative in
the interface equation (16) through
dUi
dt
=
1
τi " p
Xj=1
wE
ij EE
ij −
q
Xk=1
wI
ik EI
ik − Ui#
(24)
which yields
I D
i =
p
Xj=1(cid:18)αij −
βi
τi
wE
ij(cid:19) EE
ij +
βi
τi
q
Xk=1
wI
ik EI
ik +(cid:18) βi
τi − γi(cid:19) Ui .
10
And eventually, by virtue of Eq. (7) after multiplication with RD
field potential (DFP)
i the dendritic
Vi =
p
Xj=1
with parameters
wE
ij EE
ij +
q
Xk=1
wI
ik EI
ik + ξiUi ,
wE
ij = RD
i wE
ij(cid:18) 1
ri −
βi
τi(cid:19)
wI
ik = RD
i wI
ik
βi
τi
ξi = RD
i (cid:18) βi
τi − γi(cid:19) .
(25)
(26)
(27)
(28)
The change in sign of the inhibitory contribution from Eq. (20) to Eq. (25) has
an obvious physical interpretation: In (20), the change of membrane potential
Ui and therefore the spike rate is enhanced by EPSPs but diminished by
IPSPs. On the other hand, the dendritic shunting current I D
in (25) is large
i
for both, large EPSPs and large IPSPs.
From Eq. (20) we eventually obtain the neural network's dynamics by
taking into account that postsynaptic potentials are obtained from presy-
naptic spike trains through temporal convolution with postsynaptic impulse
response functions, i.e.
EEI
ij (t) =Z t
−∞
sEI
i (t − t′)Rj(t′) dt′
(29)
where sEI
tions, respectively, and Rj is the spike train
i (t) are excitatory and inhibitory synaptic impulse response func-
Rj(t) =Xtν
δ(t − tν − τL)
(30)
coming from presynaptic neuron j, when spikes were emitted at times tν.
The additional time constant τL is attributed to synaptic transmission delay
(Mazzoni et al. 2008). These events are obtained by integrating (20) with
initial condition
Ui(tν) = E.
11
(31)
Where E is some steady-state potential (Mazzoni et al. 2008).
t = tν the membrane reaches a threshold
If at time
Ui(t) ≥ θi(t)
(32)
(with possibly a time dependent activation threshold θi(t)) from below dUi(t)
dt >
0 then an output spike δ(t − tν) is generated, which is then followed by a
potential resetting as follows
Ui(tν+1) ← E .
(33)
Additionally, the integration of the dynamical law is restarted at time t =
tν+1 + τrp after interrupting the dynamics for a refractory period τrp.
Inserting (29) into (20) entails the evolution equation of the neural net-
work
τi
dUi
dt
+ Ui =
Xj=1
Xk=1
p
q
wE
ij sE
i (t) ∗ Rj(t) +
wI
ik sI
i(t) ∗ Rk(t) ,
(34)
where the signs had been absorbed by the synaptic weights, such that wE
for excitatory synapses and wI
ik < 0 for inhibitory synapses, respectively.
ij > 0
Following Mazzoni et al. (2008) an individual postsynaptic current I EI
ij
at a synapse between neurons i and j obeys
τ EI
d
τ EI
r
dI EI
ij
dt
dxEI
ij
dt
+ I EI
ij = xEI
ij
+ xEI
ij = F EI
ij
,
(35)
(36)
d
are decay time constants and τ EI
where τ EI
and IPSC, respectively. Auxiliary variables are denoted by xEI
prescribes presynaptic forcing
are rise time constants of EPSC
ij , while F EI
ij
r
F EI
ij = τiJijRj(t)
(37)
with spike train (30). Here, Jij = vwEI
as voltage unit.
ij denotes synaptic gain with v = 1 mV
Note that (37) is essentially a weighted sum of delta functions, such that
a single spike can be assumed as particular forcing
F = F0δ(t) ,
12
(38)
with some constant F0.
Derivating (35) and eliminating xEI
ij
transforms Eqs.
(35 -- 36) into a
linear second-order differential equation with constant coefficients
τ EI
d τ EI
r
d2I EI
ij
dt2 + (τ EI
d + τ EI
r
dI EI
ij
dt
)
+ I EI
ij = F EI
ij
.
(39)
Equation (39) with the particular forcing (38) is solved by a Green's
i (t) such that the general solution of (39) is obtained as the
function sEI
temporal convolution
I EI
ij (t) =Z t
−∞
sEI
i (t − t′)F EI
ij (t) dt′ .
(40)
For t 6= 0, (39) assumes its homogeneous form and is easily solved by
means of the associated characteristic polynomial
τ EI
d τ EI
r λ2 + (τ EI
d + τ EI
and λ2 = −1/τ EI
r
r
)λ + 1 = 0
(41)
, entailing the Green's functions
with roots λ1 = −1/τ EI
sEI
d
i (t) =(cid:16)AEIet/τ
EI
r − BEIet/τ
EI
d (cid:17) Θ(t)
(42)
with the Heaviside step function Θ(t).
The constants AEI, BEI > 0 are obtained from the initial conditions
sEI
i (t) = 0, reflecting causality, and a suitable normalization
Z ∞
0
sEI
i (t)dt = 1 .
The initial condition yields AEI = BEI ≡ SEI, while the remaining
constant
SEI =
1
,
d − τ EI
τ EI
r
due to normalization. Therefore, the normalized Green's functions are those
of Brunel and Wang (2003)
sEI
i (t) = v
τi
r (cid:16)et/τ
τ EI
d − τ EI
EI
r − et/τ
EI
d (cid:17) Θ(t) .
(43)
13
Now, we are able to compare our DFP Vi [Eq. (25)] with the estimate
of Mazzoni et al. (2008) which is given (in our notation) as the sums of the
moduli of excitatory and inhibitory synaptic currents, i.e.
V MPLB
i
=Xj
ij +Xk
I E
I I
ik
(44)
where "MPLB" refers to the authors Mazzoni et al. (2008).
From (25) and (44), respectively, we compute two models of the local
field potential (LFP). First, by summing DFP across all pyramidal neurons
(beim Graben and Kurths 2008, Mazzoni et al. 2008), and, second by taking
the DFP average (Nunez and Srinivasan 2006), which yields
L2 =
L1 = Xi
K Xi
L3 = Xi
K Xi
L4 =
Vi
1
1
V MPLB
i
V MPLB
i
Vi ,
(45)
(46)
(47)
(48)
where K is number of pyramidal neurons.
2.2. Parameter estimation
Next, we relate the electrotonic parameters of our model to the phe-
nomenological parameters of Mazzoni et al. (2008). To this end, we first
report their synaptic efficacies in Tab. 1.
Synaptic efficacies / mV on interneurons
GABA
recurrent cortical AMPA
external thalamic AMPA
2.7
0.7
0.95
on pyramidal neurons
1.7
0.42
0.55
Table 1: Parameters laid as in Mazzoni et al. (2008).
From these, we compute the synaptic weights through
wE
ij = J E
ij/v =(cid:26) 0.42 if
0.55 if
j "cortical"
j "thalamic"
(49)
14
and
wI
ik = J I
ik/v = 1.7
Next, we determine the factors ri by virtue of Eq. (23) through
ri =
wI
ik
¯gGABA
=
1.7
1 nS
= 1.7 GΩ
using the inhibitory synaptic conductivity ¯gGABA = 1 nS, Correspondingly,
Eq. (22) allows us to express αij in terms of the excitatory synaptic weights
through
αij =
wE
ij
ri
=(cid:26) 0.25 nS if
0.32 nS if
j "cortical"
j "thalamic"
From αij we can determine the total excitatory synaptic conductivities
gE
i according to Eq. (17) through
αij =
1
i (RA
i + RD
i )]
ij[1 + gE
RE
p
αij
Xj=1
gE
i "1 − (RA
i + RD
i )
p
Xj=1
αij# =
Pp
1 − (RA
i + RD
gE
i =
j=1 αij
i )Pp
1
j=1 αij
αij[1 + gE
i (RA
i + RD
i )]
and hence
RE
ij =
(50)
(51)
Inserting next (18) into (21) yields
τi = riCi
1 + gE
i (RA
i + RD
i ) + (RB
i + RC
i (RA
i ){gE
i − gI
i + RD
i )
1 + gE
i[1 + gE
i (RA
i + RD
i )]}
.(52)
Equation (52) could constraint the choice of the membrane capacitance
Ci by choosing τi = 20 ms (Mazzoni et al. 2008).
In order to also determine the DFP parameters (26) -- (28), we finally
compute the ratios
βi
τi
=
ri{1 + gE
i (RA
i + RD
gE
i (RB
i ) + (RB
i + RC
i + RC
i )
i ){gE
i − gI
i[1 + gE
i (RA
i + RD
.
i )]}}
15
The remaining electrotonic parameters RM
i are es-
timated from cell geometries as follows. The resistance R of a volume con-
ductor is proportional to its length ℓ and reciprocally proportional to its
cross-section A, i.e.
i , and RD
i , RA
i , RB
i , RC
R = ρ
ℓ
A
(53)
where ρ is the (specific) resistivity of the medium. Table 2 shows the resistiv-
ities of the three kinds of interest which then allows to evaluate the various
volume conductor resistances according to Eq. (53).
medium
cell membrane (at axon hillock)
cell plasma (cytoplasm)
extracellular space
ρ/Ωcm
5 · 107
200
333
Table 2: Resistivities of cell membrane, cell plasma and extracellular space. Parameters
from Kole and Stuart (2012), Rall (1977), Mainen et al. (1995), Gold et al. (2007). Note
that the resistivity of the cell membrane has to be related to the constant membrane
thickness (≈ 10 nm).
We consider a total dendritic length of 2ℓ = 20 µm and a dendritic radius
of a = 7 µm, that are generally subjected to variation. Equally, parameters
that were allowed to vary are the length and radius of the axon hillock,
yet herein we consider a length of 2ℓ = 20 µm and radius of a = 0.5 µm
(Destexhe 2001, Kole and Stuart 2012, Mainen et al. 1995). To evaluate the
intracellular (RA, RB) and extracellular (RD, RC) resistances, respectively,
according to Eq. (53), we consider a simple implementation where the length
ℓ is half of the dendritic length (i.e. basal and apical length are symmetrical,
but this can be broken). However, the cross sectional area for the cytoplasm
is simply A = πa2. Finally, the area of the axon hillock is simply the surface
area of a cylinder.
In order to also determine the cross-section of extracellular space between
dendritic trunks we make the following approximations. We assume that
dendritic trunks are parallel aligned cylinders of radius a and length ℓ that
are hexagonally dense packed. Then the centers of three adjacent trunks
form an equilateral triangle with side length 2a and hence area 2√3a2. The
enclosed space is then given by the difference of the triangle area and the
16
area of three sixth circle sectors, therefore
Aspace = 2√3a2 −
3
6
πa2 =(cid:18)2√3 −
1
2
π(cid:19) a2 .
Hence, the cross-section of extracellular space surrounding one trunk is
A = 6Aspace =(cid:16)12√3 − 3π(cid:17) a2 .
(54)
2.3. Simulations
Subsequently, we implement an identical network to the one considered
by Mazzoni et al. (2008) with Brian Simulator, that is a Python based en-
vironment (Goodman and Brette 2009). However, the derivations from the
previous section enables the possibility of setting a dipole observable that
measures the local dendritic field potential (DFP) on each pyramidal neurons,
given by Eq. (25). This allows then to define a mesoscopic LFP observable,
which can be equated either as averaged DFP or simply given as the sum of
DFP, given by Eqs. (45 -- 48). Primarily, we compare our LFP measure L4,
proposed as the average of DFP, with the Mazzoni et al. LFP L1 which is
defined as the sum of absolute values of GABA and AMPA currents Eq. (44).
Additionally, we also compare all possible measures, namely, mean membrane
potential 1
L2, sum of DFP L3, and the average of DFP L4.
K Pi Ui, Mazzoni et al. LFP L1, average of Mazzoni et al. DFP
For completeness, we briefly summarize the description of the network
(we refer the reader to Mazzoni et al. (2008) for details). The network mod-
els a cortical tissue with leaky integrate-and-fire neurons, composed of 1000
inhibitory interneurons and 4000 pyramidal neurons, which are described by
the evolution equation (34). The threshold crossings given by Eq. (32) is
considered static with θi = 18 mV and the reset potential E = 11 mV. The
refractory period for excitatory neurons is τrp = 2 ms while for inhibitory
neurons it is τrp = 1 ms. The network connectivity is random and sparse
with a 0.2 probability of directed connection between any pair of neurons.
The evolution of synaptic currents, fast GABA (inhibitory) and AMPA (ex-
citatory) are described via the second order evolution equations (35 -- 36),
which are activated by incoming presynaptic spikes represented by Eq. (30).
The latency of the postsynaptic currents is set to τL = 1 ms and the rise and
decay times are given by Tab. 3.
17
Synaptic times
GABA
AMPA on interneurons
AMPA on pyramidal neurons
τr / ms
0.25
0.2
0.4
τd / ms
5
1
2
Table 3: Synaptic rise (τr) and decay times (τd). Parameters laid as in Mazzoni et al.
(2008).
ij
Moreover, synaptic efficacies, J EI
, for simulation were presented in Tab. 1.
Note that relation (49) then allows to determine the synaptic weights. Addi-
tionally, all neurons receive external thalamic excitatory inputs, that is, via
AMPA-type synapses, which are activated by random Poisson spike trains,
with a time varying rate that is identical for all neurons. Specifically, the
thalamic inputs are the only source of noise, which attempts to account for
both cortical heterogeneity and spontaneous activity. This is achieved by
modeling a two level noise, where the first level is an Ornstein-Uhlenbeck
process superimposed with a constant signal and the second level is a time
varying inhomogeneous Poisson process. Thus, we have the following time
varying rate, λ(t), that feeds into inhomogeneous Poisson process:
τn
dn(t)
= −n(t) + σnr 2
dt
λ(t) = [c0 + n(t)]+
τn
η(t)
(55)
(56)
where η(t) represents Gaussian white noise, c0 represents a constant signal
(but equally could be periodic or other), and the operation [·] is the threshold-
linear function, [x]+ = x if x > 0, [x]+ = 0 otherwise, which circumvents
negative rates. The constant signal c0 can range between 1.2 to 2.6 spikes/ms.
The parameters of the Ornstein-Uhlenbeck process are τn = 16 ms and the
standard deviation σn = 0.4 spikes/ms.
For complete exposition, we note that from an implementation viewpoint
(within the Brian simulator), a copy of the postsynaptic impulse response
function (29) has to be evaluated to calculate the dendritic field poten-
tial (DFP) (25) with weights wEI
ij . This implies evaluating the second or-
der process (35) -- (36) with a different forcing term. Specifically, starting
from I EI
ij (t) ≡ wEI
and pre-multiplying both sides
with wEI
ij and subsequently re-arranging we obtain the desired forcing term
F EI
ij = wEI
ij with
equation (37) and using relation (49) we finally obtain F EI
ij τivRj(t).
ij . Note further that by expanding the term F EI
i (t) ∗ F EI
ij (t) = sEI
ij = wEI
ij F EI
ij /wEI
ij EEI
ij
18
3. Results
Following Mazzoni et al. (2008), the network simulations are run for two
seconds with three different noise levels, specifically, receiving a constant
signal with three different rates 1.2, 1.6 and 2.4 spikes/ms as depicted in
Fig. 3. Note that these input rates do not mean that a single neuron fires
at these high rates. Rather, it can be obtained from multiple neurons that
jointly fire with slower, yet desynchronized, rates converging at the same
postsynaptic cell. The Poisson process ensures that this is well represented.
19
e
t
a
r
c
i
m
a
l
a
h
t
r
e
b
m
u
n
n
o
r
u
e
n
c
e
s
/
s
e
k
p
s
i
c
e
s
/
s
e
k
p
s
i
)
V
(
1
L
P
F
L
)
V
m
(
4
L
P
F
L
Signal rate = 1.2 spikes/ms
Signal rate = 1.6 spikes/ms
Signal rate = 2.4 spikes/ms
2.0
1.5
1.0
0.5
A
2.5
2.0
1.5
1.0
0.5
0.0
C
B
4.0
3.5
3.0
2.5
2.0
1.5
1750 1800
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
1.0
1750 1800
1850
1900
1950 2000
200
150
100
50
D
200
150
100
50
E
200
150
100
50
F
0
1800
1840
1880
1920
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
3.0
2.0
1.0
0.0
Interneurons
G
2.0
1.0
0.0
H
4.0
3.0
2.0
1.0
0.0
I
1750 1800
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
20
15
10
5
J
Pyramidal neurons
16
12
8
4
K
30
20
10
L
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
350
300
250
200
150
100
50
M
350
300
250
200
150
100
50
N
600
500
400
300
200
100
O
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
14
12
10
8
6
4
P
16
14
12
10
8
6
4
Q
26
24
22
20
18
16
14
12
10
R
2
1750 1800
1850
1900
1950 2000
2
1750 1800
1850
1900
1950 2000
8
1750 1800
1850
1900
1950 2000
Time (ms)
Time (ms)
Time (ms)
Figure 3: Dynamics of the network and LFP comparisons: The three columns represent
different runs of the network for three different rates, 1.2, 1.6 and 2.4 spikes/ms. In each
column, all panels show the same 250 ms (extracted from 2 seconds simulations). The top
panels (A-C) represent thalamic inputs with the different rates. The second top panels
(D-F) corresponds to a raster plot of the activity of 200 pyramidal neurons. Panels (G-I)
depict average instantaneous firing rate (computed on a 1 ms bin) of interneurons (blue)
and panels (J-L) correspond to average instantaneous firing rate of pyramidal neurons.
Panels (M-O) show the Mazzoni et al. LFP L1 from Eq. (45). Finally, panels (P-R)
depict our proposed LFP measure L4, which is the average of dendritic field potential
(DFP) [Eq. (48)].
20
The focus is to compare our proposed measure L4, defined as mean of
the dendritic field potential (DFP) [Eq. (48)], with the Mazzoni et al. LFP
L1 from Eq. (45). In Fig. 3 one sees two main striking differences between
the two measures, namely in frequency and in amplitude. Specifically, L1
responds instantaneously to the spiking network activity by means of high
frequency oscillations. Moreover, L1 also exhibits a large amplitude.
In
contrast, our mean dendritic field potential L4 measures comparably to ex-
perimental LFP, that is, in the order of millivolts, and although it responds
to population activity, it has a relatively smoother response. Actually one
can realize that our LFP estimate represents low-pass filtered thalamic input.
The physiological relevance of this is not yet clear in our work. However,
recent work (Poulet et al. 2012) shows that desynchronized cortical state dur-
ing active behavior is driven by a centrally generated increase in thalamic
action potential firing (i.e. thalamic firing controls cortical states). Thus, it
seems that cortical synchronous activity is suppressed when thalamic input
increases, thereby suggesting that cortical desynchronized states to be re-
lated to sensory processing. This work further quantifies these observations
by applying Fast Fourier Transform (FFT) to cortical EEG and subsequently
comparing with thalamic firing rate by means of Pearson correlation coeffi-
cient. Unfortunately they do not quantify the amount of thalamic oscillations
contained within the cortical EEG.
Yet, to keep a comparable comparison between measures, we also compute
the average of the Mazzoni et al. DFP L2 [Eq. (48)] and additionally the
mean membrane potential (the standard considered in the neuroscientific
literature). These are shown in Fig. 4.
21
)
V
m
(
e
n
a
r
b
m
e
m
n
a
e
m
)
V
(
1
L
P
F
L
)
V
m
(
2
L
P
F
L
)
V
(
3
L
P
F
L
)
V
m
(
4
L
P
F
L
Signal rate = 1.2 spikes/ms
Signal rate = 1.6 spikes/ms
Signal rate = 2.4 spikes/ms
3
A
13
12
11
10
9
8
7
B
12
10
8
6
4
2
C
13
12
11
10
9
8
7
6
5
1750 1800
1850
1900
1950 2000
6
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
350
300
250
200
150
100
50
D
350
300
250
200
150
100
50
F
E
600
500
400
300
200
100
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
0
1750 1800
1850
1900
1950 2000
80
70
60
50
40
30
20
10
G
90
80
70
60
50
40
30
20
10
0
1750 1800
1850
1900
1950 2000
0
1750 1800
50
45
40
35
30
25
20
15
10
J
60
50
40
30
20
10
I
H
140
120
100
80
60
40
20
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
100
K
L
90
80
70
60
50
40
30
1750 1800
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
1750 1800
1850
1900
1950 2000
14
12
10
8
6
4
M
16
14
12
10
8
6
4
N
26
24
22
20
18
16
14
12
10
O
2
1750 1800
1850
1900
1950 2000
2
1750 1800
1850
1900
1950 2000
8
1750 1800
1850
1900
1950 2000
Time (ms)
Time (ms)
Time (ms)
Figure 4: Comparison of different LFP measures when the network receives constant signal
with three different rates (1.2, 1.6 and 2.4 spikes/ms). Again, only 250 ms is represented
(extracted from 2 sec simulation). The first plot (A-C) corresponding to the the different
rates shows the most widespread LFP measure used in the literature, namely average
membrane potential 1
Ui. The second panel (D-F) shows the Mazzoni et al. LFP
L1 from Eq. (45). The third panel (G-I) displays the average of the Mazzoni et al. DFP
L2 [Eq. (46)]. Similarly, the fourth panel (J-L) shows the total, L3, [Eq. (47)] and the
last panel (M-O) depicts the averaged, L4, [Eq. (48)] LFP measure. Note the different
amplitude scales between measures.
K Pi
22
Clearly, in terms of time profile, the summed and averaged observables
are similar within the same class of LFP measures. However, in all cases
the Mazzoni et al. LFP L1 exhibits a significantly larger order of magnitude,
which diverges substantially from experimental LFP amplitudes, typically
varying between 0.5 to 2 mV (Niedermeyer 2005, Lakatos et al. 2005). In
contrast, although the mean DFP is not contained within the interval from
0.5 to 2 mV it arguably performs better. However, we do concede further
work is required. Some gains in improving the different LFP measures can
be achieved by applying for example a weighted average, which would mimic
the distance of an electrode to a particular neuron by means of a lead field
kernel (Nunez and Srinivasan 2006). For example a convolution of either L1
or L2 with a Gaussian kernel (representing the distance to a neuron), would
yield a measure that captures better the local field potential or better the
dendritic field potential of the nearest neurons. However, further work will
be required to properly quantify the gain when space is taken into account.
In Fig. 5 we finally contrast the power spectra of the different LFP mea-
sures.
23
20
0
-20
-40
-60
-80
z
H
/
B
d
-100
-120
0
30
20
10
0
-10
-20
-30
-40
-50
-60
0
20
0
-20
-40
-60
-80
0
20
0
-20
-40
-60
-80
z
H
/
B
d
z
H
/
B
d
z
H
/
B
d
-100
0
20
0
-20
-40
-60
-80
z
H
/
B
d
-100
0
Signal rate = 1.2 spikes/ms
Signal rate = 1.6 spikes/ms
Signal rate = 2.4 spikes/ms
B
1000
2000
3000
4000 5000
E
1000
2000
3000
4000 5000
H
A
mean membrane
potential
1000
2000
3000
4000 5000
D
LFP L1
1000
2000
3000
4000 5000
G
LFP L2
20
0
-20
-40
-60
-80
-100
-120
0
30
20
10
0
-10
-20
-30
-40
-50
-60
0
20
10
0
-10
-20
-30
-40
-50
-60
-70
1000
2000
3000
4000 5000
0
1000
2000
3000
4000 5000
J
LFP L3
1000
2000
3000
4000 5000
M
LFP L4
1000
2000
3000
4000 5000
20
0
-20
-40
-60
-80
-100
0
20
0
-20
-40
-60
-80
-100
-120
0
K
1000
2000
3000
4000 5000
N
1000
2000
3000
4000 5000
Hz
Hz
0
-20
-40
-60
-80
-100
-120
0
40
20
0
-20
-40
-60
0
30
10
-10
-30
-50
-70
0
40
20
0
-20
-40
-60
-80
-100
0
20
0
-20
-40
-60
-80
-100
0
C
1000
2000
3000
4000 5000
F
1000
2000
3000
4000 5000
I
1000
2000
3000
4000 5000
L
1000
2000
3000
4000 5000
O
1000
2000
3000
4000 5000
Hz
Figure 5: Comparison of power spectra of the various LFP measures when the network
receives constant signal with three different rates (1.2, 1.6 and 2.4 spikes/ms): The first
plot (A-C) corresponding to the different rates shows the power spectrum of the average
membrane potential 1
Ui. The second plot (D-F) and third plots (G-I) show power
spectra of the total and average of L1 and L2 corresponding to Mazzoni et al. (2008),
respectively. The fourth plot (J-L) and fifth plots (M-O) display power spectra of the L3
and L4 measures from our model, respectively. Note we show the full spectrum up to
5 kHz only for convenience due to the fine sample rate.
K Pi
One interesting feature is that the power spectrum of the Mazzoni et al.
LFP measures decays much more slowly that the average membrane potential
24
for higher frequencies. This observation is true for both, L1 and L2.
In
contrast, our LFP measures L3 and L4 fare better, and in particular, L4
decays at an approximately similar rate as the average membrane potential.
4. Discussion
In this article we derived a model for cortical dipole fields, such as den-
dritic and local field potential (DFP/LFP) from biophysical principles. To
that aim we decomposed a cortical pyramidal cell, the putative generator
of those potentials, into three compartments: the apical dendritic tree as
the place of mainly excitatory (AMPA) synapses, the soma and the peri-
somatic dendritic tree as the place of mainly inhibitory (GABA) synapses
and the axon hillock as the place of wave-to-spike conversion by means of an
integrate-and-fire mechanism. From Kirchhoff's laws governing an electronic
equivalent circuit of our model, we were then able to derive the evolution
equation for neural network activity Eq. (34) and, in addition, an observa-
tion equation Eq. (25) for the dendritic dipole potential contributing to the
LFP of a cortical population.
In order to compare our approach with another model discussed in the
recent literature (Mazzoni et al. 2008, 2010, 2011) we aligned the parameters
of our model with the model of Mazzoni et al. (2008) who approximated DFP
as the sum of moduli of excitatory and inhibitory synaptic currents Eq. (44).
From both approaches, we computed four different LFP estimates: L1, the
sum of Mazzoni et al. DFP, L2, the population average of Mazzoni et al.
DFP, L3 the sum of our dipole DFP, and L4 the population average of our
dipole DFP [Eqs. (45 -- 48)].
Our results indicate two main effects between our dipole LFP measures
and those of Mazzoni et al. Firstly, the measures based on Mazzoni et al.
(2008) systematically overestimate LFP amplitude by almost one order of
magnitude. One reason for that could be attributed to the direct conversion
of synaptic current into voltage without taking extracellular conductivity into
account, as properly done in our approach. Yet, another, even more crucial
reason is disclosed by our equivalent circuit Fig. 2. In our approach there is
just one extracellular current I D flowing from the perisomatic to the apical
dendritic tree. In the model of Mazzoni et al. (2008), however, two synaptic
currents that might be of the same order of magnitude are superimposed
to the DFP. Secondly, the measures based on Mazzoni et al. (2008) also sys-
tematically overestimate LFP frequencies. This could probably be attributed
25
partly to spurious higher harmonics introduced by computing absolute val-
ues. Moreover, taking the power spectrum shows that the Mazzoni et al.
(2008) measure decays much more slowly than the average membrane poten-
tial, which is at variance with experimental data.
However, at the current stage, both models, that of Mazzoni et al. (2008)
and our own, agree with respect to the polarity of DFP and LFP. The mea-
sures based on Mazzoni et al. (2008) have positive polarity simply due to the
moduli. On the other hand, also the direction of current dipoles in our model
is constrained by the construction of the equivalent circuit Fig. 2 where cur-
rent sources are situated at the perisomatic and current sinks are situated
at apical dendritc tree. Taking this polarity as positive also entails positive
DFP and LFP that could only change in strength. However, it is well known
from brain anatomy that pyramidal cells appear in at least two layers, III
and VI, of neocortex. This is reflected in experiments when an electrode tra-
verses different layers by LFP polarity reversals, and, of course, by the fact
that LFP and EEG oscillate between positive and negative polarity. Adapt-
ing our model to this situation could be straightforwardly accomplished in
the framework of neural field theory by fully representing space and simulat-
ing layered neural fields (Amari 1977, Jirsa and Haken 1996, beim Graben
2008). By contrast such a generalization is impossible at all with the model
of Mazzoni et al. (2008) due to the presence of absolute values.
On theses grounds we have good indication that our measure is an im-
provement to the Mazzoni et al. LFP measures, and, quite importantly,
it is biophysically better motivated than the ad hoc model of Mazzoni et al.
(2008). However, much considerable effort is still required to underpin all the
relevant LFP mechanisms and to better represent experimental LFP/EEG
dynamics.
Finally, our work provides a new framework where dendritic field poten-
tials and the relationship between firing rates and local fields can be explored
without the extreme demand on computational complexity involved in multi-
compartmental modeling (Lind´en et al. 2010, Lind´en et al. 2011, Protopapas et al.
1998, Sargsyan et al. 2001) by adopting reduced compartment circuits. For
example, we envisage to extend our recent work which maps firing rate model
(derived from LIF models) to population density models (Chizhov et al.
2007), but now incorporating our observational DFP model.
In addition,
our framework is analytically amenable and thus can be applied to any lin-
ear differential equation, for instance, GIF (Gif-sur-Yvette Integrate Fire)
models, which are improvements to the LIF models and compute more ac-
26
curately spike activations (Rudolph-Lilith et al. 2012). Also resonant mem-
branes (mediated by Ca2+ and a Ca2+-activated K+ ionic currents) that
describe sub-threshold oscillations and which can be easily expressed by lin-
ear equations (Mauro et al. 1970) can be incorporated in our derivations.
We note however that our framework can be applied to non-linear equations,
with Hodgkin and Huxley (1952) type activation, but it will fall short from
explicit and analytical observation equations.
Acknowledgements
We thank Michelle Lilith, Claude B´edard, Alain Destexhe, and Jurgen
Kurths for fruitful discussion. In addition, we would like to thank Samantha
Adams for providing help with Brian Simulator installations and initial dis-
cussions of Brian usage. This research was supported by a DFG Heisenberg
grant awarded to PbG (GR 3711/1-1).
References
Amari, S.-I. (1977). Dynamics of pattern formation in lateral-inhibition type
neural fields. Biological Cybernetics 27, 77 -- 87.
B´edard, C. and A. Destexhe (2009). Macroscopic models of local field po-
tentials and the apparent 1/f noise in brain activity. Biophysical Jour-
nal 96 (7), 2589 -- 2603.
B´edard, C. and A. Destexhe (2012). Modeling local field potentials and their
interaction with the extracellular medium. In R. Brette and A. Destexhe
(Eds.), Handbook of Neural Activity Measurement, pp. 136 -- 191. Cam-
bridge: Cambridge University Press.
B´edard, C., H. Kroger, and A. Destexhe (2004). Modeling extracellular
field potentials and the frequency-filtering properties of extracellular space.
Biophysical Journal 86 (3), 1829 -- 1842.
beim Graben, P. (2008). Foundations of neurophysics.
In P. b. Graben,
C. Zhou, M. Thiel, and J. Kurths (Eds.), Lectures in Supercomputational
Neuroscience: Dynamics in Complex Brain Networks, Springer Complex-
ity Series, pp. 3 -- 48. Berlin: Springer.
27
beim Graben, P. and J. Kurths (2008). Simulating global properties of elec-
troencephalograms with minimal random neural networks. Neurocomput-
ing 71 (4), 999 -- 1007.
Berger, H. (1929). Uber das Elektroenkephalogramm des Menschen. Archiv
fur Psychiatrie 87, 527 -- 570.
Brunel, N. and X.-J. Wang (2003). What determines the frequency of fast
network oscillations with irregular neural discharges? I. synaptic dynamics
and excitation-inhibition balance. Journal of Neurophysiology 90 (1), 415
-- 430.
Chizhov, A. V., R. S., and T. J. R. (2007). A comparative analysis of a
firing-rate model and conductance-based neural population model. Physics
Letters A 369, 31 -- 36.
Creutzfeldt, O. D., S. Watanabe, and H. D. Lux (1966a). Relations be-
tween EEG phenomena and potentials of single cortical cells. I. evoked
responses after thalamic and epicortical stimulation. Electroencephalogra-
phy and Clinical Neurophysiology 20 (1), 1 -- 18.
Creutzfeldt, O. D., S. Watanabe, and H. D. Lux (1966b). Relations between
EEG phenomena and potentials of single cortical cells. II. spontaneous
and convulsoid activity. Electroencephalography and Clinical Neurophysi-
ology 20 (1), 19 -- 37.
David, O. and K. J. Friston (2003). A neural mass model for MEG/EEG:
coupling and neuronal dynamics. NeuroImage 20, 1743 -- 1755.
Destexhe, A. (2001). Simplified models of neocortical pyramidal cells pre-
serving somatodendritic voltage attenuation. Neurocomputing 38-40, 167
-- 173.
Destexhe, A., F. Mainen, and T. J. Sejnowski (1998). Kinetic models of
synaptic transmission. See Koch and Segev (1998), pp. 1 -- 25.
Gold, C., D. Henze, and C. Koch (2007). Using extracellular action potential
recordings to constrain compartmental models. Journal of Computational
Neuroscience 23 (1), 39 -- 58.
28
Goodman, D. and R. Brette (2009). The Brian simulator. Frontiers in
Neuroscience 3 (2), 192 -- 197.
Hodgkin, A. L. and A. F. Huxley (1952). A quantitative description of mem-
brane current and its application to conduction and excitation in nerve.
Journal of Physiology 117, 500 -- 544.
Holt, G. R. and C. Koch (1999). Electrical interactions via the extracellular
potential near cell bodies. Journal of Computational Neuroscience 6, 169
-- 184.
Jansen, B. H. and V. G. Rit (1995). Electroencephalogram and visual evoked
potential generation in a mathematical model of coupled cortical columns.
Biological Cybernetics 73, 357 -- 366.
Jirsa, V. K. and H. Haken (1996). Field theory of electromagnetic brain
activity. Physical Review Letters 77 (5), 960 -- 963.
Johnston, D. and S. M.-S. Wu (1997). Foundations of Cellular Neurophysi-
ology. Cambridge (MA): MIT Press.
Koch, C. and I. Segev (Eds.) (1998). Methods in Neuronal Modelling. From
Ions to Networks (2nd ed.). Computational Neuroscience. Cambridge
(MA): MIT Press.
Kole, M. and G. Stuart (2012). Signal processing in the axon initial segment.
Neuron 73 (2), 235 -- 247.
Lakatos, P., A. Shah, K. Knuth, I. Ulbert, G. Karmos, and C. Schroeder
(2005). An oscillatory hierarchy controlling neuronal excitability and stim-
ulus processing in the auditory cortex. Journal of Neurophysiology 94 (3),
1904 -- 1911.
Lind´en, H., K. Pettersen, and G. Einevoll (2010). Intrinsic dendritic filtering
gives low-pass power spectra of local field potentials. Journal of Compu-
tational Neuroscience 29, 423 -- 444.
Lind´en, H., T. T, T. C. Potjans, K. H. Pettersen, S. Grun, M. Diesmann,
and G. T. Einevoll (2011). Modeling the spatial reach of the LFP. Neu-
ron 72 (5), 859 -- 872.
29
Mainen, Z., J. Joerges, J. Huguenard, and T. Sejnowski (1995). A model
of spike initiation in neocortical pyramidal neurons. Neuron 15 (6), 1427 --
1439.
Mauro, A.,F. Conti, F., F. Dodge, and R. Schor (1970). Subthreshold be-
havior and phenomenological impedance of the squid giant axon. Journal
of General Physiology 55 , 497 -- 532.
Mazzoni, A., N. Brunel, S. Cavallari, N. K. Logothetis, and S. Panzeri (2011).
Cortical dynamics during naturalistic sensory stimulations: Experiments
and models. Journal of Physiology 105 (1-3), 2 -- 15.
Mazzoni, A., S. Panzeri, N. K. Logothetis, and N. Brunel (2008). Encoding of
naturalistic stimuli by local field potential spectra in networks of excitatory
and inhibitory neurons. PLoS Computational Biology 4 (12), e1000239.
Mazzoni, A., K. Whittingstall, N. Brunel, N. K. Logothetis, and S. Panz-
eri (2010). Understanding the relationships between spike rate and
delta/gamma frequency bands of LFPs and EEGs using a local cortical
network model. NeuroImage 52 (3), 956 -- 972.
Niedermeyer, E. (2005). The normal EEG of the waking adult. In E. Nieder-
meyer and F. L. D. Silva (Eds.), Electroencephalography: Basic principles,
clinical applications, and related fields, 5th edition, pp. 167 -- 192. Lippincott
Williams & Wilkins.
Nunez, P. L. and R. Srinivasan (2006). Electric Fields of the Brain: The
Neurophysics of EEG (2nd ed.). New York: Oxford University Press.
Omurtag, A., B. Knight, and L. Sirovich (2000). On the simulation of large
populations of neurons. Journal of Computational Neuroscience 8 (1), 51 --
63.
Poulet, J.,F.A., L.M.J. Fernandez, S. Crochet, and C.H. Petersen (2012).
Thalamic control of cortical states. Nature Neuroscience 15 (3), 370 -- 372.
Protopapas, A., M. Vanier, and J. M. Bower (1998). Simulating large net-
works of neurons. See Koch and Segev (1998), pp. 461 -- 498.
Rall, W. (1977). Core conductor theory and cable properties of neurons.
In E. R. Kandel (Ed.), Handbook of Physiology - The Nervous System,
30
Cellular Biology of Neurons, Volume 1, pp. 39 -- 97. American Physiological
Society.
Rodrigues, S., A. Chizhov, F. Marten, and J. Terry (2010). Mappings be-
tween a macroscopic neural-mass model and a reduced conductance-based
model. Biological Cybernetics 102 (5), 361 -- 371.
Rudolph-Lilith, M., M. Dubois, and A. Destexhe (2012). Analytical
integrate-and-fire neuron models with conductance-based dynamics and
realistic postsynaptic potential time course for event-driven simulation
strategies. Neural Computation 34, 1426 -- 1461.
Sargsyan, A. R., C. Papatheodoropoulos, and G. K. Kostopoulos (2001).
Modeling of evoked field potentials in hippocampal CA1 area describes
their dependence on NMDA and GABA receptors. Journal of Neuroscience
Methods 104, 143 -- 153.
Schomer, D. L. and F. H. Lopes da Silva (Eds.) (2011). Niedermayer's Elec-
troencephalography. Basic Principles, Clinical Applications, and Related
Fields (6th ed.). Philadelphia: Lippincott Williams and Wilkins.
Spruston, N. (2008). Pyramidal neurons: dendritic structure and synaptic
integration. Nature Reviews Neuroscience 9, 206 -- 221.
Wang, X.-J., J. Tegn´er, C. Constantinidis, and P. S. Goldman-Rakic (2004).
Division of labor among distinct subtypes of inhibitory neurons in a cortical
microcircuit of working memory. Proceedings of the National Academy of
Sciences of the U.S.A. 101 (5), 1368 -- 1373.
Wendling, F., J. J. Bellanger, F. Bartolomei, and P. Chauvel (2000). Rele-
vance of nonlinear lumped-parameter models in the analysis of depth-EEG
epileptic signals. Biological Cybernetics 83, 367 -- 378.
Wilson, H. R. and J. D. Cowan (1972). Excitatory and inhibitory interactions
in localized populations of model neurons. Biophysical Journal 12 (1), 1 --
24.
31
|
1902.06828 | 2 | 1902 | 2019-07-30T08:33:59 | Large-scale directed network inference with multivariate transfer entropy and hierarchical statistical testing | [
"q-bio.NC",
"cs.IT",
"cs.SI",
"cs.IT",
"physics.data-an"
] | Network inference algorithms are valuable tools for the study of large-scale neuroimaging datasets. Multivariate transfer entropy is well suited for this task, being a model-free measure that captures nonlinear and lagged dependencies between time series to infer a minimal directed network model. Greedy algorithms have been proposed to efficiently deal with high-dimensional datasets while avoiding redundant inferences and capturing synergistic effects. However, multiple statistical comparisons may inflate the false positive rate and are computationally demanding, which limited the size of previous validation studies. The algorithm we present---as implemented in the IDTxl open-source software---addresses these challenges by employing hierarchical statistical tests to control the family-wise error rate and to allow for efficient parallelisation. The method was validated on synthetic datasets involving random networks of increasing size (up to 100 nodes), for both linear and nonlinear dynamics. The performance increased with the length of the time series, reaching consistently high precision, recall, and specificity (>98% on average) for 10000 time samples. Varying the statistical significance threshold showed a more favourable precision-recall trade-off for longer time series. Both the network size and the sample size are one order of magnitude larger than previously demonstrated, showing feasibility for typical EEG and MEG experiments. | q-bio.NC | q-bio | Large-scale directed network inference with multivariate transfer entropy and
hierarchical statistical testing
Leonardo Novelli,1, ∗ Patricia Wollstadt,2, ∗ Pedro Mediano,3 Michael Wibral,4 and Joseph T. Lizier1
1Centre for Complex Systems, Faculty of Engineering and IT, The University of Sydney, Sydney, Australia
2Honda Research Institute Europe, Offenbach am Main, Germany
3Computational Neurodynamics Group, Department of Computing,
Imperial College London, London, United Kingdom
4Campus Institute for Dynamics of Biological Networks, Georg-August University, Gottingen, Germany
(Dated: July 31, 2019)
Network inference algorithms are valuable tools for the study of large-scale neuroimaging datasets.
Multivariate transfer entropy is well suited for this task, being a model-free measure that captures
nonlinear and lagged dependencies between time series to infer a minimal directed network model.
Greedy algorithms have been proposed to efficiently deal with high-dimensional datasets while avoiding
redundant inferences and capturing synergistic effects. However, multiple statistical comparisons
may inflate the false positive rate and are computationally demanding, which limited the size of
previous validation studies. The algorithm we present -- as implemented in the IDTxl open-source
software -- addresses these challenges by employing hierarchical statistical tests to control the family-
wise error rate and to allow for efficient parallelisation. The method was validated on synthetic
datasets involving random networks of increasing size (up to 100 nodes), for both linear and nonlinear
dynamics. The performance increased with the length of the time series, reaching consistently high
precision, recall, and specificity (> 98% on average) for 10 000 time samples. Varying the statistical
significance threshold showed a more favourable precision-recall trade-off for longer time series. Both
the network size and the sample size are one order of magnitude larger than previously demonstrated,
showing feasibility for typical EEG and MEG experiments.
Keywords: neuroimaging, directed connectivity, effective network, multivariate transfer entropy, information
theory, nonlinear dynamics, statistical inference, nonparametric tests
9
1
0
2
l
u
J
0
3
]
.
C
N
o
i
b
-
q
[
2
v
8
2
8
6
0
.
2
0
9
1
:
v
i
X
r
a
∗ First authors contributed equally to this work.
I.
INTRODUCTION
2
The increasing availability of large-scale, fine-grained datasets provides an unprecedented opportunity for quantitative
studies of complex systems. Nonetheless, a shift towards data-driven modelling of these systems requires efficient
algorithms for analysing multivariate time series, which are obtained from observation of the activity of a large number
of elements.
In the field of neuroscience, the multivariate time series typically obtained from brain recordings serve to infer
minimal (effective) network models which can explain the dynamics of the nodes in a neural system. The motivation
for such models can be, for instance, to describe a causal network [1, 2] or to model the directed information flow in
the system [3] in order to produce a minimal computationally equivalent network [4].
Information theory [5, 6] is well suited for the latter motivation of inferring networks that describe information
flow as it provides model-free measures that can be applied at different scales and to different types of recordings.
These measures, including conditional mutual information [6] and transfer entropy [7], are based purely on probability
distributions and are able to identify nonlinear relationships [8]. Most importantly, information-theoretic measures allow
the interpretation of the results from a distributed computation or information processing perspective, by modelling
the information storage, transfer, and modification within the system [9]. Therefore, information theory simultaneously
provides the tools for building the network model and the mathematical framework for its interpretation.
The general approach to network model construction can be outlined as follows: for any target process (element) in
the system, the inference algorithm selects the minimal set of processes that collectively contribute to the computation
of the target's next state. Every process can be separately studied as a target and the results can be combined into a
directed network describing the information flows in the system.
This task presents several challenges:
• The state space of the possible network models grows faster than exponentially with respect to the size of the
network;
• Information-theoretic estimators suffer from the "curse of dimensionality" for large sets of variables [10, 11];
• In a network setting, statistical significance testing requires multiple comparisons. This results in a high false
positive rate (type I errors) without adequate family-wise error rate controls [12] or a high false negative rate
(type II errors) with naive control procedures;
• Nonparametric statistical testing based on shuffled surrogate time series is computationally demanding but
currently necessary when using general information-theoretic estimators [13, 14].
Several previous studies [4, 15 -- 17] proposed greedy algorithms to tackle the first two challenges outlined above (see a
summary by Bossomaier et al. [14, sec 7.2]). These algorithms mitigate the curse of dimensionality by greedily selecting
the random variables that iteratively reduce the uncertainty about the present state of the target. The reduction of
uncertainty is rigorously quantified by the information-theoretic measure of conditional mutual information (CMI),
which can also be interpreted as a measure of conditional independence [6]. In particular, these previous studies
employed multivariate forms of the transfer entropy, i.e., conditional and collective forms [18, 19]. In general, such greedy
optimisation algorithms provide a locally optimal solution to the NP-hard problem of selecting the most informative
set of random variables. An alternative optimisation strategy -- also based on conditional independence -- employs a
preliminary step to prune the set of sources [20, 21]. Despite this progress, the computational challenges posed by
the estimation of multivariate transfer entropy have severely limited the size of problems investigated in previous
validation studies in the general case of nonlinear estimators, e.g., Montalto et al. [22] used 5 nodes and 512 samples;
Kim et al. [23] used 6 nodes and 100 samples; Runge et al. [21] used 10 nodes and 500 samples. However, modern
neural recordings often provide hundreds of nodes and tens of thousands of samples.
These computational challenges, as well as the multiple testing challenges described above, are addressed here by the
implementation of rigorous statistical tests, which represent the main theoretical contribution of this paper. These tests
are used to control the family-wise error rate and are compatible with parallel processing, allowing the simultaneous
analysis of the targets. This is a crucial feature, which enabled an improvement on the previous greedy algorithms.
Exploiting the parallel computing capabilities of high-performance computing clusters and graphics processing units
(GPUs) enabled the analysis of networks at a relevant scale for brain recordings -- up to 100 nodes and 10 000 samples.
Our algorithm has been implemented in the recently released IDTxl Python package (the "Information Dynamics
Toolkit xl" [24]).1
1 The "Information Dynamics Toolkit xl" is an open-source Python package available on GitHub (https://github.com/pwollstadt/IDTxl).
In this paper, we refer to the current release at the time of writing (v1.0).
We validated our method on synthetic datasets involving random structural networks of increasing size (also referred
to as ground truth) and different types of dynamics (vector autoregressive processes and coupled logistic maps).
In general, effective networks are able to reflect dynamic changes in the regime of the system and do not reflect
an underlying structural network. Nonetheless, in the absence of hidden nodes (and other assumptions, including
stationarity and the causal Markov condition), the inferred information network was proven to reflect the underlying
structure for a sufficiently large sample size [17]. Experiments under these conditions provide arguably the most
important validation that the algorithm performs as expected, and here we perform the first large-scale empirical
validation for non-Gaussian variables. As shown in the Results, the performance of our algorithm increased with the
length of the time series, reaching consistently high precision, recall, and specificity (> 98% on average) for 10 000
time samples. Varying the statistical significance threshold showed a more favourable precision-recall trade-off for
longer time series.
3
II. METHODS
A. Definitions and assumptions
Let us consider a system of N discrete-time stochastic processes for which a finite number of samples have been
recorded (over time and/or in different replications of the same experiment). In general, let us assume that the
stochastic processes are stationary in each experimental time-window and Markovian with finite memory lM.2 Further
assumptions will be made for the validation study.
The following quantities are needed for the setup and formal treatment of the algorithm and are visualised in
Figure 1 and Figure 2:
target process: Y : a process of interest within the system (where Y = {Yt t ∈ N}); the choice of the target process
is arbitrary and all the processes in the system can separately be studied as targets;
source processes: Xi: the remaining processes within the system (where i = 1, . . . , N − 1 and Xi = {Xi,t t ∈ N});
sample number (or size): T : the number of samples recorded over time;
replication number: R: the number of replications of the same experiment (e.g., trials);
target present state: Yt: the random variable (RV) representing the state of the target at time t (where t ≤ T ),
whose information contributors will be inferred;
candidate target past: Y C
i.e., Y C
<t = {Yt−1, . . . , Yt−ltarget};
<t: an arbitrary finite set of RVs in the past of the target, up to a maximum lag ltarget,
candidate sources past: X C
<t: an arbitrary finite set of RVs in the past of the sources, up to a maximum lag lsources,
i.e., X C
<t = {Xi,t−1, . . . , Xi,t−lsources i = 1, . . . , N − 1};
selected target past: Y S
<t: the subset of RVs within the candidate target past set Y C
<t that maximally reduces the
uncertainty about the present state of the target;
selected sources past: X S
<t that maximally further
reduces the uncertainty about the present state of the target, in the context of the selected target past (explained
in detail in the following section).
<t: the subset of RVs within the candidate sources past set X C
B.
Inference algorithm
For a given target process Y , the goal of the algorithm is to infer the minimal set of information contributors
<t -- in the context of the relevant information contributors from the
to Yt -- defined as the selected sources past X S
candidate target past set, defined as the selected target past Y S
<t.
The algorithm operates in four steps:
2 The present state of the target does not depend on the past values of the target and the sources beyond a maximum finite lag lM.
4
FIG. 1. Example of a possible definition of the candidate sets. The bottom row represents the time series of the target process
Y , with the present state Yt highlighted in green and the candidate target past set Y C
<t highlighted in red (up to a lag ltarget).
The remaining rows represent the time series of the source processes Xi, with the candidate sources past set X C
<t highlighted in
blue (up to a lag lsources). For simplicity, only a single trial of the experiment is represented.
1. Select variables in the candidate target past set Y C
<t to obtain Y S
<t
2. Select variables in the candidate sources past set X C
<t to obtain X S
<t
3. Prune the selected sources past variables
4. Test relevant variables collectively for statistical significance
The operations performed in the four steps are described in detail hereafter; the result is a nonuniform embedding of
the target and sources time series [15, 16, 25], as illustrated in Figure 2.3
FIG. 2. Example of a resulting nonuniform embedding of the time series relevant to Yt. The bottom row represents the time
series of the target process Y , with the present state Yt highlighted in green and the selected target past set Y S
<t highlighted in
red (as a subset of the candidate target past set shown in light red). The remaining rows represent the time series of the source
processes Xi, with the selected sources past set X S
<t highlighted in blue (as a subset of the candidate sources past set shown
in light blue). The embedding only specifies the relative lags between the variables. For simplicity, only a single trial of the
experiment is shown.
3 The term embedding refers to the property of the selected set in capturing the underlying state of the process as it relates to the target's
next value, akin to a Takens' embedding [25] yet with nonuniform delays between selected points [15, 16].
t-ltargett-1tYXN-1⋮X1t-lsourcesYtY<tCX<tCt-ltargett-1tYXN-1⋮X1t-lsourcesYtY<tSX<tS 1. Step 1: Select variables in the candidate target past set
5
The goal of the first step is to find the subset of RVs within the candidate target past set Y C
<t that maximally
reduces the uncertainty about the present state of the target while meeting statistical significance requirements. Let
Y S
<t be the selected target past set found via optimisation under these criteria.
Finding the globally optimal embedding is an NP-hard problem and requires testing all the subsets of the candidate
target past set. Since the number of subsets grows exponentially with the size of the candidate set, this is computationally
unfeasible; therefore, a greedy approximation algorithm is employed to find a locally optimal solution in the search space
of possible embeddings. This approach tackles the challenge of computational complexity by aiming at identifying a
minimal conditioning set; in doing so, it also tackles the curse of dimensionality in the estimation of information-theoretic
functionals.
<t is initialised as an empty set and it is iteratively built up via the following algorithm:
The set Y S
a For each candidate variable C ∈ Y C
b Find the candidate C∗ which maximises the CMI contribution (reduction of uncertainty) and perform a statistical
significance test against the null hypothesis of conditional independence, i.e., that the new variable does not
further reduce the uncertainty in the context of the previously included variables. If significant, add C∗ to Y S
<t
and remove it from Y C
<t. The maximum statistic is employed to control the family-wise error rate (explained in
detail in the Statistical tests section);
<t, estimate the CMI contribution I(C; YtY S
<t);
c Repeat the previous steps until the maximum CMI contribution is not significant or Y C
<t is empty.
From a distributed, intrinsic computation perspective, the goal can be interpreted as finding the embedding of the
target's past states that maximises the active information storage 4 [26] to ensure self-prediction optimality as suggested
by Wibral et al. [27]. This approach is similar to the one proposed by Garland et al. [28] but uses nonuniform
embedding and additional statistical controls.
The nonuniform embedding of the time series was introduced by Vlachos and Kugiumtzis [15] and Faes et al. [16],
who used an arbitrary threshold for the conditional mutual information. Lizier and Rubinov [4] introduced a statistical
significance test to select the candidates, which this study builds upon in proposing the maximum statistic. In addition,
they embedded the target time series before embedding the sources, i.e., the active information storage is modelled
first and the information transfer is then examined in that context, thereby taking a specific modelling perspective on
the information processing carried out by the system.
2. Step 2: Select variables in the candidate sources past set
The goal of the second step is to find the subset of RVs within the candidate sources past set X C
<t that maximally
further reduces the uncertainty about the present state of the target, in the context of the selected target past found
in the first step. Let X S
<t be the selected sources past set found via optimisation under these criteria.
As for step 1, a greedy approximation algorithm is employed and the statistical significance is tested throughout the
selection process. X S
a For each candidate variable C ∈ X C
<t is initialised as an empty set and it is iteratively built up via the following algorithm:
<t, estimate the conditional transfer entropy contribution I(C; YtY S
<t)
[18, 19, 29, 30]. When X S
<t is empty, this is simply a pairwise or bivariate transfer entropy [7]; using the
conditional form serves to prevent candidates carrying only redundant information (due to, e.g., common driver
or pathway effects) from being selected, as well as to capture synergistic interactions between C and X S
<t, X S
<t.
b Find the candidate C∗ which maximises the conditional transfer entropy contribution (reduction of uncertainty)
and perform a statistical significance test against the null hypothesis of conditional independence: if significant,
add C∗ to X S
<t. The maximum statistic is employed to control the family-wise error
rate;
<t and remove it from X C
c Repeat the previous steps until the maximum conditional transfer entropy contribution is not significant or X C
<t
is empty.
From a distributed computation perspective, the goal can be interpreted as finding the nonuniform embedding of the
source processes' past that maximises the collective transfer entropy to the target, defined as I(X S
<t) [19]. As
above, the rationale for embedding the past of the sources as a second step is to achieve optimal separation of the
storage and transfer contributions [4].
<t; YtY S
4 The active information storage is defined as the mutual information between the past and the present of the target: I(Y S
<t; Yt).
3. Step 3: Prune the selected sources past variables
6
The third step of the algorithm is a pruning procedure performed to ensure that the variables included in the early
iterations of the second step still provide a statistically-significant information contribution in the context of the final
selected sources past set X S
<t. The pruning step involves the following operations:
a For each variable C ∈ X S
<t, X S
where the the set difference operation is performed to exclude the variable C from the conditioning set;
<t, estimate the conditional mutual information contribution I(C; YtY S
<t \ {C}),
b Find the variable C∗ which minimises the CMI contribution and perform a statistical significance test: if not
<t. The minimum statistic is employed to test for significance against the null
significant, remove C from X S
hypothesis of conditional independence while controlling the family-wise error rate;
c Repeat the previous steps until the minimum CMI contribution is not significant or X S
<t is empty.
The pruning step was introduced by Lizier and Rubinov [4]; remarkably, Sun et al. [17] proved that this step is essential
for the theoretical convergence of the inferred network to the causal network in the Granger-Wiener framework; they
also rigorously laid out the mathematical assumptions needed for such convergence (see Validation tasks section).
4. Step 4: Test relevant variables collectively for statistical significance
The fourth and final step of the algorithm is the computation of the collective transfer entropy from the selected
sources past set X S
<t to the target and the performance of an omnibus test to ensure statistical significance against the
null hypothesis of conditional independence. The resulting omnibus p-value can further be used for correction of the
family-wise error rate if the inference is carried out for multiple targets. The set X S
<t is only accepted as a result if all
the statistical tests are passed. Importantly, the selected sources set X S
<t, is the final
result of the algorithm for a given target process Y . The order in which variables were inferred is not relevant.
<t, inferred in the context of Y S
The statistical tests play a fundamental role in the inference and provide the stopping conditions for the iterations
involved in the first and second steps of the algorithm. These stopping conditions are adaptive and change according to
the amount of data available (the length of the time series). Given their importance, the statistical tests are described
in detail in the following section.
C. Statistical tests
The crucial steps in the inference algorithm rely on determining whether the CMI is positive. However, due to
the finite sample size, the CMI estimators may produce non-zero estimates in the case of zero CMI and it may even
return negative estimates if the estimator bias is larger than the true CMI [10, 31]. For this reason, statistical tests
are required to assess the significance of the CMI estimates against the null hypothesis of no CMI (i.e., conditional
independence) [3, 13, 32, 33].
For certain estimators, analytic solutions exist for the finite-sample distribution under this null hypothesis (see
Lizier [34]); in the absence of an analytic solution, the null distributions are computed in a nonparametric way by
using surrogate time series [35]. The surrogates are generated to satisfy the null hypothesis by destroying the temporal
relationship between the source and the target while preserving the temporal dependencies within the sources.
Finally, the inference algorithm is based on multiple comparisons and requires an appropriate calibration of the
statistical tests to achieve the desired family-wise error rate (i.e., the probability of making one or more false discoveries,
or type I errors, when performing multiple hypotheses tests). The maximum statistic and minimum statistic tests
employed in this study were specifically conceived to tackle these challenges.
The maximum statistic test is a step-down statistical test5 used to control the family-wise error rate when selecting
the past variables for the target and sources embeddings, which involves multiple comparisons.
1. Maximum statistic test
5 A test which proceeds from the smallest to the largest p-value. When the first non-significant p-value is found, all the larger p-values are
also deemed not significant.
Let us first consider the first step of the main algorithm and assume that we have picked the single candidate variable
C∗ (from the candidate target past set Y C
<t) which maximises the CMI contribution. The maximum statistic test
mirrors this selection process by picking the maximum value among the surrogates. Specifically, let I∗ := I(C∗; YtY S
<t)
be the maximum contribution (i.e., the maximum statistic); the following algorithm is used to test I∗ for statistical
significance:
7
a For each Cj ∈ Y C
j,1, . . . , C(cid:48)
j,S and compute the corresponding surrogate
CMI values I(cid:48)
<t). More details about the surrogate generation are
provided at the end of this section. The number of surrogates S must be chosen according to the desired
significance level αmax, i.e., such that S > 1/αmax.
<t, generate S surrogates time series C(cid:48)
j,S; YtY S
j,S = I(C(cid:48)
j,1 = I(C(cid:48)
j,1; YtY S
<t), . . . , I(cid:48)
b Compute the maximum CMI value over candidates I∗
Here, n denotes the number of candidates and hence the number of comparisons. The obtained values I∗
provide the (empirical) null distribution of the maximum statistic (see Table I).
n,s) for each surrogate s = 1, . . . , S.
1 , . . . , I∗
S
s := max(I(cid:48)
1,s, . . . , I(cid:48)
c Calculate the p-value for I∗ as the fraction of surrogate maximum statistic values that are larger than I∗.
d I∗ is deemed significant if the p-value is smaller than αmax (i.e., the null hypothesis of conditional independence
for the candidate variable with the maximum CMI contribution is rejected at level αmax).
The variables and quantities used in the above algorithm are presented in Table I. The key goal in the surrogate
generation is to preserve the temporal order of samples in the target time series Yt (which is not shuffled) and preserve
the distribution of the sources Cj while destroying any potential relationships between the sources and the target [3].
This can be achieved in multiple ways. If multiple replications (e.g., trials) are available, surrogate data is generated
by shuffling the order of replications for the candidate Cj while keeping the order of replications for the remaining
variables intact. When the number of replications is not sufficient to guarantee enough permutations, the embedded
source samples within individual trials are shuffled instead (see [3, 29, 32, 33] and the summary by Lizier [34, Appendix
A.5]). Note that the generation of surrogates (steps a-c) can be avoided when the null distributions can be derived
analytically, e.g., with Gaussian estimators [36].
the main algorithm), with the only difference that Cj ∈ X C
i.e. I(cid:48)
The same test is performed during the selection of the variables in the candidate sources past set (step 2 of
<t is added to the conditioning set,
<t and that X S
<t) for each surrogate s = 1, . . . , S.
j,s; YtY S
j,s = I(C(cid:48)
<t, X S
Variable
Cj ∈ Y C
CMI
<t Ij = I(Cj; YtY S
<t)
C1
C2
...
Cn
I1
I2
...
In
I∗
max CMI
···
Surrogate variables Surrogate CMI
2 ··· S
1
S
2
1
1,2 ··· I(cid:48)
1,2 ··· C(cid:48)
1,1 I(cid:48)
1,S I(cid:48)
C(cid:48)
1,1 C(cid:48)
2,2 ··· I(cid:48)
2,2 ··· C(cid:48)
2,1 I(cid:48)
2,S I(cid:48)
2,1 C(cid:48)
C(cid:48)
...
...
...
...
...
...
n,2 ··· I(cid:48)
n,2 ··· C(cid:48)
n,S I(cid:48)
n,1 I(cid:48)
C(cid:48)
n,1 C(cid:48)
I∗
I∗
I∗
···
1,S
2,S
1
2
n,S
S
S, obtained as I∗
TABLE I. Computing the null distribution of the maximum statistic. The null distribution is empirically described by the values
I∗
1 , . . . , I∗
n,s), for each surrogate s = 1, . . . , S. Here, n denotes the number of candidates and
hence the number of comparisons. The null distribution is used to test the significance of I∗ against the null hypothesis of zero
CMI.
s := max(I(cid:48)
1,s, . . . , I(cid:48)
2. Family-wise error rate correction
How does the maximum statistic test control the family-wise error rate? Intuitively, one or more statistics will
exceed a given threshold if and only if the maximum exceeds it. This relationship can be used to obtain an adjusted
threshold from the distribution of the maximum statistic under the null hypothesis, which can be used to control the
family-wise error rate both in the weak and strong sense [37].
Let us quantify the false positive rate vFPR for a single variable when the maximum statistic at the significance level
αmax is employed. For simplicity, the derivation is performed under the hypothesis that the information contributors
to the target have been selected in the first iterations of the greedy algorithm and removed from the candidate
sources past set X C
<t. Under this hypothesis, the target is conditionally independent of the remaining n variables
in X C
<t given the selected source and target variables. Let I1, . . . , In be the corresponding CMI estimates and let
Imax := max(I1, . . . , In) be the maximum statistic. As discussed above, the estimates might be positive even under
the conditional independence hypothesis, due to finite-sample effects. Since the estimates are independently obtained
from shuffled time series, they are treated as i.i.d. RVs.
Let ithreshold be the critical threshold corresponding to the given significance value αmax, i.e., ithreshold := sup{x ∈
RP (Imax ≥ x) = αmax}. Then
8
αmax = P (Imax > ithreshold)
= 1 − P (I1 ≤ ithreshold, . . . , In ≤ ithreshold)
= 1 − n(cid:89)
P (Ij ≤ ithreshold)
Therefore,
j=1
= 1 − P (I1 ≤ ithreshold)n
= 1 − (1 − vFPR)n
vFPR = 1 − (1 − αmax)1/n
(1)
(2)
Interestingly, equation (2) shows that the maximum statistic correction is equivalent to the Dunn-Sid´ak correction [38].
Performing a Taylor expansion of (2) around αmax = 0 yields:
− j−1(cid:81)
k=0
j!
∞(cid:88)
j=1
(kn − 1)
(cid:16) αtarget
(cid:17)j
n
vFPR =
Truncating the Taylor series at j = 1 yields the first-order approximation
vFPR ≈ αmax
n
,
which coincides with the false positive rate resulting from the Bonferroni correction [12]. Moreover, since the summands
in (3) are positive for every j, the Taylor series is lower-bounded by any truncated series. In particular, the false
positive rate resulting from the Bonferroni correction is a lower bound for the vFPR (the false positive rate for a single
variable resulting from the maximum statistic test), i.e., the maximum statistic correction is less stringent than the
Bonferroni correction.
Let us now study the effect of the maximum statistic test on the family-wise error rate tFPR for a single target while
accounting for all the iterations performed during the step-down test, (i.e., tFPR is the probability that at least one of
the selected sources is a false positive). We have:
n(cid:88)
n(cid:88)
j=1
tFPR =
=
P ("the source selected on step j is false positive")
(cid:19)
(cid:18) 1 − αn
max
1 − αmax
αj
max = αmax
(3)
(4)
(5)
(6)
Therefore,
j=1
tFPR ≈ αmax
for the typical small values of αmax used in statistical testing (even in the limit of large n), which shows that αmax
effectively controls the family-wise error rate for a single target.
3. Minimum statistic test
The minimum statistic test is employed during the third main step of the algorithm (pruning step) to remove the
selected variables that have become redundant in the context of the final set of selected source past variables X S
<t,
while controlling the family-wise error rate. This is necessary because of the multiple comparisons involved in the
pruning procedure. The minimum statistic test works identically to the maximum statistic test (replacing "maximum"
with "minimum" in the algorithm presented above).
4. Omnibus test
9
Let T ∗ := I(X S
<t; YtY S
<t) be the collective transfer entropy from all the selected sources past variables X S
<t to the
target Y . The value T ∗ is tested for statistical significance against the null hypothesis of zero transfer entropy (this
test is referred to as the omnibus test). The null distribution is built using surrogates time series obtained via shuffling
of the realisations of the selected sources (see [3, 29, 32, 33] and the summary by Lizier [34, Appendix A.5]), i.e., using
a similar procedure to the one described in the Maximum statistic test section above. Testing all the selected sources
collectively is in line with the perspective that the goal of the network inference is to find the set of relevant sources
for each node.
5. Combining across multiple targets
When the inference is performed on multiple targets, the omnibus p-values can be employed in further statistical tests
to control the family-wise error rate for the overall network (e.g., via FDR-correction [12, 39], which is implemented in
the IDTxl toolbox).
It is important to fully understand the statistical questions and validation procedure implied by this approach.
Combining the results across multiple targets by reusing the omnibus test p-values for the FDR-correction yields a
hierarchical test. The test answers two nested questions: (1) 'which nodes receive any significant overall information
transfer?' and, if any, (2)'what is the structure of the incoming information transfer to each node?'. However, the
answers are computed in the reverse order, for the following reason: it would be computationally unfeasible to directly
compute the collective transfer entropy from all candidate sources to the target right at the beginning of the network
inference process. At this point, the candidate source set usually contains a large number of variables so that estimation
will likely fall prey to the curse of dimensionality. Instead, a conservative approximation of the collective information
transfer is obtained by considering only a subset of the potential sources, i.e., those deemed significant by the maximum
and minimum statistic tests described in the previous sections. Only if this approximation of the total information
transfer is also deemed significant by the omnibus test (as well as by the FDR test at the network level), then the
subset of significant sources for that target is interpreted post-hoc as the local structure of the incoming information
transfer. This way, the testing procedure exhibits a hierarchical structure: the omnibus test operates at the higher
(global) level concerned with the collective information transfer, whereas the minimum and maximum tests operate at
the lower (local) level of individual source-target connections.
Compared to a non-hierarchical analysis with a correction for multiple comparisons across all links (e.g., by network-
wide Bonferroni correction or the use of the maximum statistic across all potential links), the above strategy buys
both statistical sensitivity ('recall') and the possibility to trivially parallelise computations across targets. The
price to be paid is that a link with a relatively strong information transfer into a node with non-significant overall
incoming information transfer may get pruned, while a link with relatively weaker information transfer into a node
with significant overall incoming information transfer will prevail. This behaviour clearly differs from a correction for
multiple comparisons across all links. Arguably, this difference is irrelevant in many practical cases, although it could
become noticeable for networks with high average in-degree and relatively uniform information transfer across the
links. The difference can be reduced by setting a conservative critical threshold for the lower-level greedy analysis.
D. Validation tasks
For the purpose of the validation study, the additional assumptions of causal sufficiency 6 and the causal Markov
condition 7 were made, such that the inferred network was expected to closely reflect the structural network for a
sufficiently large sample size [17]. Although this is not always the case, experiments under these conditions allow the
evaluation of the performance of the algorithm with respect to an expected ground truth. An intuitive definition of
these conditions is provided here, while the technical details are discussed at length in [40]. Moreover, the intrinsic
stochastic nature of the processes makes purely synergistic and purely redundant interactions unlikely (and indeed
vanishing for large sample size), thus satisfying the faithfulness condition [40].
The complete network inference algorithm implemented in the IDTxl toolkit (release v1.0) was validated on multiple
synthetic datasets, where both the structural connectivity and the dynamics were known. Given the general scope of the
6 Causal sufficiency: The set of observed variables includes all their common causes (or the unobserved common causes have constant
values).
7 Causal Markov condition: A variable X is independent of every other past variable conditional on all of its direct causes.
10
toolkit, two dynamical models of broad applicability were chosen: a vector autoregressive process (VAR) and a coupled
logistic maps process (CLM); both models are widely used in computational neuroscience [41 -- 43], macroeconomics
[44, 45], and chaos theory [46].
The primary goal was to quantify the scaling of the performance with respect to the size of the network and the
length of the time series. Sparse directed random Erdos-R´enyi networks [47] of increasing size (N = 10 to 100 nodes)
were generated with a link probability p = 3/N to obtain an expected in-degree of 3 links. Both the VAR and the CLM
stochastic processes were repeatedly simulated on each causal network with increasingly longer time series (T = 100 to
10 000 samples), a single replication (or trial, i.e., R = 1), and with 10 random initial conditions. The performance was
evaluated in terms of precision, recall, and specificity in the classification of the links. Further simulations were carried
out to investigate the influence of the critical alpha level for statistical significance and the performance of different
estimators of conditional mutual information.
1. Vector autoregressive process
The specific VAR process used in this study is described by the following discrete-time recurrence relation:
Yt = βYt−1 +
αX Xt−lX + ηt
(7)
(cid:80)
where XY denotes the set of causal sources of the target process Y and a single random lag lX ∈ {1, 2, 3, 4, 5} was
used for each source X ∈ XY . A Gaussian noise term ηt with mean µ = 0 and standard deviation θ = 0.1 was
added at each time step t; the noise terms added to different variables were uncorrelated. The self-coupling coefficient
was set to β = 0.5 and the cross-coupling coefficients αX were uniform and normalised for each target such that
αX = 0.4. This choice of parameters guaranteed that the VAR processes were stable (the resulting spectral
radii were between 0.9 and 0.95) and had stationary multivariate Gaussian distributions [48]. As such, the Gaussian
estimator implemented in IDTxl was employed for transfer entropy measurements in VAR processes. Note that transfer
entropy and Granger causality [49] are equivalent for Gaussian variables [50]; therefore, using the Gaussian estimator
with our algorithm can be viewed as extending Granger causality in the same multivariate/greedy fashion.
X∈XY
2. Coupled logistic maps process
The coupled logistic maps process used in this study is described by the following discrete-time recurrence relations:
(cid:88)
X∈XY
(cid:88)
at = βYt−1 +
αX Xt−lX
Yt = (4at(1 − at) + ηt) mod 1
X∈XY
(8)
At each time step t, each node Y computes the weighted input at as a linear combination of its past value and the past
of its sources, with the same conditions used for the VAR process on the choice of the random lags lX and coupling
coefficients β and αX . The value Yt is then computed by applying the logistic map activation function f (x) = 4x(1− x)
to the weighted input at and adding the Gaussian noise ηt with the same properties used for the VAR process. Notice
that the coefficient (r = 4) used in the logistic map function corresponds to the fully-developed chaotic regime. The
modulo-1 operation ensures that Yt ∈ [0, 1] after the addition of noise. The nearest-neighbour estimators were employed
for transfer entropy measurements in the analysis of the CLM processes (in particular, Kraskov's estimator I (1) with
k = 4 nearest neighbours [31] and its extension to CMI [51 -- 53]). Nearest-neighbour estimators are model-free and are
able to detect nonlinear dependencies in stochastic processes with non-Gaussian stationary distributions; fast CPU
and GPU implementations are provided by the IDTxl package.
III. RESULTS
A.
Influence of network size and length of the time series
The aim of the first analysis was to quantify the scaling of the performance with respect to the size of the network
and the length of the time series.
11
The inferred network was built by adding a directed link from a source node X to a target node Y whenever a
significant transfer entropy from X to Y was measured while building the selected sources past set X S
<t (i.e., whenever
X ∩ X S
<t (cid:54)= ∅). The critical alpha level for statistical significance was set to αmax = 0.001 and S = 1000 surrogates
were used for all experiments unless otherwise stated. The candidate sets for the target as well as the sources were
initialised with a maximum lag of five (i.e., ltarget = lsources = 5, corresponding to the largest lag values used in the
definition of the VAR and CLM processes).
The network inference performance was evaluated in comparison to the known underlying structural network as a
binary classification task, using standard statistics based on the number of true positives (TP, i.e., correctly classified
existing links), false positives (FP, i.e., absent links falsely classified as existing), true negatives (TN, i.e., correctly
classified absent links), and false negatives (FN, i.e., existing links falsely classified as absent). The following standard
statistics were employed in the evaluation:
precision: = T P/(T P + F P )
recall: = T P/(T P + F N )
specificity: = T N/(T N + F P )
The plots in Figure 3 summarise the results in terms of precision and recall, while the specificity is additionally plotted
in the Supplementary materials. For both types of dynamics, the performance increased with the number of samples
and decreased with the size of the network.
For shorter time series (T = 100 and T = 1000), the recall was the most affected performance measure as a function
of N and T , while the precision and the specificity were always close to optimal (> 98% on average). (Note that,
while S = 1000 is minimal for αmax = 0.001, recall was unchanged using S = 10 000 for T = 100). For longer time
series (T = 10 000), high performance according to all measures was achieved for both the VAR and CLM processes,
regardless of the size of the network. The high precision and specificity are due to the effective control of the false
positives, in accordance with the strict statistical significance level αmax = 0.001 (the influence of αmax is further
discussed in the following sections). The inference algorithm was therefore conservative in the classification of the links.
VAR
CLM
FIG. 3. Precision (top) and recall (bottom) for different network sizes, sample sizes, and dynamics. Left: Vector autoregressive
process; Right: Coupled logistic maps. Each subplot shows five curves, corresponding to different time series lengths (T = 100,
300, 1000, 3000, 10 000). The results for 10 simulations from different initial conditions are shown (low-opacity markers) in
addition to the mean values (solid markers). All the random networks have an average in-degree N p = 3.
20406080100Networksize0.00.20.40.60.81.0PrecisionT=100T=300T=1000T=3000T=100001040701000.900.951.0020406080100Networksize0.00.20.40.60.81.0PrecisionT=100T=300T=1000T=3000T=100001040701000.900.951.0020406080100Networksize0.00.20.40.60.81.0RecallT=100T=300T=1000T=3000T=1000020406080100Networksize0.00.20.40.60.81.0RecallT=100T=300T=1000T=3000T=10000 B. Validation of false positive rate
12
The critical alpha level for statistical significance αmax is a parameter of the algorithm which is designed to control
the number o false positives in the network inference. As discussed in the Statistical tests section in the Methods, αmax
controls the probability that a target is a false positive, i.e., that at least one of its sources is a false positive. This
approach is in line with the perspective that the goal of the network inference is to find the set of relevant sources for
each node.
A validation study was carried out to verify that the final number of false positives is consistent with the desired
level αmax after multiple statistical tests are performed. The false positive rate (i.e. F P/(F P + T N )) was computed
after performing the inference on empty networks, where every inferred link is a false positive by definition (i.e., under
the complete null hypothesis). The rate was in good accordance with the critical alpha threshold αmax for all network
sizes, as shown in Figure 4.
The false positive rate validation was replicated in a scenario where the null hypothesis held for real fMRI data from
the Human Connectome Project resting-state dataset (see Supporting Information). The findings are presented in the
Supplementary Material, together with a note on autocorrelation. Notably, the results on fMRI data are in agreement
with the results on synthetic data shown in Figure 4.
FIG. 4. Validation of false positive rate for a single target (tFPR) on empty networks. The points indicate the average false
positive rate over 50 simulations of a vector autoregressive process (T = 10 000). The horizontal marks indicate the corresponding
5th and 95th percentiles of the expected range. These were computed empirically from the distribution of the random variable
(cid:104)Xj/N(cid:105), where Xj ∼ Binomial(N, αmax) are i.i.d. random variables, and the angular brackets indicate the finite average over
50 repetitions. The 5th percentile for N = 10 and N = 40 and αmax = 10−3 are equal to zero and therefore omitted from the
log-log plot. The identity function is plotted as a reference (dashed line).
C.
Influence of critical level for statistical significance
Given the conservative results obtained for both the VAR and CLM processes (Figure 3), a natural question is to
what extent the recall could be improved by increasing the critical alpha level αmax and to what extent the precision
would be negatively affected as a side effect.
In order to elucidate this trade-off, the analysis described above (Figure 3) was repeated for increasing values of
αmax, with results shown in Figure 5. For the shortest time series (T = 100), increasing αmax resulted in a higher
recall and a lower precision, as expected; on the other hand, for the longest time series (T = 10 000), the performance
measures were not significantly affected. Interestingly, for the intermediate case (T = 1000), increasing αmax resulted
in higher recall without negatively affecting the precision.
103102max103102FalsepositiverateN=10N=40N=70N=100Identity 13
VAR
CLM
FIG. 5. Influence of statistical significance threshold on network inference performance. Precision vs. recall for different
statistical significance levels (αmax = 0.05, 0.01, 0.001), corresponding to different colours. The plots summarise the results for
different dynamics (Top: Vector autoregressive process; Bottom: Coupled logistic maps), different time series lengths (T = 100,
1000, 10 000), and different network sizes (N = 10, 40, 70, 100, not distinguished). The arrows join the mean population values
for the lowest and highest significance levels, illustrating the average trade-off between precision loss and recall gain.
D.
Inference of coupling lags
So far, the performance evaluation focused on the identification the correct set of sources for each target node,
regardless of the coupling lags. However, since the identification of the correct coupling lags is particularly relevant in
neuroscience (see Wibral et al. [27] and references therein), the performance of the algorithm in identifying the correct
coupling lags was additionally investigated.
By construction, a single coupling lag was imposed between each pair of processes (chosen at random between
0.00.20.40.60.81.0Recall0.60.81.0PrecisionT=100T=1000T=10000max=0.001max=0.01max=0.050.00.20.40.60.81.0Recall0.50.60.70.80.91.0PrecisionT=100T=1000T=10000max=0.001max=0.01max=0.05 one and five discrete time steps, as described in the Methods). The average absolute error between the real and the
inferred coupling lags was computed on the correctly recalled sources and divided by the value expected at random
(which is the average absolute difference between two i.i.d. random integers in the [1, 5] interval). In line with the
previous results on precision, the absolute error on coupling lag is consistently much smaller than that expected at
random, even for the shortest time series (Figure 6). Furthermore, 1000 samples were sufficient to achieve nearly
optimal performance for both the VAR and the CLM processes, regardless of the size of the network. Note that as
T increases and the recall increases, the lag error can increase (c.f. T = 100 to 300 for the CLM process). This is
perhaps because while the larger T permits more weakly contributing sources to be identified, it is not large enough to
reduce the estimation error to make lag identification on these sources precise.
VAR
CLM
14
FIG. 6. Average absolute error between the real and the inferred coupling lags, relative to the value expected at random. Results
for different dynamics (Left: Vector autoregressive process; Right: Coupled logistic maps), different time series lengths (T = 100,
300, 1000, 3000, 10 000), and different network sizes (N = 10, 40, 70, 100). The error bars indicate the standard deviation over
10 simulations from different initial conditions.
E. Estimators
Given its speed, the Gaussian estimator is often used for large datasets or as a first exploratory step, even when the
stationary distribution cannot be assumed to be Gaussian. The availability of the ground truth allowed to compare
the performance of the Gaussian estimator and the nearest-neighbour estimator on the nonlinear CLM process, which
does not satisfy the Gaussian assumption. As expected, the performance of the Gaussian estimator was lower than the
performance of the nearest-neighbour estimator for all network sizes (Figure 7).
FIG. 7. Gaussian vs. nearest-neighbour estimator on the coupled logistic maps process. The precision (left) and recall (right) are
plotted against the network size and a fixed time series length (T = 10 000 samples). The results for 10 simulations from different
initial conditions are shown (low-opacity markers) in addition to the mean values (solid markers). The statistical significance
level αmax = 0.05 was employed; an even larger gap between the recall of the estimators is obtained with αmax = 0.001.
The hierarchical tests introduced in the Methods section allow running the network inference algorithm in parallel on
a high-performance computing cluster. Such parallelisation is especially needed when employing the nearest-neighbour
estimator. In particular, each target node can be analysed in parallel on a CPU (employing one or more cores) or
a GPU, which is made possible by the CPU and GPU estimators provided by the IDTxl package (custom OpenCL
102103104T0.00.10.20.3LagerrorN=10N=40N=70N=100102103104T0.00.10.20.3LagerrorN=10N=40N=70N=10020406080100Networksize0.00.51.0PrecisionNearest-neighbour estimatorGaussian estimator1040701000.940.971.0020406080100Networksize0.00.51.0RecallNearest-neighbour estimatorGaussian estimator kernels were written for the GPU implementation). A summary of the CPU and GPU run-times is provided in the
Supplementary materials.
15
IV. DISCUSSION
The algorithm presented in this paper provides robust statistical tests for network inference to control the false
positive rate. These tests are compatible with parallel computation on high-performance computing clusters, which
enabled the validation study on synthetic sparse networks of increasing size (10 to 100 nodes), using different dynamics
(linear autoregressive processes and nonlinear coupled logistic maps) and increasingly longer time series (100 to 10 000
samples). Both the network size and the sample size are one order of magnitude larger than previously demonstrated,
showing feasibility for typical EEG and MEG experiments. The results demonstrate that the statistical tests achieve
the desired false positive rate and successfully address the multiple-comparison problems inherent in network inference
tasks (Figure 4).
The ability to control the false positives while building connectomes is a crucial prerequisite for the application
of complex network measures, to the extent that [54] concluded that "specificity is at least twice as important as
sensitivity [i.e., recall] when estimating key properties of brain networks, including topological measures of network
clustering, network efficiency and network modularity". The reason is that false positives occur more prevalently
between network modules than within them and the spurious inter-modular connections have a dramatic impact on
network topology [54].
The trade-off between precision and recall when relaxing the statistical significance threshold was further investigated
(Figure 5). When only 100 samples were used, the average recall gain was more than five times smaller than the
average precision loss. In our opinion, this result is possibly due to the sparsity of the networks used in this study and
suggests a conservative choice of the threshold for sparse networks and short time series. The trade-off was reversed
for longer time series: when 1000 samples were used, the average recall gain was more than five times larger than the
average precision loss. Finally, for 10 000 samples, high precision and recall were achieved (> 98% on average) for both
the vector autoregressive and the coupled logistic maps processes, regardless of the statistical significance threshold.
For both types of dynamics, the network inference performance increased with the length of the time series and
decreased with the size of the network (Figure 3). This is to be expected since larger systems require more statistical
tests and hence stricter conditions to control the family-wise error rate (false positives). Specifically, larger networks
result in wider null distributions of the maximum statistic (i.e., larger variance), whereas longer time series have the
opposite effect. Therefore, for large networks and short time series, controlling the false positives can have a negative
impact on the ability to identify the true positives, particularly when the effect size (i.e., the transfer entropy value) is
small.
In addition, the superior ability of the nearest-neighbour estimator over the Gaussian estimator in detecting nonlinear
dependencies was quantified. There is a critical motivation for this comparison: the general applicability of the
nearest-neighbour estimators comes at the price of higher computational complexity and a significantly longer run-time,
so that the Gaussian estimator is often used for large datasets (or at least as a first exploratory step), even when the
Gaussian hypothesis is not justified. To investigate such scenario, the Gaussian estimator was tested on the nonlinear
logistic map processes: while the resulting recall was significantly lower than the nearest-neighbour estimator for all
network sizes, it was nonetheless able to identify over half of the links for a sufficiently large number (10 000) of time
samples (Figure 7).
The stationarity assumption about the time series corresponds to assuming a single regime of neuronal activity in
real brain recordings. If multiple regimes are recorded, which is typical in experimental settings (e.g., sequences of
tasks or repeated presentation of stimuli interleaved with resting time windows), different stationary regimes can be
studied by performing the analysis within each time window. The networks obtained in different time windows can
either be studied separately and compared against each other or collectively interpreted as a single evolving temporal
network. To obtain a sufficient amount of observations per window, multiple replications of the experiment under
the same conditions are typically carried out. Replications can be assumed to be cyclo-stationary and estimation
techniques exploiting this property have been proposed [53, 55]; these estimators are also available in the IDTxl Python
package. The convergence to the (unknown) causal network was only proven under the hypotheses of stationarity,
causal sufficiency, and the causal Markov condition [17]. However, conditional independence holds under milder
assumptions [56] and the absence of links is valid under general conditions. The conditional independence relationships
can, therefore, be used to exclude variables in following intervention-based causal experiments, making network
inference methods valuable for exploratory studies.
In fact, the directed network is only one part of the model and provides the scaffold over which the information-
theoretic measures are computed. Therefore, even if the structure of a system is known and there is no need for
network inference, information theory can still provide nontrivial insights on the distributed computation by modelling
16
the information storage, transfer, and modification within the system [9]. This decomposition of the predictive
information into the active information storage and transfer entropy components is one out of many alternatives within
the framework proposed by Chicharro and Ledberg [57]. Arguably, the storage-transfer decomposition reflects the
segregation-integration dichotomy that has long characterised the interpretation of brain function [58, 59]. Information
theory has the potential to provide a quantitative definition of these fundamental but still unsettled concepts [60].
In addition, information theory provides a new way of testing fundamental computational theories in neuroscience,
e.g., predictive coding [61].
As such, information-theoretic methods should not be seen as opposed to model-based approaches, but complementary
to them [62]. If certain physically motivated parametric models are assumed, the two approaches are equivalent for
network inference: maximising the log-likelihood is asymptotically equivalent to maximising the transfer entropy
[36, 63]. Moreover, different approaches can be combined, e.g., the recent large-scale application of spectral DCM
was made possible by using functional connectivity models to place prior constraints on the parameter space [64].
Networks inferred using bivariate transfer entropy have also been employed to reduce the model space prior to DCM
analysis [65].
In conclusion, the continuous evolution and combination of methods show that network inference from time series is
an active field of research and there is a current trend of larger validation studies, statistical significance improvements,
and reduction of computational complexity. Information-theoretic approaches require efficient tools to employ nearest-
neighbour estimators on large datasets of continuous-valued time series, which are ubiquitous in large-scale brain
recordings (calcium imaging, EEG, MEG, fMRI). The algorithm presented in this paper is compatible with parallel
computation on high-performance computing clusters, which enabled the study of synthetic nonlinear systems of
100 nodes and 10 000 samples. Both the network size and the sample size are one order of magnitude larger than
previously demonstrated, bringing typical EEG and MEG experiments into scope for future information-theoretic
network inference studies. Furthermore, the statistical tests presented in the Methods are generic and compatible with
any underlying conditional mutual information or transfer entropy estimators, meaning that estimators applicable to
spike trains [66] can be used with this algorithm in future studies.
V. SUPPORTING INFORMATION
The network inference algorithm described in this paper is implemented in the open-source Python software package
IDTxl [24], which is freely available on GitHub (https://github.com/pwollstadt/IDTxl). In this paper, we refer to
the current release (v1.0) at the time of writing (doi:10.5281/zenodo.2554339).
The raw data used for the experiment presented in the Supplementary Material is openly available on the MGH-USC
Human Connectome Project database (https://ida.loni.usc.edu/login.jsp).
VI. ACKNOWLEDGEMENTS
This research was supported by: Universities Australia/German Academic Exchange Service (DAAD) Australia-
Germany Joint Research Cooperation Scheme grant: "Measuring neural information synthesis and its impairment",
Grant/Award Number: PPP Australia Projekt-ID 57216857; Deutsche Forschungsgemeinschaft (DFG) Grant CRC
1193 C04; and Australian Research Council DECRA grant DE160100630 and Discovery grant DP160102742.
Data collection and sharing for the experiment presented in the Supplementary material was provided by the
MGH-USC Human Connectome Project (HCP; Principal Investigators: Bruce Rosen, M.D., Ph.D., Arthur W. Toga,
Ph.D., Van J. Weeden, MD). HCP funding was provided by the National Institute of Dental and Craniofacial Research
(NIDCR), the National Institute of Mental Health (NIMH), and the National Institute of Neurological Disorders and
Stroke (NINDS). HCP data are disseminated by the Laboratory of Neuro Imaging at the University of Southern
California.
The authors acknowledge the Sydney Informatics Hub and the University of Sydney's high-performance computing
cluster Artemis for providing the high-performance computing resources that have contributed to the research results
reported within this paper. Furthermore, the authors thank Aaron J. Gutknecht for commenting on a draft of this
paper, and Oliver Cliff for useful discussions and comments.
VII. AUTHOR CONTRIBUTIONS
Leonardo Novelli: Conceptualization; Data curation; Formal analysis; Investigation; Software; Validation; Visu-
alization; Writing -- original draft. Patricia Wollstadt: Conceptualization; Software; Writing -- review & editing.
Pedro Mediano: Software; Writing -- review & editing. Michael Wibral: Conceptualization; Funding acquisition;
Methodology; Software; Supervision; Writing -- review & editing. Joseph T. Lizier: Conceptualization; Funding
acquisition; Methodology; Software; Supervision; Writing -- review & editing.
17
[1] K. J. Friston, Human Brain Mapping 2, 56 (1994).
[2] N. Ay and D. Polani, Advances in Complex Systems 11, 17 (2008).
[3] R. Vicente, M. Wibral, M. Lindner, and G. Pipa, Journal of Computational Neuroscience 30, 45 (2011).
[4] J. T. Lizier and M. Rubinov, Multivariate construction of effective computational networks from observational data, Tech.
Rep. Preprint 25/2012 (Max Planck Institute for Mathematics in the Sciences, 2012).
[5] C. E. Shannon, Bell System Technical Journal 27, 379 (1948).
[6] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. (John Wiley & Sons, Inc., Hoboken, NJ, USA,
2005) p. 748.
[7] T. Schreiber, Physical Review Letters 85, 461 (2000).
[8] M. Palus, V. Albrecht, and I. Dvo´ak, Physics Letters A 175, 203 (1993).
[9] J. T. Lizier, The Local Information Dynamics of Distributed Computation in Complex Systems, Springer Theses (Springer
Berlin Heidelberg, Berlin, Heidelberg, 2013) p. 311.
[10] M. S. Roulston, Physica D: Nonlinear Phenomena 125, 285 (1999).
[11] L. Paninski, Neural Computation 15, 1191 (2003).
[12] T. Dickhaus, Simultaneous Statistical Inference (Springer Berlin Heidelberg, Berlin, Heidelberg, 2014) p. 180.
[13] M. Lindner, R. Vicente, V. Priesemann, and M. Wibral, BMC Neuroscience 12, 119 (2011).
[14] T. Bossomaier, L. Barnett, M. Harr´e, and J. T. Lizier, An Introduction to Transfer Entropy (Springer International
Publishing, Cham, 2016) p. 190.
[15] I. Vlachos and D. Kugiumtzis, Physical Review E 82, 016207 (2010).
[16] L. Faes, G. Nollo, and A. Porta, Physical Review E 83, 051112 (2011).
[17] J. Sun, D. Taylor, and E. M. Bollt, SIAM Journal on Applied Dynamical Systems 14, 73 (2015).
[18] J. T. Lizier, M. Prokopenko, and A. Y. Zomaya, Physical Review E 77, 026110 (2008).
[19] J. T. Lizier, M. Prokopenko, and A. Y. Zomaya, Chaos 20, 037109 (2010).
[20] J. Runge, J. Heitzig, V. Petoukhov, and J. Kurths, Physical Review Letters 108, 258701 (2012).
[21] J. Runge, P. Nowack, M. Kretschmer, S. Flaxman, and D. Sejdinovic, arXiv Preprint (2018), arXiv: 1702.07007.
[22] A. Montalto, L. Faes, and D. Marinazzo, PLoS ONE 9, e109462 (2014).
[23] P. Kim, J. Rogers, J. Sun, and E. M. Bollt, Journal of Computational and Nonlinear Dynamics 12, 011008 (2016).
[24] P. Wollstadt, J. T. Lizier, R. Vicente, C. Finn, M. Mart´ınez-Zarzuela, P. Mediano, L. Novelli, and M. Wibral, Journal of
Open Source Software 4, 1081 (2019).
[25] F. Takens, in Dynamical Systems and Turbulence, edited by D. Rand and L. Young (Springer Berlin Heidelberg, 1981) pp.
366 -- 381.
[26] J. T. Lizier, M. Prokopenko, and A. Y. Zomaya, Information Sciences 208, 39 (2012).
[27] M. Wibral, N. Pampu, V. Priesemann, F. Siebenhuhner, H. Seiwert, M. Lindner, J. T. Lizier, and R. Vicente, PLoS ONE
8, e55809 (2013).
[28] J. Garland, R. G. James, and E. Bradley, Physical Review E 93, 022221 (2016).
[29] P. F. Verdes, Physical Review E 72, 026222 (2005).
[30] V. A. Vakorin, O. A. Krakovska, and A. R. McIntosh, Journal of Neuroscience Methods 184, 152 (2009).
[31] A. Kraskov, H. Stogbauer, and P. Grassberger, Physical Review E 69, 066138 (2004).
[32] M. Ch´avez, J. Martinerie, and M. Le Van Quyen, Journal of Neuroscience Methods 124, 113 (2003).
[33] J. T. Lizier, J. Heinzle, A. Horstmann, J.-D. Haynes, and M. Prokopenko, Journal of Computational Neuroscience 30, 85
(2011).
[34] J. T. Lizier, Frontiers in Robotics and AI 1, 11 (2014).
[35] T. Schreiber and A. Schmitz, Physica D: Nonlinear Phenomena 142, 346 (2000).
[36] L. Barnett and T. Bossomaier, Physical Review Letters 109, 138105 (2012).
[37] T. Nichols and S. Hayasaka, Statistical Methods in Medical Research 12, 419 (2003).
[38] Z. Sid´ak, Journal of the American Statistical Association 62, 626 (1967).
[39] Y. Benjamini and Y. Hochberg, Journal of the Royal Statistical Society. Series B (Methodological) 57, 289 (1995).
[40] P. Spirtes, C. Glymour, and R. Scheines, Causation, Prediction, and Search, Lecture Notes in Statistics, Vol. 81 (Springer
New York, 1993).
[41] A. Zalesky, A. Fornito, L. Cocchi, L. L. Gollo, and M. Breakspear, Proceedings of the National Academy of Sciences 111,
10341 (2014).
[42] M. Rubinov, O. Sporns, C. van Leeuwen, and M. Breakspear, BMC Neuroscience 10, 55 (2009).
[43] P. A. Valdes-Sosa, A. Roebroeck, J. Daunizeau, and K. J. Friston, NeuroImage 58, 339 (2011).
[44] C. A. Sims, Econometrica 48, 1 (1980).
[45] H.-W. Lorenz, in Nonlinear Dynamical Economics and Chaotic Motion (Springer Berlin Heidelberg, 1993) pp. 119 -- 166.
[46] S. H. Strogatz, Nonlinear Dynamics and Chaos (CRC Press, Boca Raton, 2015) p. 531.
18
[47] P. Erdos and A. R´enyi, Publicationes Mathematicae Debrecen 6, 290 (1959).
[48] F. M. Atay and O. Karabacak, SIAM Journal on Applied Dynamical Systems 5, 508 (2006).
[49] C. W. J. Granger, Econometrica 37, 424 (1969).
[50] L. Barnett, A. B. Barrett, and A. K. Seth, Physical Review Letters 103, 238701 (2009).
[51] S. Frenzel and B. Pompe, Physical Review Letters 99, 204101 (2007).
[52] M. Vejmelka and M. Palus, Physical Review E 77, 026214 (2008).
[53] G. G´omez-Herrero, W. Wu, K. Rutanen, M. Soriano, G. Pipa, and R. Vicente, Entropy 17, 1958 (2015).
[54] A. Zalesky, A. Fornito, L. Cocchi, L. L. Gollo, M. P. van den Heuvel, and M. Breakspear, NeuroImage 142, 407 (2016).
[55] P. Wollstadt, M. Mart´ınez-Zarzuela, R. Vicente, F. J. D´ıaz-Pernas, and M. Wibral, PLoS ONE 9, e102833 (2014).
[56] J. Runge, Chaos 28, 075310 (2018).
[57] D. Chicharro and A. Ledberg, Physical Review E 86, 041901 (2012).
[58] S. Zeki and S. Shipp, Nature 335, 311 (1988).
[59] O. Sporns, Networks of the Brain (MIT Press, 2010) p. 424.
[60] M. Li, Y. Han, M. J. Aburn, M. Breakspear, R. A. Poldrack, J. M. Shine, and J. T. Lizier, bioRxiv Preprint , 581538
(2019).
[61] A. Brodski-Guerniero, G.-F. Paasch, P. Wollstadt, I. Ozdemir, J. T. Lizier, and M. Wibral, The Journal of Neuroscience
37, 8273 (2017).
[62] K. J. Friston, R. Moran, and A. K. Seth, Current Opinion in Neurobiology 23, 172 (2013).
[63] O. Cliff, M. Prokopenko, and R. Fitch, Entropy 20, 51 (2018).
[64] A. Razi, M. L. Seghier, Y. Zhou, P. McColgan, P. Zeidman, H.-J. Park, O. Sporns, G. Rees, and K. J. Friston, Network
Neuroscience 1, 222 (2017).
[65] J. S. Chan, M. Wibral, P. Wollstadt, C. Stawowsky, M. Brandl, S. Helbling, M. Naumer, and J. Kaiser, bioRxiv Preprint ,
178095 (2017).
[66] R. E. Spinney, M. Prokopenko, and J. T. Lizier, Physical Review E 95, 032319 (2017).
[67] D. Van Essen, K. Ugurbil, E. Auerbach, D. Barch, T. Behrens, R. Bucholz, A. Chang, L. Chen, M. Corbetta, S. Curtiss,
S. Della Penna, D. Feinberg, M. Glasser, N. Harel, A. Heath, L. Larson-Prior, D. Marcus, G. Michalareas, S. Moeller,
R. Oostenveld, S. Petersen, F. Prior, B. Schlaggar, S. Smith, A. Snyder, J. Xu, and E. Yacoub, NeuroImage 62, 2222
(2012).
[68] L. Barnett and A. K. Seth, Journal of Neuroscience Methods 201, 404 (2011).
[69] J. Theiler, Physical Review A 34, 2427 (1986).
[70] H. Kantz and T. Schreiber, Nonlinear Time Series Analysis, 2nd ed. (Cambridge University Press, 2003).
VIII. SUPPORTING INFORMATION
A. Run-time on CPU and GPU
19
The number of transfer entropy calculations scales as O(N 2dlmaxS), where N is the number of processes, d is the
average inferred in-degree, lmax is the maximum temporal search depth per process (i.e., lmax = max{ltarget, lsources}),
and S is the number of surrogates. This assumes that d is independent of the network size N ; however, note that
d = N in the worst case of a fully connected network, leading to cubic run-times.
The network inference algorithm was designed for parallelisation:
• When using the CPU estimators, it is possible to parallelise over targets, resulting in O(N dlmaxS) transfer
entropy calculations per target for the nearest-neighbour estimator and O(N dlmax) transfer entropy calculations
for the Gaussian estimator (if using analytic null distributions instead of surrogates). The complexity of each
calculation is O(kT logT ) for the nearest-neighbour estimator and O(T ) for the Gaussian estimator (where T is
number of time series samples and k is the number of nearest-neighbours).
• When using the GPU estimators, it is possible to parallelise both over targets and surrogates. Each target
requires O(N dlmax) transfer entropy calculations including surrogates, assuming that all surrogates for a source
fit into the GPU's main memory and can be processed in parallel. The complexity of each calculation is O(T 2). If
enough memory is available, it is further possible to parallelise over time samples T , resulting in faster run-times
in practice.
The practical run-time for a full network analysis that considers each process as a target depends on the number of
available computing nodes. In the worst case, where only a single computing node is available, the full run-time is
equal to the single-target run-time multiplied by N , since the target are analysed in series. In the best case, if N
computing nodes are available, the full run-time is equal to the single-target run-time, since all targets can be analysed
in parallel. Notice that there is a trade-off between run-time and memory requirements: if all the targets are analysed
in parallel, the full required memory in N times larger than the memory required in the single-node case; conversely, if
the targets are analysed in series, the full required memory is equal to the memory required in the single-node case.
In the experiments presented in this article, the algorithm was either run using a single core per target (on different
Intel Xeon CPUs with similar characteristics: 2.1-2.6 GHz), or using a whole dedicated GPU per target (NVIDIA
V100 SXM2, 16 GB RAM). These computations were performed on the Artemis computing cluster made available by
the Sydney Informatics Hub at The University of Sydney. The maximum CPU and GPU run-times for a single target
are shown in Table II, which summarises the results for different time series lengths and different network sizes. Notice
that the CPU run-time per target can be reduced if multiple cores per target are available.
Sample size Network size max GPU time max CPU time max CPU memory
per target
per target
per target
T
100
1000
10 000
N
10
40
70
100
10
40
70
100
10
40
70
100
(h)
0.005
0.01
0.02
0.03
0.05
0.20
0.25
0.45
0.85
3.30
5.00
6.70
(h)
0.01
0.05
0.10
0.15
0.63
2.30
4.30
5.50
13.00
120.00
250.00
335.00
(GB)
<0.50
<0.50
<0.50
0.63
0.65
0.70
0.70
0.70
5.50
6.60
7.40
8.00
TABLE II. Maximum CPU and GPU run-time for a single target using the nearest-neighbour estimator and 200 surrogates.
Summary of the results for different time series lengths (T = 100, 1000, 10 000) and different network sizes (N = 10, 40, 70,
100).
20
B. Validation of false positive rate on real fMRI data
The false positive rate validation (presented in Figure 4 for synthetic VAR data) was replicated in a scenario where
the null hypothesis held for real data. Once again, the aim was to verify that the false positive rate was consistent with
the desired level αmax. The Human Connectome Project resting state fMRI dataset [67] was used for this purpose
(see Supporting Information). The raw data was pre-processed by applying a 3rd order Butterworth bandpass filter
(0.01-0.08 Hz), then cutting 200 samples from the start and the end of the time series to remove potential filtering
artefacts (leaving 800 samples for the analysis). In order to build a scenario where the null hypothesis held, 10 different
random regions of interest (ROIs) were selected from different random subjects, such that the corresponding time
series were expected to be independent of each other. The network inference was performed with the same settings
used in the null test on synthetic data but employing the nearest-neighbour estimator, since the real data could not be
assumed to follow a Gaussian distribution. The results on fMRI data are presented in Figure 8 and are consistent with
the previous results on synthetic data (Figure 4).
Unless appropriate measures are taken, the strong autocorrelation typically found in real data would result in an
inflated false positive rate for short time series (an effect already observed by Barnett and Seth [68] when using Granger
causality). The issue is addressed in IDTxl by means of the dynamic correlation exclusion, also known as Theiler
window [69, 70], as originally suggested for transfer entropy estimation by Schreiber [7]. The idea is to exclude the
closest points in time from the nearest-neighbour search which is necessary for the estimation of the transfer entropy
(when using nearest-neighbour estimators). The autocorrelation decay time (i.e., the shortest time shift such that the
autocorrelation function drops by a factor of 1/e with respect to the zero-shift value [13]) is used as a heuristic to
adapt the size of the Theiler window to the data.
FIG. 8. Validation of false positive rate for a single target (tFPR) on real fMRI data. The points indicate the average false
positive rate over 50 repetitions of the experiment using random regions (N = 10) from different subjects. The horizontal
marks indicate the corresponding 5th and 95th percentiles of the expected range. These were computed empirically from the
distribution of the random variable (cid:104)Xj/N(cid:105), where Xj ∼ Binomial(N, αmax) are i.i.d. random variables, and the angular
brackets indicate the finite average over 50 repetitions. The identity function is plotted as a reference (dashed line).
C. Alternative visualisations of performance scaling by network and sample size
Figure 9 replots the precision and recall from Figure 3 using different subplots for each sample size, and additionally
shows the specificity. Similarly, Figure 10 replots the precision and recall from Figure 5 using different subplots for
each sample size.
103102max103102FalsepositiverateN=10Identity 21
FIG. 9. Precision, recall, and specificity for different network sizes, sample sizes, and dynamics. Left: Vector autoregressive
process; Right: Coupled logistic maps. Each row corresponds to a different time series length (Top: T = 100; middle: T = 1000,
bottom: T = 10 000). The results for 10 simulations from different initial conditions are shown (low-opacity circles) in addition
to the mean values (solid circles).
20406080100Networksize0.00.20.40.60.81.0T=100PrecisionRecallSpecificity20406080100Networksize0.00.20.40.60.81.0T=1000PrecisionRecallSpecificity20406080100Networksize0.00.20.40.60.81.0T=10000PrecisionRecallSpecificity20406080100Networksize0.00.20.40.60.81.0T=100PrecisionRecallSpecificity20406080100Networksize0.00.20.40.60.81.0T=1000PrecisionRecallSpecificity20406080100Networksize0.00.20.40.60.81.0T=10000PrecisionRecallSpecificity 22
FIG. 10. Precision-recall trade-off for different statistical significance levels. The plots show the results for different dynamics
(Left: Vector autoregressive process; Right: Coupled logistic maps), different time series lengths (top to bottom: T = 100, 1000,
10 000), and different network sizes (N = 10, 40, 70, 100). The error bars indicate the standard deviation over 10 simulations
from different initial conditions.
max0.00.20.40.60.81.0PrecisionT=100N=10N=40N=70N=100maxRecallT=100N=10N=40N=70N=100max0.00.20.40.60.81.0PrecisionT=1000N=10N=40N=70N=100maxRecallT=1000N=10N=40N=70N=100103102101max0.00.20.40.60.81.0PrecisionT=10000N=10N=40N=70N=100103102101maxRecallT=10000N=10N=40N=70N=100max0.00.20.40.60.81.0PrecisionT=100N=10N=40N=70N=100maxRecallT=100N=10N=40N=70N=100max0.00.20.40.60.81.0PrecisionT=1000N=10N=40N=70N=100maxRecallT=1000N=10N=40N=70N=100103102max0.00.20.40.60.81.0PrecisionT=10000N=10N=40N=70N=100103102maxRecallT=10000N=10N=40N=70N=100 |
1801.04631 | 2 | 1801 | 2019-09-23T21:55:14 | Noise Sharing and Mexican Hat Coupling in a Stochastic Neural Field | [
"q-bio.NC"
] | A diffusion-type coupling operator biologically significant in neuroscience is a difference of Gaussian functions (Mexican Hat operator) used as a spatial-convolution kernel. We are interested in pattern formation by \emph{stochastic} neural field equations, a class of space-time stochastic differential-integral equations using the Mexican Hat kernel. We explore, quantitatively, how the parameters that control the shape of the coupling kernel, coupling strength, and aspects of spatially-smoothed space-time noise, influence the pattern in the resulting evolving random field. We confirm that a spatial pattern that is damped in time in a deterministic system may be sustained and amplified by stochasticity. We find that spatially-smoothed noise alone causes pattern formation even without direct spatial coupling. Our analysis of the interaction between coupling and noise sharing allows us to determine parameter combinations that are optimal for the formation of spatial pattern. | q-bio.NC | q-bio |
Noise Sharing and Mexican Hat Coupling in a Stochastic Neural Field
1Department of Mathematics, University of Southern California, Los Angeles, CA, USA
Peter H. Baxendale1
2Department of Mathematics, University of British Columbia, Vancouver, BC, Canada
Priscilla E. Greenwood2
3Department of Psychology and Brain Research Centre, 2136 West Mall,
University of British Columbia, Vancouver, BC, V6T 1Z4 Canada∗
Lawrence M. Ward3
(Dated: September 25, 2019)
A diffusion-type coupling operator biologically significant in neuroscience is a difference of Gaussian
functions (Mexican Hat operator) used as a spatial-convolution kernel. We are interested in pattern
formation by stochastic neural field equations, a class of space-time stochastic differential-integral
equations using the Mexican Hat kernel. We explore, quantitatively, how the parameters that control
the shape of the coupling kernel, coupling strength, and aspects of spatially-smoothed space-time
noise, influence the pattern in the resulting evolving random field. We confirm that a spatial pattern
that is damped in time in a deterministic system may be sustained and amplified by stochasticity.
We find that spatially-smoothed noise alone causes pattern formation even without direct spatial
coupling. Our analysis of the interaction between coupling and noise sharing allows us to determine
parameter combinations that are optimal for the formation of spatial pattern.
Keywords: neural field equation, Mexican Hat coupling, stochastic process, spatial pattern,
spatially-smoothed noise.
PACS: 89.75.-k,, 89.75.Da, 05.45.Xt, 87.18.-h
MSC: 34C15, 35Q70
I.
INTRODUCTION
In this paper we explore the formation of pattern by a
stochastic neural field equation with simple damping as
its reaction term and with both i.i.d. and shared noise.
The existence and possible sources of shared noise cor-
relations in stochastic neural models have been studied
recently by Doiron et al. [1], and by Meyer, Ladenbauer,
and Obermayer [2]. We, on the other hand focus on the
effect of shared or correlated noise on pattern formation.
We have been shown by Hutt and colleagues [3] that
spatial pattern, embedded in deterministic space-time
dynamics but immediately damped, may be excited by
noise. Butler and Goldenfeld [4, 5], and McKane, Bian-
calani and Rogers [6] showed the existence of excitable
spatial modes that, when noise was added, were revealed
in power spectral densities. The knowledge of this noise-
facilitated source of pattern, also observed in biological
systems, motivated us to explore how certain sample path
properties of evolving stochastic neural fields depend on
the parameters in a basic example. We look at parame-
ters that control strength of coupling or local interaction
in the field, and on the extent of local sharing, or smooth-
ing, of noise.
The sample path properties we look at are: how pat-
tern grows with coupling strength, how pattern is re-
vealed and sustained by noise when field interaction has
no excitable modes, how coupling interacts with noise
smoothing, giving rise to distinct patterns for optimum
smoothing, and then yielding distorted pattern with
"too much smoothing." A striking result is that spatial
smoothing of noise, alone, without direct neural field in-
teraction, produces pattern.
Simulation of the time evolution of one-dimensional
fields yields insight about typical sample path behaviour
of the evolving random field and illustrates the analytical
conclusions we present. We use two measures of spatial
pattern. One is the spatial FFT amplitude as a stochas-
tic process in time (note, for comparison to other stud-
ies, that power spectral density is the square of the FFT
amplitude; here we compute spatial FFT amplitudes di-
rectly and from simulations). The second is a function of
space, which we call F , that allows direct observation of
dominant frequency, again a stochastic process in time.
We believe that display of sample paths of both of these
processes is innovative here, and that they will prove to
be valuable in future studies of the model we study and
of other stochastic neural fields.
What has come to be called a neural field equation is
an integro-differential equation of the form
dY (t, x) =
−Y (t, x)+
cw(x−y)S(Y (t, y))dy
dt, (1)
(cid:20)
(cid:90)
R
(cid:21)
∗ Corresponding author. Email:
[email protected], Tel.: +1
604 822 6309, Fax: +1 604 822 6923.
where Y is an R1-valued state variable for a neural sys-
tem, w is a coupling operator, for example the Mexican
Hat convolution kernel, c, is a constant called the cou-
(cid:20)
(cid:90) L
0
(cid:21)
dt
(2)
pling strength, and S is a scaling functional, typically a
sigmoid, which keeps Y bounded in the diffusion term.
In what follows we take S to be the identity within a
certain bounding range. We study a stochastic version of
(1), wrapped with length L,
dY (t, x) =
− Y (t, x) +
cw(x − y)Y (t, y)dy
+ σdG(t, x),
where σ is a constant diffusion coefficient, c ≥ 0 is a con-
stant, and G(t, x) is a Gaussian process that may depend
on both time, t, and space, x. We study the interaction
of coupling and noise smoothing in (2) theoretically in
continuous space and time, and then discretize both for
simulations.
Recently, the study of such equations, with continu-
ous state variable, has been put on a rigorous footing
by Faugeras and Inglis [7]. Conveniently for the study
of neural field equations that generate spatial pattern,
they singled out the difference-of-Gaussians coupling op-
erator (often called the Mexican Hat operator) as one
that satisfies established conditions for the existence and
uniqueness of a solution. As will be seen, this coupling
operator generates pattern when used in equations like
(2). It is sometimes preferred to the Laplacian as a cou-
pling operator [8].
In what follows we compute conditions for the inter-
action of the dominant modes produced by coupling and
noise smoothing in the stochastic model (2). When no
excitable modes are present the spatial pattern gener-
ated by the dominant modes is damped, but added noise
reveals and sustains the damped spatial pattern. Noise
smoothing can either render the revealed pattern more
clearly, or distort it, depending on the extent of the
smoothing.
New here is an analytic and graphical study of so-
lutions of (2) with smoothed noise. We are interested
in spatial pattern revealed by noise in stochastic sam-
ple paths from the evolving random field of the neural
field equation as parameters are varied. We explore the
relation between the coupling strength constant, c, and
a parameter of noise smoothing, η, the standard devia-
tion (SD) of a Gaussian smoothing kernel. We consider
separately the ranges of c where the eigenvalues are all
negative and where the maximum eigenvalue is positive.
In our computations a specific value of c separates these
ranges. Also new is the result that smoothed noise, by
itself with c = 0 in (2), can produce spatial patterns sim-
ilar to those arising from excitable spatial modes created
by the Mexican Hat coupling.
II.
INPUT NOISE
In modeling a spatially discrete stochastic neural field,
a default choice (e.g., [9]) has been to introduce an inde-
pendent Brownian component for each x. If the locations
2
are tightly packed, however, as in neural tissue, the same
input noise may be shared in a neighborhood of loca-
tions. This is because the nearer neurons are to each
other in the brain the more similar is the set of synaptic
inputs they receive. Neurons communicate more exten-
sively with each other the closer they are to each other
because the number and strength of synaptic connections
between neighboring neurons tends to decline with the
distance between them [10, 11]. Given that shared noise
would be mostly synaptic noise [11, 12], the amount of
shared noise between two neurons should decline with the
amount of effective connectivity between them. We are
interested in the effect on spatial patterns of the size of
the neighborhood in which some of the same input noise
is felt. In our exploration we have included cases where
noise is independent at each location, x, and cases where
noise is averaged over neighborhoods of various sizes. We
refer to this as spatial smoothing of noise. It has been
shown that such noise allows for (Holder) continuous so-
lutions to a broad class of stochastic integro-differential
equations, including equations such as (2) [7, 13 -- 15]. In
simulations, Section IV, for spatial smoothing we used
a Gaussian kernel. The Gaussian smoothing kernel was
convolved with i.i.d. Gaussian noises at each iteration of
the evolving spatial field.
III.
INTERACTION OF COUPLING AND
NOISE SMOOTHING IN THE NEURAL FIELD
EQUATION
We capture the interaction of coupling and noise
smoothing in terms of the Fourier component processes
associated with (2). We begin with (2), taking w and r to
be general functions. Later we specialize to Mexican Hat
and Gaussian functions. The noise G(x, t) is Gaussian
with mean 0 and covariance
E [G(t, x)G(s, y)] = min(s, t)r(x − y).
(3)
Here r is a general positive semi-definite function. We
denote the Fourier transforms of the coupling kernel w
and the noise smoothing function r by
and
W (k) =
R(k) =
e−ikxw(x) dx,
k ∈ R,
e−ikxr(x) dx,
k ∈ R.
(cid:90) ∞
(cid:90) ∞
−∞
−∞
riodic. That is, replace w(x) by w(x) =(cid:80)
and r(x) =(cid:80)
The spatial variable x is in an interval [0, L] with pe-
riodic boundary conditions, equivalently on the circle
R/LZ, so long as the functions w and r are made pe-
n∈Z w(x + nL)
n∈Z r(x + nL) and drop the 'hat' notation.
In fact the widths of the support of r and w will be less
than L.
(cid:90) L
(cid:90) L
0
0
The periodic function w(x) has Fourier coefficients
W (2πk/L) =
e−2πikx/Lw(x) dx,
k ∈ Z,
(4)
and the periodic function r(x) has Fourier coefficients
R(2πk/L) =
e−2πikx/Lr(x) dx,
k ∈ Z.
(5)
Moreover r is positive semi-definite. We consider the
model (2), where the noise G(x, t) is given by (3). A
Fourier series expansion of the solution Y (t, x) of (2)
allows us to write it in terms of a family of Ornstein-
Uhlenbeck processes indexed by spatial frequencies (wave
numbers, k). We can then compute the expected squared
amplitudes of these O-U processes.
There are standard 2-
dimensional Brownian motions {Ck(t) : t ≥ 0} for k ≥ 1
and a standard scalar Brownian motion {C0(t) : t ≥ 0},
all mutually independent, such that the solution Y (t, x)
of (2) is given by
Proposition (cf.
[17]).
(cid:16)
ak(t)e2πikx/L(cid:17)
(cid:112)R(0) dC0(t) (7)
(6)
Y (t, x) = a0(t) +
2Re
where the scalar process a0(t) satisfies
σ√
da0(t) = [−1 + cW (0)] a0(t) dt +
L
and for k ≥ 1 the complex processes ak(t) satisfy
dak(t) = [−1 + cW (2πk/L)] ak(t) dt
+
(cid:112)R(2πk/L) dCk(t).
σ√
2L
Proof. Substituting the Fourier series expansion
Y (t, x) =
ak(t)e2πikx/L
(cid:88)
k≥1
(cid:88)
k∈Z
into (2) gives(cid:88)
−(cid:88)
(cid:90) L
k∈Z
k∈Z
dak(t)e2πikx/L =
ak(t)e2πikx/Ldt
(cid:88)
k∈Z
w(x − y)
+ c
0
+ σdG(t, x).
Since
(cid:90) L
ak(t)e2πiky/Ldydt
w(x − y)e2πiky/L dy
(cid:90) L
0
= e2πikx/L
0
= e2πikx/LW (2πk/L)
w(x − y)e2πik(y−x)/L dy
3
we get(cid:88)
(cid:88)
k∈Z
=
k∈Z
dak(t)e2πikx/L
(cid:2)−1 + cW (2πk/L)(cid:3)ak(t)e2πikx/L dt + σdG(t, x).
Now multiply by e−2πi(cid:96)x/L and integrate with respect to
x. We get
Lda(cid:96)(t) = L [−1 + cW (2π(cid:96)/L)] a(cid:96)(t) dt + σdB(cid:96)(t),
equivalently
dak(t) = [−1 + cW (2πk/L)] ak(t) dt +
σ
L
dBk(t)
(9)
where
Bk(t) =
(cid:90) L
0
e−2πikx/LG(t, x) dx,
k ∈ Z.
(10)
Assuming that the initial condition Y (0, x) is real, we
have a−k(0) = ak(0). Since B−k(t) = Bk(t) for all
t, we have a−k(t) = ak(t) for all t ≥ 0. Therefore
and it suffices to study ak(t) for k ≥ 0.
ak(t)e2πikx/L + a−k(t)e−2πikx/L = 2Re(cid:0)ak(t)e2πikx/L(cid:1),
may write B0(t) = (cid:112)LR(0)C0(t) and Bk(t) =
(cid:112)(L/2)R(2πk/L)Ck(t) for k ≥ 1, where the processes
Ck are as stated in the proposition. For k ≥ 1 the
equation (9) can then be rewritten as (7) and (8).
in Appendix A, we
calculations
Using
the
(8)
Corollary 1. Suppose
Y (0, x) = A0 + 2
Ak cos(cid:0)2πkx/L + φk
(cid:1)
(cid:88)
k≥1
Let a0(t) be the solution of (7) with initial condition
a0(0) = A0, and for k ≥ 1 let ak(t) be the solution
of (8) with initial condition ak(0) = Akeiφk . Write
ak(t) = Ak(t)eiφk(t). Then
Ak(t) cos(cid:0)2πkx/L + φk(t)(cid:1).
Y (t, x) = a0(t) + 2
(cid:88)
k≥1
Thus the period k part of Y (t,·) has amplitude 2Ak(t)
and phase φk(t) determined by the Ornstein-Uhlenbeck
process ak.
Look more closely at the process ak(t). We consider
k ≥ 1 (the k = 0 case is similar). For ease of notation
write (8) as
dak(t) = λkak(t) dt + σk dCk(t).
This complex-valued SDE has solution
ak(t) = eλktak(0) + σk
≡ eλktak(0) + Nk(t),
eλk(t−s)dCk(s)
(11)
(cid:90) t
0
[16]. Here
and
λk = −1 + cW (2πk/L)
(cid:112)R(2πk/L)
σk =
σ√
2L
for k ≥ 1 (and slight modifications to the formulas if
k = 0).
We can now outline our overall strategy. Because the
integral (2) is over a bounded set, the eigenvalues of the
operator in (2) are separated, and the mode k of the
dominant eigenvalue determines the spatial frequency of
the spatial pattern. For each pair of parameters (c, η),
which will determine the coupling strength and the width
of the noise smoothing, defined in section III B, we are
able to evaluate a functional of the distribution of the
process Yc,η that is maximal at that value of (c, η) for
which kc,η is the dominant mode. This functional, the
expected squared amplitude of the kth mode of Yc,η, can
be computed in terms of the drift and diffusion coeffi-
cients, λk, σk, of the process ak(t) defined by (7), (8).
Since Ck is standard 2-dimensional Brownian motion,
then Nk(t) is 2-dimensional Gaussian with mean zero and
covariance matrix
e2λk(t−s)ds
I2 =
k(e2λkt − 1)
σ2
2λk
I2 ≡ vk(t)I2
(cid:18)(cid:90) t
0
σ2
k
(cid:19)
where I2 denotes the 2 × 2 identity matrix. Write
ak(0) = αk + iβk. Then the real part of ak is normal
with mean eλktαk and variance vk(t), and the imaginary
part of ak(t) is normal with mean eλktβk and variance
vk(t), and the real and imaginary parts are independent.
Then [16]
EAk(t)2 = Eak(t)2 = e2λkt(α2
k + β2
k) + 2v2
k =
e2λktAk(0)2 +
k(e2λkt − 1)
σ2
λk
.
(12)
This formula is valid for finite t regardless of the sign
of λk. Suppose that λk > 0 (that is, cW (2πk/L) > 1)
for some k. Then ak(t) will exhibit exponential growth.
We treat seperately the cases where all λk < 0 and those
where there exist k such that λk > 0.
A. Relationship between σ2
k and λk
Suppose that cW (2πk/L) < 1 for all k, that is all
λk < 0. Then the effect of the initial condition dies
away, and each Nk(t) converges to a stationary Ornstein-
Uhlenbeck process. The real and complex (independent)
components of Nk(t) each satisfy a scalar equation
dZt = λkZtdt + σkdWt
which has stationary distribution N (0, σ2
that, from (12),
k/(−2λk)), so
4
(13)
.
EAk(t)2 ∼ ENk(t)2
=
=
=
σ2R(2πk/L)
σ2
k−λk
2L(1 − cW (2πk/L))
σ2
2L
(1 − cW (2πk/L))
R(2πk/L)
When do we see the kth mode as dominant in a sta-
tionary solution? We should look for a parameter region
where cW (2πk/L) < 1 and thus λk = −1+cW (2πk/L) <
(1 − cW (2πk/L))
0 for all k and also the max value of
R(2πk/L)
is noticeably larger than its other values.
Notice that when λk = 0 there exists a critical point
at which center manifold theory applies [17, 18]. In this
paper we do not address center manifold theory. We first
confine ourselves to the study of the case λk < 0 for all k,
where we have an explicit expression for the relationship,
(13), between the noise smoothing and the Mexican Hat
influences on the solutions to (2). We then find that,
in spite of the increasing exponential in (12) when some
λk > 0, we can still use the ratio in (13) to predict the
optimal pairings of the width of the noise smoothing,
η, and the coupling strength, c, to obtain a clear but
transient pattern in some range of t.
B. Gaussian Noise Smoother
As stated earlier, our noise G(t, x) is Gaussian with
mean 0 and covariance
E [G(t, x)G(s, y)] = min(s, t)r(x − y).
If Z(t, x) is a Gaussian family, Brownian in time
and uncorrelated in space, then E[Z(t, x)Z(s, y)] =
min(s, t)δ(x − y), and,
if g is symmetric, then the
smoothed noise
(cid:90)
R
Z(t, y)g(x − y) dy
G(t, x) =
has
E[G(t, x)G(s, y)] = min(s, t)r(x − y)
where r(x − y) = (g ∗ g)(x − y). Here we used the sym-
metry of g. In particular if g is the density of N (0, η2)
then r is the density of N (0, 2η2), so that
exp(−x2/4η2).
r(x) =
√
1
2η
π
The Fourier transform of r(x) is thus
R(k) =
exp(−η2k2).
(14)
√
2√
π
In Section V we explore by simulation the effect on
the emerging spatial pattern of changing the standard
deviation, η, of the Gaussian noise smoother. We note
that in addition to the effect of the noise smoother on
the dominant Fourier mode, (13), there is also a variance
reduction effect that is dependent on η as well. That is,
the variance, σ2
s , of the smoothed noise for each process,
ak, is approximated by
(cid:90) +∞
(cid:20) exp(−x2/2η2)
(cid:21)2
√
η
s ≈ σ2
σ2
σ2
√
σ2
2π
−∞
dx =
√
(15)
s < σ2.
This implies that when η > 1/(2
We should thus expect that smoothed noise might be
more effective at revealing spatial patterns than would
i.i.d. noise alone, at least when η > 0.28 (c.f., [19]).
π) = 0.28, σ2
3.54η
2η
=
π
.
C. Mexican Hat Coupling Operator
We chose a form for w(x), the difference of the two
Gaussian functions, that expresses the common biological
observation that there is excitation within a small neigh-
bourhood around each location, and inhibition in a some-
what larger neighbourhood around the excitation. An-
other way to achieve this effect is to multiply by a Gaus-
sian function and then operate with a Laplacian [20];
yet another alternative operator that involves a squared
Laplacian is used in [8]. The existence and identity of
excitable spatial modes that lead to spatial pattern in
neural fields has been studied also using other approaches
(see, e.g., [3] and for a review see [21]).
To be explicit, the Mexican Hat operator [22] is defined
as
w(x) = b1 exp(−(x/d1)2) − b2 exp(−(x/d2)2)
and its Fourier transform is
√
W (k) =
π
b1d1 exp
− (d1k)2
4
(cid:20)
(cid:21)
−b2d2 exp
(cid:20)
− (d2k)2
4
(16)
(cid:21)(cid:35)
.
(cid:34)
(cid:32)
(17)
So that w(x) has indeed a Mexican Hat shape we need
b1/b2 > 1 and d2/d1 > 1. Under the additional condition
that b2d3
1, then kmax, the wave number for which
W (k) is largest, is given by
2 > b1d3
kmax =
4(d2
2 − d2
1)−1 ln
.
(18)
(cid:19)3(cid:35)(cid:33)0.5
(cid:34)
(cid:18) d2
d1
b2
b1
5
use the Mexican hat operator with b2 = d1 = 1, b1 =
1.1, d2 = 1.2. With these values kmax = 2.026 and
W (kmax) = 0.2134 so that −1 + cW (kmax) < 0 for all
c < 4.685. Note that we are treating η as a parameter of
(14) so that it can indeed take the value of 0.
It can be seen in Figure 1 that when η = 0 the value
R(k)
(1 − cW (k))
of c. It follows that the dominant mode k, that is the
is maximized at kmax = 2.026 for all values
R(2πk/L)
(1 − cW (2πk/L))
integer value of k which maximizes
, is
approximately kmaxL/(2π) = 0.322L ≈ 8 (for the value
of L used in our simulations). As the width η of the spa-
tial noise smoother increases, the dominant mode tends
to occur at lower values of k, indicating fewer spatial cy-
cles on the ring. The decline of the dominant mode with
increasing noise smoothing width occurs for lower values
of η when we consider lower values of c. For the largest
values of η the dominant mode is at k = 0, as mentioned
earlier.
IV. SIMULATIONS
An approach consistent with the preceding theory
would be to simulate each of the Fourier coefficients ak(t),
given by (11), using the calculations above. For compu-
tational purposes we would simulate only a finite number
of them, say 0 ≤ k ≤ K where K is chosen so that EA2
given by (13) is small for all k > K.
k
In fact we chose to simulate the space-time neural field
equation (2) more directly and implemented a discrete
(both space and time) approximation using the Euler-
Maruyama procedure for obtaining a numerical solution
of stochastic difference equations. This approach has
been shown to converge rapidly to a close approxima-
tion of the continuous solution [23]. For a collection of
equally spaced points {xj} write Y (t, xj) = Yj(t) and
G(t, xj) = Gj(t). The Euler-Maruyama method gives:
Yj(t + ∆t) − Yj(t)
=(cid:2) − Yj(t) + ch
(cid:88)
wj(cid:96)Y(cid:96)(t)(cid:3)∆t + σ
Gj(t + ∆t) − Gj(t)
(cid:16)
(cid:17)
(cid:96)
(19)
where wj(cid:96) represents the discretized Mexican Hat oper-
ator and h denotes the spacing between points xj. The
factor h is introduced so that the sum in (19) will be a
good approximation to the integral in (2).
D.
Interaction of k, c, η
Because of its shape-determining role in (13) we pro-
duce graphs, Figure 1, displaying the way in which the
expression
R(k)
(1 − cW (k))
varies with k, c and η. We
We used n = 128 points spaced at distance h = 0.2
apart in a 1D spatial array, and implemented a periodic
boundary so that the values of xj, and the corresponding
components Yj(t) of (19), can be thought of as forming
a discrete ring of 128 components. This corresponds to a
A. Simulation implementation
6
FIG. 1. Effects of spatial noise smoother standard deviation, η, on the ratio R(k)/(1 − cW (k)) for various values of c. Note
that the largest values of the ratio occur in the top curve in each plot, where η = 0, with curves descending in value with
increase in η. Vertical line is at k = 2, near the maximum value for all values of c when η = 0.
wrapping length L = nh = 128 × 0.2 = 25.6 in equation
(2). The 128 components were coupled by the Mexican
Hat operator (16) with parameter values b2 = d1 = 1,
b1 = 1.1, d2 = 1.2. With these values the function w(x)
had effective width from -3 to 3, being very near zero out-
side this interval, so that our discretized Mexican Hat op-
erator wj(cid:96) was effectively 31 spatial locations wide. That
is, the choice h = 0.2 was convenient for the implemen-
tation of the Mexican Hat so that it operated only on
31 of the 128 components around each position in the
wrapped 1D spatial array. We chose the width of 31 for
the Mexican Hat operator because (a) each single Mexi-
can Hat-type neural coupling in a neural system should
span only a fraction of the size of the system, and (b) 31
spatial locations, while being only a fraction of the sys-
tem size of 128, is large enough to model a Mexican Hat
operator that couples several excitatory and inhibitory
components. A "biologically" based choice would depend
on the relative size of an observed neural field to the span
of a neighborhood of a neural location that includes both
(a) (e) (c) (d) (b) (f) excitatory and inhibitory neighbors.
We implemented space-time noise as described in Sec-
tion II. In the spatially i.i.d. case the noise processes
Gj(t) were independent standard scalar Brownian mo-
tions, while for smoothed noise with parameter η the dis-
tribution of Gj(t) = G(t, xj) was as described in Section
III B. That is, on each iteration independent samples of
Gaussian noise for each component were combined for
each component by a normalized Gaussian kernel, with
standard deviation η, centered at that component.
In
our simulations, presented in Section V, we studied spa-
tially i.i.d. noise, and values for η = 0.15, 0.5, 0.67, and
1.3, corresponding roughly to kernels effectively about 1
(i.e., i.i.d., no smoothing), 5, 15, 21, and 39 spatial lo-
cations wide. These widths span a range from spatially
i.i.d. to a smoothing that combines noise from locations
over nearly 1/3 of the total ring.
We solved the stochastic difference equation (19), also
with σ = 0, iteratively and typically for 10,000 time steps
with ∆t = 0.00005, corresponding to a time interval of
length 0.5. In a few cases, to be indicated, we extended
the simulation to 500,000 iterations and t = 25, and in a
few others we used ∆t = 0.0025 in order to reach t = 25
in only 10,000 iterations. The initial values Yj(0) for
each component were independent random variables cho-
sen uniformly in the interval [0.5, 0.501]. The initial per-
turbation from the constant 0.5 is necessary to observe
pattern in the evolving coupled field in the absence of
noise [24].
For cases where all λk < 0 we calculate the amplitude
of the dominant harmonic and compare that with the am-
plitude of this harmonic in simulations. For cases where
the maximum λk > 0 we illustrate spatial pattern for
each set of parameter values for a representative realiza-
tion of the paths of all 128 components of (19) with the
ring flattened out. We also display for those realizations
the Fourier amplitudes (from a Fast Fourier Transform,
FFT) of the spatial frequencies as a stochastic process
in t, and a second measure we term F . For both the
FFT amplitudes and the computation of F we coarse-
grained time, considering 500-iteration time blocks: 1-
500, 751-1250, 1751-2250, ...,7751-8250, 8751-9250, 9501-
10000. Note that the gaps between the first two blocks
and between the final two blocks are 250 and the gaps
between the remaining contiguous timeblocks are all 500.
For the FFT amplitudes we averaged Y(t,x) for each com-
ponent over each 500-iteration block and then computed
the FFT on the resulting spatial array. F is a function
of the time-block parameter τ and (cid:96), written as
F (τ, (cid:96)) =
1
500
1
m
s ∈ timeblock τ
j=1
Yj+(cid:96)(s) − Yj(s), (20)
where (cid:96) is a spatial offset and m is the distance across the
array for which we are computing F . In our computations
m was fixed at m = 64 because the period of the spatial
pattern never exceeded this value. In the computation (cid:96)
is increased progressively across the spatial array. Thus,
(cid:88)
m(cid:88)
7
whenever the difference Yj+l(s)−Yj(s) is large, the value
of F is correspondingly increased. Local maxima in the
plot of F (τ, (cid:96)) occur wherever the spatial offset (cid:96) matches
half the period of a spatial pattern. The presence of clear
maxima in F indicates the presence of a periodic spatial
pattern, and the form of the pattern in F displays the
pattern in the Yj(t) but sometimes more clearly because
of the averaging implicit in the computation of F .
V. RESULTS
A. All λk < 0
In this section we compare representative calculations
based on the theory outlined earlier with simulations of
(19) for cases where all λk < 0, and one case where λk >
0 for illustrative purposes.
In our numerical work, the
value of c that separates the case where all λk < 0 and
where λk > 0 for some k is c ≈ 4.685.
1. No noise
Before we launch into our study of the joint effects of
smoothed noise and the Mexican Hat on the solutions to
(2), we compute the behaviour of (2) without noise.
The values for the Yj(0) are i.i.d. uniform in [0.5, 0.501].
√
Thus each Yj(0) has mean value µ = 0.5005 and standard
deviation γ = 0.001/
12 ≈ 0.0003. Since
n−1(cid:88)
Yj(0) = Y (0, jL/n) =
ak(0)e2πik(jL/n)/L
n−1(cid:88)
=
k=0
we obtain
k=0
ak(0)e2πijk/n
n−1(cid:88)
j=0
1
n
ak(0) =
Yj(0)e−2πijk/n.
128 = 0.000026.
√
12)/
√
n = (0.001/
Taking k = 0, we see that a0(0) is real with mean µ and
variance γ2/n. For 1 ≤ k ≤ n − 1 we have E[ak(0)] = 0
and Eak(0)2 = γ2/n. Thus typically Ak = ak(0) is of
√
order γ/
For c < 4.685, −1 + cW (2πk/L) < 0 for all k and
so no persistent pattern is predicted. Consider, for ex-
ample, c = 4.5. The eigenvalue −1 + 4.5W (2πk/L) is
maximized when 2πk/L ≈ kmax = 2.0263, that is for
k = 8. Thus the eigenvalue for the 8th harmonic is
−1 + 4.5W (16π/L) ≈ −1 + 4.5W (kmax) = −1 + 4.5 ×
0.22 = −1 + 0.99 = −0.01. Over the time interval [0, 0.5]
the 8th harmonic "grows" (actually damps) by a factor
of
e(−1+4.5W (16π/L))×0.5 ≈ e−0.005 = 0.995.
Given the initial amplitude of the 8th harmonic is ap-
proximately 0.000026, the typical final amplitude of the
8th harmonic would be
0.995 × 0.000026 = 0.00002587.
We simulated (19) with c = 4.5 and the average FFT am-
plitude of the 8th harmonic over the final 500 iterations
(up to t=0.5) from 10 realizations (different realizations
of the starting values) was 0.0000227, very close to the
predicted value for the dominant 8th harmonic calculated
from (12) with no noise and c = 4.5.
To illustrate calculations when λk > 0 for some k
we now consider a value of c = 15.
In this case the
eigenvalue for the 8th harmonic is −1 + 15W (16π/L) ≈
−1 + 15W (kmax) = −1 + 15 × 0.22 = −1 + 3.3 = 2.3.
Over the time interval [0, 0.5] the 8th harmonic in this
case does grow by a factor
e(−1+15W (16π/L))×0.5 ≈ e1.15 = 3.16.
Since the typical initial amplitude of the 8th harmonic is
approximately 0.000026, then the typical final amplitude
of the 8th harmonic will be
3.16 × 0.000026 = 0.000082.
We simulated (19) with c = 15 and Figure 2a displays
the average amplitudes of the spatial harmonics over 10
realizations (different realizations of the starting values).
The average value of the FFT amplitude at a spatial fre-
quency of 8 for these 10 realizations of (19) is 0.000082,
exactly the predicted value for the dominant 8th har-
monic calculated from (12) with no noise and c = 15.
2.
I.I.D. noise
For i.i.d. noise the discrete form of the Fourier ex-
pansion is valid with r(x − y) replaced by rjk = δjk.
This in turn gives a version of the Proposition (valid for
x = xj = jL/n) with R(2πk/L)/L replaced by 1/n. As
an application of the theory we can use (12) with random
initial values:
k(e2λkt − 1)
σ2
λk
E[Ak(t)2] = e2λktE[Ak(0)2] +
k = σ2/(2n) for k ≥ 1. With c = 4.5 we have
where σ2
k/(−λk)
λk < 0 for all k ≥ 1, and then E[Ak(t)2] → σ2
k/(−λk) is
as t → ∞. With σ = 1 the limiting value σ2
maximized when k = 8 and has max value 0.092. This
corresponds to A8(t) of order
0.092 = 0.30, for large t.
But this is the limiting value. The value at time t = 0.5
is much smaller. The first term contributing to E[Ak(t)2]
is at most E[Ak(0)2] = (0.000026)2. The second term is
k(eλk − 1)
σ2
which is maximized at k = 8 with max value
0.0038. Together E[Ak(0.5)2] ≤ 0.0039 for all k ≥ 1.
Thus Ak(0.5) is of order at most
0.0039 ≈ 0.06.
√
λk
√
8
FIG. 2. FFT amplitudes averaged over 10 realizations with
∆t = 0.00005. (a) No noise, c = 15, σ = 0, t = 0.5. (b) I.I.D.
noise, c = 4.5, σ = 1.0, t = 0.5. (c) I.I.D. noise, c = 4.5, σ =
1.0, t = 25. (d) Smoothed noise, c = 4.5, σ = 1.0, η = 0.5, t =
0.5. (e) Smoothed noise, c = 4.5, σ = 1.0, η = 0.5, t = 25. In
all cases the FFT amplitudes are computed on the average
Yj(t) over the final 500 iterations (up to t = 0.5 or t = 25),
kmax = 2.0263, and the number of cycles in L/2π (spatial
frequency in graphs) is 8, as described in the text.
We simulated (19) with c = 4.5 and Figure 2b displays
the average FFT amplitudes for 10 realizations (different
realizations of the starting values and the noise). The av-
erage value of the FFT amplitude at a spatial frequency
of 8 and time period 11 (average over final 500 itera-
tions up to t = 0.5) for these 10 realizations of (19) is
0.059, very close to the predicted value for the dominant
8th harmonic from the theory. Notice, however, that the
dominant frequency doesn't stand out very well from the
noise at other spatial frequencies in Figure 2b. Thus i.i.d.
noise doesn't substantially enhance the weak pattern ex-
pected from c = 4.5 at this short time interval.
To see the effect of i.i.d. noise over a longer time
interval we simulated 10 realizations of (19) with c =
4.5, ∆t = 0.00005 over a time interval from t = 0 to
t = 25 (500,000 iterations). Figure 2c shows the aver-
age FFT amplitude over the final 500 iterations. The
average FFT amplitude at a spatial frequency of 8 is ap-
proximately 0.27, close to the predicted value of 0.30 for
large t from the theory. In order to observe the stochastic
E"C"B"A"D"(a) (e) (c) (d) (b) paths of Yj(t), F and FFT amplitude, we also simulated
a single realization of (19) with c = 4.5 over a time in-
terval from t = 0 to t = 25 but with ∆t = 0.0025. Figure
3a shows the results of this simulation. After the longer
time interval the pattern more clearly stands out from
the noise, consistent with the FFT amplitude shown in
Figure 2c. There is still considerable noise, however, ev-
ident in the spatial pattern of Yj(t).
3. Coupling and noise smoothing
The interaction between smoothing width and cou-
pling strength for λk < 0 in regard to dominant modes
is described by (13). As depicted in Figure 1, the ra-
tio
R(k)
(1 − cW (k))
predicts the dominant Fourier mode for
various values of η and c. Noticeably, for any value of c,
as η increases the dominant mode eventually goes to 0.
Also, noticeably, for higher values of c this approach to
0 occurs at higher values of η. In addition, the variance
reduction property of the spatial noise smoothing, (15),
operates for values of η > 0.28, whereas for η < 0.28
variance is actually increased by the 'smoothing.' Thus,
the noise smoothing has two somewhat conflicting effects:
decreasing the variance of the noise-revealed spatial pat-
tern induced by the Mexican Hat operator, but also tend-
ing to distort it towards lower spatial frequencies. And
this tradeoff also depends on the value of the coupling
strength; higher values of c allow the variance reduction
effect to override the dominant mode reduction.
To calculate an example of the application of the the-
ory for smoothed noise we can use (12) with random ini-
tial values as in the case of i.i.d. noise. With c = 4.5.η =
k/(−λk) has max value 0.14. This
0.5, the limiting value σ2
corresponds to A8(t) of order
0.14 = 0.37, for large t.
√
But again this is the limiting value. The value at
time t = 0.5 is much smaller. The first term contribut-
ing to E[Ak(0.5)2] is at most E[Ak(0)2] = (0.000082)2.
The second term is
. For η = 0.5 this is
k(eλk − 1)
σ2
λk
maximized at k = 5 with max value 0.0083. Together
E[Ak(0.5)2] ≤ 0.0084 for all k ≥ 1. Thus Ak(0.5) is of
order at most
0.0084 ≈ 0.09.
√
We simulated (19) with c = 4.5, σ = 1.0, η = 0.5 and
Figure 2d displays the average spatial FFT amplitude re-
sults of 10 realizations (different realizations of the start-
ing values and of the noise). The average value of the
FFT amplitude at a spatial frequency of 8 and time pe-
riod 11 (average over final 500 iterations up to t = 0.5)
for these 10 reactions of (19) is about 0.06, in line with
the predicted value from the theory. Here, however, the
dominant mode from the coupling stands out better from
the surrounding spatial frequencies, except for the lower
frequencies where the dominant mode of the smoother re-
sides. The effect of the smoother is to suppress the noise
at frequencies higher than that of the dominant mode,
rather than to enhance the dominant mode, whose ampli-
9
tude is about that with i.i.d. noise. Note, however, that
both with i.i.d. noise and smoothed noise, the amplitude
of the dominant mode is substantially greater than that
with no noise, even for much larger values of c in the no
noise case. Thus, the noise amplifies the spatial pattern,
and smoothed noise makes the pattern stand out from
the noise.
Because for short simulation times and λk < 0 the am-
plitude processes of the Fourier coefficients and Y (x, t)
will be small, we also simulated 10 realizations of the case
c = 4.5, η = 0.5, σ = 1.0 over the time interval t = 25
with the same ∆t = 0.00005. For t = 25 the limiting
value of the Fourier coefficient of the 8th harmonic for
large t is approximately 0.37. Figure 2e displays the av-
erage FFT amplitude from the 10 realizations, which is
about 0.24, again in line with the theory.
In addition, we simulated the case c = 4.5, η = 0.5, σ =
1.0 over the time interval from t = 0 to t = 25 but with
∆t = 0.0025 so that we could view the sample paths of
Y (t, x), F and FFT amplitude. Figure 3b displays the
results of this simulation. The spatial pattern is clearly
evident over this longer time interval, again as predicted
by the theory. Note that the pattern is quite smooth on
the final iteration, in contrast to the pattern with i.i.d.
noise. Again, the spatial smoothing suppresses the noise
at higher frequencies, acting as a bandpass filter, rather
than enhancing the spatial pattern itself.
B. Max λk > 0
In this section we display the results of simulations
of (19) with values of c such that λk > 0 for some k.
Here (12) applies, so we expect exponential increases with
time in the amplitudes of the spatial modes, and thus in
the spatial pattern. Nonetheless, the interaction of the
noise smoothing parameter and the coupling strength are
accurately predicted by (13) and Figure 1.
1. No noise
Figure 4 displays solutions of (2) with σ = 0, i.e., solu-
tions to the deterministic version of (2) for the 128 spa-
tial replicants. With the parameters b2 = d1 = 1, b1 =
1.1, d2 = 1.2, as before, we have W (0) ≈ −0.177 and
kmax ≈ 2 and W (kmax) ≈ 0.22. For k = 0 we have
an eigenvalue −1 + cW (0) ≈ −1 − 0.177c. This eigen-
value will be negative for all c. When c = 5 the most
excitable mode has eigenvalue −1 + 5W (kmax) ≈ 0.1,
and when c = 25 the most excitable mode has eigenvalue
−1 + 25W (kmax) = 4.5. Even though these eigenval-
ues are greater than 0 we see decay both when c = 5
and when c = 25, and we do not see the effect of the
small initial perturbation away from the constant value
Y = 0.5 in time t = 0.5. Here the damping effect is
greater than the exponential in (12) initially, although
eventually the exponential causes the pattern to appear
10
is indeed the case (Figure 5a). When a small amount of
i.i.d. Gaussian noise, σ = 0.5, was added on each iteration
in (19), however, a spatial pattern is evident (Figure 5b).
More noise, σ = 1.0, makes the pattern more apparent
(Figure 5c). Thus, again we verify that noise can both
reveal weak, otherwise initially (at short time intervals)
unobservable, spatial patterns, and also sustain them at
observable levels across time. The dominant wave num-
ber (in the sense defined earlier) of the spatial pattern
does not depend on the standard deviation of the noise,
σ. It will depend, however, on the standard deviation of
the smoothing kernel, η, as predicted by the maximum
R(k)
in Figure 1. We explore this relation in
of
(1 − cW (k))
the next section.
3. Coupling and noise smoothing
As discussed earlier, noise smoothing can be expected
to affect the spatial pattern created by the Mexican Hat
coupling. When the maximum λk > 0, as here, (12) ap-
plies. The interaction of c and η in (12) is still controlled
by the ratio of σ2
k/λk, so we expect the effects of the noise
smoothing to be consistent with those for the case where
all λk < 0.
These effects are displayed in Figure 6 for c = 22.5, σ =
1.0 and values of η = 0.15, 0.5, 1.3. Figure 5c shows the
results of a simulation with these same parameter values
but no noise smoothing. When η = 0.15 < 0.28, as pre-
dicted by (15), the variance should be increased slightly
but the dominant mode is not affected, and this is appar-
ent in Figure 6a. When η = 1.3 the variance is much re-
duced but the dominant mode is shifted toward 0, reduc-
ing by one the number of cycles induced by the Mexican
Hat coupling alone (Figure 6c). When η = 0.5 > 0.28,
however, the dominant mode is unaffected and the vari-
ance also reduced, creating a more apparent spatial pat-
tern, the 'best' in our collection of solutions (Figure 6b).
The interaction between these various factors in their ef-
fects on the spatial pattern induced by the Mexican Hat
coupling thus creates optimal combinations of c, η for the
emergence of the spatial pattern.
This phenomenon of an optimal pair (c, η) is remi-
niscent of stochastic resonance (or stochastic facilitation
[25]), in which tuning of the noise strength and threshold
yields optimum performance. Of course, the greater c is,
the stronger the pattern, so this analogy only applies for
situations where damping is sufficient that the pattern is
only revealed and sustained by optimally smoothed noise.
Smoothing that is too broad imposes a lower frequency
on the array, and interferes somewhat with the pattern
created by the Mexican Hat operator, as seen in the case
where c = 22.5, η = 1.3.
FIG. 3. Spatial patterns in the case λk < 0 over interval
t = 0 to t = 25 for i.i.d. noise (a) and for smoothed noise with
η = 0.5 (b). Here ∆t = 0.0025, c = 4.5, σ = 1.0. Top row:
amplitude Y (t, x) ; second row Y (t, x) at t = 25; third row:
F ; bottom row: FFT amplitude. Here the expected number
of cycles in L/2π is again 8, as described in the text.
(not shown). When c = 75, however, the most excitable
mode has eigenvalue −1 + 75W (kmax) ≈ 15.5. Moreover,
again, arg max{W (2πk/L : k ∈ Z≥0} ≈ (L/2π) × 2 ≈ 8,
and so (with no noise) the most excitable mode should
have period 8. By time t = 0.5 we begin to see the expo-
nential growth and the pattern of 8 spatial cycles for the
Yj(t) amplitude, FFT amplitude, and F , produced by
the Mexican Hat operator. See the top graph of Figure
4c.
2.
I.I.D. noise
Figure 5 displays the effects of adding i.i.d. Gaussian
noise to the neural field equation with the same param-
eters for w(x) as in Figure 4 except that in Figure 5
c = 22.5 so λk > 0 for some k. We would expect, based
on Figure 4b,where c = 25, that little or no indication of
spatial pattern would be apparent when σ = 0, and that
η=0.5&i.i.d.&FFT&Amplitude&FFT&Amplitude&(a) i.i.d. (b) η =0.5 11
FIG. 4. Damping of spatial patterns for which cW (kmax) > 1. Top row: amplitude Y (t, x) ; middle row: F ; bottom row: FFT
amplitude. As c (indicated at top of figure) increases the spatial pattern becomes more apparent. Here σ = 0 and thus noise
smoothing is irrelevant. Here kmax = 2.0263 and the expected number of cycles in L/2π is 8, as described in the text.
C. Spatial smoothing of noise without coupling
As described earlier, there are spatial modes of the
smoothed noise itself. For small η these are spread over
a range of values of k, whereas for large η they are con-
centrated near k = 0, as shown in Figure 1, although
the dominant mode is always 0. Because the dominant
mode of 0 creates no spatial pattern, however, we might
expect that when there are non-zero spatial modes of
significant amplitude, the noise alone, in the absence of
Mexican Hat coupling, might induce a spatial pattern.
Figure 7 displays solutions to the stochastic neural field
equation (2) with c = 0 and Gaussian-smoothed noise.
That is, there is no coupling via the Mexican Hat opera-
tor. In these cases, however, η has been varied, from i.i.d.
spatio-temporal noise to η = 0.67 and to η = 1.3. We
observe that smoothing the noise itself creates a spatial
pattern with FFT amplitude depending on η, effecting
what we could term a 'coupling through partially shared
noise.' The FFT amplitude decays exponentially toward
the higher frequencies, as expected. With η = 2.4 the
FFT amplitude is concentrated close to spatial frequency
0 (not shown). Indeed, when the Gaussian smoother has
significant weighting over the entire ring, η > 2, there is
only one large bump in the typical simulated path, and
the FFT amplitude is concentrated at a spatial frequency
of one cycle over L (not shown). Finally, the spatial pat-
tern created by the smoothing kernel can be expected to
interact with that created by the Mexican Hat operator
to create a sustained and powerful standing wave when
the noise smoothing is optimal, as in Figure 6b when
η = 0.5 or to overwhelm the Mexican Hat pattern when
noise smoothing is too great, as in Figure 6c when η = 1.3
.
Pattern produced by Gaussian noise smoothing is,
however, significantly different from pattern produced by
Mexican Hat coupling. Gaussian noise smoothing with
c = 0 tends to produce stochastic paths that resem-
ble irregular bumps over space, and these bumps move
around in the spatial array as the field evolves, making
the evolving neural field resemble a chimera of emerging
and fading pattern [26]. The Mexican Hat coupling with
or without smoothing tends to result in stochastic paths
that resemble stripes rather than bumps (Figures 5, 6),
although the stripes do move around somewhat as the
field evolves, especially when the noise is not optimally
smoothed.
VI. DISCUSSION
We have illustrated, in the context of a "standard" [7]
stochastic neural field equation, (2), that a Mexican Hat
convolution kernel produces spatial patterns that (even-
tually) can be revealed and sustained by noise, even when
all eigenvalues are negative, whereas without noise the
damping tends to dominate the pattern. Moreover, over
long time intervals, Gaussian-smoothed noise alone also
c=5$c=25$c=75$FFT$Amplitude$FFT$Amplitude$FFT$Amplitude$(a) (c) (b) 12
FIG. 5. Added noise reveals and sustains the spatial pattern. Rows are the same as in Figure 4, noise level σ is indicated at
the top of each column. Here c = 22.5, with no noise smoothing.
produces a spatial pattern, and the two sources of pattern
interact. It has been known for some time that a Mexican
Hat convolution kernel produces spatial patterns, similar
to Turing patterns, in a variety of contexts (e.g., [22, 24]).
It was known previously that noise can reveal and sus-
tain such patterns which are otherwise damped, much as
quasicycles are revealed and sustained by noise in tem-
poral stochastic processes. We studied major features of
this process, exploring the dependence of the pattern on
the strength of the Mexican Hat coupling and the width
of the noise smoothing.
First, we found a parametric measure of the interaction
between noise sharing/smoothing and Mexican Hat cou-
pling, based on the Fourier expansion of the neural field
equation. We then demonstrated, for the case where all
eigenvalues of the operator are negative, a close corre-
spndence between the predictions of the continuous the-
ory and an Euler-Maruyama discretization of the the-
ory. Next we explored, for the case where at least one
eigenvalue of the operator is positive, the relationship be-
tween the coupling strength, c and the width of the noise
smoothing, η. We found a family of 'best combinations'
of parameters controlling coupling and noise smoothing
that produced the strongest patterns. A particular find-
ing is that Gaussian noise smoothing can itself, without
coupling, produce spatial patterns in the context of neu-
ral field equations, and likely in other contexts as well.
A. Development of spatial pattern in time
We explored solutions to (2) and (19) over both short,
t = 0.5, and longer, t = 25, time intervals. The re-
sults over the longer time intervals demonstrate that
even when all eigenvalues are negative the Mexican Hat
coupling can produce spatial patterns in the presence of
added noise, and these patterns can be quite clear when
noise is shared/smoothed over a local neighbourhood.
The brain, however, does not remain in a single state
for such long time intervals. A more likely scenario for
application of our results to real brains is that the brain's
state changes over shorter time intervals, typically every
few hundred milliseconds or faster. Thus, our results
for the shorter time intervals are probably most relevant.
Importantly, it would be inefficient for the brain to have
evolved a system of local coupling, the Mexican Hat cou-
pling, for which all eigenvalues are negative, so that very
long time intervals are required for the coupling to cre-
σ=0$σ=1.0$σ=0.5$FFT$Amplitude$FFT$Amplitude$FFT$Amplitude$(a) (c) (b) 13
FIG. 6. Smoothed noise reveals and sustains the spatial pattern as well as making it more regular. Rows are the same as in
Figure 4 except for addition of row 2, the amplitude after the 10,000th iteration. Smoothing kernel standard deviation, η, is
indicated at the top of each column. Here the standard deviation of the Gaussian noise is σ = 1.0, and the coupling strength is
c = 22.5. The two somewhat contrasting effects of the smoothed noise interact with the spatial pattern created by the Mexican
Hat operator.
ate spatial patterns. To illustrate this, compare Figure
2E with Figure 6 middle column. The FFT amplitudes
of the spatial patterns are roughly equivalent in the two
cases. The time required to reach this state, however, is
t = 25 for Figure 2E, where c = 4.5 and λk < 0 for all k,
whereas the time required to reach the state in Figure 6
is t = 0.5, where c = 22.5 and λk > 0 for at least one k.
Thus, even though we don't know in terms of scaling how
this space-time model might relate to real systems, the
scenario where at least one λk > 0 is the most likely to
be present. Combined with optimal noise sharing, e.g.,
with η = 0.5 in our simulations and in the example just
described, this scenario would implicate a functional role
for the local Mexican Hat coupling, such as enhancing
edges of neural representations of visual stimuli.
produce spatial patterns has, we believe, been unappre-
ciated until now. In fact, the Fourier transform of the
process (2) predicts its spatial modes, including the case
in which, because c = 0, the Mexican Hat kernel has
no effect. In this case, where the smoothing kernel acts
alone, a greater spread of noise smoothing leads to a nar-
rower range of k with significant power. Smoothing that
is significant over the entire lattice leads to a range of k
that is close to 0. The pattern that results from the in-
teraction of the Mexican Hat coupling and the coupling
by smoothing of noise is a combination of their respec-
tive spatial modes, as reflected in Figure 1. Which modes
dominate in a particular implementation depends on the
weighting of the respective operators and their extents
with respect to the size of the system.
B. Spatially-smoothed, noise-induced patterns
C. Relation to other work
The fact that spatially-smoothed noise, i.e. noise that
has non-zero correlation length in space, can by itself
There is an extensive literature on stochastic neural
field equations. We have chosen a representative, as-
η=0.15'η=1.3'η=0.5'FFT'Amplitude'FFT'Amplitude'FFT'Amplitude'(a) (c) (b) 14
FIG. 7. Applying the Gaussian noise-smoothing kernel across neighbors creates a spatial pattern. Rows are the same as in
Figure 4. Here c = 0 (so no coupling via Mexican Hat operator), σ = 0.5, and standard deviations of the Gaussian smoothing
kernel, η, are indicated at top of figure and i.i.d. indicates no noise smoothing.
sorted, sample from this literature and point out similar-
ities and differences to the present work, and directions
for future studies.
Meyer, Ladenbauer and Obermayer [2] produced a grid
array of spiking neurons with a Mexican Hat coupling
structure and measured the covariance pattern of the
resulting spike counts while varying the parameters of
the Mexican Hat coupling. They found an oscillating
pattern of correlation decay around particular fixed neu-
rons, produced by Mexican Hat coupling, with wider in-
hibitory than excitatory coupling. This pattern did not
appear when an inverse Mexican Hat coupling, in which
the excitatory coupling extended further than did the
inhibitory coupling, was imposed on the grid. The sys-
tem was driven by external i.i.d. synaptic noise to each
neuron, so that the output spike pattern correlation was
necessarily produced by the coupling. The approach of
the study [2] was complimentary, or dual, to the approach
and objective of the present paper.
In [17, 27] there is a cubic reaction term that suc-
ceeds in keeping the process stochastically bounded. This
is different from the thresholded identity function used
in the present work.
In [27] the coupling operator is
(1 + (∂2/∂x2))2, which has an effect similar to a Mexican
Hat, whereas in [17] an effectively Mexican Hat opera-
tor, written differently, is used. Both in [27] and in [17]
the noise is either uncorrelated spatially or, the other ex-
treme, 'global fluctuations,' in which the same noise is
added to all components of the neural field at each time
point. This is in contrast to locally spatially smoothed
noise, which we studied here.
In these papers the an-
alytic method of center manifold theory together with
adiabatic elimination is used to obtain solutions to the
neural field equation.
The paper [24] studied a model that creates a mov-
ing front between states 0 and 1 using a cubic reaction
term as in [27]. At the same time a Mexican Hat kernel
together with a diffusion term creates a Turing pattern.
Homogeneoous solutions coexist with spatially periodic
states. There is no stochastic term, however, and the
effects of noise in this model are unknown.
In [8], Gaussian white noise, as in [28] together with
o + ∇2)2 and, again, a
spatial coupling of the form (K 2
cubic reaction term, in a Stratonovich SDE, create pat-
terns in R2. These patterns resemble various highly reg-
ular patterns of vegetation that occur on slopes in semi-
deserts around the world.
In Touboul's paper [9] space-dependent delays are in-
troduced. Again the noise is not smoothed across space.
A relevant result is convergence-in-law of network equa-
tions. These are in continuous time and discrete in num-
bers of neurons and of populations, both of which in-
η=0.001&η=0.1&η=0.2&!!!!!!i.i.d.!!!!!!!η=1.3!!!!!!!η=0.67!FFT!Amplitude!FFT!Amplitude!FFT!Amplitude!(a) i.i.d. (c) (b) crease to infinity, the 'neural-field limit.' The limit is a
particular McKean-Vlasov equation, a stochastic neural
mean field equation with delays.
These other works suggest various additional ways to
pursue the questions studied here. For example, one
could insert a cubic reaction term in (2) instead of us-
ing the function S to maintain stochastic boundedness
of the process. Additional analytic results might well be
obtained using the center manifold theory for the case
k = 0 as in [17, 27].
Finally, the present paper can also be seen in the
context of the broader field of pattern formation aris-
ing from stochastic differential equations. A fairly re-
cent text summarizing many problems and results in this
field is that of Blomker [29].
In that work methods
are described for the approximation of stochastic par-
tial differential equations near a change of stability using
amplitude equations. Blomker focuses on rigorous er-
ror estimates for such approximations with the aim of
enabling their use in physics and applied mathematics.
This text provides many useful clues about how to ex-
tend the present work, including a detailed description
of approximative center manifold theory.
15
Appendix A
Proposition: Consider the processes Bk = {Bk(t) : t ≥
0} defined in (10) for non-negative integers k.
(i) The processes B0, B1, B2, . . . are independent.
(ii) B0 is scalar Brownian motion with variance LR(0).
(iii) For k ≥ 1, Bk is 2-dimensional Brownian motion
with variance (L/2)R(2πk/L).
Proof: The family {Bk(t) : k ≥ 0, t ≥ 0} is a complex
valued mean 0 Gaussian process, so it enough to calculate
covariances. Each process Bk is a complex-valued Gaus-
sian process with independent increments in time, so it is
enough to calculate covariances of the complex random
variables {Bk(1) : k ≥ 0}. Write Bk(1) = B1
k. Since
r(x) = r(−x) we have
k +iB2
(cos 2πkx/L)r(x) dx = R(2πk/L),
(sin 2πkx/L)r(x) dx = 0.
0
We calculate some covariances for k, (cid:96) ≥ 0. First the
real parts:
(cid:90) L
(cid:90) L
0
kB1
(cid:96)
= E
E(cid:2)B1
(cid:3)
(cid:34)(cid:32)(cid:90) L
(cid:90) L
(cid:90) L
(cid:90) L
=
0
0
0
=
cos
0
= R(2π(cid:96)/L)
= R(2π(cid:96)/L)
kB2
(cid:96)
= E
E(cid:2)B2
(cid:3)
(cid:34)(cid:32)(cid:90) L
(cid:90) L
(cid:90) L
(cid:90) L
=
0
0
0
=
sin
0
= R(2π(cid:96)/L)
= R(2π(cid:96)/L)
Similarly for the imaginary parts:
(cid:33)(cid:32)(cid:90) L
(cos 2πkx/L)G(1, x) dx
(cid:33)(cid:35)
(cos 2π(cid:96)y/L)G(1, y) dy
0
(cos 2πkx/L)(cos 2π(cid:96)y/L)r(x − y) dxdy
2π(cid:96)(y − x)
L
− sin
2π(cid:96)x
L
sin
2π(cid:96)(y − x)
L
cos
2πkx
L
0
2π(cid:96)x
(cid:32)(cid:90) L
(cid:0) cos
(cid:90) L
L if k = (cid:96) = 0
L/2 if (cid:96) = k
else.
L
0
0
(cos 2πkx/L)(cos 2π(cid:96)x/L) dx
(cid:33)(cid:32)(cid:90) L
(cid:33)(cid:35)
(sin 2π(cid:96)y/L)G(1, y) dy
0
2π(cid:96)(y − x)
L
cos
+ cos
2π(cid:96)x
L
sin
2π(cid:96)(y − x)
L
(sin 2πkx/L)(sin 2π(cid:96)y/L)r(x − y) dxdy
(sin 2πkx/L)G(1, x) dx
2πkx
L
2π(cid:96)x
(cid:0) sin
(cid:32)(cid:90) L
(cid:90) L
(cid:26) L/2 if (cid:96) = k (cid:54)= 0
L
0
0
0
else.
(sin 2πkx/L)(sin 2π(cid:96)x/L) dx
(cid:33)
(cid:1)r(x − y) dy
dx
(cid:33)
(cid:1)r(x − y) dy
dx
Finally for the 'mixed' terms:
kB2
(cid:96)
= E
(cid:3)
E(cid:2)B1
(cid:34)(cid:32)(cid:90) L
(cid:90) L
(cid:90) L
(cid:90) L
=
0
0
=
cos
0
0
(cid:32)(cid:90) L
(cid:90) L
0
L
2πkx
(cid:0) sin
= R(2π(cid:96)/L)
= 0.
0
(cos 2πkx/L)(sin 2π(cid:96)x/L) dx
(cid:33)(cid:32)(cid:90) L
(cid:33)(cid:35)
(sin 2π(cid:96)y/L)G(1, y) dy
0
(cos 2πkx/L)G(1, x) dx
(cos 2πkx/L)(sin 2π(cid:96)y/L)r(x − y) dxdy
2π(cid:96)x
L
cos
2π(cid:96)(y − x)
L
+ cos
2π(cid:96)x
L
sin
2π(cid:96)(y − x)
L
16
(cid:33)
(cid:1)r(x − y) dy
dx
Since orthogonality implies independence for Gaussian random variables, the results follow directly.
COMPETING INTERESTS
AUTHOR'S CONTRIBUTIONS
All authors contributed to the conceptualization and
writing of the paper. The numerical simulations were
accomplished by LMW.
ACKNOWLEDGEMENTS
The authors declare that they have no competing in-
Lawrence M. Ward was supported by Discovery Grant
terests.
A9958 from NSERC of Canada.
[1] B. Doiron, A. Litwin-Kumar, R. Rosenbaum,
and K. Josic, Nature Neuroscience
G.K. Ocker,
19, 383 (2016).
[2] R. Meyer, J. Ladenbauer, and K. Obermeyer, Frontiers
in Computational Neuroscience 11, 34 (2017).
[3] A. Hutt, M. Bestehorn,
and T. Wenneker, Network:
Computation in Neural Systems 14, 351 (2003).
[4] T. Butler and N. Goldenfeld, Physical Review E 80,
030902R (2009).
[14] M. Ferrante and M. Sanz-Sol´e, ESAIM Probability and
Statistics 10, 380 (2006).
[15] M. Sanz-Sol´e and M. Sarr`a, in Seminar on Stochastic
Analysis, Random Fields and Applications, III (Ascona,
1999), Progress in Probability, Vol. 52 (Birkhauser, Basel,
2002) pp. 259 -- 268.
[16] C.W. Gardiner, Handbook of stochastic methods for
physics, chemistry and the natural sciences 2nd ed
(Springer, N.Y., 1990).
[5] T. Butler and N. Goldenfeld, Physical Review E 84,
[17] A. Hutt, A. Longtin, and L. Schimansky-Geier, Physica
011112 (2011).
D 237, 755 (2008).
[6] A. McKane, T. Biancalani, and T. Rogers, Bulletin of
[18] P. Boxler, Probability Theory and Related Fields 83,
Mathematical Biology 76, 895 (2014).
509-545 (1989).
[7] O. Faugeras and J. Inglis, Journal of Mathematical Biol-
ogy 71, 259 (2015).
[8] S. Scarsaglio, F. Laio, P. D'Odorico,
and L. Ridolfi,
Mathematical Biosciences 229, 174 (2011).
[9] J. Touboul, The Annals of Applied Probability 24, 1298
(2014).
[19] M.S. Nixon and A.S. Aguado,Feature Extraction and Im-
age Processing for Computer Vision: Third Edition (El-
sevier, 2008).
[20] D. Marr, Vision (The MIT Press, 1982).
[21] P. C. Bressloff, Journal of Physics A : Mathematical and
Theoretical 45, 033001 (2012).
[10] S. Song, P.J.Sjostrom, M. Reigl, S. Nelson,
and
[22] J. D. Murray, Mathematical Biology (Springer, Berlin,
D. Chklovskii, PLoS Biology 3(3), e68 (2005).
1989).
[11] B. Haider and D. McCormick, Neuron 62, 171 (2009).
[12] A. A. Faisal, L. P. J. Selen, and D. Wolpert, Nature
Reviews Neuroscience 9, 292 (2008).
[23] P. E. Kloeden and E. Platen, Numerical Solution
of Stochastic Differential Equations (Springer-Verlag,
Berlin, 1992).
[13] R. Dalang and N. Frangos, The Annals of Probability 26,
[24] J. Siebert and E. Scholl, Europhysics Letters 109, 40014
187 (1998).
(2015).
[25] M. D. McDonnell and L. M. Ward, Nature Reviews Neu-
Berlin, 1986) pp. 265 -- 439.
roscience 12, 415 (2011).
[26] Y. Kuramoto and D. Battogtokh, Nonlinear Phenomena
in Complex Systems 5, 380 (2002).
[27] A. Hutt, A. Longtin, and L. Schimansky-Geier, Physical
Review Letters 98, 230601 (2007).
[28] J. Walsh, in ´Ecole d'´et´e de probabilit´es de Saint-Flour,
XIV -- 1984, Lecture Notes in Mathematics. (Springer,
[29] D. Blomker, Amplitude Equations for Stochastic Par-
tial Differential Equations (World Scientific, Singapore,
2007)
.
17
|
1709.03000 | 3 | 1709 | 2018-10-30T15:34:12 | Network constraints on learnability of probabilistic motor sequences | [
"q-bio.NC"
] | Human learners are adept at grasping the complex relationships underlying incoming sequential input. In the present work, we formalize complex relationships as graph structures derived from temporal associations in motor sequences. Next, we explore the extent to which learners are sensitive to key variations in the topological properties inherent to those graph structures. Participants performed a probabilistic motor sequence task in which the order of button presses was determined by the traversal of graphs with modular, lattice-like, or random organization. Graph nodes each represented a unique button press and edges represented a transition between button presses. Results indicate that learning, indexed here by participants' response times, was strongly mediated by the graph's meso-scale organization, with modular graphs being associated with shorter response times than random and lattice graphs. Moreover, variations in a node's number of connections (degree) and a node's role in mediating long-distance communication (betweenness centrality) impacted graph learning, even after accounting for level of practice on that node. These results demonstrate that the graph architecture underlying temporal sequences of stimuli fundamentally constrains learning, and moreover that tools from network science provide a valuable framework for assessing how learners encode complex, temporally structured information. | q-bio.NC | q-bio | Network constraints on learnability of probabilistic
motor sequences
Ari E. Kahn1,2,3, Elisabeth A. Karuza4, Jean M. Vettel3,5,2, and Danielle S. Bassett2,6,7,8*
1Department of Neuroscience, University of Pennsylvania, Philadelphia, PA 19104 USA
2Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104 USA
3Human Research and Engineering Directorate, U.S. Army Research Laboratory, Aberdeen, MD 21001 USA
4Department of Psychology, University of Pennsylvania, Philadelphia, PA 19104 USA
5Department of Psychological and Brain Sciences, University of California, Santa Barbara, CA 93106 USA
6Department of Electrical & Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104 USA
7Department of Neurology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104 USA
8Department of Physics & Astronomy, College of Arts and Sciences, University of Pennsylvania, Philadelphia, PA
19104 USA
*To whom correspondence should be addressed: [email protected].
8
1
0
2
t
c
O
0
3
]
.
C
N
o
i
b
-
q
[
3
v
0
0
0
3
0
.
9
0
7
1
:
v
i
X
r
a
Keywords:
graph learning, statistical learning, motor sequence learning, graph theory, probabilistic sequences
Abstract
Human learners are adept at grasping the complex relationships underlying incoming sequential input1. In the present work,
we formalize complex relationships as graph structures2 derived from temporal associations3,4 in motor sequences. Next,
we explore the extent to which learners are sensitive to key variations in the topological properties5 inherent to those graph
structures. Participants performed a probabilistic motor sequence task in which the order of button presses was determined
by the traversal of graphs with modular, lattice-like, or random organization. Graph nodes each represented a unique button
press and edges represented a transition between button presses. Results indicate that learning, indexed here by participants'
response times, was strongly mediated by the graph's meso-scale organization, with modular graphs being associated with
shorter response times than random and lattice graphs. Moreover, variations in a node's number of connections (degree) and a
node's role in mediating long-distance communication (betweenness centrality) impacted graph learning, even after accounting
for level of practice on that node. These results demonstrate that the graph architecture underlying temporal sequences of
stimuli fundamentally constrains learning, and moreover that tools from network science provide a valuable framework for
assessing how learners encode complex, temporally structured information.
Main
Our ability to interact with our environment necessitates that we parse complex stimuli into smaller units, such as words and
phrases in language input, or events in streams of visual stimuli. This essential process relies at least in part on the statistical
regularities present around us, and often operates automatically and without any explicit, verbalizable knowledge of underlying
rules1. Statistical regularities can be inferred from various sources of information, including but not limited to the temporal
order in which stimuli are experienced. As early as infancy, humans reliably detect the probabilities with which one stimulus
transitions to another (transition probabilities, such as one syllable following another in spoken language) and the frequencies
with which stimuli temporally co-occur (co-occurrence frequencies)6. Similar forms of pattern sensitivity have been observed
beyond the language domain, including motor learning7,8 and visual event segmentation9,10.
Ongoing research examines which types of statistics induce learning -- including statistical associations between movements11
and between non-linguistic sounds12. Moreover, evidence suggests that, depending on context, learners can extract both adjacent
2
and non-adjacent dependencies between stimuli13,14. Taken together, these studies suggest that second- or third-order statistical
relationships may be encoded implicitly, and furthermore, that higher-level organizational principles themselves might be
implicitly learned. Indeed, recent work has shown that temporal ordering of visual stimuli can convey the organizational
principle of modularity3. This observation opens up the possibility of studying whether certain organizational principles are
more or less facilitative of learning, and whether information embedded in certain organizational structures might be easier to
learn than information embedded in others.
An ideally suited language in which to define such higher-order principles is network science2, an emerging interdisciplinary
field that addresses the architecture, dynamics, and design of complex systems composed of many connected parts5. The
set of parts (network nodes) and connections (network edges) are often parsimoniously encoded in a mathematical object
called a graph15. In the context of learning, we can construct a graph that encodes the pattern of relationships between objects,
movements, or sounds. Prior theoretical work16 has addressed the relationship between graph structure and artificial grammars
(such as in Cleeremans & McClelland 11), and we build on this work by empirically addressing the impact of graph-based
properties. Recently Karuza et al. 4 capitalized on this approach to define a graph from which the temporal ordering of visual
stimuli was drawn. Learners exhibited a robust on-line measure of learned graph structure: a surprisal effect defined as an
increase in reaction time when transitioning between modules. Importantly, this surprisal effect was dependent on the type of
traversal through the graph, and was more strongly pronounced when traversals through the graph provided redundancy in local
information.
While the manner in which a graph is traversed can influence learning, the nature of the graph itself may serve as an even
more fundamental constraint on the potential for humans to learn organizational principles of information. Many real-world
systems including language17, conceptual knowledge18,19, social groups20, and societies21 display non-trivial higher-order
structure such as clustering or hub-and-spoke architecture that is relevant for how humans can optimally communicate, interact,
and ultimately survive in their environment. Moreover, our knowledge about these systems unfolds and grows over time as we
experience new parts (nodes) and their relations (edges). Understanding the impact of such higher-order structure on learning
could help to explain why knowledge of some (natural or man-made) systems may be more easily acquired than others, and
why individuals differ in their capacity to learn them. It may also shed light on the question of how humans can generalize from
local statistical information to develop representations of broad-scale organizational patterns22.
Here, we examined whether the higher-order regularities of three graph structures influenced implicit learning of statistical
relationships among temporally ordered stimuli. Specifically, we trained subjects on a self-paced Serial Reaction Time (SRT)
3/29
task, where each trial was drawn from a traversal through a graph. Each node represented a stimulus, and each edge represented
a possible transition between two nodes. Based on prior work demonstrating learners' sensitivity to higher-order statistics
in SRT-like tasks11, we hypothesized that learners would display sensitivity to graph structure as evidenced by a surprisal
effect. Next, we systematically varied graph structure to examine the impact of graph topology on the acquisition of complex,
multi-element motor sequences. We hypothesized that learners would display increasingly rapid execution of button presses
when presented with modular graph structures in comparison to either random or lattice graphs. The predicted preference
for modular graphs is based on evidence across disparate fields of scientific inquiry. Modular topologies are more frequently
observed in real-world systems than either random or lattice-like topologies23. Moreover, the clustering and hierarchy of
modular graphs in natural systems can emerge in response to constraints on network wiring costs24, and similar constraints
on complexity may impact learnability itself. Indeed, we predicted that learning mechanisms should be tuned to statistical
features of natural stimuli25. Finally, we hypothesized that learning performance would vary over different regions of the graph
based on both local and global properties. Local properties are those inherent in a single node and its immediate neighbors, and
global properties encompass organizational features of the entire graph such as clustering or repeated structure. Together these
properties can be used to assess learners' sensitivity to variations in the graph across topological scales. Collectively, the results
we report below suggest that learners extract and exploit the graph topology defining temporal sequences of stimuli, and that
topological features impact speed of acquisition.
Results
Setup: We recruited 381 unique participants: 109 participants for a first experiment, 59 participants for a second experiment,
and 223 participants for a third experiment. Subjects performed a self-paced SRT motor response task using a keyboard. Stimuli
were five grey squares in a horizontal row; to indicate a target key or pair of keys that the subject was meant to press, the
corresponding square(s) would be outlined in red (Fig. 1A). The squares corresponded spatially with keys 'Space', 'H', 'J', 'K'
and 'L', such that the left square represented 'Space' and the right square represented 'L' (Fig. 1B). These keys were chosen
so that the subject's hand could ergonomically rest over all five keys at once. Sequence Generation: The order in which
stimuli were presented to the subject was prescribed by either a random or a Hamiltonian walk on a graph of N = 15 nodes
connected by E = 30 edges (Supplementary Table 1). Random Walk: For any two nodes that were not connected by an edge,
the transition probability was equal to zero. For any two nodes that were connected by an edge, the transition probability was
equal to 1 divided by the number of edges emanating from the pre-transition node. Hamiltonian Walk: A walk composed of
4/29
Figure 1. Experimental Setup. (A) An example of the first few steps of a graph traversal defined by a walk on the graph.
Top: Each node is uniquely associated with a key combination, and the sequence of key combinations is determined by a walk
on the graph. Bottom: A series of trials are presented to the participant. The red squares indicate which keys to press on that
trial. Colored arrows illustrate the edge from the graph at top being traversed. However, the participant only is shown the five
squares. (B) The mapping between fingers and keys, and average reaction times for each key press. Top: A schematic of the
mapping between visual stimuli (squares) and response effectors (fingers). Bottom: The average reaction time (RT) for each
key or pair of keys across all data. The diagonal elements of the matrix represent trials in which a single key was pressed, and
the off-diagonal elements of the matrix represent trials in which a pair of keys was pressed. (C) The three graph structures that
we examine in this study. From left to right, we show a modular graph, a lattice graph, and a random graph with N = 15 nodes
connected by E = 30 edges.
a series of cycles, each of which visited every node on the graph exactly once. Each cycle was randomly generated starting
from a node adjacent from the endpoint of the previous cycle. Graphs: We compared learning rates across 3 different graph
topologies, each consisting of 15 nodes and 30 edges: a modular graph, a lattice graph, and a random graph (Fig.1C). Briefly,
5/29
HJKLSpaceAC721859107410739647658971043109480494611818471036809LKHJSpaceFinger 1RT (ms)HKJLSpaceB1200950700TrialNode3143512345. . .. . .51423 the modular graph consists of three clusters of five interconnected nodes. The lattice can be thought of as a grid, wrapping
around at its boundary. The random graphs differed between subjects and had no consistent organizational principles besides
constituting a single connected component and maintaining the same number of nodes and edges as the other two graphs.
(See Methods for formal definitions.) For both the modular and lattice graphs, the equal degree distribution coupled with a
random walk leads to uniform pairwise probabilities for all possible transitions from a given node in the graph. Experiment:
We ran three experiments, each consisting of two back-to-back stages that differed in which graph and walk type was used to
generate the stimulus sequence. The first experiment considered learning on one of three distinct topologies (modular, lattice,
or random) instantiated as sparse graphs containing only a minority of possible edges between nodes; learning from one of the
structured topologies was followed by learning from a fully connected graph structure, in an effort to identify any changes in
learning driven by the addition of novel edges. Pragmatically, the experiment was operationalized with a within-subjects design.
Specifically, the first stage of the experiment used either a modular, lattice, or random graph to generate a sequence of 1500
stimuli via a random walk on the graph. The second stage of the experiment used a fully connected graph to generate a sequence
of 500 stimuli via a random walk on the graph (which amounted to a random stimulus order). With every node connected to
every other node (eliminating any informative structure of the input), the fully connected graph allowed a comparison between
previously trained edges and novel edges. The second experiment consisted of 1500 trials of a random walk on a modular graph,
followed by 500 trials of a Hamiltonian walk on the same graph, allowing us to confirm that learned differences transferred to a
different walk type on the same graph structure. The third experiment directly compared learning effects between graph types,
accounting for individual variability in baseline reaction times and learning rates using a within-subjects design. Similar to the
first experiment, the first stage of the experiment consisted of a sequence of 1500 stimuli via a random walk on either a modular
graph, a lattice graph, or a random graph. However, unlike the first experiment, the second stage was another sequence of 1500
stimuli via a random walk on one of the remaining two sparse structured graph types. Analysis: We verified the surprisal
effect using all modular graph traversals from stage 1 (Experiment 1). The effect of modifying the graph topology through the
addition of novel edges was assessed using a sparse graph followed by a full graph in a between-subjects design (Experiment 1).
Our assessment of the sensitivity of the surprisal effect to a switch from a random walk to a Hamiltonian walk was based on
Experiment 2. Our comparison of behavioral performance between the sparse graph types (modular, lattice, and random) was
based on a within-subjects design (Experiment 3). Finally, our comparison of local and global graph properties was based on
data acquired during the random walk traversal of a sparse random graph (stage 1 of Experiments 1 and 3).
After determining differences in reaction time by key or key combination (Fig. 1B) (which was then added as a regressor in
6/29
all subsequent models; see Methods), we asked whether participants displayed evidence of learning the probabilistic motor
sequence. A commonly studied marker of motor skill learning is an exponential drop-off in movement time with trials
practiced26. We observed this drop-off, despite our task being probabilistic rather than deterministic. Learners exhibited large
decreases in reaction time on average over the course of the first experiment. Across the initial structured graph stage, we
observed mean reaction time decreasing by nearly 500 ms (Fig. 2A, black line. See Supplementary Figures 1 & 2 for additional
information).
The observed overall decrease in reaction time with training suggested general improvement in task performance. To
explicitly evaluate learning based on graph structure, we tested for the cross-cluster surprisal effect, a previously reported
measure of graph learning on modular graphs4. Briefly, the cross-cluster surprisal effect is a significant increase in reaction
time when a participant traverses an edge located between clusters, in comparison to when a participant traverses edges within
clusters. Using a mixed effects model (see Methods), we observed a statistically significant increase in reaction time across
edges that connected two clusters (Fig. 2A,B; linear mixed effects model; t(29) = 3.61, p < 0.002, expected increase of 63.6
ms; 95% confidence interval: 29.07 to 98.10, see Supplementary Table 2). To provide an intuitive visualization of this finding,
we first note that the graph was symmetric across the three clusters; because the starting position and traversal direction within
the graph varied for each subject, the distinction between the three clusters was arbitrary when comparing across subjects.
We therefore remapped each edge to the equivalent edge within a single canonical cluster, which visually highlights the clear
difference in reaction times for between- versus within-cluster edges (Fig. 2C).
The observed surprisal effect suggests that participants are sensitive to graph structure. However, an alternative explanation
is that the surprisal effect reflects a difference in processing cost inherent to local repetitions associated with a random walk on
the modular graph. To either support or dismiss this alternative explanation, we asked whether the surprisal effect would persist
when we modified the walk to sample sparsely from each module in time, eliminating within-module repetitions of button
presses. Prior work4 suggests that learners are unlikely to show a surprisal effect when exposed to a Hamiltonian walk, where
each node is only visited once per cycle. However, if clusters correspond to a learned feature of the graph, then we expect that
learners first trained on a random walk (where we expect a surprisal effect) on a graph, and then exposed to a Hamiltonian walk
(where we do not expect a surprisal effect) on the same graph, will continue to show the surprisal effect.
In this experiment, subjects were first exposed to 1500 trials of a random walk on the modular graph, followed by 500 trials
of a Hamiltonian walk on the same graph. We were interested to determine whether a subject's surprisal effect in the first stage
was correlated with their surprisal effect in the second stage. Using a mixed effects model, we found that our estimate for the
7/29
Figure 2. Modular Graph Learning Effects. (A) Mean reaction times (RTs) as a function of trial number for stage 1 of
Experiment 1, among participants exposed to the modular graph. The red line indicates the mean for cross-cluster trials, and the
black line indicates the mean for all other trials, each binned in sets of 30 trials (n=30 subjects). (B) Mean reaction times on
correct trials for the modular graph. An increase in reaction time across cluster boundaries can be seen, here visualized by
yellower colors in the matrix elements that sit between the larger blocks. (C) Mean reaction times collapsed across the
symmetric structure of the modular graph. All three clusters were structurally identical and starting position was randomized
between subjects, so we combine reaction times across the three clusters into one 'canonical' cluster for visualization purposes
only. The mean increase in reaction time between clusters is more apparent, here visualized by yellower colors on the edges
that connect the top cluster with the two bottom clusters. (D) Relationship between surprisal effect on stage 1 (random walk)
and surprisal effect on stage 2 (Hamiltonian walk) for each subject of Experiment 2. Subjects that displayed a strong surprisal
effect in stage 1 likewise do so when the walk structure is changed (n=59).
surprisal effect was significantly reduced and not significant in the Hamiltonian walk, with an estimated increase in RT of 7.16
ms (Supplementary Table 3; linear mixed effects model; t(58) = 0.69, p = 0.49, 95% confidence interval: -13.17 to 27.49).
However, a subject's coefficient for cross-cluster surprisal in the random walk was significantly correlated with that in the
8/29
0100012001400050010001500TrialMean RT (ms)Cross−Cluster EdgeFALSETRUEAPearson's ρ = 0.8p < 0.001−1000100200300400−1000100200300400Surprisal Effect in Stage 1 (ms)Surprisal Effect in Stage 2 (ms)Dn=301357911131513579111315NodePrevious Node10501150RT (ms)Bn=162n=59C5151423 Hamiltonian walk (Fig. 2D; t(57) = 10.089, p < 0.001, Pearson's correlation coefficient r = 0.8, 95% confidence interval: 0.69
to 0.88), suggesting a diminished yet persistent effect of topological edge role after eliminating local repetitions. We verified
that this result was not the result of idiosyncrasies in assignment of motor actions to nodes by performing a permutation test
where the identity of cross-cluster edges was randomly assigned (see Supplementary Fig. 3).
While the cross-cluster surprisal effect is a useful measure of how well a modular graph is being learned, it is not a measure
that generalizes to non-modular structures. To quantitatively examine the learnability of graph structure across many graph
topologies, it would be useful to develop a generalizable measure of the learnability of single transitions from one button
press to another. In developing such a measure, it is important to note that two potential explanations exist for improvement
in response to a given target: (i) improvement is node-dependent (for all edges leading to that node), or (ii) improvement is
edge-dependent, with the rate of improvement depending on the preceding node. Notably, the inclusion of additional edges
(with the same set of nodes) in stage 2 of Experiment 1 led to a large increase in mean reaction time (Fig. 3A). To verify that
subjects had learned the edges versus nodes of the graph, we examined stage 2 of Experiment 1: when subjects were shown
(and asked to respond to) a sequence of stimuli drawn from a fully connected graph. We labeled each edge as either previously
learned or not previously learned, based on the respective motor action assigned to the endpoint nodes and whether the edge
(sequence of motor actions) was present in the graph learned by that subject in the first stage. We then estimated learner
sensitivity to new edges as measured by a change in reaction time, which we referred to as the novel edge effect (see Methods).
Using a mixed effects model, we found that subjects were significantly slower when responding to edges that had not previously
been seen, with an expected increase of 25.5 ms (linear mixed effects model; Supplementary Table 4; t(131) = 4.35, p < 0.001,
95% confidence interval: 14.01 to 37.0). This finding supports the notion that subjects learn single edges in a graph. However,
altering the number of edges in the graph decreases predictability of transitions from a single node.
While subjects were faster on previously seen edges, this difference in reaction time could be attributable to improvements
in compound motor movements, rather than to the learning of any higher-order structure. We therefore asked whether reaction
time improvements for sequences generated by each graph type might modulate this novel edge effect. In other words, was it
the case that subjects with greater sensitivity to graph structure would be more affected by disruptions to it? Further, might
this association between learning measures differ by graph type? We estimated each subject's learning rate27, and we also
estimated the novel edge effect for each subject. We found that faster learners showed a significantly greater novel edge effect
in the second stage of the experiment than slower learners (Fig. 3B; t(107) = 3.79, p < 0.001 Pearson's correlation coefficient
r = 0.34, 95% confidence interval: 0.17 to 0.50). When subdivided by graph type (Fig. 3C), this effect was significant for the
9/29
modular graph (t(28) = 2.93, p = 0.007, Pearson's correlation coefficient r = 0.48, 95% confidence interval: 0.15 to 0.72) and
the lattice graph (t(41) = 2.28, p = 0.027, Pearson's correlation coefficient r = 0.34, 95% confidence interval: 0.04 to 0.58),
but not significant for the random graph (t(34) = 1.31, p = 0.2, Pearson's correlation coefficient r = 0.22, 95% confidence
interval: -0.11 to 0.51). Intriguingly, this pattern of results suggests that subjects can more easily learn the regular structure
of modular and lattice graphs, and display slower reaction times when expectations are violated. We note, however, that the
difference between the modular and lattice conditions and the random condition was not in itself significant (modular and
random: Fisher's z = 1.18, one-sided p = 0.12, lattice and random: Fisher's z = 0.55, one-sided p = 0.29.)
Next, we tested whether certain graph structures facilitate learning more than others. We predicted that sequences generated
by the modular graph would be the easiest for participants to learn, due to the graph's segregated meso-scale structure. As
subject groups were exposed to different pairs of graph topologies, we performed three separate within-subject analyses using
the data from Experiment 3. Each analysis examined a pair of graph types, with the order of exposure counterbalanced between
subjects. For example, the first group was composed of (i) subjects first exposed to a stream of stimuli produced by a random
walk on the lattice graph, followed by a stream of stimuli produced by a random walk on the random graph, as well as (ii)
subjects first exposed to a stream of stimuli produced by a random walk on the random graph, followed by a stream of stimuli
produced by a random walk on the lattice graph. In the same manner, the second group corresponded to modular/lattice, and
the third group corresponded to modular/random. We separately fit a mixed effects model to each group. We found that the
modular graph elicited significantly quicker responses than both the lattice (linear mixed effects model; t(70) = 2.35, p = 0.022;
expected difference of 34.89 ms; 95% confidence interval: -44.49 to -4.02) and random (t(69) = 3.429, p = 0.001; expected
difference of -34.89 ms; 95% confidence interval: -54.82 to -14.95) graphs (Fig. 3D,E). We did not find a significant difference
between the lattice and random graphs (t(68) = 1.48, p = 0.14; expected difference of 12.85 ms; 95% confidence interval:
-29.88 to 4.17). Models are summarized in Supplementary Table 5. These findings support the hypothesis that the presence of
meso-scale structure in modular graphs impacts learnability.
In a final set of analyses, we investigated the extent of the influence of graph structure on learning. More specifically, we
tested whether smaller scale topological features or larger scale topological features might also impact learning, in addition to the
meso-scale features studied in the previous section. First, we studied smaller scale topological features using degree, a summary
statistic of a node's neighborhood defined by the number of edges emanating from a node. Second, we examined large scale
topological features using betweenness centrality, which intuitively captures a node's role in mediating long distance traversals
through the graph, and which is defined by the fraction of shortest paths that pass through a given node (Fig. 4A). We studied the
10/29
Figure 3. Learning Rate and Edge Surprisal. Impact of new edges on reaction time. (A) Mean RT increases in stage 2
when new edges are added to the graph (trials 1501-2000). Included for comparison are Experiments 2 and 3, where --
respectively -- only the walk or a subset of edges were changed. In both cases the increase in RT is much smaller. (B)
Per-subject learning rate correlated with the novel edge effect, defined as the mean difference in reaction time for a subject
learning the second graph when responding to a novel edge versus a familiar edge (see Methods; n=109). Learning rate, the
model coefficient for log(trial), was scaled amongst all subjects to the range [0,1]. The blue line is the least squares fit, with the
gray envelope indicating the 95% confidence interval. (C) Individual correlations shown for the three types of graphs trained on
in the first stage. Subjects exposed to the modular and lattice graphs show a significant relationship (p < 0.01,n = 30 and
p < 0.03,n = 43, respectively), while those exposed to the random graph do not (p < 0.2,n = 36). Solid lines represent least
squares fits, and gray envelopes represent the respective 95% confidence intervals. (D) Differences in reaction time by graph
type, across graphs learned in sequence. Each bar shows the number of milliseconds by which the modeled effect for the top
listed graph is faster. The increase in RT from lattice to modular, and from random to modular graphs, are both significant to
p = 0.02 and p = 0.001, respectively (See Supplementary Table 5). Error bars indicate standard error as estimated in the mixed
effects model. Asterisks indicate significance in the mixed effects model. Group sizes: Lattice-Random: n=70,
11/29
Modular-Lattice: n=72, Modular-Random: n=71. (E) Examples of the graph types.
Stage 210001200140010001250150017502000TrialMean RT (ms)Experiment1 (n=109)2 (n=59)3 (n=213)ABC010203040***Significant in Mixed Effects ModelLattice vs.RandomModular vs.LatticeGraph ComparisonModular vs.Random=RT Difference (ms)DLatticeModularRandomE0Stage 10Pearson's ρ = 0.34p < 0.0010501001500.000.250.500.751.00Scaled Learning RateNovel Edge Effect (ms)n=1090501001500.000.250.500.751.00Scaled Learning RateNovel Edge Effect (ms)Stage 1 GraphModular (n=30)Lattice (n=43)Random (n=36)n=70n=72n=71 relation between reaction time and these statistics specifically in the random graph exposures from the first and third experiment,
where we observed the greatest variability in degree and betweenness centrality over nodes. In all cases, we regressed out
the visits to a node, to separate the influence of network topology from the influence of increased exposure. We found that
node degree highly predicted the mean response time on a node (Fig. 4B,D; Kendall's τ = 0.072,n = 2655, p < 0.001), as
did node betweenness centrality (Fig. 4C,E; Kendall's τ = 0.044,n = 2655, p < 0.001). These results indicate that not only
does meso-scale graph organization affect learnability, but so do smaller scale topological features quantifying the number of
edges in a node's immediate neighborhood, and larger scale topological features quantifying a node's role in long-distance
traversals through the graph. Intriguingly, we observed an inverted relationship when we refrained from regressing out the
number of visits to a node, a fact that highlights the complex relationship between topology and learnability. (For full results,
and additional findings related to other graph metrics, see Supplementary Fig. 4-7. Also note that degree and node betweenness
centrality were correlated in the graphs analyzed in this experiment -- Kendall's τ = 0.669,n = 2655, p < 0.001 -- indicating
that nodes with dense local connectivity also play an important role in long-distance traversals.)
In turning to a discussion of our results, we begin by grounding our experimental setup and findings in the context of prior
literature. Then in later sections, we turn our attention to a discussion of more specific implications of graph architecture for
learning. In this study, we capitalize on a rich history of using motor response times as a learning measure (e.g., Cleeremans &
McClelland 11; for reviews see: Cleeremans et al. 28, Robertson 29). While those studies (and much of the artificial grammar
learning work since Reber 30) have clear parallels to the present findings, our experiments diverge in the fundamental question
they address. While we similarly generate sequences by "walking" along the edges of a graph, we additionally systematically
manipulate the topological properties of the graph underlying motor responses. In contrast to the bulk of the artificial grammar
learning literature, which largely relies on an arbitrary finite-state grammar, we instead apply organizational principles informed
by graph theory to study biases in human learning.
In the current study, we employed tools from the field of network science to determine whether and how graph structure
influences sequence learning. More specifically, our experiment built upon previous work demonstrating that modular graph
organization influences visual statistical learning3,4. We extended this line of inquiry to the motor domain by assigning unique
button presses to nodes of a modular graph, generating sequences via a random walk on that graph, and asking whether learners
exhibit similar behavioral effects at the boundaries between clusters of nodes. Indeed, we found that learners displayed a
sharp increase in reaction time when transitioning from one cluster of button presses to another, indicating that they developed
implicit expectations about the underlying topology of incoming sequential input. The sum of these results indicates that
12/29
Figure 4. Relation Between Small and Large Scale Graph Statistics and Reaction Time. (A) Illustration of node
betweenness centrality. We show shortest paths from a number of nodes on the far left to a node on the far right, which all pass
through the blue node. (B) Relationship between node degree and reaction time (RT), after regressing out the number of visits
to a node, with each point representing a separate node, e.g., 15 points per subject. The regression line shows least squares fit,
and the gray envelope is the 95% confidence interval. Reported correlation is based on Kendall's τ (n=177 subjects). (C)
Relationship between node betweenness centrality and reaction time using the same approach as used with node degree. (D)
Mean reaction time shown as a function of degree, where the mean was z-scored across the 15 nodes for a given subject. Error
bars represent bootstrapped 95% confidence intervals. (E) Mean reaction time as a function of node betweenness centrality
using the same approach as used with node degree.
13/29
τ=0.072p<0.001−0.50.00.51.00369Node DegreeRT (Node Visits Regressed Out)Bτ=0.044p<0.001−0.50.00.51.00.00.10.20.30.4Node Betweenness CentralityRT (Node Visits Regressed Out)C−0.20.00.20369Node DegreeRT (Node Visits Regressed Out)Least Squares FitMeanD−0.20.00.20.00.10.20.30.4Node Betweenness CentralityRT (Node Visits Regressed Out)Least Squares FitMeanEAn=177 subjects learners capitalize on modular structure in developing expertise in executing complex motor sequences. We further generalized
our observations to other graph topologies, and showed that certain local-, meso-, and global-scale features of graph architecture
are associated with higher learning rates than others, suggesting a critical impact of graph topology on motor skill acquisition.
One important and outstanding question is the following: What exactly forms the basis for the observed increase in RT when
switching between clusters in the modular graph? One could imagine a scenario in which the importance of the topological role
played by the cross-cluster edges would lead to an optimization of the associated motor movements, and a decrease (rather than
an increase) in RT. Our expectation of an increase in RT was primarily based on prior observations in similar tasks, particularly
the visual perception task reported in Karuza et al. 4. The fact that we observe the same increase in RT at cross-cluster edges in
a motor task suggests that the surprisal effect is a general (rather than modality-specific) property of probabilistic sequential
learning of modular structures. Yet, its presence across task modalities does not equate to a cause. Here we describe several
possible mechanisms for the surprisal effect at a number of different explanatory levels.
One level of explanation begins with the underlying neural processes. Using a visual item rotation detection task based on a
graph traversal, Schapiro et al. 3 measured the distinguishability of clusters by asking subjects to mark natural 'segmentation'
points. By studying fMRI data acquired during the performance of the task, the authors found that patterns of BOLD activity
were more similar for items within clusters than for items between clusters, despite the fact that visual stimuli were randomly
assigned to nodes. Based on this work, it is natural to ask whether a similar mechanism might apply to the processing of
targets for our motor response task. If so, then similar neural representations of nodes within a cluster might allow for quicker
responses to targets that are nearby in representational similarity space. It would be interesting to test this possibility in a future
study that combined neuroimaging with the behavioral assay we provide in our study.
Neural processes aside, we have referred to the increase in RT at cross-cluster edges as a surprisal effect based on prior
work4. Nevertheless, alternative interpretations -- beyond surprise -- also exist. For instance, the increase in RT when entering a
new cluster might be construed as analogous to the sequence initiation cost commonly observed in SRT8 and DSP31 tasks,
where increases in RT are observed at the beginning of separable chunks of responses. The idea that our observed surprisal
effect may in fact represent a preparatory cost is certainly plausible. However, it remains an open question how a sequence
initiation cost might generalize to a situation such as the current task in which subsequences are less structured. Indeed, it
would be interesting in the future to assess evidence for a type of sequence initiation cost that reflects preparation for a regime
of likely (rather than fixed) responses.
The surprisal effect also shows strong similarities to the so-called switch cost observed in the task switching literature32.
14/29
In common parlance, the switch cost is an increase in response time when learners are prompted to switch tasks between
trials, where a task might be reporting the color of a stimulus, or the magnitude of a number. Typically each task is conceived
of as having its own stimulus-response mapping, which is similar to the discrete set of responses anticipated within each of
our clusters. While many studies have demonstrated the persistence of the switch cost even when subjects anticipate task
order, a few studies have importantly shown that a switch cost is still observed when participants are led to implicitly learn a
task sequence33,34, similar to the implicit presentation of modular structure in our current study. Moreover, the relationship
between task and sequence has been shown to be hierarchical35, where both task and sequence interact with one another. This
observation provides a link to our current work where not only do the two interact, but the task (in the form of a cluster) is
defined by the sequence. In many real-world situations, tasks do not have explicitly defined subtasks nor clear boundaries
between those subtasks, and natural divisions are only learned through experience. Thus, a comprehensive exploration of the
flexible interplay between task and task sequence might be a promising direction for future research.
One final framing of interest relates to the dependence of RT on the local topology as measured by the degree and
betweenness centrality of the random graph. Both of these metrics capture the distinctiveness of a node's role within the graph.
The degree reflects the number of accessible nodes from a single node; the betweenness centrality reflects the likelihood that a
node will be necessary to traverse when moving between any pair of nodes in the graph. Importantly, the modular graph also
has important local structure, and the nodes connecting two clusters in the modular graph (boundary nodes) serve a distinct
role from the nodes existing within a cluster. One possibility is that a similar mechanism underlies the modulation of RT by
the cluster structure and the modulation of RT by the degree and betweenness centrality structure. For instance, boundary
nodes in the modular graph exhibit much higher betweenness centrality (0.22) than within-cluster nodes (0.04). And indeed,
in both the random graphs and the modular graphs, higher betweenness centrality is significantly associated with higher RTs
when differences in degree are accounted for. Thus one could conceive of the surprisal effect as a direct result of these graph
properties. However, our current experiment does not allow us to further explore this relationship, as subjects only learned
a single modular graph with two distinct classes of nodes. A potential future direction would be to explore whether this is a
general relationship across other graphs with modular structure.
The primary aim of this study was to examine differences in learners' sensitivity to distinct graph structures, while holding
constant the process through which these graphs were traversed (i.e., via a random walk). By exposing participants to modular,
lattice, and random graphs, we sought to ascertain how higher-level structure might aid or impede motor skill acquisition. Given
that modularity is an essential organizational principle underlying such varied systems as music36 and social networks21,37,
15/29
we anticipated a privileged role for this form of information structure, relative even to sequences generated from the highly
structured lattice graphs. In an initial between-subjects experimental design, we provided evidence that learners tracked pairwise
statistics, or edges linking nodes across all three graph structures. However, the extent to which learners displayed sensitivity to
novel edges was predicted by learning rate only in the lattice and modular graphs. In other words, the overall timecourse of
learning throughout exposure to sequences generated by a random graph was not associated with sensitivity to local transition
statistics on subsequent measures. We therefore propose that graphs featuring regular structural organization (i.e., lattice and
modular) might serve to boost knowledge of local regularities.
Further, by capitalizing on a complementary within-subjects experimental design, we directly contrasted the acquisition
of sequences generated by distinct graph structures. Compellingly, learning rates for the modular graph condition were
significantly faster relative to both the lattice and random graph conditions. While the differences between the highly structured
modular condition and the relatively unstructured random condition were perhaps to be expected, the differences between the
modular condition and the lattice condition were uniquely insightful. In particular, nodes within the lattice and modular graphs
were precisely equated in degree. They only differed in their meso-scale architecture, wherein neighboring nodes within the
modular graph were densely interconnected with one another. While the lattice graph was also highly structured, it lacked
this key organizational property, to the detriment of the learner, which demonstrates that the learned pairwise associations
do not capture the full scope of learners' pattern sensitivity. Instead, we provide evidence that learners clearly benefit from
modularity when it underpins the generation of complex motor sequences. Notably, our performance measure associated with
modularity, the surprisal effect, persisted even when altering graph topology and the walk taken upon that topology. However,
the effect observed when we considered an altered transition structure was significantly weaker than the effect observed when
we considered the original transition structure. We believe that we have ruled out simple confounds, particularly in having
shown that the relationship between reaction time on the random and Hamiltonian walks is specific to those edges that bridge
clusters, and not an artifact of our analysis methods. However, ruling out these simple confounds is not wholly satisfying,
and it remains an open question whether the weakening of the surprisal effect reflects limited training in the first stage of
the experiment, or whether learners discard their previous response biases as they adapt to the new statistical structure of the
second stage of the experiment. Regarding the first point, while sensitivity to statistical structure emerges in a short time frame,
persistence and generalization to new contexts may require more extensive training. Likewise, it would interesting to retest our
current experimental setup, but without a shift to the Hamiltonian walk, to ask whether the break itself disrupts previously
learned statistics. We note that learned statistics are particularly sensitive to contextual shifts38, and therefore it is possible
16/29
that the division between the two sections of the experiment was too explicit given the short timeframe. Fully addressing this
possibility will require further data collection in future.
The graph-specific effects that we observed indicated that meso-scale graph architecture impacts learning. Yet, these data
do not address whether meso-scale architecture alone is privileged, or whether both smaller and larger scale topological features
also play a role in the learning process. Using the random graph exposures, which had the greatest variability in multiscale
network architecture, we studied (i) a measure of small scale topological structure in the node degree, which captures the extent
to which a given node is connected to other nodes in a network, and (ii) a measure of large scale topological structure in the
node betweenness centrality, which incorporates information about the role of a node within the entire graph by measuring its
importance in shortest paths between other node pairs. We found that nodes of higher degree and betweenness centrality were
associated with lower reaction times, as might be expected simply from the fact that learners would be more frequently exposed
to these nodes via a random walk through the graph. Unexpectedly, however, after regressing out the number of times each
node was visited, these relationships were inverted such that nodes of higher degree and betweenness centrality were associated
with higher reaction times. This finding has important implications for how we understand the impact of smaller and larger
scale architecture on learning. Specifically, at small topological scales, we propose that when learners are exposed to sequences
generated by a heterogeneous topology such as is present in random graphs, a trade-off exists between repeated exposure to a
given node (i.e., due to high degree) leading to a lower reaction time, and the complex representation introduced by the high
number of its neighbors leading to a higher reaction time. At large topological scales, we similarly propose that a trade-off
exists between repeated exposure to a given node due to its location along shortest paths in the graph, and the complexity of the
possible paths along which it could be visited.
In our current study, degree and betweenness centrality were significantly correlated with stimulus exposure. These nodes
thus displayed a tradeoff between (i) familiarity, as modulated by the frequency of exposure, and (ii) uncertainty, as modulated
by the probability of the future state being narrowly versus widely spread amongst motor actions. Understanding this potential
tradeoff remains an important area for future work. One tractable strategy could be to consider other, non-random walks on the
graph that would allow a node to exhibit both low familiarity and high uncertainty, or vice versa. It could also be useful to
consider graph topologies that would allow variation in betweenness centrality without impacting degree, so as to disambiguate
the effect of one versus the other. For example, graphs could be constructed in which low-degree nodes serve as bridges between
disconnected areas of the graph, thus having high betweenness centrality. Separately modulating different local statistics on the
graph, as well as the type of walk used, could allow for an expanded understanding of the topological drivers of the observed
17/29
RT variation.
The separability of perceptual and motor learning in this and similar tasks continues to be a matter of debate. For example,
Deroost & Soetens 39 found that, in the context of an SRT task, stimulus-stimulus (i.e., perceptual) learning was limited to
simple deterministic sequences. After training participants on more complex probabilistic sequences, they did not find evidence
of perceptual learning. The separability of perceptual and motor learning systems is a challenging issue and one that is not yet
fully resolved (e.g., it is likely that one system bolsters the other). Unfortunately, because the present set of experiments was
not designed to address this issue, we cannot make strong claims one way or the other. Thus, we elect to maintain the most
conservative interpretation of the data possible: our task involves sequences of motor responses, so we frame our results as
evidence of motor skill learning. However, teasing apart perceptual and motor learning under this framework is a fascinating
area for future study. Especially when considering closely related prior work3,4, we suggest that our observed pattern of results
is likely reflective of domain-general processes.
We note a few methodological considerations that are particularly pertinent to this work. First, there exists a broad literature
on the theory of graph structure as well as on structures found in natural stimuli. Here we sample only a small portion of
possible graphs representing stimulus relationships. We study two stereotyped graphs, one with meso-scale clustering (modular)
and one with no meso-scale clustering (lattice). However, graphs can exhibit other diverse topologies such as core-periphery
structure, as well as other configurations of both high and low clustering beyond those tested. Moreover, the graphs we examine
are all comprised of 15 nodes. While impractical for the current study, the relationships between natural stimuli might best be
represented by graphs composed of hundreds or even thousands of nodes. Thus an open question is how these results generalize
to both larger and more diverse networks. Second, we collected no personal information on Mechanical Turk participants.
While we screened for eligibility using location and browser, we collected no information on handedness, age, or prior typing
experience. However, as our regression models all incorporate per-subject baseline and learning rate effects, we expect minimal
impact on our results from any between-subject differences. Third, there exist several limitations to using participants from
Mechanical Turk, who might each be viewing the experiment on different browsers and with different levels of accuracy and
speed in their internet connection. Fourth, it is possible that reaction time differences might be driven by recency priming. We
know that processing times are reduced for a stimulus recently viewed by the learner, and that different nodes within a graph
may be differentially affected. This is primarily a concern for modeling the cross-cluster surprisal effect, where transition nodes
may have not been seen as recently as pre-transition nodes. However, Karuza et al. 4 found that low level perceptual priming
effects did not fully account for the observed cross-cluster surprisal effect. Moreover, all between-graph comparisons are
18/29
dependent on reaction time across the entire graph rather than between different classes of nodes. We also note that excluding
priming effects may be overly conservative, given that they may serve as a local cue to graph organization. Fifth and finally, our
study does not address the question of how the exploration of graph structures in real stimuli is instantiated in the brain. For
future work, it would interesting to consider recent evidence that co-occurrence of visual stimuli leads to increasingly similar
neural representations in particular areas of human neocortex40,41, and that these same areas can encode associative distances
between objects3,42. These data suggest that an understanding of the sensitivity to the topological properties of graph structures
may have important implications in future for an understanding of neural coding.
In conclusion, we note that our results highlight the importance of topological structure in learning from a complex
environment. We first demonstrate that learning of higher-level statistics operates in the context of an SRT task. Related
previous paradigms instead focused on perceptual learning tasks without this complex motor component. Thus, our results
suggest that graph-based statistical learning mechanisms are unlikely to be modality-specific. Second, we have examined the
impact of systematic differences in graph organization on learning. We find significant differences in subject performance on
different graph types, despite identical numbers of stimuli and possible transitions in each graph. In particular, subjects show
overall faster reaction times on sequences drawn from the modular graph. Lastly, we begin to explore why sequences drawn
from these graphs may be easier or harder to learn, based on node-level statistics that topologically classify the node in relation
to both its immediate neighbors and its global role within the graph. Understanding the degree to which these results generalize
to other graph structures remains an important direction for future research, as well as understanding whether these results can
provide insight on the organization of naturally occurring complex systems. Additionally, our current study does not address
the impact of long-term learning on the surprisal effect. Whether the increase in RT persists or disappears as learners become
better trained on the task may help distinguish between competing causes.
Methods
Participants
All participants provided informed consent as specified by the Institutional Review Board (IRB) of the University of Penn-
sylvania, and study methods and experimental protocols were approved by the IRB. We recruited 381 unique participants to
complete our study on Amazon's Mechanical Turk, an online marketplace for crowdsourced work. Worker IDs were used to
exclude any duplicate participants, both within and between the three experiments. The entire sample included 109 participants
for the first experiment, 59 participants for the second experiment, and 213 participants for the third experiment. No statistical
19/29
methods were used to pre-determine sample sizes but our sample sizes are similar to those reported in previous publications3,4
All participants were financially remunerated for their time. In the first experiment, participants were paid up to $7 for an
estimated 40 minutes: $2 for completing each of the two stages, $1 for completing the entire task, and an extra $1 on each stage
on which they correctly answered at least 90% of trials. Experiment 2 provided the same payment as Experiment 1, but with an
estimated duration of 40 minutes. In the third experiment, subjects were paid up to $11 for an estimated 60 minutes: $3 per
stage, $1 for completing the entire task, and $2 for >90% performance on each stage.
Experimental Setup
Subjects performed a self-paced SRT motor response task using a keyboard. Stimuli were represented as a horizontal row of
five gray squares; all five squares were shown at all times during the main phase of the experiment. To indicate a target key or
pair of keys that the subject was meant to press, the corresponding squares would be outlined in red (Fig. 1A). When subjects
pressed the correct key combination, the squares on the screen would immediately display the next target. If an incorrect key
was pressed, or a key was left out of a two-key combination, the message "Error!" was displayed on the screen below the
stimuli, and remained until the subject pressed the correct key(s). The squares corresponded spatially with keys 'Space', 'H',
'J', 'K' and 'L', such that the left square represented 'Space' and the right square represented 'L'. These keys were chosen to
ergonomically rest underneath the subject's right hand on a QWERTY keyboard with their thumb above 'Space', index finger
above 'H', and so on (Fig. 1B).
The order in which stimuli were presented to the subject in Experiments 1 and 3 was prescribed by a random walk on a
graph of N = 15 nodes connected by E = 30 edges. In each graph, one of the 15 one- or two-finger key combinations was
randomly assigned to each node (Fig. 1A). A different graph (mapping of key presses to nodes) was generated for each random
walk. For any two nodes that were not connected by an edge, the transition probability was equal to zero. For any two nodes
that were connected by an edge, the transition probability was equal to 1 divided by the number of edges emanating from
the pre-transition node. In Experiment 2, stage 1 consisted of a random walk as previously described. The order of stimulus
presentation in stage 2 was generated by a series of Hamiltonian cycles through the graph. A single cycle consisted of every
node in the graph being visited exactly once, and each cycle was followed by another Hamiltonian cycle beginning from a node
adjacent from where the last cycle ended.
We studied the learning of 3 different graph topologies: a modular graph, a lattice graph, and a random graph (Fig. 1C). The
modular graph was characterized by 3 clusters of 5-nodes each, and a greater number of edges between nodes in a cluster than
20/29
between nodes in different clusters. Importantly, each node in the graph had exactly 4 edges, or a degree of k = 4. Thus, for any
two nodes that were connected by an edge, the transition probability was equal to 25%. The lattice graph was similar to a ring
lattice, in which nodes near one another on the ring tended to be connected to one another. As with the modular graph, each
node had exactly 4 edges, or a degree k = 4, thereby creating a flat transition probability of 25% between any two connected
nodes. Random graphs were selected out of a possible pool formed by creating 1500 instantiations of the Erdos -- Rényi graph
model, all of which we guaranteed were fully connected and had radius of at least 3, meaning there was at least one pair
of nodes with a shortest path involving three edges. We then sorted the ensemble by their estimated modularity43, and we
discarded graphs with the highest and lowest 2.5% of modularity values. The same pool of random graphs was used across all
subjects within a given experiment, though the pool differed between the experiments. In these random graphs, the degree
of each node varied from k = 1 to k = 9, and the transition probabilities varied accordingly. Finally, in the second stage of
the first experiment, we used a fully connected graph as a point of comparison, in which any node can transition to any other
node. In this case, the degree of each node was k = 14, and the probability of transitioning between any two nodes was 1/14 or
approximately 7.14%.
We ran three separate experiments (see Supplementary Table 1). Each one consisted of two stages that differed in which
graph was used to generate the stimulus sequence. The first experiment examined learning as the underlying structure
transitioned from a regular graph to a fully connected graph, and did so by comparing reaction times in response to novel versus
previously learned edges. The first stage of the first experiment used either a modular graph (n=30), a lattice graph (n=43), or a
random graph (n=36) to generate a 1500-node random stimulus walk, while the second stage used a fully connected graph
to generate a 500-node walk (random stimulus order). In the second experiment, both stages consisted of a walk over the
modular graph. However, the first stage was a 1500-node random walk, while the second stage was a 500-node Hamiltonian
walk (n=59). The third experiment employed a within-subjects design to directly compare learning effects between the graph
types, accounting for individual variability in baseline reaction times and learning rates. Similar to the first experiment, the first
stage of the second experiment consisted of a 1500-node random stimulus walk from either a modular graph, a lattice graph,
or a random graph; however, unlike the first experiment, the second stage was a 1500-node random stimulus walk on one of
the remaining two graph types. For example, if a subject was shown a sequence of stimuli produced from the modular graph
in the first stage, then in the second stage the subject would be shown a sequence of stimuli produced from either the lattice
graph or the random graph. For each of the six possible pairs of graphs we collected reaction time data from at least 30 subjects
(modular/lattice: n=36, modular/random: n=37, lattice/modular: n=36, lattice/random: n=38, random/modular: n=34, and
21/29
random/lattice: n=32). In all cases, subjects were randomly assigned to experimental conditions. Assignment was done in
the experiment code, blinding experimenters to the condition assignment for each individual participant. Subjects were only
excluded if they failed to complete the study.
Experimental Procedures
At the beginning of the first experiment, subjects were provided with the following instructions: "In a few minutes, you will see
five squares shown on the screen, which will light up as the experiment progresses. These squares correspond with keys on
your keyboard, and your job is to watch the squares and press the corresponding key when that square lights up. This part
will take around 30 minutes, followed by a similar task which will take about 5 minutes." To incentivize accuracy on the task,
subjects were informed that if they answered more than 90% of trials correctly, they would receive a $2 bonus. Subjects were
also instructed "The amount of time the segments take is not fixed, but the number of responses you have to make is. Therefore,
you should make your responses both quickly and accurately." While the reward was solely based on accuracy, workers had an
implicit incentive to finish quickly due to the fixed reward, allowing more time for other tasks on Mechanical Turk.
Before the full experiment began, subjects were given a short quiz to verify that they had read and understood the instructions.
If any questions were answered incorrectly, subjects were shown the instructions again and asked to repeat the quiz until they
answered all questions correctly. Next, all subjects were shown a 10-trial segment that did not count towards their performance;
this segment also displayed text on the screen explicitly telling the subject what keys to press on their keyboard. Subjects then
began the 1500-trial stage. A brief reminder was presented before the second stage, but no new instructions were given. After
completing the second stage, subjects were presented with performance information and their bonus earned, as well as the
option to provide feedback. The second experiment used the same instructions as the first experiment, though the estimated
time for the second task was changed to 10 minutes.
The third experiment was nearly identical, except that the initial text was changed to reflect the second structured walk:
"In a few minutes, you will see five squares shown on the screen, which will light up as the experiment progresses. These
squares correspond with keys on your keyboard, and your job is to watch the squares and press the corresponding key when that
square lights up. This part will take around 30 minutes, followed by a similar task which will take another 30 minutes." This
phrasing in the instructions ensured that learners would differentiate between stages of the experiment, reducing the potential
of carry-over effects between learning of graph structures while still preserving the benefit of a within-subject comparison.
Both the quiz and 10-trial practice session were still present, and stimulus presentation was identical to that used in the first
22/29
experiment.
Mixed effects models were fit using the lme4 library in R (R v3.5.0, lme4 v1.1-17), using the lmer function. Predictors were
centered to reduce multicollinearity. All contrasts were orthogonally coded. The observed correlation between fixed effects was
less than 0.7. Random effects were chosen as the maximal structure that allowed model convergence, as specified in the next
section. All tests were two-sided unless otherwise noted.
Analytical Approach
For every trial, we computed the reaction time based on the elapsed time from the last button press. We only excluded trials for
two reasons: subjects answered incorrectly on their first attempt or the reaction time was implausible (under 100ms or over
5000ms, or more than 3 SDs from their mean reaction time).
Effect of Targets on Reaction Time: Since performance was measured based on a key press, it was important to determine
whether biomechanical factors related to the use of different fingers, or to different combinations of finger pairs, influenced
reaction time. We predicted that one-finger responses would show shorter reaction times than two-finger responses, and that
response time would vary based on the finger needed in the response. We calculated the average reaction time for each key
press or combination of key presses across subjects and training stages (Fig. 1C). While we observed complex differences in
reaction time by finger, we also found robust differences in reaction time driven by the number of fingers required, with one-key
presses displaying a shorter reaction time than two-key presses (paired two-tailed t-test for one- and two-finger means for each
subject: t(321) = 35.56, p < 0.001; mean difference: 228.86 ms, 95% confidence interval: 216.20 to 241.52 ms). We observed
that the relative ordering of keys or key combinations by reaction time was remarkably well-preserved across subjects, and the
difference was independent of which graph was seen in the first stage of the experiment (random, lattice, or modular). This
observation suggests unique motor timing associated with each target (Supplementary Fig. 8). To ensure that our findings were
not systematically biased by differences in reaction time across key presses, we included the target key press as a regressor in
all statistical models.
Learning Rate: Based on a subject's reaction time profile across the session, we estimated each subject's learning rate from
a linear mixed effects model. The learning rate estimate was the per-subject random effect for trial number. A faster learning
rate was captured as a negative coefficient, which indicated that a subject's reaction time decreased more rapidly over time. The
model was fit to the first stage of the first experiment, with formula RT ∼ log(Trial)∗Graph+Target+(1+log(Trial) Subject),
where Target represented the key combination of the target node, controlling for biomechanical differences in motor response,
23/29
Graph was one of random/lattice/modular, and Trial was the sequential trial number (1 to 1500). We verified experimentally
that log(Trial) provided a substantially better fit than Trial (See Supplementary Fig. 9). The log transformation served to
increase the normality of the data, although formal testing of the degree of normality was not performed.
Surprisal Effect: One measure reflective of meso-scale graph structure is the surprisal effect, defined as an increase in
reaction time when transitioning to a new cluster as compared to any previous within-cluster reaction times4. To measure
this effect in our task, we fit a linear mixed effects model of the form RT ∼ log(Trial)∗ EdgeType + Target + (1 + log(Trial)∗
EdgeType Subject) to data acquired during the first stage of the experiment on the modular graph where RT is reaction time,
and where EdgeType indicated whether an edge was within or between clusters. From this model, we examined the model
coefficient for EdgeType.
Surprisal Transfer: We fit the surprisal effect model separately to the stage 1 (random walk) and stage 2 (Hamiltonian
walk) data from Experiment 2. We estimated the per-subject effect of EdgeType as a measure of surprisal, and additionally
examined the Pearson correlation coefficient between each subject's surprisal effect in stage 1 versus surprisal effect in stage 2.
Novel Edge Effect: Next, we sought to examine whether subjects were specifically improving their performance at
transitions present in the graph and therefore sensitive to violations of the learned structure of the graph. We defined a
LearnedEdge variable for all edges in stage 2 that was true if an edge between the set of finger combinations had been
present in the first stage, and false otherwise. We then computed a novel edge effect measure as the coefficient for learned
versus unlearned edges in a linear mixed effects model fit to the fully connected graph data using RT ∼ log(Trial)∗ Graph∗
LearnedEdge + Target + (1 + log(Trial)∗ LearnedEdge Subject).
Graph Effects: In the third experiment, we investigated whether differences in meso-scale structure affected learnability
by quantifying reaction time differences due to graph type. Since not all subjects were exposed to all graph types, we
split the data into three groups based on exposure: modular/lattice, random/modular, and random/lattice, with each group
roughly split by which graph was used for the first stage versus for the second stage. We fit a linear mixed effects model
to each of the three data subsets, in order to estimate whether graph type was a significant effect in each model: RT ∼
log(Trial)∗ Graph∗ Stage + Target + (1 + log(Trial)∗ GraphSubject), where Trial varied between 1 and 1500, Graph was
the current graph of the two graphs that the subject saw, and Stage corresponded to either the first stage or the second stage.
Node Effects: After determining the effect of graph type, we finally turned to quantifying the impact of node-level statistics,
or those that could explain reaction time differences within a single graph. We examined several traditional graph metrics:
degree, clustering coefficient, node betweenness centrality, and edge betweenness centrality. The degree is defined as the
24/29
number of edges connecting to a given node, given by ki = ∑ j Ai j where A is the adjacency matrix. The clustering coefficient
can be defined as the fraction of possible edges between a node's neighbors, given by Ci = 2Li
ki(ki−1) where Li is the number of
edges between any two neighbors of node i. The node betweenness centrality is defined as the fraction of shortest paths in
the graph that pass through node v, given by CB(v) = ∑s(cid:54)=v(cid:54)=t
σst (v)
σst where σst (v) is the number of shortest paths from node s
to node t that pass through node v, and σst is the total number of shortest paths in the graph from node s to node t. The edge
betweenness centrality is defined as the fraction of shortest paths passing through an edge e, given by CB(e) = ∑s(cid:54)=e(cid:54)=t
σst (e)
σst
,
where now e refers to an edge rather than a node.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Code Availability
The code that supports the findings of this study is available from the corresponding author upon reasonable request.
25/29
References
1. Aslin, R. N. & Newport, E. L. Statistical Learning: From Acquiring Specific Items to Forming General Rules. Current
Directions in Psychological Science 21, 170 -- 176. ISSN: 0963-7214 (2012).
2. Newman, M. E. J. Networks: An Introduction (Oxford University Press, 2010).
3. Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B. & Botvinick, M. M. Neural representations of events
arise from temporal community structure. Nature Neuroscience 16, 486 -- 492. ISSN: 1097-6256 (Feb. 2013).
4. Karuza, E. A., Kahn, A. E., Thompson-Schill, S. L. & Bassett, D. S. Process Reveals Structure: How a Network Is
Traversed Mediates Expectations About Its Architecture. Scientific Reports 7, 12733 (2017).
5. Newman, M. E. J. Complex Systems: A Survey. Am. J. Phys. 79, 800 -- 810 (2011).
6. Saffran, J. R., Aslin, R. N. & Newport, E. L. Statistical Learning by 8-Month-Old Infants. Science 274, 1926 -- 1928. ISSN:
0036-8075 (1996).
7. Nissen, M. J. & Bullemer, P. Attentional requirements of learning: Evidence from performance measures. Cognitive
Psychology 19, 1 -- 32. ISSN: 00100285 (Jan. 1987).
8. Hunt, R. H. & Aslin, R. N. Statistical learning in a serial reaction time task: Access to seperable statistical cues by
individual learners. Journal of experimental psychology. General 130, 658 -- 680. ISSN: 0096-3445 (2001).
9. Fiser, J. & Aslin, R. N. Statistical learning of higher-order temporal structure from visual shape sequences. Journal of
Experimental Psychology: Learning, Memory, and Cognition 28, 458 -- 467. ISSN: 0278-7393 (2002).
10. Turk-Browne, N. B., Jungé, J. A. & Scholl, B. J. The Automaticity of Visual Statistical Learning. Journal of Experimental
Psychology: General 134, 552 -- 564. ISSN: 1939-2222 (2005).
11. Cleeremans, A. & McClelland, J. L. Learning the structure of event sequences. Journal of Experimental Psychology:
General 120, 235 -- 253. ISSN: 1939-2222 (1991).
12. Furl, N. et al. Neural prediction of higher-order auditory sequence statistics. NeuroImage 54, 2267 -- 2277. ISSN: 10538119
(Feb. 2011).
13. Newport, E. L. & Aslin, R. N. Learning at a distance I. Statistical learning of non-adjacent dependencies. Cognitive
Psychology 48, 127 -- 162. ISSN: 00100285 (2004).
26/29
14. Gómez, R. L. Variability and Detection of Invariant Structure. Psychological Science 13, 431 -- 436. ISSN: 0956-7976
(Sept. 2002).
15. Bollobas, B. Random Graphs (Cambridge University Press, 2001).
16.
Jarvis, J. P. & Shier, D. R. in Applied mathematical modeling: a multidisciplinary approach (1999). doi:10.1.1.39.
1937.
17. Goldstein, R. & Vitevitch, M. S. The influence of clustering coefficient on word-learning: how groups of similar sounding
words facilitate acquisition. Frontiers in Psychology 5, 2009 -- 2014. ISSN: 1664-1078 (Nov. 2014).
18. Bales, M. E. & Johnson, S. B. Graph theoretic modeling of large-scale semantic networks. J Biomed Inform 39, 451 -- 464
(2006).
19. Vitevitch, M. S. What can graph theory tell us about word learning and lexical retrieval? J Speech Lang Hear Res 51,
408 -- 422 (2008).
20. Palla, G., Barabasi, A. L. & Vicsek, T. Quantifying social group evolution. Nature 446, 664 -- 667 (2007).
21. Girvan, M. & Newman, M. E. J. Community structure in social and biological networks. Proceedings of the National
Academy of Sciences 99, 7821 -- 7826. ISSN: 0027-8424 (June 2002).
22. Karuza, E. A., Thompson-Schill, S. L. & Bassett, D. S. Local Patterns to Global Architectures: Influences of Network
Topology on Human Learning. Trends Cogn Sci 20, 629 -- 640 (2016).
23. Steyvers, M. & Tenenbaum, J. B. The large-scale structure of semantic networks: Statistical analyses and a model of
semantic growth. Cognitive Science 29, 41 -- 78. ISSN: 03640213 (2005).
24. Mengistu, H., Huizinga, J., Mouret, J. B. & Clune, J. The Evolutionary Origins of Hierarchy. PLoS Comput Biol 12,
e1004829 (2016).
25. Hermundstad, A. M. et al. Variance predicts salience in central sensory processing. eLife 3, 1 -- 28. ISSN: 2050084X (2014).
26. Heathcote, A., Brown, S. & Mewhort, D. J. The power law repealed: the case for an exponential law of practice.
Psychonomic bulletin & review 7, 185 -- 207 (June 2000).
27. Karuza, E. A., Farmer, T. A., Fine, A. B., Smith, F. X. & Jaeger, T. F. On-line measures of prediction in a self-paced
statistical learning task. Proceedings of the 36th Annual Meeting of the Cognitive Science Society (2014).
27/29
28. Cleeremans, A., Destrebecqz, A. & Boyer, M. Implicit Learning: News From the Front. Trends in Cognitive Sciences 2,
406 -- 416 (1998).
29. Robertson, E. M. The Serial Reaction Time Task: Implicit Motor Skill Learning? Journal of Neuroscience 27, 10073 --
10075 (2007).
30. Reber, A. S. Implicit Learning of Artificial Grammars. Journal of Verbal Learning and Verbal Behavior 6, 855 -- 863
(1967).
31. Verwey, W. B., Abrahamse, E. L. & de Kleine, E. Cognitive Processing in New and Practiced Discrete Keying Sequences.
Frontiers in Psychology 1, 1 -- 13 (2010).
32. Kiesel, A. et al. Control and Interference in Task Switching-A Review. Psychological Bulletin 136, 849 -- 874 (2010).
33. Koch, I. Automatic and Intentional Activation of Task Sets. Journal of Experimental Psychology: Learning, Memory, and
Cognition 27, 1474 -- 1486 (2001).
34. Gotler, A., Meiran, N. & Tzelgov, J. Nonintentional Task Set Activation: Evidence From Implicit Task Sequence Learning.
Psychonomic Bulletin & Review 10, 890 -- 896 (2003).
35. Schneider, D. W. & Logan, G. D. Hierarchical control of cognitive processes: Switching tasks in sequences. Journal of
Experimental Psychology 135, 623 -- 640 (4 2006).
36. Gleiser, P. & Danon, L. Community structure in jazz. Adv. Compl. Syst. 6, 565 -- 573 (2003).
37. Tompson, S. H., Kahn, A. E., Falk, E. B., Vettel, J. M. & Bassett, D. S. Individual differences in learning social and
nonsocial network structures. Journal of Experimental Psychology: Learning, Memory, and Cognition. ISSN: 1939-1285
(July 2018).
38. Gebhart, A. L., Aslin, R. N. & Newport, E. L. Changing structures in midstream: Learning along the statistical garden
path. Cognitive Science 33, 1087 -- 1116. ISSN: 03640213 (2009).
39. Deroost, N. & Soetens, E. Perceptual or motor learning in SRT tasks with complex sequence structures. Psychological
Research 70, 88 -- 102 (2006).
40. Messinger, A., Squire, L. R., Zola, S. M. & Albright, T. D. Neuronal representations of stimulus associations develop in
the temporal lobe during learning. Proceedings of the National Academy of Sciences 98, 12239 -- 12244. ISSN: 0027-8424
(Oct. 2001).
28/29
41. Li, N. & DiCarlo, J. J. Unsupervised Natural Experience Rapidly Alters Invariant Object Representation in Visual Cortex.
Science 321, 1502 -- 1507. ISSN: 0036-8075 (2008).
42. Garvert, M. M., Dolan, R. J. & Behrens, T. E. A map of abstract relational knowledge in the human hippocampal -- entorhinal
cortex. eLife 6, 1 -- 20. ISSN: 2050-084X (Apr. 2017).
43. Newman, M. E. Modularity and community structure in networks. Proc Natl Acad Sci U S A 103, 8577 -- 8582 (2006).
Acknowledgments
We thank David M. Lydon-Staley and Steven H. Tompson for feedback on earlier versions of this manuscript. This work
was supported by the National Science Foundation CAREER award to DSB (PHY-1554488), the Army Research Laboratory
through contract number W911NF-10-2-0022, and the Army Research Office through contract number Grafton-W911NF-16-
1-0474 and contract number DCIST- W911NF-17-2-0181. We also acknowledge additional support from the John D. and
Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the ISI Foundation, the Paul Allen Foundation, the
Army Research Office (Bassett-W911NF-14-1-0679), the Office of Naval Research, the National Institute of Mental Health
(2-R01-DC-009209-11, R01 MH112847, R01-MH107235, R21 MH-106799, R01 MH113550), the National Institute of Child
Health and Human Development (1R01HD086888-01), the National Institute of Neurological Disorders and Stroke (R01
NS099348), and the National Science Foundation (BCS-1441502, BCS-1631550, and CNS-1626008). The content is solely the
responsibility of the authors and does not necessarily represent the official views of any of the funding agencies. The funders
had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Competing Interests
The authors declare no competing interests.
Author Contributions
A.E.K., E.A.K., J.M.V., and D.S.B. conceived the project. A.E.K., E.A.K., J.M.V., and D.S.B. planned the experiments and
analyses. A.E.K. performed the experiments and analyses. A.E.K., E.A.K., and D.S.B. wrote the manuscript and Supplementary
Information. E.A.K., J.M.V., and D.S.B. edited the manuscript and Supplementary Information.
29/29
|
1210.2140 | 1 | 1210 | 2012-10-08T03:41:45 | Natural electromagnetic waveguide structures based on myelin sheath in the neural system | [
"q-bio.NC",
"cond-mat.soft",
"physics.bio-ph"
] | The saltatory propagation of action potentials on myelinated axons is conventionally explained by the mechanism employing local circuit ionic current flows between nodes of Ranvier. Under this framework, the myelin sheath with up to 100 layers of membrane only serves as the insulating shell. The speed of action potentials is measured to be as fast as 100 m/s on myelinated axons, but ions move in fluids at just 100 nm/s in a 1 V/m electric field. We show here the action potentials, in the form of electromagnetic (EM) pulses, can propagate in natural EM waveguide structures formed by the myelin sheath merged in fluids. The propagation time is mainly cost on the duration for triggering EM pulses at nodes of Ranvier. The result clearly reveals the evolution of axons from the unmyelinated to the myelinated, which has remarkably enhanced the propagation efficiency by increasing the thickness of myelin sheath. | q-bio.NC | q-bio | Natural electromagnetic waveguide structures based on myelin sheath in the
neural system
Jiongwei Xue and Shengyong Xu
Department of Electronics, and Key Laboratory for the Physics & Chemistry of
Nanodevices, School of Electronics Engineering and Computer Science, Peking
University, Beijing 100871, People’s Republic of China.
Jiongwei Xue, [email protected]
Shengyong Xu, [email protected]
The saltatory propagation of action potentials on myelinated axons is conventionally
explained by the mechanism employing “local circuit ionic current flows” between
nodes of Ranvier. Under this framework, the myelin sheath with up to 100 layers of
membrane only serves as the insulating shell. The speed of action potentials is
measured to be as fast as 100 m/s on myelinated axons, but ions move at just 100
nm/s in a 1 V/m electric field. We show here the action potentials, in the form of
electromagnetic (EM) pulses, can propagate in natural EM waveguide structures
formed by the myelin sheath merged in fluids. The propagation time is mainly cost on
the duration for triggering EM pulses at nodes of Ranvier. The result clearly reveals
the evolution of axons from the unmyelinated to the myelinated, which has
remarkably enhanced the propagation efficiency by increasing the thickness of myelin
sheath.
Key words: action potential; waveguide; myelinated axon; the node of Ranvier;
electric synapse; muscle fiber
In nature, a high speed for information transmission among different parts of a
live creature is one of the crucial factors that determine its survival chance. The
information in a biosystem is transmitted either chemically or physically, or by a
hybrid combination of both 1. The velocities for diffusion of molecules in a fluid and
for directional movement of ions in an external electric field are both very slow 2. For
instance, in a dilute electrolyte, experimental data show that under an electric field
intensity of 1 V/m ions can only move at a speed about 100 nm/s. By sharp contrast,
an EM wave propagates in vacuum at a speed of 3×108 m/s. The transmission of an
EM wave can be regarded as transmission of energy, not transport of mass. With
limited scattering effects, an EM wave can travel in vacuum or in a uniform dielectric
over a long distance without remarkable loss in energy. This is the reason that Hubble
Space Telescope can receive images of galaxies at the edge of the universe, and a
high-quality optical fiber can efficiently transfer data across the Pacific. Engineers
have also made a variety of transmission lines and waveguides for high-efficiency
transmission of EM waves 3,4.
It is well known that the action potential is generated by the transient ion flows
passing through ion channels. Maxwell equations show that such transient ion flows
emit EM waves in the forms of electrical pulses in the way similar to dipole antennas.
And, the electrical pulses can propagate along certain transmission path without
employing physical movement of charge carriers. Therefore, it is a wise choice for the
nature to transmit information through EM waves.
However, an EM wave is usually strongly scattered in a complex biosystem made
of materials with varied phases, densities, dielectric constants, shapes and sizes. The
strong scattering processes are sometimes useful for precise imaging 5-7. The inner
part of an axon contains many sub-cellular organelles and cytoskeleton structures
such as microtubule, microfilament and neurofilament, thus it has a large resistivity
for dc current and strongly scatters an EM wave due to its non-uniformity in dielectric
constant, shape, size and density of sub-cellular organelles. As a result, the inner part
of an axon is not favorable for transmission of electric signals via either dc currents or
EM waves.
But fortunately, the phospholipid bilayer has a uniform structure. It is the basic
framework of the membrane, and such a membrane covers each and every single cell
in a biosystem. In this work, we show that by using the membrane as the main frame,
EM waveguide structures can be naturally formed in the inner fluidic environment in
biosystems, and such EM waveguide structures are efficient for the propagation of
action potentials.
Results
The EM waveguide structure formed by electrolytes and dielectrics
Figure 1a schematically shows the way of transmission for a transverse
electromagnetic (TEM) wave, in vacuum or a uniform dielectric. Here the electrical
field intensity E, the magnetic induction intensity B and wave vector k are
perpendicular to each other. Along the transmission path, the energy density for a
TEM wave, often characterized by E2, keeps constant in vacuum and decreases
slightly in a waveguide over a long distance.
a), Schematic diagram of a TEM wave with wave vector of k
Figure 1
travelling in vacuum or a uniform dielectric along z direction. b), A model of a
fluid-dielectric-fluid sandwich EM waveguide structure formed from one layer
of dielectric media between two layers of electrolyte. c), The calculated REM
value versus Cbulk, where the frequency of the EM wave is set at 10 kHz, and
the permittivity of the central dielectric layer at 3.0.
It is well known that a conductor-dielectric (or vacuum)-conductor sandwich
structure can form a parallel plate waveguide. Figure 1b shows the model of an
electrolyte-dielectric-electrolyte sandwich EM waveguide structure, which is formed
by one layer of dielectric between two layers of electrolyte (see Supplementary 1).
The calculated reflection coefficient REM value versus the concentration of bulk
electrolyte Cbulk is plotted in Figure 1c, where the frequency of the EM wave is
arbitrarily set at 10 kHz, and the permittivity of the central dielectric layer at 3.0. In
the range of bulk concentration from 0.01 to 1.0 mol/l, REM increases from 0.58 to
0.95. It suggests that the electrolyte-dielectric-electrolyte sandwich structure works
well as an EM waveguide, though not as perfect as a metal waveguide. An EM
waveguide can also be formed within a confined fluid channel, as shown in
Supplementary 2.
The fluid-membrane-fluid sandwich structure as an EM waveguide
EM waveguide structures can be formed naturally in biosystems with soft
materials only. Figure 2a shows a typical example, where a membrane, shown with
phospholipid bilayer, is sandwiched between two layers of fluid full of ions at both
sides. As shown in Figure 1c, the reflection coefficient REM is calculated to be 0.87 at
the concentration of 0.15 mol/l, a typically ion concentration in a live biosystem.
Therefore the fluid-membrane-fluid waveguide structure shown in Figure 2a can
be considered as a parallel plate waveguide. Ideally a parallel plate waveguide does
not have dispersion on frequency for propagation of TEM waves. The characteristic
impedance of such a waveguide is a constant dependent only on the geometry and
material parameters of the structure. For the TEM mode, as showed in Figure 2b, in
the coordinates where the dielectric layer and the fluid layers are perpendicular to x
direction and the EM wave is transmitted along z direction, the two fluid layers work
as the conductor cladding plates, thus Ey = Ez = 0, By =0, where Ey and Ez are the
electrical field intensities along y direction and z direction, respectively; By is the
magnetic field intensity along y direction. Figure 2c schematically shows the field
lines of E and B in x-y plane when viewing along the z direction. It gives another
reason that why the ion channels are embedded in membrane, generating ion current
flows perpendicular to the membrane, for the parallel plate waveguide structure prefer
to transmit polarized TEM wave with E perpendicular to the membrane.
a), A membrane merged in fluids full of ions at both sides is a
Figure 2
typical example of electrolyte-dielectric- electrolyte EM waveguide structure.
The dashed purple line indicates a propagating EM wave, but the wavelength of
the EM wave is not scaled to the thickness of the membrane. Such an EM wave
can be generated by a transient ion current passing through an ion channel
embedded in the membrane. b), A schematic three dimensional view of a TEM
propagating in the fluid-membrane-fluid waveguide structure along z direction,
where E is perpendicular to the membrane. c), A schematic diagram of the field
lines of E (represented by arrows) and B (represented by dash lines) in x-y plane
viewing along z direction.
EM coupling between connected membranes
Next we discuss the propagation of EM waves in the membrane network. In a
biosystem when two membranes are close or connected by proteins, a kind of
dielectric softmaterial, as shown in Figure 3a, an EM wave can inject into the
“X”-shaped structure from one side of any membrane, and comes out from both
membranes at the other side. This is an important concept showing that an EM wave
can travel through a network of coupled membrane network, as long as the membrane
network is surrounded by electrolytes. Therefore, the propagation of dendritic
impulses in a neural system is dictated by this principle.
Figure 3
images of coupled membrane waveguide
Cross-sectional
structures. a), When two membranes are connected in a biosystem, an EM
wave can inject into the “X”-shaped structure from one of the two channels at
the left side, and comes out through both channels at the right side, when the
surrounded with electrolytes. b), Schematic
membrane network
is
cross-sectional structure of a myelinated axon near a node of Ranvier, where
the axon is firmly wrapped with a myelin sheath consisted of multilayered
membranes. An EM pulse generated at the node of Ranvier can be transmitted
in both the membrane of axon and the myelin sheath. c), An electric synapse,
where at the central region there is a gap junction, and the two membranes are
connected by connexons. The inset shows the membrane.
Figure 3b shows the schematic cross-sectional structure of a myelinated axon
near a node of Ranvier, where the axon is firmly wrapped with the myelin sheath,
multilayered membranes. An axon is conventionally considered as a transmission line
and the myelin sheath acts as the insulating sheath. However it is hard to understand
that myelin sheath use even more than one hundred layers of membrane for insulation.
Dictated as above, an EM pulse generated at the node of Ranvier can be transmitted
both in the membrane of axon and the myelin sheath. Therefore, the myelin sheath
acts in fact as the dielectric core of the waveguide structure by adding the thickness of
the total membrane layers. Figure 3c shows the cross-sectional structure of an electric
synapse, where the ends of two opposite axons are coupled with a gap junction. At the
gap, with a narrow spacing around 3 nm, the local membranes of two axons are
connected by connexons. No ion channels can generate action potentials existing in
the gap junction, but experimentally it has been found that action potentials can
propagate through the gap junction from one axon to the other without any
measurable delay in time.
The mechanism of such a transmission cannot be explained clearly by
conventional model with ionic current flows. However, following the model presented
in this paper, it is a natural result that the action potential is transmitted in the form of
the EM pulse which propagates inside the fluid-membrane-fluid waveguide structure
of one axon and passes across the connexon-coupled gap junction and continuously
propagates inside the fluid-membrane-fluid waveguide of the other axon. The time
delay for such an EM-wave propagation across the coupling junction is in the order of
10 fs, therefore is not measurable.
Action potentials “saltatory” between nodes of Ranvier on myelinated axons
Now we can apply the fluid-membrane-fluid waveguide structure to explain the
propagation of action potential between neighboring nodes of Ranvier along a
myelinated axon, so-called “saltatory” mechanism, which is not clearly explained
under conventional framework with local circuit ionic current flows.
Figure 4 shows schematic pictures of such a process in four frames, where the
myelinated axon structure is viewed in its cross-sectional cartoon, and the scales of
the axon, axon membrane, myelin sheath and spacing between two nodes of Ranvier
do not strictly follow the real scales. The center of the axon is colored in light blue,
where three nodes of Ranvier are labeled as A, B, and C. The array of narrow arrows
schematically represents the electrical field of the EM pulses transmitted through the
membrane of the axon and myelin sheath.
Figure 4 Schematic cross-sectional cartoon pictures showing the propagation
of action potentials along a myelinated axon among three nodes of Ranvier
labeled as “A”, “B” and “C”. a), An EM pulse, E is represented by the array of
narrow arrows, propagates from left side passing through Node A, via the path
of fluid-membrane-fluid waveguide structure, causing excitation of ion
channels at Node A, as indicated with a thick red arrow. b), This excitation
causes a transient flow of ions thus generates a fresh EM pulse, and triggers the
excitation of ion channels at Node B. c) & d), The same process is repeated so
that fresh, EM pulses are generated at Node B and Node C, sequentially, and as
a result the EM pulse is transmitted from left to right through the membrane of
the axon and myelin sheath.
In the first frame, Figure 4a, an EM wave propagates from left side passing
through Node A, causing excitation of ion channels at this node. The excitation,
indicated with a thick red arrow, causes a transient flow of ions perpendicular to the
membrane. Thus according to Maxwell Equations, as shown in the second frame
Figure 4b, it generates a fresh pulse of EM wave, which is similar to the function of a
dipole antenna. The power of this EM pulse is determined by the number and
acceleration of ions passing through the channel in the transient process. This fresh
EM pulse propagates through the membrane and myelin sheath of the axon, at a speed
of ~ 108 m/s, passes through Node B and triggers excitation of ion channels in the
membrane at this node, too. Then, as shown in the third frame Figure 4c, a fresh EM
wave pulse is generated at Node B, which propagates shortly along the axon and
triggers excitation at Node C. Again, the fourth frame Figure 4d shows the repeated
EM wave pulses generated at Node C. If the structure of node of Ranvier, the number
and distribution of ion channels, and the triggering mechanism at each ion channel are
roughly the same 8-10, we can expect that the fresh EM wave pulses generated at
different nodes have the same strength and shape. And this has been proven by many
experimental reports.
The Propagation efficiency for EM waves in different types of axons
Next we examine the propagation efficiency for EM waves in myelinated and
unmyelinated axons of various dimensions.
In a live cell, the skeleton of membrane, the phospholipid bilayer, is usually
embedded with certain density of proteins and other molecules, and it consists of
intrinsic structural defects. As a result, an EM wave propagating inside the membrane
is remarkably scattered, leading to a limited propagation distance. As well as the fluid,
full of ions and molecules, is not a conductor as good as the metal.
Now we introduce L0 as the efficient propagation length for an EM wave
transmitted along an axon waveguide, defined as in z direction. Assume the original
described by 𝐸𝐸 = 𝐸𝐸0 𝑒𝑒 −𝑧𝑧 /𝐿𝐿0 . When z = L0, 2L0, and 3L0, E ≈ 37% E0, 14% E0, and 5%
electric field intensity E0 attenuates along the propagation path in the way that can be
E0, respectively. Therefore, 1-3 L0 is the distance that an EM wave can be transmitted
efficiently, beyond which, the energy strength of the propagating EM wave is
assumed to be too small to be detected by natural bio-sensors in a biosystem.
We have mentioned that the model shown in Figure 1b is similar to that of a
parallel plate waveguide. While wrapped on an axon surface, the membrane forms a
co-axial geometry, by which the side-leaking effect of EM energy is remarkably
eliminated. The unmyelinated axons are usually thicker than myelinated ones. And,
experiments show that the thicker the unmyelinated axons, the faster the transmission
speed for action potential.
Figure 5
The dependence of effective propagation length L0 on the axon
structure and diameter. a) & b), Schematic cross-sections of an unmyelinated
axon and a myelinated axon, respectively. Their membrane structure,
highlighted with a purple frame in each figure, is a phospholipid bilayer with
of calculated L0 versus √r of an unmyelinated axon (shown in solid black line)
thickness of d0. The inner and outer diameters are defined as r and R. c), Plots
and three myelinated axons with Ns of 10 (shown in red line), 20 (shown in
blue line) and 40 (shown in green line). d), Plots of calculated L0 versus Ns of
two myelinated axons with r = 1 μm (shown in green line) and r = 4 μm
(shown in blue line).
Figure 5a and Figure 5b show schematically the cross-sections of these two
kinds of axons, respectively. Both of their membranes are basically made of
phospholipid bilayer with thickness of d0 and two protein layers attached to both sides
of the phospholipid bilayer. From reported experimental data, d0 is measured to be
around 3 nm, and the thickness of the two protein layers, d1, is estimated to be around
7 nm. In the figures the inner and outer diameters of these axons are defined as r and
R, respectively. Hence R ≈ r+d0+d1 for an unmyelinated axon, and R ≈
r+(2Ns+1)·(d0+d1) for myelinated axon, where Ns is the number of sheath layers 11.
Based on the waveguide theory, and assuming the protein layer and the
phospholipid bilayer have the same dielectric constant, we have calculated the values
of L0 versus square root of r for both unmyelinated and myelinated axons. The results
speed for action potential linearly increases with √𝑟𝑟, matching excellently with the
are plotted in Figure 5c and Figure 5d. It qualitatively shows that the transmission
well known experimental data that reveal a square root dependence of “length
constant” for neural signal transmission on the diameter of axons 12-15. It also shows,
for EM wave transmission a myelinated axon is much more efficient than an
unmyelinated axon of the same inner diameter. For instance, the L0 of an
unmyelinated axon with r = 400 μm is roughly the same as that of a myelinated axon
with 40 layers of sheath, r = 4.0 μm and overall diameter ~ 10 μm.
Discussions
With a geometric structure similar to that of a metallic parallel plate waveguide,
the fluid-membrane-fluid waveguide is in favor of propagating TEM waves. In a
metallic parallel plate waveguide, the component of a TEM wave with electrical field
parallel to x direction can propagate for a long distance, but its component with
electrical field parallel to y direction cannot. However, as EM waves propagate in
fluids in a way different from that in a metal, so the characteristic nature of a metallic
parallel plate waveguide is not fully applicable in the waveguides shown in Figure 2.
In fluids and membranes, EM waves are expected to show frequency dispersion.
As shown in Figure 4, the “saltatory” phenomenon observed in myelinated axons
is a natural result of propagation of EM pulses between neighboring nodes of Ranvier.
It is important to note that, the propagation of an EM pulse along the membrane of
axon and the myelin sheath does not require any external bias voltage or directional
current of ions along the axon axis either inside or outside the axon. And, this process
does not cause the reversal of membrane potential in the internode region. As this
region could be as long as 2 mm, indeed the local electric field generated at any node
is screened at this distance, therefore the bias for the local circuit ion current proposed
in conventional theory does not exist.
The results shown in Figure 5c can be considered as a snap shot of such an
evolution. In nature the myelinated axons in complex biosystems are developed from
unmyelinated axons in simpler biosystems. One layer of membrane for the
unmyelinated axon is developed into multilayer of membrane for the myelinated. In
the conventional framework of neuroscience, it is hard to explain why the number of
membrane layers for the myelin sheath is observed as many as a hundred or even
more. However, as shown in Figure 5c and Figure 5d, in the view of EM waveguide,
the reason is clear: The thicker the myelin sheath the better the efficiency of the
waveguide. As a result, due to this mechanism, the number of myelinated axons a
vertebrate can contain in its spine is a thousand times more than that of unmyelinated
ones having the same transmission efficiency, and this is much favorable for
advanced communications in the whole body.
Note that in Figure 5c and Figure 5d, the plotted efficient propagation length L0
is scaled to arbitrary unit (a.u.) due to lack of the exact experimental data. In 1960s,
Kanno & Loewenstein conducted experiments on electrical coupling between gland
cells. The results showed attenuation of an EM pulse inside a biosystem with a
to L0) of 0.8-1.2 mm 16. The well-constructed
“length constant” (similar
fluid-membrane-fluid waveguides described in this work is a better path for
transmission of EM waves than the random gland cell samples Kanno used in the
experiment. Therefore, we can expect an L0 be larger than 1 mm in the best
myelinated axons.
The transmission speed, va, of an action potential signal propagating along an
axon is commonly measured to be 1-100 m/s in different kinds of axons. But an EM
wave needs only 10 ns to propagate over a 1 m. Therefore almost all the transmission
time is allocated to the generation of EM pulses at the ion channels. We define here τ
as the time for the rising edge of a single action potential pulse, which measures
0.1-0.5 ms in various experiments. Hence for myelinated axons when va = 10 m/s, it
leads to an average spacing Λ of 1-5mm between two neighboring excitation sources.
This rough estimation is well consistent with experimental results showing spacing of
0.2-2 mm between neighboring nodes of Ranvier 17. And clearly, as the ion channels
on one axon are almost identical so that τ is a constant, in order to increase the exact
transmission speed va, the best strategy for an axon is to reduce the number of nodes
of Ranvier per unit length, i.e., to increase the average spacing between two
neighboring nodes. As a result, it leads to a simple correlation: va ≈ Λ/τ.
Finally, in the following we discuss a strong supporting evidence for the proposed
model that fluid-membrane-fluid EM waveguide structures are the signal transmission
paths in a biosystem. In a live human muscle tissue, when the muscle fibers are
ordered to work (e.g. to shrink), an action potential signal is transferred through an
axon to the muscle fibers. At the end of an axon, a synapse is directly contacted to the
sarcolemma, an excitable membrane of a muscle fiber. Interestingly, the final
sub-cellular structures inside the muscle fiber that reacts with the action potential are
a number of myofibrils. Each myofibril is firmly surrounded with a sarcoplasmic
reticulum (SR), and each SR is connected via a complex network of T tubules of
varied lengths with the sarcolemma, as schematically shown in Figure 6a and Figure
6b. The action of each myofibril is triggered by voltage-dependent calcium channels,
which are special proteins embedded in the SR membrane and can releases a large
number of Ca2+ ions to the myofibril once triggered.
According to the conventional theories, the action potential can be applied to the
sarcolemma via a synapse. However, if this is the case, the transmission of electrical
signals inside a muscle fiber faces difficulties: (1), there is no axon or excitable
membrane connecting the myofibrils and the sarcolemma; (2), there is no loop for
local current flow or pressure gradient, i.e., there is no driving force for the motion of
ions or molecules in a T tubule; (3), the number of synapse connected to the muscle
fiber is much less than the number of myofibril inside each muscle fiber, so that the
length of T tubules from each myofibril to the sarcolemma is different from each other,
and (4), it is known that the diffusion speed of molecules in solution, and the motion
of ions under an external dc field, are both very slow.
Figure 6 a), Schematic diagram of a muscle fiber, where a synapse is
connected to the sarcolemma. The myofibrils inside the muscle fiber are
connected to the sarcolemma via a complex network of T tubules, and each
(SR). b), A
myofibril
reticulum
is surrounded with a sarcoplasmic
cross-sectional diagram of the black circle in a), consisting of synapse,
sarcolemma, T tubule, SR and myofibril. c), A cross-sectional diagram of the
black circle in b), showing an electrolyte fluid-phospholipid bilayer-fluid
waveguide structure.
But in a live biosystem a bundle of myofibrils in a single muscle fiber do react
together to the same external action potential signal within a time scale of 20-100 ms
18. This strongly indicates that the same single action potential signal is received by
each and every myofibril almost simultaneously. So a consisted explanation is that the
signal is not transmitted by molecule diffusion or ion transportation in the T tubules,
but by EM waves through the membrane of T tubule. The membrane of T tubule
serves as a good EM waveguide, and as a result, although the absolute distance of
different myofibrils away from the same synapse-sarcolemma junction varies much,
the difference in the transmission times of EM wave is much less than 1 ns, therefore
this delay is negligible.
Conclusion
Clearly, we have shown that to transmit information in a biosystem by EM waves
through the fluid-membrane-fluid waveguide structures is physically feasible and an
energy-efficient approach. For transmitting an EM wave in a uniform membrane, it
needs only a proper source of the EM pulses, such as a transient ionic current flow
passing through an ion channel; it does not require any external bias voltage or
directional current. The energy dissipation during the propagating of an EM wave is
much smaller than that in case of a mass transport.
References
1. Debanne, D. Information Processing in the Axon. Nature Reviews Neuroscience
5, 304-316 (2004).
2. Vincent, C. A. The Motion of ions in Solution under the Influence of an Electric
Field. Journal of Chemical Education 53, 490-493 (1976).
Pozar, D. M. Microwave Engineering (John Wiley & Sons, 3rd Ed., MA, USA,
2005).
Pfeiffer, F., David, C., Burghammer, M., Riekel, C. & Salditt, T.
Two-Dimensional X-ray Waveguides and Point Sources. Science 297, 230-234
(2002).
3.
4.
7.
8.
9.
5. Chapman, D., Thomlinson, W., Johnston, R. E., Washburn, D., Pisano, E., Gmür,
N., Zhong, Z., Menk, R., Arfelli, F. & Sayers, D. Diffraction enhanced X-ray
imaging. Physics in Medicine and Biology 42, 2015-2025 (1997).
6. Momose, A., Takeda, T., Itai, Y. & Hirano, K. Phase-contrast X-ray computed
tomography for observing biological soft tissues. Nature Medicine 2, 473-475
(1996).
Bilderback, D. H., Hoffman, S. A. & Thiel, D. J. Nanometer spatial resolution
achieved in hard X-ray imaging and Laue diffraction experiments. Science 263,
201-203 (1994).
Ellisman, M. H. Molecular Specializations of the Axon Membrane at Nodes of
Ranvier are not Dependent Upon Myelination. Journal of Neurocytology 8,
719-735 (1979).
Poliak, S. & Peles, E. The Local Differentiation of Myelinated Axons at Nodes
of Ranvier. Nature Reviews Neuroscience 4, 968-980 (2003).
10. Lai, H. C. & Jan, L. Y. The Distribution and Targeting of Neuronal
Voltage-gated Ion Channels. Nature Reviews Neuroscience 7, 548-562 (2006).
11. Michailov, G. V., Sereda, M. W., Brinkmann, B. G., Fischer, T. M., Haug, B.,
Birchmeier, C., Role, L., Lai, C., Schwab, M. H., Nave, K. –A., Axonal
Neuregulin-1 Regulates Myelin Sheath Thickness. Science 304, 700-703 (2004).
12. A. L. Hodgkin & A. F. Huxley, The Components of Membrane conductance in
the Giant axon of Loligo, Journal of Physiology 116 , 473-496, 1952
13. Hodgkin, A. L. & Huxley, A. F. The Dual Effect of Membrane Potential on
Sodium Conductance in the Giant Axon of Loligo. Journal of Physiology 116 ,
497-506 (1952).
14. Hodgkin, A. L. & Huxley, A. F. A Quantitative Description of Membrane
Current and its Application to Conduction and Excitation in Nerve. Bulletin of
Mathematical Biology 52, 25-71 (1990).
15. Goldstein, S. S. & Rall, W. Changes of Action Potential Shape and velocity for
changing core conductor Geometry. Biophysical Journal 14, 731-757 (1974).
16. Kanno, Y. & Loewenstein, W. R. Low-resistance Coupling between Gland Cells.
Some Observations on Intercellular Contact Membranes and Intercellular Space.
Nature 201, 194-195 (1964).
17. Bear, M. F., Cannors, B. W. & Paradiso, M. A. Neuroscience: Exploring the
Brain (Lippincott Williams & Wilkins Inc., 2nd Ed., Philadelphia, USA, 2001).
18. Burke, R. E., Levine, D. N., Tsairis, P. & Zajac, III., F. E. Physiological types
and histochemical profiles in motor units of the cat gastrocnemius. Journal of
Physiology 234, 723-748 (1973).
Acknowledgements
We thank W.Q. Sun, J. Tang, C.K. Ong, Bonnie W. Lu, and R.J. Dai for valuable
discussions. This work is financially supported by Peking University, the NSFC
(Grant No 11074010) and MOST of China (Grant No 2011CB933002).
Affiliations
Department of Electronics, and Key Laboratory for the Physics & Chemistry of
Nanodevices, School of Electronics Engineering and Computer Science, Peking
University, Beijing 100871, People’s Republic of China.
Contributions
S.Y.X. proposed the concept that an axon works like a fluid EM waveguide and the
EM pulse model for the saltatory mechanism on myelinated axons. J.W.X. found the
fluid-phospholipid bilayer-fluid waveguide structure and completed the theoretical
calculation. J.W.X. and S.Y.X. wrote the manuscript together.
Competing financial interests
The authors declare no competing financial interests.
Corresponding author
Correspondence to: Shengyong Xu, E-mail: [email protected]
Supplementary 1: The reflection coefficient REM of EM waves reflected at the
smooth interface of in a uniform dielectric
In a uniform media, the Maxwell equations for an EM wave can be written in the
∇ × 𝑬𝑬 = − 𝜕𝜕𝑩𝑩𝜕𝜕𝜕𝜕 = 𝑖𝑖𝜔𝜔𝜔𝜔𝜔𝜔0𝑯𝑯 (1)
form of the following:
∇ × 𝑯𝑯 = 𝜕𝜕𝑫𝑫𝜕𝜕𝜕𝜕 + 𝑱𝑱 = −𝑖𝑖𝜔𝜔𝜔𝜔𝜔𝜔0𝑬𝑬 + 𝑱𝑱 (2)
∇ ∙ 𝑬𝑬 = 𝜌𝜌𝜔𝜔𝜔𝜔0 (3)
∇ ∙ 𝑯𝑯 = 0 (4)
where E, B, H, D and J are vectors for the electric field intensity, magnetic induction
intensity, magnetic field intensity, electric displacement and conduction current
density; ρ, t, and ω are the free charge density, time and circular frequency of the EM
wave; ε0, ε, μ0 and μ are the permittivity of vacuum and the media, the permeability of
media with ρ = 0 can be written as 𝑬𝑬(𝒛𝒛, 𝜕𝜕) = 𝑬𝑬𝑒𝑒 𝑖𝑖 (𝒌𝒌∙𝒛𝒛−𝜔𝜔𝜕𝜕 ) = 𝑬𝑬𝑒𝑒 −𝜶𝜶∙𝒛𝒛𝑒𝑒 𝑖𝑖 (𝜷𝜷∙𝒛𝒛−𝜔𝜔𝜕𝜕 ) , where k
vacuum and the uniform media, respectively. In general, an EM wave propagating in a
�𝛽𝛽2 − 𝛼𝛼 2 = 𝜔𝜔2 𝜔𝜔𝜔𝜔0 𝜔𝜔𝜔𝜔0 (5)
= iα + β, with α and β being the attenuation constant and phase constant described as:
𝜶𝜶 ∙ 𝜷𝜷 = 12 𝜔𝜔𝜔𝜔𝜔𝜔0𝜎𝜎 (6)
For a plane EM wave incident perpendicularly to a conductor,
𝛼𝛼 2 = 12 𝜔𝜔2 𝜔𝜔𝜔𝜔0 𝜔𝜔𝜔𝜔0 ��1 + 𝜎𝜎 2
𝜔𝜔 2 𝜔𝜔02𝜔𝜔2 − 1� (7)
𝛽𝛽2 = 12 𝜔𝜔2 𝜔𝜔𝜔𝜔0 𝜔𝜔𝜔𝜔0 ��1 + 𝜎𝜎 2
𝜔𝜔 2 𝜔𝜔02𝜔𝜔2 + 1� (8)
where σ is the conductivity of the conductor. In vacuum, α = 0.
𝒆𝒆𝒏𝒏 × (𝑬𝑬𝑩𝑩 − 𝑬𝑬𝑨𝑨) = 0 (9)
For an EM wave incident to a surface from material A to material B,
𝒆𝒆𝒏𝒏 × (𝑯𝑯𝑩𝑩 − 𝑯𝑯𝑨𝑨) = 𝑱𝑱𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒔𝒆𝒆 (10)
𝒆𝒆𝒏𝒏 ∙ (𝑫𝑫𝑩𝑩 − 𝑫𝑫𝑨𝑨 ) = 𝛴𝛴𝑠𝑠𝑠𝑠𝑟𝑟𝑠𝑠𝑠𝑠𝑠𝑠𝑒𝑒 (11)
𝒆𝒆𝒏𝒏 ∙ (𝑩𝑩𝑩𝑩 − 𝑩𝑩𝑨𝑨) = 0 (12)
where EA, HA, DA and BA, are the electrical field intensity, magnetic field intensity,
electric displacement and magnetic induction intensity in material A, respectively, and
EB, HB, DB and BB are the parameters for material B. Jsurface is the surface free current
density, Σsurface the surface free charge density, and en the normal vector pointing to
the inside of material B.
Defined E, E’ and E’’ are the electrical field intensities for the incident, reflected
and transmitted EM waves, respectively. So EA = E+E’, and EB = E’’. The reflection
coefficient REM is defined as:
𝑅𝑅𝐸𝐸𝐸𝐸 = �𝑬𝑬′𝑬𝑬 �2 (13)
An ideal waveguide has REM of 1, while for practical waveguides REM < 1.0, i.e.,
part of the EM wave energy is lost in the conductor by Joule heating, so that the
strength of the EM wave decreases along the transmission path. And, the EM wave
energy is also dissipated via scattering process in the dielectric.
It has been proven that an EM wave can only penetrate a thin layer of metal. The
thickness of this thin layer is defined as the penetration depth δ, and for a good
conductor 𝛿𝛿 ≈ � 2𝜔𝜔𝜔𝜔 𝜔𝜔 0 𝜎𝜎 . As a result, an incident EM wave completely reflects at the
surface of a perfect conductor, similar to an incident light reflected by a mirror. Thus
an EM wave can be trapped between two walls of conductors. A conductor-dielectric
(or vacuum)-conductor sandwich structure can form a waveguide, so that an EM wave
can travel inside the dielectric layer with little energy leakage. The efficiency of a
waveguide for confining the energy of an EM wave in its inner space can be
characterized by REM. An ideal waveguide has REM of 1, while for practical
waveguides REM < 1.0, i.e., part of the EM wave energy is lost in the conductor by
Joule heating, so that the strength of the EM wave decreases along the transmission
path. And, the EM wave energy is also dissipated via scattering process in the
dielectric media. Along the transmission path, the energy density for a plane EM wave,
often characterized by E2, keeps constant in vacuum and decreases slightly in a
waveguide over a long distance.
According to the Nernst-Planck Equation, the conduction current density J of the
𝐽𝐽(𝑧𝑧) = � 𝑁𝑁𝐴𝐴 𝑒𝑒0 𝑍𝑍𝑖𝑖 � −𝐷𝐷𝑖𝑖 𝜕𝜕𝐶𝐶𝑖𝑖 (𝑧𝑧)𝜕𝜕𝑧𝑧 − 𝐹𝐹𝑅𝑅𝑅𝑅 𝐷𝐷𝐶𝐶𝑖𝑖 𝜕𝜕𝜑𝜑𝜕𝜕𝑧𝑧 + 𝐶𝐶𝑖𝑖 𝑉𝑉(𝑧𝑧)� (14)
electrolyte fluid layer along z direction can be written as:
i
where NA is Avogadro’s constant, e0 being the proton charge; Di, Zi, and Ci are
parameters for the ith kind of ions, namely the coefficient of diffusion, valence and
concentration, V(z) is the fluid velocity along z direction; F being Faraday constant, R
the gas constant, φ the electrical potential, and T the absolute temperature. We
consider an electrolyte solution in which the concentrations of cations and anions are
the same, defined as Cbulk. If there is no gradient of concentration in z direction and
𝜎𝜎𝑏𝑏𝑠𝑠𝑏𝑏𝑏𝑏 (0), is estimated to be 𝜎𝜎𝑏𝑏𝑠𝑠𝑏𝑏𝑏𝑏 (0) = 2𝜔𝜔𝐸𝐸 𝑁𝑁𝐴𝐴 𝑒𝑒0𝐶𝐶𝑏𝑏𝑠𝑠𝑏𝑏𝑏𝑏 , where μM is the mobility of
V(z) = 0, in a uniform applied electric field E, the dc conductivity of bulk electrolyte,
ions 1,2.
𝜎𝜎ac (𝜔𝜔) = 𝜎𝜎1 (𝜔𝜔) + 𝑖𝑖𝜎𝜎2 (𝜔𝜔) (15)
The ac conductivity σ(ω) of a bulk electrolyte fluid can be described as 3:
𝜎𝜎(0)
𝜎𝜎1 (𝜔𝜔) =
1 + (𝜔𝜔𝐷𝐷𝐷𝐷/𝑏𝑏𝐵𝐵 𝑅𝑅)2 − [2𝜔𝜔2 − 𝜔𝜔4 /𝛺𝛺2 ]/𝛺𝛺2 (16)
𝜎𝜎2 (𝜔𝜔) = 𝜔𝜔𝜎𝜎(0)[𝐷𝐷𝐷𝐷/𝑏𝑏𝐵𝐵 𝑅𝑅 − 𝑏𝑏𝐵𝐵 𝑅𝑅(𝛺𝛺2 − 𝜔𝜔2 )/𝐷𝐷𝐷𝐷(𝛺𝛺2 )2 ]
(17)
1 + (𝜔𝜔𝐷𝐷𝐷𝐷/𝑏𝑏𝐵𝐵 𝑅𝑅)2 − [2𝜔𝜔2 − 𝜔𝜔4 /𝛺𝛺2 ]/𝛺𝛺2
where Ω is the Einstein Frequency, kB being Boltzmann’s constant. Only the
conduction current, which has the same phase with that of the electric field, consumes
the energy.
Supplementary 2: An EM waveguide structure consisting of two heavily charged
dielectric layers and a fluid with a large gradient of ion concentration
Figure S1 a), Schematic model of another possible fluid EM waveguide, where a
fluid layer is trapped between two dielectric layers, and a large gradient of ion
concentration exists from the edge of the fluid towards the center. b), A plot of
calculated REM value versus surface charge density Σs for the waveguide model
proposed in a), where Cbulk is set at 1.0×10-7 mol/l.
Briefly, in a practical system, when a solid surface is in contact with electrolyte, it
is often charged with a surface charge density of Σs. In a dilute electrolyte solution,
the average concentration of counterions in the electrical double layer (EDL) 4 is
much larger than that in the bulk solution, so a large gradient of ion concentration
exists from the edge of the channel towards the center. The thickness of an EDL can
be estimated to be the Debye length, 1/κ, where:
𝜅𝜅 2 = 2𝑒𝑒02𝑁𝑁𝐴𝐴 𝐶𝐶𝑏𝑏𝑠𝑠𝑏𝑏𝑏𝑏
𝜔𝜔0 𝜔𝜔𝑏𝑏𝐵𝐵 𝑅𝑅 (18).
As a result, the EDL has a much higher conductivity when compared to the dilute
bulk solution, and it can serve as the conducting cladding layer of a waveguide. The
conductivity of EDL, σEDL can be revealed from Σs or the zeta-potential, ζ 5. Figure
S1b shows that, this kind of fluidic waveguide only works well under the condition
when the surface charge density is high and the ion concentration of bulk solution is
low. By contrast, our calculation shows that when Cbulk is bigger than 0.1mol/l, the
EDL phenomenon has negligible effect on REM for the structure shown in Figure 1b.
In this waveguide structure, a large gradient of ion concentration exists from the
edge of the channel towards the center. In practical systems, when a solid surface is in
contact with electrolyte, it is often charged with a surface charge density of Σs. In a
dilute electrolyte solution, the average concentration of counterions in the EDL is
much larger than that in the bulk solution, as shown in Figure S1a, where the channel
surface facing the electrolyte is assumed negatively charged. The thickness of an EDL
can be estimated to be the Debye length, 1/κ. Compared to the dilute bulk solution,
the EDL has a much higher conductivity, and it can serve as the conducting cladding
layer of a waveguide. Thus the fluidic channel structure shown in Figure 1c is indeed
an EM waveguide. The conductivity of EDL, σEDL can be revealed from the surface
charge density Σs or the zeta-potential, ζ. When the surface charge density is high and
the ion concentration of bulk solution is low, the waveguide works well. Figure S1b
shows an extreme example, where Cbulk is set at 1.0×10-7 mol/l and REM is plotted
versus Σs. In this case, one sees that REM increases with Σs, from 0.54 at 0.01 e/nm2 to
0.95 at 0.1 e/nm2.
References for the Supplementary
1.
3.
Levine, S., Marriott, J. R. & Robinson, K. Theory of Electrokinetic Flow in a
Narrow Parallel-plate Channel. Journal of the Chemical Society, Faraday
Transaction 2: 71, 1-11 (1975).
Bocquet, L. & Charlaix, E. Nanofluidics, from bulk to interfaces. Chemical
Society Review 39, 1073-1095 (2010).
Chandra, A., Wei, D. Q. & Patey, G. N. The frequency dependent conductivity
of electrolyte solutions. The Journal of Chemical Physics 99, 2083-2094 (1993).
Schoch, R. B., Han, J. Y. & Renaud, P. Transport phenomena in nanofluidics.
Review of Modern Physics 80, 839-883 (2008).
5. Kirby, B. J. & Hasselbrink Jr., E. F. Zeta potential of microfluidic substrates: 1.
Theory, experimental techniques, and effects on separations. Electrophoresis 25,
187-202 (2004).
2.
4.
|
1905.02007 | 1 | 1905 | 2019-05-06T12:57:12 | Axonal Computations | [
"q-bio.NC"
] | Axons functionally link the somato-dendritic compartment to synaptic terminals. Structurally and functionally diverse, they accomplish a central role in determining the delays and reliability with which neuronal ensembles communicate. By combining their active and passive biophysical properties, they ensure a plethora of physiological computations. In this review, we revisit the biophysics of generation and propagation of electrical signals in the axon, their complex interplay, and their rich dynamics. We further place the computational abilities of axons in the context of intracellular and intercellular coupling. We discuss how, by means of sophisticated biophysical mechanisms, axons expand the repertoire of axonal computation, and thereby, of neural computation. | q-bio.NC | q-bio | Axonal Computations
Pepe Alcami1,2,∗, Ahmed El Hady3,4,∗
1. Neurobiology, Ludwig-Maximilians University, Martinsried, Munich, Germany
2. Behavioral Neurobiology, Max Planck Institute for Ornithology, Seewiesen, Germany
3. Princeton Neuroscience Institute, Princeton University, Princeton , New Jersey , USA
4. Howard Hughes Medical Institute, Princeton University, Princeton , New Jersey , USA
*correspondence should be addressed to Pepe Alcami ([email protected])or Ahmed El Hady ([email protected])
Abstract
Axons functionally link the somato-dendritic compartment to synaptic terminals. Structurally and func-
tionally diverse, they accomplish a central role in determining the delays and reliability with which neuronal
ensembles communicate. By combining their active and passive biophysical properties, they ensure a plethora of
physiological computations. In this review, we revisit the biophysics of generation and propagation of electrical
signals in the axon, their complex interplay, and their rich dynamics. We further place the computational abilities
of axons in the context of intracellular and intercellular coupling. We discuss how, by means of sophisticated
biophysical mechanisms, axons expand the repertoire of axonal computation, and thereby, of neural computation.
1
Introduction
Neurons are compartmentalized into input compartments formed by dendrites and somas, and an output com-
partment, the axon. However, in the current era, there is a widespread tendency to consider that the biophysics
of single neurons do not matter to understand neuronal dynamics and behavior. Neurons are often treated as
point processes with disregard of the complex biophysical machinery that they have evolved. Moreover, neuronal
computations are assumed to be mostly performed by dendrites or at synapses (Sudhof and Malenka, 2008; Stuart
et al., 2016), and axons are reduced to simple, static and reliable devices. However, a wealth of literature supports
that this is not the case: axons form complex structures that ensure a variety of sophisticated functions. Here,
we aim to review evidence that axons perform complex computations which depend on a myriad of biophysical
details and ensure the generation and propagation of neuronal outputs. We will not discuss biophysics of synaptic
release, reviewed elsewhere (Sudhof and Malenka, 2008).
More than six decades after seminal discoveries in experimentally accessible invertebrate axons (Hodgkin and
Huxley, 1952), axonal research has unraveled previously-unsuspected, rich and dynamical electrical signaling in
axons, which consists of a hybrid of analog and digital signaling. Contrary to the giant invertebrate axons studied
in the early days (Fig. 1A; (Hodgkin and Huxley, 1952; Furshpan and Potter, 1959)), most invertebrate and
vertebrate axons are thin and present complex extended arborizations (e.g., Fig. 1B-D), making it difficult to
record from them. Electron and optical microscopy have made it possible to deepen our understanding of fine
axonal structures (Rash et al., 2016; D'Este et al., 2017). Moreover, the combination of new structural imaging
techniques with electrophysiology has made it possible to study how structural changes at the sub-micrometer
scale impact function (Ch´ereau et al., 2017).
We will first illustrate general signaling principles in axons, before delving into the generation and propagation
of action potentials (APs) and their dynamical regulation. Finally, we will position axons in the context of their
interactions with other compartments and with each other, by virtue of axo-axonal coupling (Katz and Schmitt,
1940; Furshpan and Potter, 1959) and finally, indirectly via glial cells.
9
1
0
2
y
a
M
6
]
.
C
N
o
i
b
-
q
[
1
v
7
0
0
2
0
.
5
0
9
1
:
v
i
X
r
a
2 Review of Axonal Computations
2.1 GENERAL PRINCIPLES: SIGNALING IN AXONS
2.1.1 Of axons and brains
The propagation of electrical signals along axons controls the reliability and the timing at which neural networks
communicate. The spatial extent of axonal trees introduces delays between the generation of an electrical signal
(APs or subthreshold signals) in a neuron and the arrival of this signal to the presynaptic site, where information
is relayed to a postsynaptic cell. Thereby, the axonal propagation delay influences the temporal relationship of
presynaptic and postsynaptic activity (Izhikevich et al., 2004).
The delays in AP propagation encode information (Seidl et al., 2010) and contribute to re-configuring neural
circuits through plastic mechanisms such as spike timing dependent plasticity (Bi and Poo, 1998; Izhikevich et al.,
1
2004). Each synapse has its own 'critical window', given by the delay between presynaptic and postsynaptic
activity, to induce synaptic plasticity. The speed at which signals travel along axons can vary over four orders
of magnitude, from tens of centimeters per second in thin unmyelinated axons to hundreds of meters per second
in giant myelinated fibers (Schmidt-Hieber et al., 2008; Xu and Terakawa, 1999)); highlighting the ability of
biophysical specializations characterizing different axons to conduct APs at different speeds.
Axons have evolved different intricate geometries (Fig.1), revealing specific structure-function specializations.
Constraints to axonal morphology include (1) spatial constraints: axons occupy large volumes of nervous system,
yet they require internal components to their function such as mitochondria, which constrain their minimal
functional size; (2) energetic requirements may have evolutionary exerted a selection pressure on axonal properties
(Perge et al., 2009; Harris and Attwell, 2012); (3) efficiency of electrical information processing in relation to
structure-function specializations. The last point on electrical information processing will be the main focus of
the current review.
2.1.2 Propagation of electrical signals in the axon
Propagation of electrical signals in the axon results from a combination of specialized active and passive mech-
anisms. Active properties are shaped by neuronal voltage-gated ion channels recruited as a function of the
dynamics of membrane potential whereas passive properties are determined by the axonal membrane at rest and
by axonal geometry.
As increasingly appreciated, APs are not just an on/off switch, and electrical signaling in the axon should be
regarded as a hybridization of analog and digital signalling (Shu et al., 2006). Furthermore, APs are strongly
regulated by the background analog activity provided by subthreshold postsynaptic events. For example, the
inactivation of potassium Kv1 channels in the axon initial segment broadens the axonal AP waveform and
increases unitary excitatory postsynaptic potentials (EPSP) amplitude in layer V pyramidal neurons (Kole et
al. 2007). Note that some neurons do not encode information with APs but only with graded signals (graded
potential neurons)(Borst and Haag, 1996) .
2.1.3 Non-electrical signaling
Signaling in axons and neurons is usually regarded as purely electrical. However, electrical signals are accom-
panied by other biophysical changes in the neuronal membrane. The AP is accompanied by changes in many
biophysical properties such as temperature, mechanical membrane properties and optical birefringence (Cohen
et al., 1970; Tasaki and Byrne, 1992; Howarth, 1975; Howarth et al., 1975; Abbott et al., 1965; Tasaki and Iwasa,
1982b). For example the changes in the optical properties of the axon during propagation have been at the basis
of the label free interferometric imaging of APs in in-vitro systems (Oh et al., 2012; Akkin et al., 2007; Batabyal
et al., 2017).
Mechanical displacements associated with the AP have been measured in many experimental contexts (Tasaki
and Iwasa, 1982a,b; Iwasa and Tasaki, 1980; Hill et al., 1977). These mechanical displacements can be regarded
as propagating surface modes that are elicited via the large electrostatic force produced by the AP. In this
regard, the AP is an electro-mechanical pulse (El Hady and Machta, 2015). Converging evidence suggests that
mechanics play a role in electrical signalling (Tyler, 2012). Moreover, there is mounting evidence that some
voltage sensitive channels such as sodium and potassium channels are mechanically modulated locally and that
many neurons express mechanically-activated channels (Ranade et al., 2015; Schmidt and MacKinnon, 2008;
Schmidt et al., 2012). In addition, mechanical pressure at the soma of high-resistance neurons can induce APs
(Alcami et al., 2012), an effect likely to be even stronger at axons, which are highly resistive. Theoretical and
biophysical evidence of non-electrical signaling has surprisingly not yet led to a thorough investigation of the
functional relevance of these mechanical displacements and other non-electrical signaling modalities to axonal
computation.
2.2 BIOPHYSICS OF ACTION POTENTIAL GENERATION
2.2.1 A brief historical perspective on the action potential
At the resting, non-excited state, mostly potassium channels are open and the resting potential is, as a conse-
quence, close to the reversal potential for potassium, maintained around -70 mV. In 1939, Hodgkin and Huxley
published the first trace of an AP recorded from the squid axon using an intracellular electrode where one can
see a very clear overshoot. Following this in 1949, the proposal that sodium ions are the main mediator of AP
generation was put forward by Hodgkin and Katz (Hodgkin and Katz, 1949). They studied the effect of system-
atically varying the concentration of sodium ions and measured its impact on the amplitude of the AP recorded.
Subsequently, Keynes managed to show that nerve excitation leads to an increase in the transmembrane flow
of sodium ions by tracing the movement of the radioactive isotope N a24 in repeatedly stimulated squid axons
(Keynes, 1951). These seminal findings confirmed that sodium ions are the main contributors to AP generation.
2
2.2.2 Variations of action potentials
APs are characterized by a width, height and an overshoot magnitude (i.e. referring to the AP height beyond 0
mV). Despite sharing a general biophysical mechanism for initiation, the shape of APs varies in different types of
neurons. For example, there are cells that exhibit very narrow APs of a few hundreds of microseconds in width,
reflecting a fast spiking behavior, such as some GABAergic interneurons, Purkinje neurons which are GABAergic
projection neurons, glutamatergic neurons of the subthalamic nucleus and Medial nucleus of the trapezoid body
(MNTB) cells in the superior olivary complex. CA1 pyramidal neurons in the hippocampus have, in contrast,
relatively wide AP and dopaminergic neurons have an even wider AP of up to 4 ms ( reviewed in (Bean, 2007)).
Apart from the shape of the single spike, neurons can exhibit a diversity of firing patterns. For example, they
can be bursting or non-bursting on one hand and they can be adapting or non-adapting on the other hand.
Adaptation during a train of APs refers to the process by which AP properties, typically rate and amplitude,
decrease within the spike train. There are many biophysical mechanisms by which spike frequency adaptation can
happen. The most prominent mechanisms are an increase in outward current flowing through calcium-activated
potassium channels and an increasing outward current produced by the electrogenic sodium-potassium pump
(Powers et al., 1999). Moreover, there is also a substantial sodium channel inactivation induced by a long lasting
depolarization (Sawczuk et al., 1997). The diversity of AP shapes along with the firing patterns are a result of
the different combinations of ion channels that the neuron expresses (reviewed in (Marder and Goaillard, 2006)).
2.2.3 The axon initial segment
Before delving into the mechanistic details of AP generation, it is crucial to appreciate the complex anatomy of
the Axon Initial Segment (AIS), the site of AP initiation (Fig. 3A). Generally speaking, the AIS refers to the
first 20-60 µm of the axon but the specific length of the AIS is variable across neurons as it has to adapt it to
its own excitability properties. The excitability of the AIS is modulated by varying compositions and subtypes
of sodium and potassium channels clustered at the AIS. As an example, motor alpha and gamma neurons have
an AIS length of 28 - 29 µm but alpha and gamma motor neurons differ in the percentage of N av 1.6 and N av
1.1 channels subtypes distribution (Duflocq et al., 2011) which underlies the different excitability properties of
both subtypes of motor neurons. It is important to note that the different excitable properties are owed to the
fact that N av 1.6 and N av 1.1 channels have different kinetics.
The AIS contains a highly-specialized protein machinery that gives it a distinct character. One such protein
is Ankyrin G, a scaffolding protein that is also present at the nodes of Ranvier (Kordeli et al., 1995). Ankyrin
G anchors voltage-gated channels such as voltage-gated sodium (N av) and potassium channels (Kv) to the
membrane along with other adhesion molecules (Davis et al., 1996). Beta IV spectrin is another protein expressed
in the axon initial segment. Beta IV spectrin's main function is to cluster sodium channels at the axon initial
segment while simultaneously binding to the actin cytoskeleton (reviewed in (Rasband, 2010))( Fig.3 A).
The AIS in mammalian neurons has been established as the site of the AP initiation following a series of
seminal studies that began in the mid 1950s (Araki and Otani, 1955; Coombs et al., 1957; Fatt, 1957). The
location of AP initiation has been further confirmed by combining precise electrophysiological measurements and
imaging technologies. These allow to precisely identify the locus of AP initiation which was established to be
in the distal part of the AIS, 20 to 40 µm from the soma (Atherton et al., 2008; Foust et al., 2010; Kole et al.,
2007; Meeks and Mennerick, 2007; Palmer et al., 2010; Palmer and Stuart, 2006). The AIS is considered to
be the site of AP initiation also because of several key properties. It contains a high sodium channel density.
Furthermore, AIS sodium channels show a voltage dependence shifted to lower voltages, favoring their activation
at less depolarized voltages than at the soma. Finally, the relatively large electrotonic distance of the AIS
from the capacitive load of the soma renders distal sodium influx more efficient in evoking a local membrane
depolarization, compared to an AIS that would start at the soma. Note that electrophysiological recordings from
invertebrates have shown that the initiation of the AP can happen at multiple locations acting in an independent
manner (Calabrese and Kennedy, 1974; Maratou and Theophilidis, 2000; Meyrand et al., 1992). Along with its
specialized protein machinery and acting as the site of AP initiation, the AIS also acts as a diffusion barrier
between the somatodendritic and axonal compartments, filtering transport materials passing from soma to axon
(Song et al., 2009; Brachet et al., 2010). Note however that the AIS length can be regulated in an activity-
dependent manner (Kuba et al., 2014) and that in some neurons, AP initiation has been reported to occur at
the first node of Ranvier (Lehnert et al., 2014; Clark et al., 2005) , as will be developped later.
2.2.4 The sodium ionic dynamics and action potential initiation
The extent to which the density of sodium channels is higher in the AIS and the contribution of this high
density of sodium channels to AP initiation is a matter of active investigation. There is a consensus that sodium
channel density is higher in the AIS but the order of magnitude is still unclear.
Immunostaining of sodium
channels consistently indicates that there is higher channel density in the AIS of various neuron types (Meeks
and Mennerick, 2007; Wollner and Catterall, 1986; Boiko et al., 2003). Lorincz and Nusser in (Lorincz and Nusser,
3
2010) counted around 200 N av 1.6 sodium channels per micrometers square in the AIS of hippocampal pyramidal
cells using electron microscopy. Given that the conductance of a single N av channel is around 15 pS ( Colbert
and Johnston, 1996), one would expect a conductance density of about 3000 pS per micrometers square. On
the contrary, electrophysiological measurements from membrane patches pinpoint that sodium channel density
is about 3 4 channels per micrometers square in the AIS, which is the same as the somatic density (Colbert
and Johnston 1996, Colbert and Pan 2002). A similar conclusion of equal densities was reached on the basis of
recordings from blebs that form when cortical axons are cut. This might be due to the inability to draw AIS N a+
channels into the patch-clamp recording pipette due to their tight coupling to the actin cytoskeleton (Kole et al.,
2008). In this article, the authors record a much larger sodium current after disruption of the actin cytoskeleton
(Kole et al., 2008).
Kinetics of sodium currents underlying AP generation have been extensively studied. Hodgkin and Huxley
have proposed that the activity of sodium channels can be fitted by m3 activation kinetics. Baranauskas and
Martina (2006) (Baranauskas and Martina, 2006) found that sodium currents in three types of central neurons
(prefrontal cortical cells, dentate gyrus granule cells and CA1 pyramidal cells) activate faster than predicted by
Hodgkin-Huxley type kinetics following m2 activation kinetics. Moreover, it was found that the half activation
voltage of voltage-gated sodium channels in layer 5 pyramidal neurons is 7 -14 mV lower in the distal AIS
compared to the soma and decreases further with increasing distance from the soma (Colbert and Pan, 2002;
Hu et al., 2009). N av1.6 channels, which are predominantly expressed in the axon initial segment and have a
lower half activation voltage (Rush et al., 2005), are proposed to be primarily responsible for the initial slope of
the AP. Sodium channels at the AIS are more capable of producing a persistent sodium current (Astman et al.,
2006; Stuart and Sakmann, 1995). The persistent sodium current has a significant influence on the AP threshold
(Kole et al., 2008). Moreover, it is implicated in the generation of the AP afterdepolarization and it therefore
contributes directly to the generation of high frequency AP bursts (Azouz et al., 1996).
Using sodium imaging to follow sodium influx in axon, soma and basal dendrites (Fleidervish et al., 2010), the
authors suggest that the ratios of N av channel densities in these regions are approximately 3:1:0.3. Interestingly
in another study by (Lazarov et al., 2018), APs were initiated in the AIS, even when axonal N av channel density
was reduced to about 10 percent in a beta IV spectrin mutant mouse. This experimental finding indicates to
a great extent that AP initiation in the AIS does not require such a high local channel density. However in
that study, the precision of AP timing was substantially compromised when axonal channel density was reduced.
Likewise, Kopp-Scheinpflug and Tempel found a decrease in the temporal accuracy of AP generation from MNTB
cells in a beta IV spectrin mutant mouse (Kopp-Scheinpflug and Tempel, 2015)
The aforementioned section highlights the complexity of sodium ion dynamics and begs for novel imaging
modalities that allow tying ultrasfast dynamics with the axonal ultrastructures in order to get insights into the
intricate biophysical mechanisms underlying the very first microseconds of AP initiation.
2.2.5 The action potential rapidness
A simple but very informative way to study the properties of the APs is to plot the time derivative of the voltage
(dV/dt) versus the voltage. This is called phase-plane plot. The spike threshold can be easily visualized in such
a representation, where it corresponds to the voltage at which dV/dt rises abruptly. Coombs et al. (Coombs
et al., 1957) noticed that the main spike is preceded by a smaller earlier component (referred to as a kink). This
component is interpreted as reflecting initiation of the spike in the initial segment of the axon. One can record
such a component in somatic spikes in many central neurons, including neocortical pyramidal neurons which we
will focus our discussion on here. As mentioned above, one of the most striking features of the AP recorded from
cortical neurons has been the existence of a kink at the initiation of the AP (Fig. 3B1). One can define the
rapidness of AP onset in this cell as the slope of the phase-plane plot at dV/dt = 10 mV /ms. In (Naundorf
et al., 2006), the AP rapidness was measured from the cat visual cortical cells. AP rapidness varied between
around 20 to 60 ms−1. It is important to mention that sharp, step-like onsets of APs have been recorded in vivo
in many preparations (cat visual cortex: (Azouz and Gray, 1999), and cat somatosensory cortex: in (Yamamoto
et al., 1990)).
Several hypothesis have been proposed to explain the origin of the AP kink: 1) the backpropagation to the
soma of a smoother AP generated at the AIS; 2) an abrupt opening of sodium channels due to the biophysics
of neuronal compartmentalization; 3) a decreased membrane time constant due to the loading of the dendritic
compartment and 4) the cooperativity of sodium channels at the AIS.
The lateral current hypothesis states that the kink at spike onset reflects lateral current coming from the axon
which becomes sharper through backpropagation from the initiation site to the soma while initiation is smooth
at the initiation site (Yu et al., 2008; McCormick et al., 2007) . In Yu et al. 2008, authors perform simultaneous
recordings from axon blebs and soma, finding a smoother AP onset in the axon, additionally reproducing these
results in a model. It is important to note that this study supporting the lateral current hypothesis was done in
blebs recordings which are injured axons that may have undergone severe cytoskeletal reorganization (Spira et al.,
2003) affecting sodium channels dynamics. This reorganization might alter the true dynamics of AP initiation.
4
Although the kink indeed might reflect the lateral current coming from the axon (Milescu et al., 2010),
this hypothesis fails to account for the ability of neurons to follow 200 - 300 Hz frequency inputs. In order to
account for this discrepancy, Romain brette in (Brette, 2013) proposed that the compartmentalization and the
distance between the soma and the AIS leads to spike initiation sharpness. In his proposal, Brette suggests that
the rapidness arises from the geometrical discontinuity between the soma and the AIS, rather than from the
backpropagation of axonal APs. When sodium channels are placed in a thin axon, they open abruptly rather
than gradually as a function of somatic voltage, as an all-or-none phenomenon.
Another proposal that takes into account the geometry of neurons is Eyal et al. (Eyal et al., 2014) in which
the authors propose that increasing the dendritic membrane surface area (the dendritic impedance load) both
enhances the AP onset in the axon and also shifts the cutoff frequency of the modulated membrane potential to
higher frequencies. This dendritic size effect is the consequence of the decrease in the effective time constants of
the neuron with increasing dendritic impedance load. The authors have shown this in a computational model
of reconstructed layer 2 / 3 pyramidal neurons of humans and rats. The firing pattern at the axon is strongly
shaped by the size of the dendritic tree. Authors predict that neurons with larger dendritic trees have a faster
AP onset.
Another proposal to interpret AP onset rapidness is that sodium channels, which are assumed to be opening
independently within the Hodgkin-Huxley framework, are gated cooperatively. The cooperativity model proposed
that the half-activation voltage of the channels becomes dependent on the probability of the opening of the
neighboring channels. These cooperative effects might happen mechanistically on very fast timescales either
through a purely electrical, mechanical or electro-mechanical coupling. Though there is no direct experimental
test of the cooperativity of neuronal sodium channels at the AIS, it can theoretically account for the observed
discrepancy between the sodium channel density and the very rapid rise of the AP at the site of initiation in
the initial segment (Naundorf et al., 2006). It is important to note that cooperative gating has been previously
observed in calcium, potassium and HCN channels (Marx et al., 2001; Kim et al., 2014; Dekker and Yellen, 2006).
Apart from its mechanistic underpinnings, the fast rise of APs has attracted both experimental and theoretical
approaches to study its functional implications on the biophysics of neuronal populations. Before detailing those
functional implications, it is important here to go through a useful theoretical abstraction: a typical cortical
neuron, embedded in a cortical network in vivo, receives about 10.000 synaptic inputs. Assuming that each of
these synaptic inputs is active with a rate on the order of 1 to 10 Hz, incoming signals arrive at a rate of 10
kHz. As a result, the membrane voltage exhibits strong, temporally irregular fluctuations. To understand the
computational capabilities of e.g. cortical circuits, it is essential to characterize single neuron computation under
such realistic operating conditions. To control the activity of entire neuronal circuits while preserving their natural
firing characteristics, it would be advantageous to introduce artificial input components mimicking intrinsically
generated synaptic input under precise experimental control. In order to study the dynamical properties of cortical
neurons, experimenters have mimicked synaptic bombardment in vitro by injecting stochastic inputs modeled
as an Ornstein-Uhlenbeck process in which sinusoidal inputs are embedded (Neef et al., 2013; Tchumatchenko
et al., 2010). This experimental setting has allowed the measurement of the dynamic gain of neurons, which
means how much neurons attenuate their input in the frequency domain and how fast they are able to follow
a rapidly-fluctuating input. This has led to the establishment of the ability of cortical neurons to follow high
frequency inputs up to 200-300 Hz (Higgs and Spain, 2009; Higgs et al., 2006; Tchumatchenko et al., 2011).
Moreover, there has been a series of theoretical studies exploring the dependence of encoding capacity on the
active properties of the AP initiation (Fourcaud-Trocm´e et al., 2003; Huang et al., 2012; Wei and Wolf, 2011) .
2.2.6 AP trajectories beyond the rapidness of initiation
Although we have mostly concentrated initial spike rapidness, note that the phase plot can display a variety
of trajectories after the onset of the AP (Fig.3). These trajectories are determined by additional ion channels
in conjunction with sodium channels. Potassium channels are typically responsible for the repolarizing phase
of the AP. The relative temporal profiles of activation of sodium and potassium channels and their subunit
composition determine the width of the AP (Lien and Jonas, 2003). Furthermore, the spatial location of ion
channels also contributes to the shape of AP trajectories (Yang et al., 2016)(Fig. 3B2). It has also been shown
that the AP shape affects calcium currents and transmitter release in a calyx-type synapse in the rat auditory
brainstem (Borst and Sakmann, 1999). Therefore, the AP shape beyond the initial spike rapidness provides an
extra dimension for information encoding on a variety of timescales.
2.2.7 The action potential at nodes of Ranvier
Although the AP is generated at the AIS, the nodes of Ranvier contain a machinery to regenerate the AP. It
is important to note that there are striking similarities between axon initial segment and nodes of Ranvier. A
great deal of the protein machinery in the axon initial segment is also present in the nodes of Ranvier where
the regeneration of the AP is performed. Ankyrin G, the major scaffolding protein in the AIS and the ring-like
arrangement of actin and beta IV spectrin are also found in the nodes of Ranvier (DEste et al., 2017). Nodes
5
of Ranvier are distributed spatially along the axon to guarantee the faithful propagation of the AP. In addition,
the first node of Ranvier was found to be crucial for high bandwidth bursting activity in neocortical layer 5
pyramidal neurons (Kole, 2011). In that study, authors show that nodal persistent sodium currents at the first
node of Ranvier hyperpolarize AP threshold and amplify the afterdepolarization. The study opens up the space
for a computational role of the first node of Ranvier beyond conduction of the propagating AP. The first node
faithfully follows spike frequencies with a 100 µs delay (Foust et al., 2010; Khaliq and Raman, 2006; Palmer et al.,
2010; Palmer and Stuart, 2006). This had led to the speculation that the nodes of Ranvier, and in particular the
first node, may have an active computational role in modulating the AP initiation itself.
It is worth noting that nodes of Ranvier also express N aV 1.6, the same sodium channel subtype that is ex-
pressed in the AIS. It is even more striking that the sodium channels at the nodes also undergo the developmental
changes from N aV 1.2 to N aV 1.6 during postnatal development, following a similar developmental trajectory as
those found in the AIS (Rios et al., 2003). N aV is highly clustered at the nodes of Ranvier in the order of 1200
channels per micrometers square (Rosenbluth, 1976), while internodes contain 2025 channels per micrometers
square (Ritchie and Rogart, 1977)). This very high density ensures high-fidelity conduction of the AP which will
be further developed in the review. Given the close proximity of the first node to the cell body and high density
of N aV channels, it has been postulated that, in addition to securing propagation, it could potentially generate
the AP (Clark et al., 2005; Colbert and Pan, 2002; Lehnert et al., 2014) Although this might be happening, the
overwhelming evidence favors that the AP initiation is happening at the AIS.
2.2.8 Ectopic spiking
AP initiation at the AIS and its subsequent orthodromic propagation have been extensively investigated. How-
ever, a number of studies demonstrates that distally generated spikes or ectopic APs co-exist in invertebrate
and vertebrate neurons (Mulloney and Selverston, 1972; Maranto and Calabrese, 1984; Meyrand et al., 1992;
Sheffield et al., 2010; Dugladze et al., 2012; Lehnert et al., 2014). Marked differences in somatic AP recordings
are observed when those are generated at the AIS or in a distal part of the axon. In particular, APs have more
negative thresholds when they are generated in distal parts of the axon. This is likely due to the fact that they
do not show the strong coupling of soma to AIS which is responsible for the inactivation of sodium channels
at the AIS and for a more depolarized AP generation threshold. Interestingly, given the different potential at
which ectopic spikes are generated, the conductances activated during AP generation may differ between distal
and proximal APs (Meyrand et al., 1992).
In invertebrates, numerous examples of ectopic spikes and of their functional relevance have been described
(Mulloney and Selverston, 1972; Maranto and Calabrese, 1984; Meyrand et al., 1992). Ectopic spikes in neurons
that target the hearts of the leech have been proposed to control the heart firing frequency (Maranto and
Calabrese, 1984). In the somatogastric ganglion of the crab, ectopic spikes are typically observed in the lateral
gastric motor neuron only when the muscles remain attached to the preparation, ectopic spikes being induced by
motor contraction (Meyrand et al., 1992). Remarkably, ectopic spikes in (Meyrand et al., 1992) fail to depolarize
terminals onto interneurons located close to the soma, whereas they efficiently excite a distal postsynaptic
target, the muscle. Thus, orthodromically-propagating spikes generated close to the soma and antidromically-
propagating spikes generated distally co-exist and they can reach synapses that target different postsynaptic
neurons.
Ectopic spikes have also been observed in vertebrates. In principal cells from CA3 area in the hippocampus
during rhythmic activity in the gamma range (Dugladze et al., 2012), axons fire APs at five times the firing
frequency detected at the soma. This is due to the activity of a specific interneuron, the axo-axonic cell,
that inhibits the initial segment, thereby avoiding the backpropagation of axonal spikes to the soma. Another
noteworthy case of co-ocurrence of AIS and ectopic spikes has been reported in (Lehnert et al., 2014) in auditory
medial superior olive (MSO) neurons. A realistic model of MSO neurons which takes into account their axonal
structure and ion channel composition together with a physiological synaptic bombardment regime generated
APs at both the AIS and at the first node of Ranvier. Additionally, under certain pathological conditions, e.g. in
epilepsy, hyperexcitable axons have been reported to generate ectopic spikes in the hippocampus (Stasheff et al.,
1993).
We would like to make at this point a clarification: although 'action potential' and 'spike' are two terms used
to refer to the sodium AP generated at the axon initial segment, the term spike seems preferentially used in the
literature to refer to spikes that can differ in their location (dendritic or ectopic spikes) and in their underlying
ionic mechanism (e.g., calcium spike, see below).
2.2.9 Beyond the sodium spike
The spike initiated through sodium influx is not the only spike propagating in the axon. There are spikes
that are generated through the influx of calcium.
In the giant axon of the jellyfish Aglantha digitale (order
Hydromedusae), both sodium and calcium spikes propagate in the axon (Mackie and Meech, 1985). Sodium-
6
dependent spikes are responsible for fast swimming and calcium spikes mediate slow swimming. Also in the axon
giant neurone R2 of Aplysia, the propagating AP has a mixed Na/Ca dependency (Horn and Miller, 1978) .
In vertebrates, calcium spikes are typically restricted to the dendritic compartment where there is a calcium
spike generation mechanism. The biophysical mechanisms contributing to the dendritic spikes initiation in the
distal apical trunk and proximal tuft of hippocampal CA1 pyramidal neurons (Gasparini et al., 2004) are (1)
dendritic sodium and potassium channels that set the threshold and determine the shape and forward-propagation
of dendritic spikes, (2) the highly synchronized inputs (approximately 50 synapses activated within 3 msec) and
(3) the spatial clustering (all inputs arriving in 100 µm of the dendrite). Moreover, NMDA receptors enhance
spike initiation by counteracting the shunting of AMPA synaptic conductances. In the olfactory bulb mitral cells
and in hippocampal and cortical pyramidal cells, dendritic spikes can trigger one or more axonal APs (Chen
et al., 2002; Stuart et al., 1997; Golding and Spruston, 1998; Larkum et al., 1999, 2001; Ariav et al., 2003) . In
addition, backpropagating axonal APs can themselves promote dendritic spikes, a reciprocal interaction that can
lead to a burst of axonal APs (Pinsky and Rinzel, 1994; Mainen and Sejnowski, 1996; Larkum et al., 1999, 2001;
Doiron et al., 2002).
2.3 BIOPHYSICS OF ACTION POTENTIAL PROPAGATION
2.3.1 An equivalent electrical circuit for axons
The passive properties of axons can be modeled by an equivalent electrical circuit (Fig. 3A, top). The axonal
membrane can be reduced to two circuit elements: the lipid bilayer, modeled by a capacitor and ion channels,
by a resistor. The resistance to axial current flow can be modeled by an additional resistance.
How fast signals propagate is critically controlled by the capacitance. Electrical currents need to first charge
the membrane capacitance, that opposes the flow of electric current, before electrically-charged membranes can
undergo voltage changes. Capacitances are proportional to the membrane capacity (capacitance per surface
area) and to the membrane surface area. The capacity of neuronal membranes is in the order of 1 µF per
cm2 (Gentet et al., 2000). Capacitance measurements from small axons are in the range of tens of picoFahrads
(pF) (Mejia-Gervacio et al., 2007). Capacitances of larger neurons (i.e., strongly-ramifying interneurons and
projection neurons) are therefore expected to be in the range of hundreds of pF to the nF range. Dynamic
changes in capacitance are suggested by activity-dependent changes in the size of axons (e.g.,(Ch´ereau et al.,
2017)). Little is known about dynamical changes in capacitance due to temperature, lipid composition, and
whether these may significantly impact the propagation of electrical signals along the axon.
The membrane resistance plays a major role in controlling how far signals spread in space along the axon
before membrane potential changes become imperceptible. This is typically measured by the space constant,
defined as the distance over which the membrane potential decays to 37 percent of its initial value. The more
open channels are available at the membrane, the more the axial current is attenuated in space along the axon
due to current leakage through membrane channels. Remarkably, the membrane resistance can change (reviewed
in (Debanne et al., 2019)), providing a potential source of variation in the conduction of electrical signals along
axons.
The axonal axial resistance to current flow results from the combination of the resistivity of the axoplasm
(that is, the cytoplasm of the axon), given in Ohms*cm, and the diameter of the cylinder. Axial resistivity values
are in the range of 100 Ω.cm (Carpenter et al., 1975; Cole, 1975). Computational models of the calyx of Held
suggest that the latency and amplitude of signals propagating between release sites is highly sensitive to changes
in axial resistivity, and that these changes may have a large impact on synaptic release (Spirou et al., 2008).
However, changes in resistivity have not been reported so far experimentally. The axon diameter can vary in three
orders of magnitude, from around hundred nanometers to hundreds of micrometers in diameter. Giant fibers
found in invertebrates are an example of specialized large-diameter structures that propagate electrical signals
with high speed over long distances ( in (Hodgkin and Huxley, 1952; Xu and Terakawa, 1999), Fig. 4A). The
larger the diameter of the axon, the faster the electrical signal propagates. Taking into account cable theoretic
considerations, the propagation speed increases with the square root of the diameter. Remarkably, the axon
diameter, as well as the diameter of synaptic boutons, are not static properties of axons. In fact, axon diameter
has been shown to be regulated in hippocampal principal cells in an activity-dependent manner. Plasticity
protocols induced changes in axonal diameter which were accompanied by significant changes in AP conduction
velocity along CA3 pyramidal cell axons (Ch´ereau et al., 2017)(Fig. 5A).
2.3.2 Propagation of subthreshold signals
Subthreshold signals propagate passively in axons, reaching synaptic terminals, where they influence synaptic
release (Alle and Geiger, 2006). Subthreshold membrane fluctuations consist of synaptic events, typically in the
range of a 100s of µVs to mVs, which in cortical neurons approximate highly stochastic background dynamics in
vivo (Rudolph and Destexhe, 2003). Their propagation, referred to as analog, in comparison to the digital nature
of the AP, modulates the efficiency of APs to induce neurotransmitter release in synaptic terminals (Shu et al.,
7
2006). Given that subthreshold signals are not regenerated along the axon, analog signaling is more prominent
in proximal portions of the axon. The presence of combined analog and digital signals in axons has been termed
hybrid analog-digital signaling.
Of particular interest is that the propagation of slow analog subthreshold signals does not only reach chemical
synapses. Subthreshold signals also reach axonal electrical synapses, at which the signal is expected to be conveyed
with high efficiency to the postsynaptic site due to the continuous and low-pass filtering properties of electrical
transmission (reviewed in (Alcami and Pereda, 2019)).
2.3.3 Reliability of propagation and action potential failures
Before we consider how signals propagate along axons, let us discuss the reliability of propagation. APs do not
always efficiently propagate, and in fact, failures in AP propagation are observed in all neuron types at high
firing frequencies (Krnjevic and Miledi, 1959; Grossman et al., 1979; Monsivais et al., 2005).
Although geometry can generate failures (e.g. at branch points, see next section), failures typically involve
active mechanisms during repetitive activation of axons as observed in a number of vertebrate and invertebrate
preparations (Krnjevic and Miledi, 1959; Monsivais et al., 2005; Mar and Drapeau, 1996). Repetitive activity
typically leads to extracellular potassium accumulation and subsequently to the depolarization of the axon,
inducing failures of conduction (Grossman et al., 1979). In other cases however, repetitive activation induces AP
propagation failures via hyperpolarization of the membrane (Mar and Drapeau, 1996). Note that propagation
failures likely affect only a fraction of APs at physiological firing rates, and that most APs succeed in propagating.
in response
to synaptic inputs. Indeed, specific potassium channels of the type IA underlie a hyperpolarization-mediated
conduction block in hippocampal pyramidal cells (Debanne et al., 1997). Remarkably, the activation and de-
inactivation kinetics of IA allow for a history-dependent conduction block of APs.
Interestingly, changes in AP failure can also occur in response to membrane fluctuations, e.g.
2.3.4 More than linear cables: Impact of axonal branching and inhomogeneities
Axons are typically formed by complex trees and spatial heterogeneities. These include branching points, vari-
cosities (local enlargements of the axon containing the release machinery of chemical presynaptic sites) and large
structures specialized in the interaction with other cells. Examples of such structures are the basket formed by a
specific type of interneuron, basket cells, around principal cells in many brain regions; the glomerular collateral
of the climbing fiber in the cerebellum; the pinceau structure surrounding cerebellar Purkinje cells (Palay and
Chan-Palay, 2012) ; Fig. 1B, C) or the calyx of Held in the auditory system.
Changes in axon diameter at varicosities or branching points are characterized by an impedance mismatch,
that is, a need of a larger current to flow in one of the two directions to electrically load axonal branches,
provoking changes in AP propagation (Goldstein and Rall, 1974; Manor et al., 1991). Axonal branches indeed
represent a local challenge to the propagation of electrical signals, due to the impedance mismatch between the
mother and daughter branches. That is, if APs need to load a larger impedance as they propagate from one
branch into two branches, their propagation will be delayed relative to the speed that they would have had if no
branching was present.
In the most extreme case, the electrical signal fails to load one or two of the branches, resulting in AP
propagation failure. Propagation failures have been shown to occur in branches of a number of neurons (Yau,
1976; Gu et al., 1991; Grossman et al., 1979). An interesting example is provided by the medial pressure sensory
neuron in the leech, where the failure of APs to propagate can differentially affect postsynaptic cells contacted
by distinct presynaptic branches (Gu et al., 1991).
Finally, branching points can also provoke a surprising effect: APs can be slowed down in the ms range, up
to a level that allows the mother branch to overcome the refractory period for AP generation. As a consequence,
the AP can 'reflect' or travel backwards, increasing synaptic release at synapses present in the branch where the
AP reflects (Baccus, 1998; Baccus et al., 2000).
The impact of complex morphologies on AP propagation is likely to be relevant in the complex axonal
ramification patterns of many neurons, including vertebrate interneurons (Ofer et al., 2019). The usage of
voltage sensitive-dyes that track membrane voltage with sub-millisecond precision (Palmer and Stuart, 2006)
should allow following large portions of axons both in vitro and in vivo, and characterize the propagation of APs
along complex axonal structures.
2.3.5 Biophysical properties of myelinated fibers
Many vertebrate and some invertebrate fibers are myelinated, a specialization endowed by the glial ensheathment
of axons (Castelfranco and Hartline, 2015). Myelin, formed by compact lipidic layers produced by the membranes
of glial cells (oligodendrocytes in the central nervous system and Schwann cells in the peripheral nervous system),
strongly impacts the propagation of electrical signals.
8
In his seminal study, Lillie (Lillie, 1925) wrapped an iron wire placed in an acidic solution with an insulating
glass cylinder. This increased the speed of propagation of electrical signals along the wire, which he postulated
to occur in myelinated fibers. His prediction was confirmed decades later in axons by recording electrical signals
at nodes, leading to the concept of the saltatory nature of transmission of electrical signals in myelinated fibers
(Huxley and Stampfli, 1949). Saltatory comes from the latin verb 'saltare' (to jump), an analogy describing the
very fast propagation of electrical signals between nodes of Ranvier.
Myelin modifies the electrical circuit that models the passive properties of axons described above. It increases
the effective radial resistance and decreases the effective capacitance of the axon (Fig. 4). This effect is due to
the different properties of series resistors and series capacitors added to the circuit: series resistors sum their
resistances and the inverse of capacitances from capacitors in series sums. The increase in effective membrane
resistance (Bakiri et al., 2011) and the decrease in effective membrane capacitance by myelin have two major
consequences. On the one hand, the increase in the effective axonal resistance by myelin increases their length
constant. On the other hand, as a consequence of the decrease in the effective axonal capacitance, the time
required to effectively load axons decreases, dramatically accelerating the propagation of electrical signals. This
speeding of electrical propagation by a reduced effective capacitance underlies saltatory conduction between nodes
of Ranvier (Huxley and Stampfli, 1949; Castelfranco and Hartline, 2015) .
2.3.6 Geometry of axons and myelin
Specific myelination parameters have been shown to maximize conduction velocity of electrical signals along
axons. In particular, the ratio of the axonal diameter d to the fiber diameter D (the summed diameter of axon
and myelin sheath), defined as the 'g-ratio', has been shown to critically control the conduction of electrical
signals. Rushton (Rushton, 1951) developed a biophysical formalism to model current flow in a myelinated
axon. He deduced the relation between the ratio l/D (internode length l over D) and the g-ratio. He further
demonstrated that the ratio l/D is maximal when g = 0.6, a value also found analytically to maximize the space
constant. In an independent approach, Deutsch (Deutsch, 1969) mathematically derived the geometry of axonal
and myelin properties that maximize, this time, conduction velocity. He deduced that the propagation velocity
is inversely proportional to the internode length and to the RC time constant given by the internal resistance
of the axon and the capacitance of the membrane. Maximizing conduction speed requires minimizing the time
constant of the circuit. This resulted in the same geometrical properties as those derived by Rushton: a g-ratio of
0.6. Additional modeling studies including a more complete description of myelinated fibers converged to similar
conclusions (Goldman and Albus, 1968). Therefore, due to the biophysical constraints posed by the thickness of
myelin and axonal size, both conduction speed and efficient spatial propagation of electrical signals are maximized
by specific axonal and myelin geometries. It is noteworthy that the geometry of axons and myelin does not only
control AP propagation speed (Rushton, 1951; Deutsch, 1969) but also their temporal jitter (Kim et al., 2013).
It seems difficult to imagine that neurons have fine-tuned their dendritic computations in a cell-type specific
manner (Stuart et al., 2016), but that axons would, on the contrary, be invariant and homogeneously optimizing
speed by their geometry. Additionally, we know that nervous systems adapt and fine tune a plethora of properties
to accomplish specific functions (Marder and Taylor, 2011). Axons indeed adjust different parameters which
impact the speed of propagation of electrical signals (Seidl et al., 2010; Ford et al., 2015; Arancibia-Carcamo
et al., 2017) to obtain different speeds of computation.
An illustrative example of how the geometry of axons and myelin is tuned to adjust AP propagation speed,
deviating from a g-ratio of 0.6, is provided by the axon properties which encode spatial location in the avian
brain (Seidl et al., 2010). The longer contralateral fibers have larger-diameter axons and longer internodal
distances, compensating in this manner for an otherwise slower conduction delay relative to the shorter fibers on
the ipsilateral side. In this manner, axons ensure a coincident arrival of contralateral and ipsilateral signals to
the synaptic terminals. Another interesting example is provided by fibers specialized in carrying information for
low-frequency sounds, which show larger diameters than those carrying high-frequency sounds, but also shorter
internodes.
In doing so, these fibers deviate from the classical dependence of both variables established by
Rushton. This specialization is proposed to ensure a proper function of the circuit (Ford et al., 2015). Moreover,
internode length and node diameter show graded properties as they approach the terminal in the auditory
granular bushy cell axon, implementing an efficient invasion of the axon terminal by APs (Ford et al., 2015).
We have seen that a large number of structural myelination parameters act in concert to control the speed
of AP propagation (Goldman and Albus, 1968). It is noteworthy to mention that myelin has introduced new
degrees of freedom in the regulation of AP speed: speed depends on distance between nodes of Ranvier, on node
length, node composition, and thickness of myelin (Rushton, 1951; Ford et al., 2015; Arancibia-Carcamo et al.,
2017; Wu et al., 2012). These parameters can be modulated independently or in combination, thereby increasing
the number of available mechanisms by which nervous systems tune the axonal propagation of electrical signals.
Furthermore, dynamical changes in axons and myelin (Sampaio-Baptista et al., 2013; McKenzie et al., 2014;
Sinclair et al., 2017; Fields, 2015) suggest that nervous systems dynamically adjust the properties of their fibers
to achieve the specific behaviors that they control.
9
2.3.7 Active contributions to variations of action potentials along the axon at nodes and terminals
Myelinated fibers are characterized by the presence of non-myelinated sections, which concentrate the machinery
to generate APs. These include nodes, the heminode (last unmyelinated portion before the terminal) and synaptic
terminals. Remarkably, AP shape can suffer modifications along the axon by being modulated at each one of these
loci by the active channels that they express. Each active AP regeneration site along the axon can potentially
generate AP variants due to their specific state and composition of ion channels and membrane potential.
An example of local variations in AP shape through active mechanisms is found at presynaptic mossy fiber
terminals (present in the axon of dentate gyrus granule cells), where local potassium channels modulate spike
shape (Alle et al., 2011). Another example is the activation of calcium-dependent potassium channels in Purkinje
cell nodes of Ranvier. These channels repolarize membranes, de-inactivating sodium channels which can then
generate fast frequency spikes, and in this manner prevent AP failure at these high frequencies (Grundemann and
Clark, 2015). Additionally, changes in membrane voltage in the axon induced by active mechanisms can impact
the efficiency of AP propagation. For example, in the leech touch cells, adaptation in response to repetitive AP
firing consists of a hyperpolarization, resulting in the blockade of AP propagation (Van Essen, 1973).
We previously described how actively-generated signals can differently be passed on by two bifurcating
branches, illustrating how passive properties can filter actively-generated signals (Gu et al., 1991). Remark-
ably, active properties of axonal branches can prevent failures by potentiating signals in specific branches (Cho
et al., 2017). Failures that would occur as a consequence of the passive filtering of electrical signals can be
prevented in a branch-specific manner by the presence of the sodium channel subunit NavβII, which potentiates
AP propagation.
2.3.8 The geometries of axons and myelin are plastic
Geometrical properties of both axons and myelin have been shown to be highly plastic. Indeed, the diameter of
axons can change as a function of neuronal activity (Sinclair et al., 2017; Ch´ereau et al., 2017). The understanding
of the mechanisms of myelination has revealed that properties of nodes and internodes are subject to plasticity
(Young et al., 2013), reviewed in (Kaller et al., 2017). Interestingly, major macroscopic changes in myelination
occur in response to training paradigms in mice, allowing the acquisition of motor skills (Sampaio-Baptista et al.,
2013; McKenzie et al., 2014; Xiao et al., 2016), and models of injury induce strong remodelling of myelination
patterns in the auditory brainstem (Sinclair et al., 2017). Changes in myelination can involve several mechanisms
at time scales of hours to days: changes in internode length (Etxeberria et al., 2016), in myelin thickness (Sinclair
et al., 2017) and in node length (suggested in (Arancibia-Carcamo et al., 2017), (Fig.5).
2.3.9 Nanostructures with unknown function
The evolutionary appearance of myelin has led to new specialized microstructures or microdomains (namely
nodes, paranodes, juxtaparanodes, heminodes). Electron microscopy and recently superresolution imaging have
revealed additional structures whose contribution to axonal computation remains mysterious. These are organized
at the nanometer scale in a highly regular spatial organization (D'Este et al., 2017), rising the question of the
function of these periodic structures. For example, d'Este et al. show that the voltage-dependent potassium
channel subunit Kv1.2 channels, found at the juxtaparanodes, correlates in space with the underlying actin
cytoskeleton. An additional structure that has attracted our attention is the Schmidt-Lanterman incisure, a
spiral cytoplasmic expansion from the outer tongue of myelin to the inner tongue. Incisures express gap junctions
and their contribution to the electrical properties of myelin remain mysterious (Kamasawa et al., 2005). Likewise,
exquisite arrangements of structures apposing glial and neuronal membranes and also with other glial membranes
in the form of rosettes formed by ion channels at the paranode are still poorly understood (Rash et al., 2016).
These findings open up the following question: are these highly-ordered structures an epiphenomenon resulting
from the necessary anchoring to regular cytoskeletal structures, or is this spatial regularity impacting function?
One would expect that further perturbing and studying those nanoscale structures in the future will further our
understanding of how they contribute to axonal computations.
2.4 AXONS ARE NOT ALONE: INTRACELLULAR AND INTERCELLULAR
COUPLING
2.4.1 Crosstalk of axons with other compartments
Although we have done a treatment of the axon as an isolated cable as has classically been performed in the
early days (Lillie, 1925; Huxley and Stampfli, 1949; Hodgkin and Huxley, 1952), the axonal cable is coupled to
the somatic and dendritic compartment at one end and, additionally, directly or indirectly to electrically-coupled
cells, which also functionally act like an electrical compartment (Furshpan and Potter, 1959; Alcami and Marty,
2013; Eyal et al., 2014). The electrical coupling between the axon and the soma, and indirectly to dendrites
and dendritically-coupled cells, influences AP generation in the axon (Bekkers and Hausser, 2007; Eyal et al.,
10
2014; Amsalem et al., 2016; Alcami, 2018; Goldwyn et al., 2019). These compartments create parallel pathways
for current flow that compete with the flow of currrent in the axial resistance of the axon. Loading of these
non-axonal compartments was shown to impact AP threshold and speed (Bekkers and Hausser, 2007; Eyal et al.,
2014; Amsalem et al., 2016). Their contribution to the effective membrane time constant is substantial. The
somato-dendritic compartment has been shown to, by this mechanism, modify both the threshold and the initial
rise of the AP (Eyal et al., 2014) .
It is further interesting to consider the axon in the context of synaptic integration, that is, the computation
of information received at synapses. Similar to the somato-dendritic and junctional compartments, which act as
current sinks, influencing the effective kinetics and strength of excitatory inputs recorded at the soma, before they
reach the AIS (Norenberg et al., 2010; Alcami, 2018), the axonal membrane also behaves as a current sink, leaking
current generated at synapses in the somatic and dendritic compartments with its charging time constant. This
phenomenon has been shown to, through a passive mechanism, accelerate the time-course of somatically-recorded
excitatory synaptic events (Mejia-Gervacio et al., 2007). It has additionally been suggested to also accelerate
the time-course of spikelets generated by electrically-coupled cells (Alcami and Marty, 2013). Mejia-Gervacio
and collaborators (Mejia-Gervacio et al., 2007) show that the capacitive loading of axons introduces in cerebellar
molecular layer interneurons a computational time constant of about 3 ms, which is one order of magnitude slower
than the faster time constant to load the somatodendritic compartment. This relatively slowly-charging process
of axons accelerates the decay of excitatory postsynaptic potentials, reducing the time window for AP generation.
Therefore, electrical signals that arrive to the axon are not only passively influenced by other compartments, but
they also influence passive computation by the non-axonal compartments, as part of a system formed by coupled
compartments.
Let us now turn our attention onto the signals propagating in axons as cells detect events in their somas and
dendrites, which is a consequence of the current flow of synaptic events into the axon. These signals propagate
at long distances before their complete attenuation (contributing to the analog signaling in the axon that was
previously introduced). The space constant at the hippocampal unmyelinated granule cell axon is in the range
of hundreds of micrometers (Alle and Geiger, 2006) and the subthreshold propagation of voltage depolarizations
has been shown to increase AP evoked release (Shu et al., 2006). Interestingly, these signals do not only travel
orthodromically towards the terminals, but signals generated in the axon, e.g. by synaptic receptors present on
the presynaptic membrane, also travel antidromically to the soma (Trigo et al., 2010). As a consequence, these
axonal PSPs depolarize the soma and influence AP generation (de San Martin et al., 2015).
2.4.2 Direct coupling between axons
Direct coupling between axons was suggested in early work, bringing up the concept that networks of axons may
directly interact with each other (Katz and Schmitt, 1940). These interactions can be explained by two forms
of electrical transmission (Fig. 6A): ephaptic transmission due to the generation of an electric field by an axon,
affecting the excitability of a neighboring axon (Katz and Schmitt, 1940; Blot and Barbour, 2014; Han et al.,
2018), and transmission mediated by gap junction-mediated electrical synapses (Schmitz et al., 2001; Furshpan
and Potter, 1959; Watanabe and Grundfest, 1961; Bennett et al., 1963; Robertson et al., 1963). As a matter of
fact, electrical synapses were first discovered in axons in a variety of preparations (Furshpan and Potter, 1959;
Watanabe and Grundfest, 1961; Bennett et al., 1963; Robertson et al., 1963) .
Axonal electrical synapses are expected to strongly affect electrical signals by adding a conductance pathway
directly in the axon. Remarkably, axonal gap junctions allow inputs to arrive at the axon directly, blurring the
pure output role of axons: an AP in a presynaptic axon induces a spikelet (a low-pass filtered version of the
presynaptic AP) in the postsynaptic axon. Spikelets excite distal parts of the axon where the gap junctions
are located, and when the spikelet depolarizes the membrane potential sufficiently, they are able to evoke APs
(Chorev and Brecht, 2012).
Interestingly, in vivo recordings from hippocampal principal cells in CA regions
from rodents revealed two types of APs: one type was preceded by a shoulder which was identical to the rising
phase of the spikelet, and the other was a full-blown AP lacking the shoulder (Epsztein et al., 2010). Spikelets
correlate with AP firing from nearby cells, confirming that they are generated by electrically-coupled cells, and
not by spontaneous AP firing of distal axons (Chorev and Brecht, 2012). Altogether, these studies suggest that
hippocampal pyramidal cell axons are excited through electrical synapses in physiological conditions in behaving
animals, as it had been previously shown in slices (Schmitz et al., 2001) .
Axonal gap junctions have additionally been proposed to underlie fast synchronization of neuronal ensembles
in both physiological and pathophysiological conditions (Schmitz et al., 2001; Roopun et al., 2010). Interestingly,
axo-axonal gap junctions, by coupling axons, that is, two output structures, can have a strong impact on spike
coordination and coding efficiency (Wang et al., 2017) .
Last and not least, direct axo-axonal coupling can be mediated by chemical synapses onto axons (Fig. 6A)
which are typically performed by a specific type of interneuron targeting the AIS of principal cells (Somogyi et
al., 1998). It has been shown that the presynaptic activity of chandelier cells modulates the sodium channel
dynamics on the axon initial segment (Inda et al., 2006). Terminals found in the neocortex and hippocampus
11
were found to exert inhibitory control over AP initiation (Dugladze et al., 2012).
2.4.3 Axons interact indirectly via glial cells
Finally, let us consider the contribution of glial cells to axonal function as an indirect coupling pathway between
axons (Fig. 6B). The biophysical properties of myelin are achieved by glial cells that produce myelin. Since
several axons are typically contacted by a myelinating cell, this introduces de facto an indirect coupling pathway
between axons. A number of transmission mechanisms has been described between axons and myelin (reviewed
in ((Micu et al., 2017))), whose dynamic properties depend on signaling from neurons (Hines et al., 2015).
3 Conclusion
Here, we have reviewed the biophysical nature of computations performed by the axon. We have shown that
the axon is not just a cable on which electrical pulses propagate but rather a computational device that mod-
ulates signaling and adds to the complexity of information processing in the brain. It should be appreciated
that these computations are sophisticated enough to contribute to network level phenomena up to behavior in
vivo. Relating axonal computation to behavioral phenomenology is still a nascent area but it will complement
studies that have focused almost exclusively on somas, dendrites or a coarse grained view of the neurons. The
study of axonal computation is hurdled by technical challenges, but there is already an emerging interest in
developing technologies that will allow to electrophysiologically and structurally study axonal processes. Further
complication arises when one realizes that axons do not act alone but in concert, exhibiting collective modes of
computation conveyed by inter-axonal signaling. We propose that axonal biophysics are of vital importance for
understanding not only how single neurons process information but also how neural networks coordinate their
activity.
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
Both authors contributed equally to literature search and writing.
Funding
PA is funded by the Munich Center for Neurosciences. AEH is funded by the Howard Hughes Medical Institute.
Acknowledgments
We are grateful to Conny Kopp-Scheinflug for her feedback on previous versions of this manuscript.
References
Abbott, B., Howarth, J., and Ritchie, J. (1965). The initial heat production associated with the nerve impulse
in crustacean and mammalian non-myelinated nerve fibres. The Journal of physiology, 178(2):368 -- 383.
Akkin, T., Joo, C., and De Boer, J. F. (2007). Depth-resolved measurement of transient structural changes
during action potential propagation. Biophysical journal, 93(4):1347 -- 1353.
Alcami, P. (2018). Electrical synapses enhance and accelerate interneuron recruitment in response to coincident
and sequential excitation. Frontiers in cellular neuroscience, 12.
Alcami, P., Franconville, R., Llano, I., and Marty, A. (2012). Measuring the firing rate of high-resistance neurons
with cell-attached recording. Journal of Neuroscience, 32(9):3118 -- 3130.
Alcami, P. and Marty, A. (2013). Estimating functional connectivity in an electrically coupled interneuron
network. Proceedings of the National Academy of Sciences, 110(49):E4798 -- E4807.
12
Alcami, P. and Pereda, A. E. (2019). Beyond plasticity: the dynamic impact of electrical synapses on neural
circuits. Nature reviews. Neuroscience.
Alle, H. and Geiger, J. R. (2006). Combined analog and action potential coding in hippocampal mossy fibers.
Science, 311(5765):1290 -- 1293.
Alle, H., Kubota, H., and Geiger, J. R. (2011). Sparse but highly efficient kv3 outpace bkca channels in action
potential repolarization at hippocampal mossy fiber boutons. Journal of Neuroscience, 31(22):8001 -- 8012.
Amsalem, O., Van Geit, W., Muller, E., Markram, H., and Segev, I. (2016). From neuron biophysics to orientation
selectivity in electrically coupled networks of neocortical l2/3 large basket cells. Cerebral Cortex, 26(8):3655 --
3668.
Araki, T. and Otani, T. (1955). Response of single motoneurons to direct stimulation in toad's spinal cord.
Journal of neurophysiology, 18(5):472 -- 485.
Arancibia-Carcamo, I. L., Ford, M. C., Cossell, L., Ishida, K., Tohyama, K., and Attwell, D. (2017). Node of
ranvier length as a potential regulator of myelinated axon conduction speed. Elife, 6:e23329.
Ariav, G., Polsky, A., and Schiller, J. (2003). Submillisecond precision of the input-output transformation
function mediated by fast sodium dendritic spikes in basal dendrites of ca1 pyramidal neurons. Journal of
Neuroscience, 23(21):7750 -- 7758.
Astman, N., Gutnick, M. J., and Fleidervish, I. A. (2006). Persistent sodium current in layer 5 neocortical
neurons is primarily generated in the proximal axon. Journal of Neuroscience, 26(13):3465 -- 3473.
Atherton, J. F., Wokosin, D. L., Ramanathan, S., and Bevan, M. D. (2008). Autonomous initiation and propaga-
tion of action potentials in neurons of the subthalamic nucleus. The Journal of physiology, 586(23):5679 -- 5700.
Azouz, R. and Gray, C. M. (1999). Cellular mechanisms contributing to response variability of cortical neurons
in vivo. Journal of Neuroscience, 19(6):2209 -- 2223.
Azouz, R., Jensen, M. S., and Yaari, Y. (1996). Ionic basis of spike after-depolarization and burst generation in
adult rat hippocampal ca1 pyramidal cells. The Journal of physiology, 492 ( Pt 1)(Pt 1):211 -- 223.
Baccus, S. A. (1998). Synaptic facilitation by reflected action potentials: enhancement of transmission when
nerve impulses reverse direction at axon branch points. Proceedings of the National Academy of Sciences,
95(14):8345 -- 8350.
Baccus, S. A., Burrell, B. D., Sahley, C. L., and Muller, K. J. (2000). Action potential reflection and failure at axon
branch points cause stepwise changes in epsps in a neuron essential for learning. Journal of Neurophysiology,
83(3):1693 -- 1700.
Bakiri, Y., K´arad´ottir, R., Cossell, L., and Attwell, D. (2011). Morphological and electrical properties of oligoden-
drocytes in the white matter of the corpus callosum and cerebellum. The Journal of physiology, 589(3):559 -- 573.
Baranauskas, G. and Martina, M. (2006). Sodium currents activate without a hodgkin and huxley-type delay in
central mammalian neurons. Journal of Neuroscience, 26(2):671 -- 684.
Batabyal, S., Satpathy, S., Bui, L., Kim, Y.-T., Mohanty, S., Bachoo, R., and Dav´e, D. P. (2017). Label-free
optical detection of action potential in mammalian neurons. Biomedical optics express, 8(8):3700 -- 3713.
Bean, B. P. (2007). The action potential in mammalian central neurons. Nature Reviews Neuroscience, 8(6):451.
Bekkers, J. M. and Hausser, M. (2007). Targeted dendrotomy reveals active and passive contributions of the
dendritic tree to synaptic integration and neuronal output. Proceedings of the National Academy of Sciences,
104(27):11447 -- 11452.
Bennett, M. V., Aljure, E., Nakajima, Y., and Pappas, G. D. (1963). Electrotonic junctions between teleost
spinal neurons: electrophysiology and ultrastructure. Science, 141(3577):262 -- 264.
Bi, G.-q. and Poo, M.-m. (1998). Synaptic modifications in cultured hippocampal neurons: dependence on spike
timing, synaptic strength, and postsynaptic cell type. Journal of neuroscience, 18(24):10464 -- 10472.
Blot, A. and Barbour, B. (2014). Ultra-rapid axon-axon ephaptic inhibition of cerebellar purkinje cells by the
pinceau. Nature neuroscience, 17(2):289.
13
Boiko, T., Van Wart, A., Caldwell, J. H., Levinson, S. R., Trimmer, J. S., and Matthews, G. (2003). Functional
specialization of the axon initial segment by isoform-specific sodium channel targeting. Journal of Neuroscience,
23(6):2306 -- 2313.
Borst, A. and Haag, J. (1996). The intrinsic electrophysiological characteristics of fly lobula plate tangential
cells: I. passive membrane properties. Journal of computational neuroscience, 3(4):313 -- 336.
Borst, J. G. G. and Sakmann, B. (1999). Effect of changes in action potential shape on calcium currents and
transmitter release in a calyx -- type synapse of the rat auditory brainstem. Philosophical Transactions of the
Royal Society of London. Series B: Biological Sciences, 354(1381):347 -- 355.
Brachet, A., Leterrier, C., Irondelle, M., Fache, M.-P., Racine, V., Sibarita, J.-B., Choquet, D., and Dargent, B.
(2010). Ankyrin g restricts ion channel diffusion at the axonal initial segment before the establishment of the
diffusion barrier. The Journal of cell biology, 191(2):383 -- 395.
Brette, R. (2013). Sharpness of spike initiation in neurons explained by compartmentalization. PLOS Computa-
tional Biology, 9(12):1 -- 10.
Calabrese, R. L. and Kennedy, D. (1974). Multiple sites of spike initiation in a single dendritic system. Brain
Research, 82(2):316 -- 321.
Carpenter, D. O., Hovey, M. M., and Bak, A. F. (1975). Resistivity of axoplasm. ii. internal resistivity of giant
axons of squid and myxicola. The Journal of general physiology, 66(2):139 -- 148.
Castelfranco, A. M. and Hartline, D. K. (2015). The evolution of vertebrate and invertebrate myelin: a theoretical
computational study. Journal of computational neuroscience, 38(3):521 -- 538.
Chen, W. R., Shen, G. Y., Shepherd, G. M., Hines, M. L., and Midtgaard, J. (2002). Multiple modes of action
potential initiation and propagation in mitral cell primary dendrite. Journal of Neurophysiology, 88(5):2755 --
2764.
Ch´ereau, R., Saraceno, G. E., Angibaud, J., Cattaert, D., and Nagerl, U. V. (2017). Superresolution imaging
reveals activity-dependent plasticity of axon morphology linked to changes in action potential conduction
velocity. Proceedings of the National Academy of Sciences, 114(6):1401 -- 1406.
Cho, I. H., Panzera, L. C., Chin, M., and Hoppa, M. B. (2017). Sodium channel β2 subunits prevent action
potential propagation failures at axonal branch points. Journal of Neuroscience, 37(39):9519 -- 9533.
Chorev, E. and Brecht, M. (2012). In vivo dual intra-and extracellular recordings suggest bidirectional coupling
between ca1 pyramidal neurons. Journal of neurophysiology, 108(6):1584 -- 1593.
Clark, B. A., Monsivais, P., Branco, T., London, M., and Hausser, M. (2005). The site of action potential
initiation in cerebellar purkinje neurons. Nature Neuroscience, 8(2):137 -- 139.
Cohen, L., Hille, B., and Keynes, R. (1970). Changes in axon birefringence during the action potential. The
Journal of physiology, 211(2):495 -- 515.
Colbert, C. M. and Pan, E. (2002).
Ion channel properties underlying axonal action potential initiation in
pyramidal neurons. Nature Neuroscience, 5(6):533 -- 538.
Cole, K. (1975). Resistivity of axoplasm. i. resistivity of extruded squid axoplasm. The Journal of general
physiology, 66(2):133 -- 138.
Coombs, J., Curtis, D., and Eccles, J. (1957). The generation of impulses in motoneurones. The Journal of
physiology, 139(2):232 -- 249.
Davis, J. Q., Lambert, S., and Bennett, V. (1996). Molecular composition of the node of ranvier: identification
of ankyrin-binding cell adhesion molecules neurofascin (mucin+/third fniii domain-) and nrcam at nodal axon
segments. The Journal of cell biology, 135(5):1355 -- 1367.
Debanne, D., Guerineau, N. C., Gahwiler, B. H., and Thompson, S. M. (1997). Action-potential propagation
gated by an axonal i a-like k+ conductance in hippocampus. Nature, 389(6648):286.
Debanne, D., Inglebert, Y., and Russier, M. (2019). Plasticity of intrinsic neuronal excitability. Current Opinion
in Neurobiology, 54:73 -- 82.
Dekker, J. P. and Yellen, G. (2006). Cooperative gating between single hcn pacemaker channels. The Journal of
general physiology, 128(5):561 -- 567.
14
D'Este, E., Kamin, D., Balzarotti, F., and Hell, S. W. (2017). Ultrastructural anatomy of nodes of ranvier in the
peripheral nervous system as revealed by sted microscopy. Proceedings of the National Academy of Sciences,
114(2):E191.
Deutsch, S. (1969). The maximization of nerve conduction velocity. IEEE Transactions on Systems Science and
Cybernetics, 5(1):86 -- 91.
Doiron, B., Laing, C., Longtin, A., and Maler, L. (2002). Ghostbursting: A novel neuronal burst mechanism.
Journal of Computational Neuroscience, 12(1):5 -- 25.
Duflocq, A., Chareyre, F., Giovannini, M., Couraud, F., and Davenne, M. (2011). Characterization of the axon
initial segment (ais) of motor neurons and identification of a para-ais and a juxtapara-ais, organized by protein
4.1 b. BMC biology, 9(1):66.
Dugladze, T., Schmitz, D., Whittington, M. A., Vida, I., and Gloveli, T. (2012). Segregation of axonal and
somatic activity during fast network oscillations. Science, 336(6087):1458 -- 1461.
DEste, E., Kamin, D., Balzarotti, F., and Hell, S. W. (2017). Ultrastructural anatomy of nodes of ranvier in the
peripheral nervous system as revealed by sted microscopy. Proceedings of the National Academy of Sciences,
114(2):E191 -- E199.
El Hady, A. and Machta, B. B. (2015). Mechanical surface waves accompany action potential propagation. Nature
communications, 6:6697.
Epsztein, J., Lee, A. K., Chorev, E., and Brecht, M. (2010). Impact of spikelets on hippocampal ca1 pyramidal
cell activity during spatial exploration. Science, 327(5964):474 -- 477.
Etxeberria, A., Hokanson, K. C., Dao, D. Q., Mayoral, S. R., Mei, F., Redmond, S. A., Ullian, E. M., and Chan,
J. R. (2016). Dynamic modulation of myelination in response to visual stimuli alters optic nerve conduction
velocity. Journal of Neuroscience, 36(26):6937 -- 6948.
Eyal, G., Mansvelder, H. D., de Kock, C. P., and Segev, I. (2014). Dendrites impact the encoding capabilities of
the axon. Journal of Neuroscience, 34(24):8063 -- 8071.
Fatt, P. (1957). Electric potentials occurring around a neurone during its antidromic activation. Journal of
neurophysiology, 20(1):27 -- 60.
Fields, R. D. (2015). A new mechanism of nervous system plasticity: activity-dependent myelination. Nature
Reviews Neuroscience, 16:756 EP -- .
Fleidervish, I. A., Lasser-Ross, N., Gutnick, M. J., and Ross, W. N. (2010). Na+ imaging reveals little difference
in action potential -- evoked na+ influx between axon and soma. Nature neuroscience, 13(7):852.
Ford, M. C., Alexandrova, O., Cossell, L., Stange-Marten, A., Sinclair, J., Kopp-Scheinpflug, C., Pecka, M.,
Attwell, D., and Grothe, B. (2015). Tuning of ranvier node and internode properties in myelinated axons to
adjust action potential timing. Nature Communications, 6:8073 EP -- .
Fourcaud-Trocm´e, N., Hansel, D., Van Vreeswijk, C., and Brunel, N. (2003). How spike generation mechanisms
determine the neuronal response to fluctuating inputs. Journal of Neuroscience, 23(37):11628 -- 11640.
Foust, A., Popovic, M., Zecevic, D., and McCormick, D. A. (2010). Action potentials initiate in the axon
initial segment and propagate through axon collaterals reliably in cerebellar purkinje neurons. Journal of
Neuroscience, 30(20):6891 -- 6902.
Furshpan, E. J. and Potter, D. D. (1959). Transmission at the giant motor synapses of the crayfish. The Journal
of physiology, 145(2):289 -- 325.
Gasparini, S., Migliore, M., and Magee, J. C. (2004). On the initiation and propagation of dendritic spikes in
ca1 pyramidal neurons. Journal of Neuroscience, 24(49):11046 -- 11056.
Gentet, L. J., Stuart, G. J., and Clements, J. D. (2000). Direct measurement of specific membrane capacitance
in neurons. Biophysical Journal, 79(1):314 -- 320.
Golding, N. L. and Spruston, N. (1998). Dendritic sodium spikes are variable triggers of axonal action potentials
in hippocampal ca1 pyramidal neurons. Neuron, 21(5):1189 -- 1200.
Goldman, L. and Albus, J. S. (1968). Computation of impulse conduction in myelinated fibers; theoretical basis
of the velocity-diameter relation. Biophysical Journal, 8(5):596 -- 607.
15
Goldstein, S. S. and Rall, W. (1974). Changes of action potential shape and velocity for changing core conductor
geometry. Biophysical Journal, 14(10):731 -- 757.
Goldwyn, J. H., Remme, M. W., and Rinzel, J. (2019). Soma-axon coupling configurations that enhance neuronal
coincidence detection. PLoS computational biology, 15(3):e1006476.
Grossman, Y., Parnas, I., and Spira, M. E. (1979). Differential conduction block in branches of a bifurcating
axon. The Journal of physiology, 295:283 -- 305.
Grundemann, J. and Clark, B. A. (2015). Calcium-activated potassium channels at nodes of ranvier secure axonal
spike propagation. Cell Reports, 12(11):1715 -- 1722.
Gu, X. N., Muller, K. J., and Young, S. R. (1991). Synaptic integration at a sensory-motor reflex in the leech.
The Journal of physiology, 441:733 -- 754.
Han, K.-S., Guo, C., Chen, C. H., Witter, L., Osorno, T., and Regehr, W. G. (2018). Ephaptic coupling promotes
synchronous firing of cerebellar purkinje cells. Neuron, 100(3):564 -- 578.e3.
Harris, J. J. and Attwell, D. (2012). The energetics of cns white matter. Journal of Neuroscience, 32(1):356 -- 371.
Higgs, M. H., Slee, S. J., and Spain, W. J. (2006). Diversity of gain modulation by noise in neocortical neurons:
regulation by the slow afterhyperpolarization conductance. Journal of Neuroscience, 26(34):8787 -- 8799.
Higgs, M. H. and Spain, W. J. (2009). Conditional bursting enhances resonant firing in neocortical layer 2 -- 3
pyramidal neurons. Journal of Neuroscience, 29(5):1285 -- 1299.
Hill, B. C., Schubert, E. D., Nokes, M. A., and Michelson, R. P. (1977). Laser interferometer measurement of
changes in crayfish axon diameter concurrent with action potential. Science, 196(4288):426 -- 428.
Hines, J. H., Ravanelli, A. M., Schwindt, R., Scott, E. K., and Appel, B. (2015). Neuronal activity biases axon
selection for myelination in vivo. Nature Neuroscience, 18:683 EP -- .
Hodgkin, A. L. and Huxley, A. F. (1952). A quantitative description of membrane current and its application to
conduction and excitation in nerve. The Journal of physiology, 117(4):500 -- 544.
Hodgkin, A. L. and Katz, B. (1949). The effect of sodium ions on the electrical activity of the giant axon of the
squid. The Journal of physiology, 108(1):37 -- 77.
Horn, R. and Miller, J. J. (1978). Calcium-dependent increase in spike duration during repetitive firing of aplysia
axon in the presence of tea. Journal of Neurobiology, 9(5):341 -- 352.
Howarth, J. (1975). Heat production in non-myelinated nerves. Philosophical Transactions of the Royal Society
of London. B, Biological Sciences, 270(908):425 -- 432.
Howarth, J., Keynes, R., Ritchie, J., and Von Muralt, A. (1975). The heat production associated with the passage
of a single impulse in pike olfactory nerve fibres. The Journal of physiology, 249(2):349 -- 368.
Hu, W., Tian, C., Li, T., Yang, M., Hou, H., and Shu, Y. (2009). Distinct contributions of nav1.6 and nav1.2 in
action potential initiation and backpropagation. Nature Neuroscience, 12:996 EP -- .
Huang, M., Volgushev, M., and Wolf, F. (2012). A small fraction of strongly cooperative sodium channels boosts
neuronal encoding of high frequencies. PLoS One, 7(5):e37629.
Huxley, A. F. and Stampfli, R. (1949). Evidence for saltatory conduction in peripheral myelinated nerve fibres.
The Journal of physiology, 108(3):315 -- 339.
Inda, M. C., DeFelipe, J., and Munoz, A. (2006). Voltage-gated ion channels in the axon initial segment of human
cortical pyramidal cells and their relationship with chandelier cells. Proceedings of the National Academy of
Sciences, 103(8):2920 -- 2925.
Iwasa, K. and Tasaki, I. (1980). Mechanical changes in squid giant axons associated with production of action
potentials. Biochemical and biophysical research communications, 95(3):1328 -- 1331.
Izhikevich, E. M., Gally, J. A., and Edelman, G. M. (2004). Spike-timing Dynamics of Neuronal Groups. Cerebral
Cortex, 14(8):933 -- 944.
Kaller, M. S., Lazari, A., Blanco-Duque, C., Sampaio-Baptista, C., and Johansen-Berg, H. (2017). Myelin
plasticity and behaviour -- connecting the dots. Current Opinion in Neurobiology, 47:86 -- 92.
16
Kamasawa, N., Sik, A., Morita, M., Yasumura, T., Davidson, K., Nagy, J., and Rash, J. (2005). Connexin-47
and connexin-32 in gap junctions of oligodendrocyte somata, myelin sheaths, paranodal loops and schmidt-
lanterman incisures: Implications for ionic homeostasis and potassium siphoning. Neuroscience, 136(1):65 --
86.
Katz, B. and Schmitt, O. H. (1940). Electric interaction between two adjacent nerve fibres. The Journal of
physiology, 97(4):471 -- 488.
Keynes, R. (1951). The ionic movements during nervous activity. The Journal of physiology, 114(1-2):119 -- 150.
Khaliq, Z. M. and Raman, I. M. (2006). Relative contributions of axonal and somatic na channels to action
potential initiation in cerebellar purkinje neurons. Journal of Neuroscience, 26(7):1935 -- 1944.
Kim, G. E., Kronengold, J., Barcia, G., Quraishi, I. H., Martin, H. C., Blair, E., Taylor, J. C., Dulac, O.,
Colleaux, L., Nabbout, R., and Kaczmarek, L. K. (2014). Human slack potassium channel mutations increase
positive cooperativity between individual channels. Cell Reports, 9(5):1661 -- 1672.
Kole, M. H., Letzkus, J. J., and Stuart, G. J. (2007). Axon initial segment kv1 channels control axonal action
potential waveform and synaptic efficacy. Neuron, 55(4):633 -- 647.
Kole, M. H. P. (2011). First node of ranvier facilitates high-frequency burst encoding. Neuron, 71(4):671 -- 682.
Kole, M. H. P., Ilschner, S. U., Kampa, B. M., Williams, S. R., Ruben, P. C., and Stuart, G. J. (2008). Action
potential generation requires a high sodium channel density in the axon initial segment. Nature Neuroscience,
11:178 EP -- .
Kopp-Scheinpflug, C. and Tempel, B. L. (2015). Decreased temporal precision of neuronal signaling as a candidate
mechanism of auditory processing disorder. Hearing Research, 330:213 -- 220.
Kordeli, E., Lambert, S., and Bennett, V. (1995). Ankyrin a new ankyrin gene with neural-specific isoforms
localized at the axonal initial segment and node of ranvier. Journal of Biological Chemistry, 270(5):2352 -- 2359.
Krnjevic, K. and Miledi, R. (1959). Presynaptic failure of neuromuscular propagation in rats. The Journal of
physiology, 149(1):1 -- 22.
Kuba, H., Adachi, R., and Ohmori, H. (2014). Activity-dependent and activity-independent development of the
axon initial segment. Journal of Neuroscience, 34(9):3443 -- 3453.
Larkum, M. E., Zhu, J. J., and Sakmann, B. (1999). A new cellular mechanism for coupling inputs arriving at
different cortical layers. Nature, 398(6725):338 -- 341.
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(Pt 2):447 -- 466.
Lazarov, E., Dannemeyer, M., Feulner, B., Enderlein, J., Gutnick, M. J., Wolf, F., and Neef, A. (2018). An
axon initial segment is required for temporal precision in action potential encoding by neuronal populations.
Science advances, 4(11):eaau8621.
Lehnert, S., Ford, M. C., Alexandrova, O., Hellmundt, F., Felmy, F., Grothe, B., and Leibold, C. (2014).
Action potential generation in an anatomically constrained model of medial superior olive axons. Journal of
Neuroscience, 34(15):5370 -- 5384.
Lien, C.-C. and Jonas, P. (2003). Kv3 potassium conductance is necessary and kinetically optimized for high-
frequency action potential generation in hippocampal interneurons. Journal of Neuroscience, 23(6):2058 -- 2068.
Lillie, R. S. (1925). Factors affecting transmission and recovery in the passive iron nerve model. The Journal of
general physiology, 7(4):473 -- 507.
Lorincz, A. and Nusser, Z. (2010). Molecular identity of dendritic voltage-gated sodium channels. Science,
328(5980):906 -- 909.
Mackie, G. O. and Meech, R. W. (1985). Separate sodium and calcium spikes in the same axon. Nature,
313(6005):791 -- 793.
Mainen, Z. F. and Sejnowski, T. J. (1996). Influence of dendritic structure on firing pattern in model neocortical
neurons. Nature, 382(6589):363 -- 366.
17
Manor, Y., Koch, C., and Segev, I. (1991). Effect of geometrical irregularities on propagation delay in axonal
trees. Biophysical Journal, 60(6):1424 -- 1437.
Mar, A. and Drapeau, P. (1996). Modulation of conduction block in leech mechanosensory neurons. Journal of
Neuroscience, 16(14):4335 -- 4343.
Maranto, A. R. and Calabrese, R. L. (1984). Neural control of the hearts in the leech,hirudo medicinalis. Journal
of Comparative Physiology A, 154(3):381 -- 391.
Maratou, E. and Theophilidis, G. (2000). An axon pacemaker: diversity in the mechanism of generation and
conduction of action potentials in snail neurons. Neuroscience, 96(1):1 -- 2.
Marder, E. and Goaillard, J.-M. (2006). Variability, compensation and homeostasis in neuron and network
function. Nature Reviews Neuroscience, 7(7):563.
Marder, E. and Taylor, A. L. (2011). Multiple models to capture the variability in biological neurons and
networks. Nature Neuroscience, 14:133 EP -- .
Marx, S. O., Gaburjakova, J., Gaburjakova, M., Henrikson, C., Ondrias, K., and Marks, A. R. (2001). Coupled
gating between cardiac calcium release channels (ryanodine receptors). Circulation research, 88(11):1151 -- 1158.
McCormick, D. A., Shu, Y., and Yu, Y. (2007). Hodgkin and huxley model -- still standing? Nature, 445:E1 EP
-- .
McKenzie, I. A., Ohayon, D., Li, H., Paes de Faria, J., Emery, B., Tohyama, K., and Richardson, W. D. (2014).
Motor skill learning requires active central myelination. Science, 346(6207):318 -- 322.
Meeks, J. P. and Mennerick, S. (2007). Action potential initiation and propagation in ca3 pyramidal axons.
Journal of Neurophysiology, 97(5):3460 -- 3472.
Mejia-Gervacio, S., Collin, T., Pouzat, C., Tan, Y. P., Llano, I., and Marty, A. (2007). Axonal speeding: Shaping
synaptic potentials in small neurons by the axonal membrane compartment. Neuron, 53(6):843 -- 855.
Meyrand, P., Weimann, J., and Marder, E. (1992). Multiple axonal spike initiation zones in a motor neuron:
serotonin activation. Journal of Neuroscience, 12(7):2803 -- 2812.
Micu, I., Plemel, J. R., Caprariello, A. V., Nave, K.-A., and Stys, P. K. (2017). Axo-myelinic neurotransmission:
a novel mode of cell signalling in the central nervous system. Nature Reviews Neuroscience, 19:49 EP -- .
Milescu, L. S., Yamanishi, T., Ptak, K., and Smith, J. C. (2010). Kinetic properties and functional dynamics of
sodium channels during repetitive spiking in a slow pacemaker neuron. Journal of Neuroscience, 30(36):12113 --
12127.
Monsivais, P., Clark, B. A., Roth, A., and Hausser, M. (2005). Determinants of action potential propagation in
cerebellar purkinje cell axons. Journal of Neuroscience, 25(2):464 -- 472.
Mulloney, B. and Selverston, A. (1972). Antidromic action potentials fail to demonstrate known interactions
between neurons. Science, 177(4043):69 -- 72.
Naundorf, B., Wolf, F., and Volgushev, M. (2006). Unique features of action potential initiation in cortical
neurons. Nature, 440(7087):1060.
Neef, A., Hady, A. E., Nagpal, J., Broking, K., Afshar, G., Schluter, O. M., Geisel, T., Bamberg, E., Fleischmann,
R., Stuhmer, W., et al. (2013). Continuous dynamic photostimulation-inducing in-vivo-like fluctuating con-
ductances with channelrhodopsins. arXiv preprint arXiv:1305.7125.
Norenberg, A., Hu, H., Vida, I., Bartos, M., and Jonas, P. (2010). Distinct nonuniform cable properties optimize
rapid and efficient activation of fast-spiking gabaergic interneurons. Proceedings of the National Academy of
Sciences, 107(2):894.
Ofer, N., Shefi, O., and Yaari, G. (2019). Axonal tree morphology and signal propagation dynamics improve
interneuron classification. bioRxiv.
Oh, S., Fang-Yen, C., Choi, W., Yaqoob, Z., Fu, D., Park, Y., Dassari, R. R., and Feld, M. S. (2012). Label-free
imaging of membrane potential using membrane electromotility. Biophysical journal, 103(1):11 -- 18.
Palay, S. L. and Chan-Palay, V. (2012). Cerebellar cortex: cytology and organization. Springer Science & Business
Media.
18
Palmer, L. M., Clark, B. A., Grndemann, J., Roth, A., Stuart, G. J., and Husser, M. (2010). Rapid report:
Initiation of simple and complex spikes in cerebellar purkinje cells. The Journal of Physiology, 588(10):1709 --
1717.
Palmer, L. M. and Stuart, G. J. (2006). Site of action potential initiation in layer 5 pyramidal neurons. Journal
of Neuroscience, 26(6):1854 -- 1863.
Perge, J. A., Koch, K., Miller, R., Sterling, P., and Balasubramanian, V. (2009). How the optic nerve allocates
space, energy capacity, and information. Journal of Neuroscience, 29(24):7917 -- 7928.
Pinsky, P. F. and Rinzel, J. (1994).
Intrinsic and network rhythmogenesis in a reduced traub model for ca3
neurons. Journal of Computational Neuroscience, 1(1):39 -- 60.
Powers, R. K., Sawczuk, A., Musick, J. R., and Binder, M. D. (1999). Multiple mechanisms of spike-frequency
adaptation in motoneurones. Journal of Physiology-Paris, 93(1-2):101 -- 114.
Ranade, S. S., Syeda, R., and Patapoutian, A. (2015). Mechanically activated ion channels. Neuron, 87(6):1162 --
1179.
Rasband, M. N. (2010). The axon initial segment and the maintenance of neuronal polarity. Nature Reviews
Neuroscience, 11(8):552.
Rash, J. E., Vanderpool, K. G., Yasumura, T., Hickman, J., Beatty, J. T., and Nagy, J. I. (2016). Kv1 channels
identified in rodent myelinated axons, linked to cx29 in innermost myelin: support for electrically active myelin
in mammalian saltatory conduction. Journal of Neurophysiology, 115(4):1836 -- 1859.
Rios, J. C., Rubin, M., Martin, M. S., Downey, R. T., Einheber, S., Rosenbluth, J., Levinson, S. R., Bhat, M.,
and Salzer, J. L. (2003). Paranodal interactions regulate expression of sodium channel subtypes and provide
a diffusion barrier for the node of ranvier. Journal of Neuroscience, 23(18):7001 -- 7011.
Ritchie, J. M. and Rogart, R. B. (1977). Density of sodium channels in mammalian myelinated nerve fibers and
nature of the axonal membrane under the myelin sheath. Proceedings of the National Academy of Sciences of
the United States of America, 74(1):211 -- 215.
Robertson, J. D., Bodenheimer, T. S., and Stage, D. E. (1963). The ultrastructure of mauthner cell synapses
and nodes in goldish brains. The Journal of cell biology, 19(1):159 -- 199.
Roopun, A. K., Simonotto, J. D., Pierce, M. L., Jenkins, A., Nicholson, C., Schofield, I. S., Whittaker, R. G.,
Kaiser, M., Whittington, M. A., Traub, R. D., and Cunningham, M. O. (2010). A nonsynaptic mechanism
underlying interictal discharges in human epileptic neocortex. Proceedings of the National Academy of Sciences,
107(1):338 -- 343.
Rosenbluth, J. (1976). Intramembranous particle distribution at the node of ranvier and adjacent axolemma in
myelinated axons of the frog brain. Journal of Neurocytology, 5(6):731 -- 745.
Rudolph, M. and Destexhe, A. (2003). Characterization of subthreshold voltage fluctuations in neuronal mem-
branes. Neural Computation, 15(11):2577 -- 2618.
Rush, A. M., Dib-Hajj, S. D., and Waxman, S. G. (2005). Electrophysiological properties of two axonal sodium
channels, nav1.2 and nav1.6, expressed in mouse spinal sensory neurones. The Journal of physiology, 564(Pt
3):803 -- 815.
Rushton, W. A. H. (1951). A theory of the effects of fibre size in medullated nerve. The Journal of physiology,
115(1):101 -- 122.
Sampaio-Baptista, C., Khrapitchev, A. A., Foxley, S., Schlagheck, T., Scholz, J., Jbabdi, S., DeLuca, G. C.,
Miller, K. L., Taylor, A., Thomas, N., Kleim, J., Sibson, N. R., Bannerman, D., and Johansen-Berg, H. (2013).
Motor skill learning induces changes in white matter microstructure and myelination. Journal of Neuroscience,
33(50):19499 -- 19503.
Sawczuk, A., Powers, R. K., and Binder, M. D. (1997). Contribution of outward currents to spike-frequency
adaptation in hypoglossal motoneurons of the rat. Journal of neurophysiology, 78(5):2246 -- 2253.
Schmidt, D., del M´armol, J., and MacKinnon, R. (2012). Mechanistic basis for low threshold mechanosensitivity
in voltage-dependent k+ channels. Proceedings of the National Academy of Sciences, 109(26):10352 -- 10357.
Schmidt, D. and MacKinnon, R. (2008). Voltage-dependent k+ channel gating and voltage sensor toxin sensitivity
depend on the mechanical state of the lipid membrane. Proceedings of the National Academy of Sciences,
105(49):19276 -- 19281.
19
Schmidt-Hieber, C., Jonas, P., and Bischofberger, J. (2008). Action potential initiation and propagation in
hippocampal mossy fibre axons. The Journal of Physiology, 586(7):1849 -- 1857.
Schmitz, D., Schuchmann, S., Fisahn, A., Draguhn, A., Buhl, E. H., Petrasch-Parwez, E., Dermietzel, R.,
Heinemann, U., and Traub, R. D. (2001). Axo-axonal coupling: A novel mechanism for ultrafast neuronal
communication. Neuron, 31(5):831 -- 840.
Seidl, A. H., Rubel, E. W., and Harris, D. M. (2010). Mechanisms for adjusting interaural time differences to
achieve binaural coincidence detection. Journal of Neuroscience, 30(1):70 -- 80.
Sheffield, M. E. J., Best, T. K., Mensh, B. D., Kath, W. L., and Spruston, N. (2010). Slow integration leads to
persistent action potential firing in distal axons of coupled interneurons. Nature Neuroscience, 14:200 EP -- .
Shu, Y., Hasenstaub, A., Duque, A., Yu, Y., and McCormick, D. A. (2006). Modulation of intracortical synaptic
potentials by presynaptic somatic membrane potential. Nature, 441(7094):761 -- 765.
Sinclair, J. L., Fischl, M. J., Alexandrova, O., He, M., Grothe, B., Leibold, C., and Kopp-Scheinpflug, C.
(2017). Sound-evoked activity influences myelination of brainstem axons in the trapezoid body. Journal of
Neuroscience, 37(34):8239 -- 8255.
Song, A.-h., Wang, D., Chen, G., Li, Y., Luo, J., Duan, S., and Poo, M.-m. (2009). A selective filter for
cytoplasmic transport at the axon initial segment. Cell, 136(6):1148 -- 1160.
Spira, M. E., Oren, R., Dormann, A., and Gitler, D. (2003). Critical calpain-dependent ultrastructural alterations
underlie the transformation of an axonal segment into a growth cone after axotomy of cultured aplysia neurons.
Journal of Comparative Neurology, 457(3):293 -- 312.
Spirou, G., Chirila, F., von Gersdorff, H., and Manis, P. (2008). Heterogeneous ca2+ influx along the adult calyx
of held: A structural and computational study. Neuroscience, 154(1):171 -- 185.
Stasheff, S. F., Hines, M., and Wilson, W. A. (1993). Axon terminal hyperexcitability associated with epilepto-
genesis in vitro. i. origin of ectopic spikes. Journal of Neurophysiology, 70(3):961 -- 975.
Stuart, G. and Sakmann, B. (1995). Amplification of epsps by axosomatic sodium channels in neocortical
pyramidal neurons. Neuron, 15(5):1065 -- 1076.
Stuart, G., Spruston, N., and Hausser, M. (2016). Dendrites. Oxford University Press.
Stuart, G., Spruston, N., Sakmann, B., and Hausser, M. (1997). Action potential initiation and backpropagation
in neurons of the mammalian cns. Trends in Neurosciences, 20(3):125 -- 131.
Sudhof, T. C. and Malenka, R. C. (2008). Understanding synapses: Past, present, and future. Neuron, 60(3):469
-- 476.
Tasaki, I. and Byrne, P. M. (1992). Heat production associated with a propagated impulse in bullfrog myelinated
nerve fibers. The Japanese journal of physiology, 42(5):805 -- 813.
Tasaki, I. and Iwasa, K. (1982a). Further studies of rapid mechanical changes in squid giant axon associated
with action potential production. The Japanese journal of physiology, 32(4):505 -- 518.
Tasaki, I. and Iwasa, K. (1982b). Rapid pressure changes and surface displacements in the squid giant axon
associated with production of action potentials. The Japanese journal of physiology, 32(1):69 -- 81.
Tchumatchenko, T., Malyshev, A., Geisel, T., Volgushev, M., and Wolf, F. (2010). Correlations and synchrony
in threshold neuron models. Physical review letters, 104(5):058102.
Tchumatchenko, T., Malyshev, A., Wolf, F., and Volgushev, M. (2011). Ultrafast population encoding by cortical
neurons. Journal of Neuroscience, 31(34):12171 -- 12179.
Trigo, F. F., Bouhours, B., Rostaing, P., Papageorgiou, G., Corrie, J. E., Triller, A., Ogden, D., and Marty, A.
(2010). Presynaptic miniature gabaergic currents in developing interneurons. Neuron, 66(2):235 -- 247.
Tyler, W. J. (2012). The mechanobiology of brain function. Nature Reviews Neuroscience, 13(12):867.
Van Essen, D. C. (1973). The contribution of membrane hyperpolarization to adaptation and conduction block
in sensory neurones of the leech. The Journal of Physiology, 230(3):509 -- 534.
Wang, S., Borst, A., Zaslavsky, N., Tishby, N., and Segev, I. (2017). Efficient encoding of motion is mediated by
gap junctions in the fly visual system. PLOS Computational Biology, 13(12):1 -- 22.
20
Watanabe, A. and Grundfest, H. (1961). Impulse propagation at the septal and commissural junctions of crayfish
lateral giant axons. The Journal of General Physiology, 45(2):267 -- 308.
Wei, W. and Wolf, F. (2011). Spike onset dynamics and response speed in neuronal populations. Physical Review
Letters, 106(8):088102.
Wollner, D. A. and Catterall, W. A. (1986). Localization of sodium channels in axon hillocks and initial segments
of retinal ganglion cells. Proceedings of the National Academy of Sciences, 83(21):8424 -- 8428.
Wu, L. M. N., Williams, A., Delaney, A., Sherman, D. L., and Brophy, P. J. (2012). Increasing internodal distance
in myelinated nerves accelerates nerve conduction to a flat maximum. Current Biology, 22(20):1957 -- 1961.
Xiao, L., Ohayon, D., McKenzie, I. A., Sinclair-Wilson, A., Wright, J. L., Fudge, A. D., Emery, B., Li, H.,
and Richardson, W. D. (2016). Rapid production of new oligodendrocytes is required in the earliest stages of
motor-skill learning. Nature Neuroscience, 19:1210 EP -- .
Xu, K. and Terakawa, S. (1999). Fenestration nodes and the wide submyelinic space form the basis for the
unusually fast impulse conduction of shrimp myelinated axons. Journal of Experimental Biology, 202(15):1979 --
1989.
Yamamoto, T., Samejima, A., and Oka, H. (1990). The mode of synaptic activation of pyramidal neurons in the
cat primary somatosensory cortex: an intracellular hrp study. Experimental Brain Research, 80(1):12 -- 22.
Yang, Y., Ramamurthy, B., Neef, A., and Xu-Friedman, M. A. (2016). Low somatic sodium conductance enhances
action potential precision in time-coding auditory neurons. Journal of Neuroscience, 36(47):11999 -- 12009.
Yau, K. W. (1976). Receptive fields, geometry and conduction block of sensory neurones in the central nervous
system of the leech. The Journal of physiology, 263(3):513 -- 538.
Young, K. M., Psachoulia, K., Tripathi, R. B., Dunn, S.-J., Cossell, L., Attwell, D., Tohyama, K., and Richardson,
W. D. (2013). Oligodendrocyte dynamics in the healthy adult cns: evidence for myelin remodeling. Neuron,
77(5):873 -- 885.
Yu, Y., Shu, Y., and McCormick, D. A. (2008). Cortical action potential backpropagation explains spike threshold
variability and rapid-onset kinetics. Journal of Neuroscience, 28(29):7260 -- 7272.
21
Figures
Figure 1: Diversity of axons: morphology and function. A. Giant fibers in the crawfish contact each other
by electrical synapses. From Furshpan and Potter, 1958. B. Cerebellar basket cells collaterals contact several
Purkinje cells, which they inhibit at the basket structure via chemical synapses, and the pinceau via ephaptic
inhibition. From Palay and Chan-Palay, Springer Verlag Berlin, 1974. C. The climbing fiber in the cerebellum
shows different axonal specializations: the glomerular collateral in the granule cell layer and the climbing fiber
wrapping around Purkinje cells. From Palay and Chan-Palay, Springer Verlag Berlin, 1974. D. The honeybee
axon of the BC cell projects onto the left and right hemispheres, targeting a number of regions (one of the target
regions is labelled in yellow). From Zwaka et al., 2016.
22
Figure 2: Overview of axonal electrical signaling. Axonal signaling is formed by a mixture of analog and
digital signaling.
Insets: membrane potential traces. Action potentials can propagated orthodromically or
antidromically. Right, in blue, active sites with the machinery to generate or regenerate action potentials. AIS,
axon initial segment.
Figure 3: Overview of Action Potential Diversity. A Ultrastructure of the axon initial segment. Green is sodium
channels, orange circles are Ankryin G molecules, red lines are beta IV spectrin and orange lines are the actin
rings. B1. Left, example of fast rising action potential from a neocortical pyramidal cell. Right, phase plane
plot of this cell. B2. Left, example of a stellate cell from the cochlear nucleus with two components: a fast
rising phase of the action potential contributed by the AIS and a second phase contributed by sodium by the
somatodendritic compartment. Left, phase plane plot of this cell.
23
Figure 4: Equivalent circuits of axons and structural biophysical specializations that improve conduction Top
left. An axon can be reduced to an equivalent electrical circuit involving an axial resistor (parallel to the
membrane) conveyed by the intracellular medium of the axon (axoplasm), and a parallel RC circuit formed by
the membrane resistance and membrane capacitance. Top right, conduction can be improved by decreasing axon
diameter, and thereby longitudinal resistance to current flow, allowing for a larger current flow in the axon relative
to the membrane resistor, and thereby a larger space constant and faster conduction. Bottom, an alternative
circuit modification occurs with myelination: additional capacitances conveyed by myelin in series with axonal
capacitance reduce the effective capacitance and additional resistance increase the effective resistance, increasing
propagation speed and space constant, respectively. Note that the battery associated to each membrane resistor
has been omitted for simplification purposes. Orange arrows represent the current flow, and their thickness is
indicative of the relative current flow in the axoplasm and in the radial direction out of the axon.
24
Figure 5: Structure-function plasticity A. A1. Axon structural plasticity. A2. Functional implications: action
potential latency, measured by antidromically stimulating axons before and after a plasticity protocol, decreases
several minutes after the protocol. From Chereau et al. (2015) B. B1. Myelin structural plasticity can involve
changes in internode length (Etxebarria et al., 2016), in myelin thickness (Sinclair et al., 2017) or in node length
(suggested in Arancibia-Carcamo, 2017). B2. Functional implications of changes in myelination: MBP optical
density, putatively proportional to the amount of myelin positively correlates in the learning rate in a motor skill
learning paradigm. From Sampaio-Baptista et al. (2013).
Figure 6: Coupling modalities between axons. A. Three types of synapse-mediated coupling: through axo-axonal
gap junctions (left), ephaptic coupling (middle), and chemical axonic synapses (right). B. Indirect coupling via
myelinating cells.
25
|
1910.10559 | 1 | 1910 | 2019-10-23T13:42:53 | Working memory facilitates reward-modulated Hebbian learning in recurrent neural networks | [
"q-bio.NC",
"cs.NE"
] | Reservoir computing is a powerful tool to explain how the brain learns temporal sequences, such as movements, but existing learning schemes are either biologically implausible or too inefficient to explain animal performance. We show that a network can learn complicated sequences with a reward-modulated Hebbian learning rule if the network of reservoir neurons is combined with a second network that serves as a dynamic working memory and provides a spatio-temporal backbone signal to the reservoir. In combination with the working memory, reward-modulated Hebbian learning of the readout neurons performs as well as FORCE learning, but with the advantage of a biologically plausible interpretation of both the learning rule and the learning paradigm. | q-bio.NC | q-bio |
Working memory facilitates reward-modulated
Hebbian learning in recurrent neural networks
Roman Pogodin1,∗, Dane Corneil2, Alexander Seeholzer2, Joseph Heng2, Wulfram Gerstner2
1Gatsby Computational Neuroscience Unit, University College London, London, UK
2 School of Computer and Communication Sciences and School of Life Sciences,
Brain Mind Institute, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland
∗[email protected]
Abstract
Reservoir computing is a powerful tool to explain how the brain learns temporal se-
quences, such as movements, but existing learning schemes are either biologically
implausible or too inefficient to explain animal performance. We show that a net-
work can learn complicated sequences with a reward-modulated Hebbian learning
rule if the network of reservoir neurons is combined with a second network that
serves as a dynamic working memory and provides a spatio-temporal backbone sig-
nal to the reservoir. In combination with the working memory, reward-modulated
Hebbian learning of the readout neurons performs as well as FORCE learning, but
with the advantage of a biologically plausible interpretation of both the learning
rule and the learning paradigm.
1
Introduction
Learning complex temporal sequences that extend over a few seconds -- such as a movement to
grab a bottle or to write a number on the blackboard -- looks easy to us but is challenging for
computational brain models. A common framework for learning temporal sequences is reservoir
computing (alternatively called liquid computing or echo-state networks) [1, 2, 3]. It combines a
reservoir, a recurrent network of rate units with strong, but random connections [4], with a linear
readout that feeds back to the reservoir. Training of the readout weights with FORCE, a recursive
least-squares estimator [1], leads to excellent performance on many tasks such as motor movements.
The FORCE rule is, however, biologically implausible: update steps of synapses are rapid and large,
and require an immediate and precisely timed feedback signal. A more realistic alternative to FORCE
is the family of reward-modulated Hebbian learning rules [5, 6, 7], but plausibility comes at a price:
when the feedback (reward minus expected reward) is given only after a long delay, reward-modulated
Hebbian plasticity is not powerful enough to learn complex tasks.
Here we combine the reservoir network with a second, more structured network that stores and
updates a two-dimension continuous variable as a "bump" in an attractor [8, 9]. The activity of the
attractor network acts as a dynamic working memory and serves as input to the reservoir network
(fig. 1). Our approach is related to that of feeding an abstract oscillatory input [10] or a "temporal
backbone signal" [11] into the reservoir in order to overcome structural weaknesses of reservoir
computing that arise if large time spans need to be covered.
In computational experiments, we show that a dynamic working memory that serves as an input
to a reservoir network facilitates reward-modulated Hebbian learning in multiple ways: it makes
a biologically plausible three-factor rule as efficient as FORCE; it admits a delay in the feedback
signal; and it allows a single reservoir network to learn and perform multiple tasks.
NeurIPS 2019 workshop "Real Neurons & Hidden Units: Future directions at the intersection of neuroscience
and artificial intelligence", Vancouver, Canada.
2 Model
Our architecture is simple: the attractor network (the "memory") receives some task-specific input
and produces a robust two-dimensional neural trajectory; the reservoir network (the "motor cortex")
shapes its dynamics with this trajectory, and produces a potentially high-dimensional output (fig. 1).
Figure 1: Model architecture: a moving 2D bump (left: activity bump (red) surrounded by inactive
neurons (blue)) in an attractor network (left circle with two bump trajectories) projects to a reservoir
(right circle); the output z(t) is read out from the reservoir and approximates the target function.
Attractor network. Following [9], the bump attractor consists of 2500 neurons evolving as
τm x = −x + [Jx + e − h]+ ,
τa
h = −h + s x ,
(1)
where x is the vector of firing rates evolving with time constant τm, e is the task-specific external
input, h is an adaptation variable with time constant τa and s is the strength of adaptation. The weight
matrix J = Js + Jh has two parts. The symmetric part Js creates a two-dimensional translation-
invariant structure resulting in bump-like stable activity patterns, whereas Jh represents structural
noise. Due to the adaptation h, the bump moves across a path defined by the initial conditions and
structural noise, creating long -- lasting reliable activity patterns which also depend on the input e.
Reservoir network. The reservoir learns to approximate a target function f (t) with the output z(t)
by linearly combining the firing rate r with readout weights Wro: z = Wror + ηηη ≡ z + ηηη with
readout noise ηηη. We use the same number of neurons (1000) and parameters as [1, 6],
τ u = −u + λWrecr + Wfbz + c Wattrx ,
r = tanh(u) + ξξξ ,
(2)
where u is the membrane potential, ξξξ is the firing rate noise, Wattr scales attractor input with
coupling c, Wrec and λ regulate chaotic activity [4], and Wfb implements the feedback loop.
Learning rule. We use the reward-modulated Hebbian rule of [6] for the readout weights Wro,
∆Wro(t) = η(t)M (t)(z(t) − ¯z(t))r(cid:62)(t) ,
(3)
where ¯x denotes low-pass filtering of x, such that z(t) − ¯z(t) ≈ ηηη(t). The reward modulation M (t)
tracks performance P (t) as
η(t) = η0/(1 + t/τη) ,
(cid:26)1, P (t) > ¯P (t) ,
0, P (t) ≤ ¯P (t) .
P (t) = −(cid:107)f (t) − z(t)(cid:107)2 , M (t) =
(4)
The update rule is an example of a NeoHebbian three-factor learning rule [12] and mimics gradient
descent if we ignore the feedback loop [5]. For model details, see appendix A.
3 Experiments
In fig. 2, the learning rules are compared on 50 target functions sampled from a Gaussian Process
(GP) with exponential squared kernel (σ2 = 104 to match the complexity of hand-picked functions
from [1, 6]). After each training period, we measure performance with normalized cross-correlation
between the output and the target (ranging from -1 to 1, where 1 is a perfect match) on a single trial
with frozen weights. Details are provided in appendix A; code: https://github.com/neuroai-workshop-
anon-1224113/working-memory-facilitating-reservoir-learning.
3.1 Reward-modulated Hebbian learning with attractor input reaches FORCE performance
When tested on one-second signals similar to those of [1, 6] (two insets in fig. 2), the full network
with attractor input and reward-modulated Hebbian learning learns faster and more reliably than
2
reward-modulated Hebbian learning without the input from the attractor network. After about 90
training trials, the full network achieves the performance of the FORCE rule (for which training error
approaches one in the first trial, [1], while test error does so after 30-50 trials, [6]; fig. 2A).
For target signals that extend over 10 seconds (same smoothness of the target functions, two insets in
fig. 2B), the reward-modulated Hebbian rule achieves a performance of 1 after 200 trials if combined
with input from the attractor network (fig. 2B) but fails completely without the attractor network
(tuning of the hyperparameters on a logarithmic scale did not help; data not shown). Thus a three-
factor learning rule succeeds to learn complex tasks if combined with a temporally structured input
from the attractor network.
Figure 2: A. Mean test performance as a function of trials with a reward-modulated Hebbian (rmHebb)
learning rule combined with (red) or without (green) input from attractor. Inset: target function over
1s and examples of the output after training. B. Same, but for a target signal that spans over 10s.
Shaded area shows standard deviation. The FORCE rule achieves a performance of 1 in all trials.
C. Learning performance for different couplings c of the attractor output to reservoir input (c = 1
normal coupling, c = 0 no coupling) when updates are delayed to the end of every one-second trial.
3.2 Attractor input allows realistic frequency of weight updates
FORCE learning needs a feedback signal at every time step. Standard reward-modulated Hebbian
learning can support very small delays, but fails if updates are less frequent than every few ms [6]. In
our approach (fig. 2C), proposed updates are summed up in the background, but applied only at the
end of a one-second trial. We find that even with such a temporally sparse update, learning is still
possible.
The input from the dynamic working memory is necessary to achieve this task: when the strength of
the input from the attractor network gradually decreases, performance drops; in the total absence of
attractor input (c = 0.0; note that the reservoir still receives weak input noise) learning completely
fails. Strikingly, delayed updates do not hurt performance, and the system achieves high (> 0.9)
cross-correlation in fewer than 100 training trials if the input from the attractor network is strong
enough. The transient drop in performance shortly after the start in fig. 2C is likely due to Wro = 0
in the beginning, meaning that the output is uncorrelated with the firing rates, and therefore the
cumulative weight update does not approximate gradient information.
3.3 Working memory translates to efficient reservoir learning of multiple signals
It is well known that reservoir networks can learn multiple tasks given different initial conditions
with both FORCE [1] and the reward-modulated Hebbian rule [6]. We want to check whether this
also holds for our approach. We conjecture that different inputs to the attractor network generate
unique neural trajectories [9] that can be exploited by the reservoir network.
To test this hypothesis, we train the network to produce hand-written digits. The static input to the
attractor comes from the pre-processed MNIST dataset (network inputs are taken from one of the last
layers of a deep network trained to classify MNIST) in order to provide a realistic input to the attractor
network which transforms the static input into dynamic trajectories (noiseless, fig. 3B, and noisy,
fig. 3D). We record 50 attractor trajectories used for training (used 4 times each, resulting in 2000
training trials) and 50 for testing of each digit (1 second each), where each trajectory corresponds
to a distinct input pattern. The reservoir learns a single drawing for each class. The variance of the
structural noise in the attractor network is 3 times larger compared to the previous experiments in
order to produce more robust bump trajectories (fig. 3D).
The reward-modulated Hebbian rule masters 10 out of 10 digits when driven by a noiseless input
from the attractor network (fig. 3A). In the presence of noise in the attractor network (fig. 3D), the
3
050100150200250300350400period0.00.20.40.60.81.0cross-corr ± STDrmHebbrmHebb + attractorAtargetoutput050100150200250300350400period0.00.20.40.60.81.0cross-corr ± STD020406080100period0.00.20.40.60.81.0cross-corr ± SEMBCc=0.0c=0.1c=0.2c=0.3c=0.4c=0.5-1 performance is imperfect for "five" and "six" (fig. 3C). We checked that FORCE learning with the
same noisy input did not improve the performance (data not shown). Note that a linear readout of the
attractor (without the reservoir) would be insufficient: first, sometimes single digit trajectories are
very dissimilar (e.g. the different zero's in fig. 3D); second, at points where trajectories cross each
other, a delay-less linear readout must produce the same output, no matter what the digit is.
Figure 3: A. Targets (green) and test outputs (red) for reward-modulated Hebbian rule with noiseless
attractor input. B. Noiseless attractor trajectories for each digit. C. Same as A but with noisy attractor
trajectories. D. Same as B but static input's noise resulted in noisy trajectories (individual lines).
4 Discussion
We showed that a dynamic working memory can facilitate learning of complex tasks with biologically
plausible three-factor learning rules. Our results indicate that, when combined with a bump attractor,
reservoir computing with reward-modulated learning can be as efficient as FORCE [1], a widely used
but biologically unrealistic rule. The proposed network relies on a limited number of trajectories in
the attractor network. To increase its capacity, a possible future direction would be to combine input
from the attractor network with another, also input-specific, but transient input that would bring the
reservoir into a different initial state. In this case the attractor network would work as a time variable
(as in [9]), and the other input as the control signal (as in [1]).
Apart from the biological relevance, the proposed method might be used for real-world applications of
reservoir computing (e.g. wind forecasting [13]) as it is computationally less expensive than FORCE.
It might also be an interesting alternative for learning in neuromorphic devices.
Acknowledgments
This research was supported by the Gatsby Charitable Foundation, Swiss National Science Foundation
(no. 200020 - 184615) and by the European Union Horizon 2020 Framework Program under grant
agreement no. 785907 (HumanBrain Project, SGA2). The authors thank Moritz Deger for an earlier
version of the reservoir code and Jorge Aurelio Menendez for useful comments on the manuscript.
References
[1] D. Sussillo and L.F. Abbott. Neuron, 63:544 -- 557, 2009.
[2] W. Maass, T. Natschläger, and H. Markram. Neural Computation, 14:2531 -- 2560, 2002.
[3] H. Jaeger and H. Haas. Science, 304:78 -- 80, 2004.
[4] H. Sompolinsky, A. Crisanti, and H. J. Sommers. Physical Review Letters, 61:259 -- 262, 1988.
[5] R. Legenstein, S. M. Chase, A. B. Schwartz, and W. Maass. Journal of Neuroscience, 30:8400 -- 8410, 2010.
[6] G.M. Hoerzer, R. Legenstein, and W. Maass. Cerebral cortex, 24:677 -- 90, 2014.
[7] N. Frémaux and W. Gerstner. Frontiers in Neural Circuits, 9:85, 2016.
[8] S. Wu, K. Hamaguchi, and S. Amari. Neural computation, 20:994 -- 1025, 2008.
[9] V. Itskov, C. Curto, E. Pastalkova, and G. Buzsaki. Journal of Neuroscience, 31:2828 -- 2834, 2011.
[10] P. Vincent-Lamarre, G. Lajoie, and J.P. Thivierge. J. Comput. Neurosci., 41:305 -- 322, 2016.
[11] W. Nicola and C. Clopath. Nat. Comm., 8:2208, 2017.
[12] W. Gerstner, M. Lehmann, V. Liakoni, D. Corneil, and J. Brea. Front. Neural Circ., 12:53, 2018.
[13] E. López, C. Valle, H. Allende, E. Gil, and H. Madsen. Energies, 11:526, 2018.
4
DABC A Supplementary Materials
Simulation details. Both networks were simulated with the Euler method with the step size dt = 1
ms. The attractor network dynamics was recorded after a 100 ms warm up period to allow creation of
the bump solution, during which it received additional input from the images in section 3.3. Training
was done consequently, without breaks in the dynamics between trials. For testing, the network started
from the preceding training state and continued with frozen weights. After testing, the pre-training
activity was restored. The code for experiments is available at https://github.com/neuroai-workshop-
anon-1224113/working-memory-facilitating-reservoir-learning.
Test functions. Gaussian process test function were drawn from
(5)
Forcing both ends of the function to be zero and denoting x = (0, T − 1)(cid:62), z = (1, . . . , T − 2)(cid:62),
we sample test functions as
fGP ∼ GP (0, K) , K(x, y) = exp(cid:0)−(x − y)2/(2σ2)(cid:1) .
fGP (1, . . . , T − 2) ∼ N(cid:0)0, K(z, x)K−1(x, x)K(x, z)(cid:1) ,
(6)
where T is either 103 (short tasks) or 104 (long tasks). We chose σ2 to roughly match the complexity
of targets from [6] (σ2 = 104). 50 random functions were tested on 50 random reservoirs that
nevertheless received the same attractor input (Wattr was not resampled). In section 3.3, the same
reservoir was used for all runs. The noisy input for section 3.3 was taken from an intermediate layer
of a deep network trained to classify MNIST, and the noiseless input stimulated only a 5 by 5 square
of neurons (unique for each digit).
Attractor network parameters. The time constants were τm = 30 ms, τa = 400 ms. Adaptation
strength was s = 1.5. The external input e was drawn independently for each neuron from a Gaussian
distribution N (1, 0.00252). In section 3.3, the task-specific input was added to the noisy one.
For the connectivity matrix J = Js + Jh, the noisy part was drawn independently as (Jh)ij ∼
N (0, σ2/Nattr), with Nattr = 2500 and σ = 2 in all experiments except for section 3.3, where we
used σ = 6 for more robust trajectories. The symmetric part arranged the neurons on a 2D grid, such
that every neuron i had its coordinates xi and yi ranging from 0 to 49. The connectivity led to mutual
excitation of nearby neurons and inhibition of the distant ones,
(Js)ij = −0.375 +
exp(cid:0)−d(i, j)2/2(cid:1) ,
(7)
(cid:113)
1√
2π
d(i, j) =
π
L
(min(xi − xj, L − xi − xj))2 + (min(yi − yj, L − yi − yj))2, L = 50. (8)
The bump center (used in fig. 3B and D) corresponded to the mean of the activity on the torus.
Denoting activity of each neuron as r(x, y), the center on the x axis was calculated as
xcenter =
L
2π
angle
1
L
r(xn, ym)e2πin/L
, L = 50,
(9)
where "angle" computes the counterclockwise angle of a complex variable (ranging from 0 to 2π).
Reservoir network parameters. The time constant was τ = 50 ms, and total coupling strength
was λ = 1.5. The readout weights Wro were initialized to zero. The feedback weights were
drawn independently from a uniform distribution as (Wfb)ij ∼ U(−1, 1). Both the recurrent
connections and the weights from the attractor to the reservoir were drawn independently as
(Wrec)ij, (Wattr)ij ∼ N (0, 1/pNres) · Be(p), with p = 0.1, Nres = 1000, and Be being the
Bernoulli distribution. A new reservoir, and thus Wrec, was sampled for each new test function. The
matrix Wrec was the same for all tasks except the last one in section 3.3.
State noise ξξξ and exploratory noise ηηη were generated independently from the uniform distribution
as ξi ∼ U(−0.05, 0.05), ηi ∼ U(−0.5, 0.5). When attractor was present, the reservoir neurons also
received weak independent noise drawn from N (0, 0.0025).
Learning rule. Low-pass filtering was done as
¯x(t + dt) = ¯x(t) + dt (x(t) − ¯x(t))/τf , τf = 5 ms, ¯x(0) = 0.
(10)
The learning rate η(t) was computed as η(t) = η0/(1 + t/τl) (η0 = 5 · 10−4, τl = 2 · 104 ms) and
held at η0 in section 3.2 to make conclusions independent of the decay.
5
(cid:32)L−1(cid:88)
L−1(cid:88)
n=0
m=0
(cid:33)
|
1901.00708 | 1 | 1901 | 2019-01-03T13:17:19 | Decoding Hand Kinematics from Local Field Potentials Using Long Short-Term Memory (LSTM) Network | [
"q-bio.NC"
] | Local field potential (LFP) has gained increasing interest as an alternative input signal for brain-machine interfaces (BMIs) due to its informative features, long-term stability, and low frequency content. However, despite these interesting properties, LFP-based BMIs have been reported to yield low decoding performances compared to spike-based BMIs. In this paper, we propose a new decoder based on long short-term memory (LSTM) network which aims to improve the decoding performance of LFP-based BMIs. We compare offline decoding performance of the proposed LSTM decoder to a commonly used Kalman filter (KF) decoder on hand kinematics prediction tasks from multichannel LFPs. We also benchmark the performance of LFP-driven LSTM decoder against KF decoder driven by two types of spike signals: single-unit activity (SUA) and multi-unit activity (MUA). Our results show that LFP-driven LSTM decoder achieves significantly better decoding performance than LFP-, SUA-, and MUA-driven KF decoders. This suggests that LFPs coupled with LSTM decoder could provide high decoding performance, robust, and low power BMIs. | q-bio.NC | q-bio | Decoding Hand Kinematics from Local Field Potentials Using
Long Short-Term Memory (LSTM) Network
Nur Ahmadi1,2, Student Member, IEEE, Timothy G. Constandinou1,2, Senior Member, IEEE, and
Christos-Savvas Bouganis1, Senior Member, IEEE
9
1
0
2
n
a
J
3
]
.
C
N
o
i
b
-
q
[
1
v
8
0
7
0
0
.
1
0
9
1
:
v
i
X
r
a
Abstract -- Local field potential (LFP) has gained increas-
ing interest as an alternative input signal for brain-machine
interfaces (BMIs) due to its informative features, long-term
stability, and low frequency content. However, despite these
interesting properties, LFP-based BMIs have been reported
to yield low decoding performances compared to spike-based
BMIs. In this paper, we propose a new decoder based on long
short-term memory (LSTM) network which aims to improve
the decoding performance of LFP-based BMIs. We compare
offline decoding performance of the proposed LSTM decoder
to a commonly used Kalman filter (KF) decoder on hand
kinematics prediction tasks from multichannel LFPs. We also
benchmark the performance of LFP-driven LSTM decoder
against KF decoder driven by two types of spike signals: single-
unit activity (SUA) and multi-unit activity (MUA). Our results
show that LFP-driven LSTM decoder achieves significantly
better decoding performance than LFP-, SUA-, and MUA-
driven KF decoders. This suggests that LFPs coupled with
LSTM decoder could provide high decoding performance,
robust, and low power BMIs.
I. INTRODUCTION
Brain machine interfaces (BMIs) hold the promise to
restore lost motor function in persons with neurological
disorders (e.g. spinal cord injury) by enabling interaction
with the environment through their neural activity. A key
component of a BMI system is the decoder which translates
neural activity (i.e.
intentions) into motor commands to
control external devices such as computer cursors and robotic
arms, or the subject's own muscles. To date, high decod-
ing performance BMIs have predominantly utilized spikes
as the decoder's input signals which are recorded using
intracortical microelectrode arrays [1]. Despite compelling
results reported in several studies [2], [3], spike-based BMIs
face two major challenges towards their clinically viable
translation [4]. First, the number of recorded spike signals
declines over time, which can degrade the performance of
BMI decoder. This long-term instability is thought to be
caused by scar tissue formation around the electrodes and
micromotion of the electrodes [4], [5]. Second, due to high
*Nur Ahmadi
sampling rate (>10 kHz), detecting and sorting spikes in an
implanted circuitry or transmitting the raw data require high
power consumption, which will in turn increase heat and
limit the device's battery lifetime [4].
One approach to address these challenges is to utilize
another signal modality within neural activity, namely local
field potential (LFP), as an alternative input signal for BMI
decoder. LFP is a low frequency extracellular voltage thought
to be mainly generated from postsynaptic currents and reflect
activity of population of neurons in the vicinity of the
electrodes. LFP is thus believed to be less sensitive to scar
formation and micromotion of the electrodes [4]. Several
studies have shown that LFPs are more stable than spikes and
that considerable amount of movement-related information
contained within LFPs are still present even after spikes are
lost [1], [6], [7]. In addition, LFPs can be processed at signif-
icantly lower sampling rate than spikes, which translates into
lower power consumption and lower complexity (i.e. memory
requirements). However, despite these interesting properties,
LFP-based BMIs have been shown to yield low decoding
performances compared to spike-based BMIs [1], [5], [6],
[8], [9]. Most of LFP-based BMIs employ Kalman filter
(KF) decoder which assumes linear and Gaussian distribution
on both the observation and state dynamics. However, since
LFPs exhibit nonlinear, non-stationary, and non-Gaussian
characteristics [10], aforementioned assumptions could lead
to suboptimal results. Therefore, a more flexible and effec-
tive decoding method is required to improve the decoding
performance of LFP-based BMIs.
One promising solution to this issue is recurrent neural
networks with long short-term memory (LSTM) architecture,
which do not require any assumption on the data. In fact,
LSTMs have become the state-of-the-art methods for various
applications ranging from speech recognition and synthesis,
language modeling and translation, to audio and video anal-
ysis [11]. Despite the successes in a variety of fields, to our
knowledge, however, LSTMs have not been applied to LFP-
based BMI decoding.
is supported by the graduate scholarship awarded by
Indonesia Endowment Fund for Education (LPDP), Republic of Indonesia.
1Nur Ahmadi, Timothy G. Constandinou, and Christos-Savvas Bouganis
are with the Department of Electrical and Electronic Engineering, Imperial
College London, SW7 2BT, UK. Email: {n.ahmadi16,
t.constandinou,
christos-savvas.bouganis}@imperial.ac.uk
In light of this, in the present paper, we propose an LSTM-
based decoder for decoding hand kinematics from multichan-
nel LFPs recorded intracortically from non-human primate
(NHP) during free-reaching tasks. We then compare offline
decoding performance of the proposed LSTM decoder to a
commonly used Kalman filter (KF) decoder. Additionally,
we also benchmark the decoding performance between LFP-
This preprint is the "accepted" version by IEEE NER. c(cid:13)2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained
for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new
collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
2Nur Ahmadi and Timothy G. Constandinou are additionally with the
Centre for Bio-Inspired Technology, Institute of Biomedical Engineering,
Imperial College London, SW7 2AZ, UK
Fig. 1. Block diagram of LFP-based BMI decoding using an LSTM network. The larger gray-shaded block illustrates a detailed schematic of an LSTM.
(HT). We computed Hilbert envelope only from delta band
driven LSTM decoder and KF decoder driven by two types
(0-4 Hz) since other frequency bands in our data did not
of spike signals: single-unit activity (SUA) and multi-unit
contain significant information as observed through time-
activity (MUA).
frequency decomposition. Both LMP and DHE features were
computed within overlapping windows that slide every 4 ms
to match the timescale of hand kinematics. We used window
widths ranging from 32 ms to 512 ms with an increment of
32 ms. For performance evaluation, we selected a window
width and LFP features that yielded the best decoding
performance on validation set. The processing steps of our
LFP-based BMI decoding is illustrated in Fig. 1.
A. Neural Recording and Behavioral Task
II. METHODS
(dataset
Spikes were extracted from each channel by band pass
filtering (from 500 to 5000 Hz) raw neural data and then
detecting the filtered signal amplitudes that crossed a prede-
termined threshold value. All the detected spikes within sin-
gle channel are referred to as multi-unit activity (MUA). To
extract single-unit activity (SUA), the detected spikes were
sorted (i.e. classified) into distinct putative single units. More
detailed information on spike detection and sorting processes
can be found in [13]. In this study, units with spike rates
below 0.5 Hz were excluded [13] thus leaving only 130 (124)
SUAs and 91 (94) MUAs for dataset I (II). We computed
spike counts within different overlapping bin/window widths
(same range as of LFPs) sliding every 4 ms (corresponding
to the kinematics time scale). Overlapping bin was used
because it has been demonstrated to yield better decoding
performance than non-overlap bin [15], [16]. Similar to that
of LFP-based decoding, we selected a bin width that resulted
in the best decoding performance on validation set.
C. Long Short-Term Memory (LSTM) Network
Long short-term memory (LSTM) is a type of recurrent
neural networks (RNNs) that was developed by Hochreiter
and Schmidhuber in 1997 [17]. It has successfully addressed
the vanishing or exploding gradient problem encountered
when training traditional RNNs. LSTM is very well suited
to many sequential data problems as it is capable of learning
long-term temporal dependencies through gating mechanism.
A commonly used variant of LSTM units contains three gates
(forget, input, and output) and a memory cell that control the
In this
study, we used two datasets
I:
I 20170124 01; dataset II: I 20170127 03) from a public
neural data recorded from primary motor cortex (M1) area
of an adult male Rhesus macaque monkey (Macaca mulata)
by Sabes Lab [12]. Recording duration for dataset I and
II were 9 and 12 minutes, respectively. The neural data
was recorded using 96-channel silicon microelectrode array
(Utah array with 400 µm inter-electrode distance and 1 mm
electrode length) while the monkey was performing self-
paced reaching tasks without inter-trial intervals within 8-by-
8 square grid. The recording was referenced to a silver wire
placed under the dura (several cm away from the electrodes).
The neural data was sampled at 24.4 kHz and filtered with
a 4th oder low-pass filter at 7.5 kHz. Fingertip position in
x-y coordinates was sampled at 250 Hz. Velocity was then
computed from position by using discrete derivative. A more
detailed description of the experiment setup is given in [13].
B. Neural Signal Processing
LFPs were obtained by low pass filtering raw neural data
using 4th order Butterworth filter at 300 Hz and then down-
sampling them to 1 kHz. LFPs have been shown to contain
common noises, partly arising from the use of a single,
distal reference (unipolar) [14]. To remove these common
noises, we performed common average reference (CAR) and
subtracted it from LFP signal
in each channel. At each
time instant, CAR was computed by averaging LFP signals
across all channels. Two LFP features, local motor potential
(LMP) [1], [6], [9] and delta Hilbert envelope (DHE), were
extracted. LMPs were computed from LFPs by using time-
domain moving average filter. DHEs were calculated firstly
by low pass filtering LFPs with 4th order Butterworth filter
at 4 Hz and then computing the envelope (i.e. instantaneous
amplitude) of their analytic signals through Hilbert transform
TABLE I
HYPERPARAMETERS FOR LSTM DECODER
Hyperparameter
Number of units
Number of epochs
Batch size
Dropout rate
Learning rate
Values
{50, 75, 100, · · · , 200}
{2, 3, 4, · · · , 8}
{32, 64, 96, 128}
{0, 0.1, 0.2, 0.3, 0.4}
{0.001, 0.0015, 0.002, · · · , 0.003}
flow of information. In this study, we used this LSTM variant
where its components' state at time instant t can be described
by:
ft = σ(Wf xt + Uf ht−1 + bf )
it = σ(Wixt + Uiht−1 + bi)
ct = tanh(Wcxt + Ucht−1 + bc)
ct = ft (cid:12) ct−1 + it (cid:12) ct
ot = σ(Woxt + Uoht−1 + bo)
ht = tanh(ct) (cid:12) ot
(1)
where x, h, f, i, o, c represent the input, output, forget gate,
input gate, output gate, and memory cell, respectively; σ
and tanh denote sigmoid and hyperbolic tangent activation
functions, respectively; (cid:12) denotes element-wise multiplica-
tion. W, U, b are weight matrices and bias vector parameters
that need to be learned during training.
Training an LSTM decoder requires setting up configura-
tion parameters called hyperparameters, which in this study
are given in Table I. The hyperparameter configuration was
tuned through a Bayesian optimization library (Hyperopt)
[18] with 200 iterations. The LSTM decoder was imple-
mented using Keras framework with TensorFlow backend
and trained using RMSprop optimizer and root mean squared
error (RMSE) loss function.
D. Performance Evaluation and Metrics
To evaluate the performance of the LSTM decoder, we
used two common metrics, namely, root mean squared error
(RMSE) and Pearson's correlation coefficient. These metrics
were computed between decoded and actual velocities on
x- and y- coordinates. We split each dataset into k = 10
subdatasets which were then categorized into a training
set (concatenation of k − 2 subdatasets), validation set (1
subdataset), and testing set (1 subdataset). The training set
was used to find values of the LSTM parameters (W, U, b).
The validation set (from dataset I) was used to find the opti-
mal configuration of LSTM hyperparameters, LFP features,
window widths, and lags between LFP/spike features and
the kinematics that minimized average RMSE across x-y
coordinates. This configuration was then used for evaluating
the performance of LSTM decoder on testing sets. We
iterated the performance evaluation on k different testing sets
for each dataset.
III. RESULTS
According to our evaluation on validation set (dataset I),
the following configuration of LSTM hyperparameters led
to smallest average RMSE: number of layer = 1, number of
timesteps = 2, number of units = 100, number of epochs = 6,
Fig. 2.
Comparison of decoding performance of LSTM decoder under
different LFP features and window widths. Red stars indicate window widths
that yielded smallest average RMSEs on validation set of dataset I.
batch size = 32, dropout rate = 0.2, and learning rate = 0.001.
This configuration was then used to find the optimal LFP
features and window width using the validation set (dataset I)
and to benchmark the performance of LSTM decoder against
KF decoder using the testing sets (dataset I and II).
A. LFP Feature and Window Width Selection
The LFP features used for comparison were LMP, DHE,
and combination of both, which were extracted from two
reference schemes (unipolar and CAR). According to our
results, CAR-based LFP features led to superior decoding
performance than unipolar-based LFP features. LMP, on
average, resulted in better performance than DHE. Combi-
nation of both LMP and DHE features yielded improved
performance than that of single LFP feature. In regards
to different window widths ({32, 64, 96,··· , 512} ms), we
observed a similar trend across different LFP features. The
decoding performance increased (smaller average RMSE) as
the window width increased up to a particular point where
above this point the performance would decrease. This par-
ticular point corresponds to a window width within a range of
256− 320 ms. The summary of LSTM decoder performance
with respect to different LFP features and window widths
is depicted in Fig. 2. For performance benchmark against
KF decoder on testing sets, we selected a combination of
LMP and DHE features extracted from CAR-based LFPs
with 320 ms window width. Through the same procedure, we
selected a combination of LMP and DHE features extracted
from unipolar-based LFPs with 256 ms for KF decoder
(figure is not shown due to page limitation).
B. Window Width Selection for Spike-driven Decoders
We also investigated the impact of different window
widths to decoding performance of LSTM and KF decoders
driven by spike signals (SUA and MUA). A similar trend
as in LFP-driven decoders was also observed in the cases
of spike-driven LSTM and KF decoders. The decoding
performance improved as the window width increased and
at a particular point the performance would decline. The
326496128160192224256288320352384416448480512Window Width (ms)37.540.042.545.047.550.052.5Average RMSE (mm/s)UNI[LMP]CAR[LMP]UNI[DHE]CAR[DHE]UNI[LMP+DHE]CAR[LMP+DHE] Fig. 3. Comparison of decoding performance of LSTM decoder driven by
SUA and MUA under different window widths. Red stars indicate window
widths that yielded smallest average RMSEs on validation set of dataset I.
summary of spike-driven LSTM decoder performance with
respect to different window widths is shown in Fig. 3. Due
to page limitation, the summary of spike-driven KF decoder
is not shown. For performance comparison on testing sets,
we selected a window width of 256 ms for both SUA- and
MUA-driven LSTM decoders. Window widths of 256 ms
and 288 ms were selected for SUA- and MUA-driven KF
decoders, respectively.
C. Decoding Performance Comparison
Firstly, we compared the decoding performance of our
proposed LFP-driven LSTM decoder to LFP-, SUA-, and
MUA-driven KF decoder (referred to as LFP-LSTM, LFP-
KF, SUA-KF, and MUA-KF respectively). As shown in
Fig. 4, LFP-LSTM achieved significantly better decoding
performance than LFP-KF, SUA-KF, and MUA-KF mea-
sured in both metrics (RMSE and correlation) for both
datasets (I and II). LFP-LSTM yielded average RMSE of
36.90 ± 2.85 (43.60 ± 2.31) mm/s and average correlation
of 0.83 ± 0.02 (0.82 ± 0.02) for dataset I (II). This average
RMSE value corresponded to performance improvement of
35.59±4.77 (32.93±3.78)% for dataset I (II) with respect to
LFP-KF as shown in Fig. 5. In term of average correlation,
the performance improvement made by LFP-LSTM over
LFP-KF was 17.32 ± 2.85 (14.98 ± 2.65)% for dataset I
(II).
In addition, we also compared the decoding performance
of LFP-LSTM to SUA- and MUA-driven LSTM decoder
(referred to as SUA-LSTM and MUA-LSTM). The results
show that LFP-LSTM outperformed SUA-LSTM but was
outperformed by MUA-LSTM for both datasets. The overall
performance comparison among different decoding meth-
ods is shown in Fig. 4. We plotted performance improve-
ment/decrease of different decoders with respect to LFP-
KF as can be seen in Fig. 5. MUA-LSTM yielded largest
performance improvement of 40.15±5.07 (36.15±4.42)% in
average RMSE and 20.91± 2.40 (17.81± 3.78)% in average
correlation for dataset I (II). SUA-KF showed improved
performance than LFP-KF in dataset I but worse performance
in dataset II. Examples of actual and decoded velocities (in
x and y directions) from LFP-LSTM and LFP-KF are given
Fig. 4. Decoding performance comparison of LSTM and KF decoders
under different input signals. Black horizontal lines inside the boxes repre-
sent the medians; red stars represent the means; colored boxes denote the
interquartile ranges; whiskers extend 1.5× from upper and lower quartiles.
in Fig. 6. For the sake of readability, only decoded velocities
from these two decoders are shown.
IV. DISCUSSION
In this study, we propose a new decoder based on LSTM
network to predict (offline) hand kinematics from multichan-
nel LFPs of an NHP subject. We have shown that LFP-driven
LSTM decoder significantly outperforms KF decoder driven
by LFP, SUA, and MUA. This demonstrates the effectiveness
of LSTM decoder in capturing nonlinear and complex rela-
tionship between LFPs and hand kinematics. On the other
hand, KF decoder with its inherent assumptions results in
poor performance, especially for LFPs that exhibit high spa-
tial correlation [14]. LFPs recorded from unipolar reference
are often contaminated by spatially correlated noises, which
affects the decoding performance. These noises can be elimi-
nated by using common average reference (CAR) employed
prior to LFP feature extraction process. Our results show
that CAR can lead to improved decoding performance. Our
results also show that LMP contains significant amount of
movement-related information as indicated by good decoding
performance compared to other LFP features, which is in
good agreement with [1]. However, our results differs from
[1] in that delta band of LFP is more informative than the
higher bands. By combining LMP with DHE as the input
for LSTM decoder, the performance of LSTM decoder can
be improved. Prior studies demonstrated that LFP-driven
decoder performed slightly worse or comparable to spike-
driven decoder [1], [5]. Using hybrid LFP-spike signals,
they could achieve better decoding performance than that
of spikes only. Here, we show that using only LFP signals,
our proposed LFP-driven LSTM decoder can significantly
outperform spike-driven KF decoder.
Performance comparison using the same LSTM decoder
with different input signals (LFP, SUA, and MUA) show
that MUA-LSTM achieves the best decoding performance,
326496128160192224256288320352384416448480512Window Width (ms)3540455055Average RMSE (mm/s)SUA-LSTMMUA-LSTMDataset I30.040.050.060.0Average RMSE (mm/s)LFP-LSTMLFP-KFSUA-LSTMSUA-KFMUA-LSTMMUA-KFDataset II40.050.060.070.0Average RMSE (mm/s)Dataset I0.700.750.800.85Average CorrelationDataset II0.600.700.80Average Correlation Fig. 5. Performance comparison of different decoders relative to LFP-KF.
A Positive (negative) value indicates performance improvement (decrease).
Black error bars represent 95% confidence intervals.
followed by LFP-LSTM and SUA-LSTM, respectively. This
may indicate that MUA contains more movement-related
information than that of SUA. Better performance of MUA
over SUA for both KF and LSTM decoders obtained from
this study offers a new perspective to the debate whether
or not spikes should be sorted [19], [20]. However,
the
changes of spike amplitude and waveform over time pose
critical challenges in adjusting the threshold crossing value
for spike detection, sorting and tracking the spikes into the
same putative single units. Furthermore, the high sampling
rate required for these spike processing demands high power
consumption that hinders the implementation of wireless,
scalable, and implantable BMIs. LFP offers attractive proper-
ties to address these challenges: long-term stability and low-
frequency content (i.e. low sampling rate). In this study, we
only use two datasets corresponding two recording sessions
(3 days gap between the sessions). For future work, we will
investigate the stability and robustness of LFP-, SUA-, and
MUA-driven LSTM decoders over long period of time.
Overall, our results suggest that LFP indeed contains a rich
movement-related information, which corroborates the idea
that LFP is a promising alternative signal input for BMIs.
Along with their stable and low-frequency properties, LFPs
coupled with LSTM decoder could potentially provide high
decoding performance, robust, and low power BMIs.
ACKNOWLEDGEMENT
We thank J. E. O'Doherty and P. N. Sabes for making
their data publicly available.
REFERENCES
[1] S. D. Stavisky et al., "A high performing brain -- machine interface
driven by low-frequency local field potentials alone and together with
spikes," J. Neural Eng., vol. 12, no. 3, p. 036009, 2015.
[2] M. M. Shanechi, "Brain -- machine interface control algorithms," IEEE
Trans. Neural Syst. Rehabil. Eng, vol. 25, no. 10, pp. 1725 -- 1734,
2017.
[3] D. M. Brandman et al., "Human intracortical recording and neural
decoding for brain-computer interfaces," IEEE Trans. Neural Syst.
Rehabil. Eng, vol. 25, no. 10, pp. 1687 -- 1696, 2017.
Fig. 6. Examples of actual and decoded velocities in x- and y-directions
from dataset II.
[4] A. Jackson and T. M. Hall, "Decoding local field potentials for neural
interfaces," IEEE Trans. Neural Syst. Rehabil. Eng, vol. 25, no. 10,
pp. 1705 -- 1714, 2017.
[5] E. J. Hwang and R. A. Andersen, "The utility of multichannel local
field potentials for brain -- machine interfaces," J. Neural Eng., vol. 10,
no. 4, p. 046005, 2013.
[6] R. D. Flint et al., "Long term, stable brain machine interface per-
formance using local field potentials and multiunit spikes," J. Neural
Eng., vol. 10, no. 5, p. 056005, 2013.
[7] D. Wang et al., "Long-term decoding stability of local field potentials
from silicon arrays in primate motor cortex during a 2D center out
task," J. Neural Eng., vol. 11, no. 3, p. 036009, 2014.
[8] A. K. Bansal et al., "Decoding 3D reach and grasp from hybrid signals
in motor and premotor cortices: spikes, multiunit activity, and local
field potentials," J. Neurophysiol., vol. 107, no. 5, pp. 1337 -- 1355,
2011.
[9] R. D. Flint et al., "Accurate decoding of reaching movements from
field potentials in the absence of spikes," J. Neural Eng., vol. 9, no. 4,
p. 046006, 2012.
[10] D. L. Menzer et al., "Characterization of trial-to-trial fluctuations in
local field potentials recorded in cerebral cortex of awake behaving
macaque," J. Neurosci. Methods, vol. 186, no. 2, pp. 250 -- 261, 2010.
[11] K. Greff et al., "LSTM: A search space odyssey," IEEE Trans. Neural
Syst. Rehabil. Eng, vol. 28, no. 10, pp. 2222 -- 2232, 2017.
[12] J. E. O'doherty et al.,
multichannel
sensorimotor
doi:10.5281/zenodo.583331.
"Nonhuman primate
cortex
electrophysiology,"
reaching with
2017,
[13] J. G. Makin et al., "Superior arm-movement decoding from cortex
with a new, unsupervised-learning algorithm," J. Neural Eng., vol. 15,
no. 2, p. 026010, 2018.
[14] L. Xinyu et al., "Adaptive common average reference for in vivo
multichannel local field potentials," Biomed. Eng. Letters, vol. 7, no. 1,
pp. 7 -- 15, 2017.
[15] S. Koyama et al., "Comparison of brain -- computer interface decoding
algorithms in open-loop and closed-loop control," J. Comput. Neu-
rosci., vol. 29, no. 1-2, pp. 73 -- 87, 2010.
[16] N. Ahmadi et al., "Spike rate estimation using bayesian adaptive kernel
smoother (BAKS) and its application to brain machine interfaces," in
Proc. of the Ann. Int. Conf. of the IEEE Engineering in Medicine and
Biology Society.
IEEE, 2018, pp. 2547 -- 2550.
[17] S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural
Computation, vol. 9, no. 8, pp. 1735 -- 1780, 1997.
[18] J. Bergstra et al., "Hyperopt: a python library for model selection and
hyperparameter optimization," Comput. Sci. Discovery, vol. 8, no. 1,
p. 014008, 2015.
[19] S. Todorova et al., "To sort or not to sort: the impact of spike-sorting
on neural decoding performance," J. Neural Eng., vol. 11, no. 5, p.
056005, 2014.
[20] B. P. Christie et al., "Comparison of spike sorting and thresholding
of voltage waveforms for intracortical brain -- machine interface perfor-
mance," J. Neural Eng., vol. 12, no. 1, p. 016009, 2014.
Dataset I−10010203040Δ Average RMSE (%)LFP-KFLFP-LSTMSUA-LSTMSUA-KFMUA-LSTMMUA-KFDataset II−10010203040Δ Average RMSE (%)Dataset I−1001020Δ Correlation (%)Dataset II−1001020Δ Correlation (%)0246810−2000200x-velocity (mm/s)ActualLFP-KFLFP-LSTM0246810Time (s)−400−2000200y-velocity (mm/s) |
1907.11570 | 1 | 1907 | 2019-07-26T13:31:26 | Decoding the circuitry of consciousness: from local microcircuits to brain-scale networks | [
"q-bio.NC"
] | Identifying the physiological processes underlying the emergence and maintenance of consciousness is one of the most fundamental problems of neuroscience, with implications ranging from fundamental neuroscience to the treatment of patients with disorders of consciousness (DOC). One major challenge is to understand how cortical circuits at drastically different spatial scales, from local networks to brain-scale networks, operate in concert to enable consciousness, and how those processes are impaired in DOC patients. In this review, we attempt to relate available neurophysiological and clinical data with existing theoretical models of consciousness, while linking the micro- and macro-circuit levels. First, we address the relationships between awareness and wakefulness on the one hand, and cortico-cortical, and thalamo-cortical connectivity on the other hand. Second, we discuss the role of three main types of GABAergic interneurons in specific circuits responsible for the dynamical re-organization of functional networks. Third, we explore advances in the functional role of nested oscillations for neural synchronization and communication, emphasizing the importance of the balance between local (high-frequency) and distant (low-frequency) activity for efficient information processing. The clinical implications of these theoretical considerations are presented. We propose that such cellular-scale mechanisms could extend current theories of consciousness. | q-bio.NC | q-bio | Decoding the circuitry of consciousness:
from local microcircuits to brain-scale networks
Julien Modolo1, Mahmoud Hassan1, Fabrice Wendling1,*, and Pascal Benquet1
1: Univ Rennes, INSERM, LTSI -- U1099, F-35000 Rennes, France
*Corresponding author: Fabrice Wendling -- [email protected]
Abstract
Identifying the physiological processes underlying the emergence and maintenance of
consciousness is one of the most fundamental problems of neuroscience, with implications
ranging from fundamental neuroscience to the treatment of patients with disorders of
consciousness (DOC). One major challenge is to understand how cortical circuits at drastically
different spatial scales, from local networks to brain-scale networks, operate in concert to enable
consciousness, and how those processes are impaired in DOC patients. In this review, we attempt
to relate available neurophysiological and clinical data with existing theoretical models of
consciousness, while linking the micro- and macro-circuit levels. First, we address the
relationships between awareness and wakefulness on the one hand, and cortico-cortical, and
thalamo-cortical connectivity on the other hand. Second, we discuss the role of three main types
of GABAergic interneurons in specific circuits responsible for the dynamical re-organization of
functional networks. Third, we explore advances in the functional role of nested oscillations for
neural synchronization and communication, emphasizing the importance of the balance between
local (high-frequency) and distant (low-frequency) activity for efficient information processing.
1
The clinical implications of these theoretical considerations are presented. We propose that such
cellular-scale mechanisms could extend current theories of consciousness.
Keywords
Disorders of consciousness, functional connectivity, micro-circuitry, communication through
coherence, gating by inhibition, electroencephalography.
Introduction
Understanding how consciousness arises from communication among brain regions is a
question of the utmost importance in the field of neuroscience in general, and for the diagnosis
and treatment of patients suffering from disorders of consciousness (DOC) in particular. The
problem of consciousness can be seen as fundamental (e.g.: "What is consciousness? Why do we
have subjective, conscious experiences?", such questions are referred to as the "hard" problem of
consciousness (Harnad, 1998)) or more empirical (e.g.: "What are the processes associated with
the emergence and maintenance of consciousness?", this forms the "soft" problem of
consciousness (Harnad, 1998)). In this review, we aim at understanding 1) how brain networks at
different scales are involved in enabling and maintaining conscious processes of information
transmission and processing related to awareness and wakefulness, and 2) how these
mechanisms are related to the disruptions of consciousness in DOC patients.
Many theories have been proposed to explain how consciousness originates, ranging from
abstract and informational concepts to neurophysiology-based theories. The most widespread
theories of consciousness have a fundamental assumption in common: information processing in
2
the human brain networks is inextricably linked with consciousness. A recent paper by Dehaene
and colleagues summarizes this principle as follows (Dehaene et al., 2017):
"What we call "consciousness" results from specific types of information-processing
computations, physically realized by the hardware of the brain."
The three main theories of consciousness include the Integrated Information Theory (IIT)
(Tononi, 2004), the Dynamic Core Hypothesis (Tononi and Edelman, 1998) (DCH), and the
Global Workspace Theory (Baars, 1988; Dehaene et al., 1998; Dehaene et al., 2003) (GWT).
Historically, DCH theory has been the first to refer to the notion of information processing
involved in consciousness (Tononi and Edelman, 1998). This theory is based on the central role
of functional clusters in the thalamo-cortical system and re-entrant interactions, with high
integration and differentiation of neuronal activity being crucial in the emergence of conscious
phenomena. IIT, which is an evolution and generalization of DCH, is based on a set of axioms
from which postulates are derived. IIT also provides a computable quantity, F, also called
integrated information, that quantifies the level of consciousness. In this framework, if
combining sub-elements increases information processing capability more than linearly adding
these elements, then integrated information increases. Global Workspace Theory (GWT) is a
theory of consciousness theory that is more directly connected with neurophysiology and
neuroanatomy. The main hypothesis of GWT is that conscious information is globally available
within the brain, and that two fundamentally different computational systems co-exist: 1) a
network of distributed "local" processors operating in parallel in the brain ("unconscious"), and
2) a "global" workspace formed by a network of distributed interconnected cortical areas
involved in conscious perception (Baars, 1988). The key concept here is that the global
workspace is composed of distant regions densely connected through glutamatergic cortico-
3
cortical connections as opposed to the network of local processors operating in "isolation" (in
parallel). It is worth noting that this distinction between unconscious and conscious processes has
been recently challenged, and might be an oversimplification (Melnikoff and Bargh, 2018). In
GWT, conscious perception is associated with "ignition", a large-scale brain activation pattern
induced by exposure to a stimulus (Dehaene et al., 2003). If the stimulus does not trigger
ignition, and if the induced brain response remains spatially confined and is brief, then the
perception will not reach consciousness. In other terms, a stimulus has to be sufficiently long and
strong to reach consciousness, which suggests a form of filtering mechanism that is consistent
with the view that only a limited amount of information effectively enters in the global
workspace. Despite these successes in accounting for experimental data in humans regarding
subliminal (unconscious) and conscious perception (Sergent and Dehaene, 2004; King et al.,
2016), one drawback of GWT is that it does not explicitly relate the large-scale recruitment of
brain regions during conscious access with cellular mechanisms. More precisely, what prevents
ignition for short, irrelevant stimuli; and conversely, what enables ignition for strong stimuli?
The neuroanatomical, neurophysiological and dynamical mechanisms behind ignition are of
fundamental importance to understand how we become conscious of a stimulus, or how
alterations of brain networks can result in impaired consciousness in DOC patients.
If one accepts that consciousness is associated with a sufficiently complex (in the algorithmic
sense of "less compressible") information processing, then the emergence of consciousness is
critically dependent on three factors: 1) a physical network enabling interactions between its
components; 2) the flexibility to re-organize transiently sub-networks to achieve greater
computation capabilities by increasing the number of possible configurations and input-output
functions, through functional connectivity; and 3) dynamic communications between its
4
components. These three critical components have the potential to be altered, for example in
lesions following traumatic brain injury. While the physical large-scale network linking brain
regions is well defined and known as the connectome, there are still unresolved questions
regarding the transient organization of clusters performing specific computations (functional
networks), the associated means of communication (neural coding) and how large-scale
functional brain networks and information routing can reconfigure rapidly depending on micro-
circuits regulation.
In this review, the objective is therefore to propose a bridging between cortical microcircuits
on the one hand (cellular scale), and brain-scale activity associated with the two main
dimensions of conscious perception (awareness and wakefulness) on the other hand. Such multi-
scale understanding is a pre-requisite to understand how brain networks become dysfunctional in
DOC, and might contribute to reconcile GWT and DCH into a unified framework.
The review is organized as follows. First, we examine the relationship between the two
dimensions of consciousness, namely wakefulness and awareness, with functional connectivity
between cortical regions and the thalamus. Second, we review the "means of communication"
enabling complex information processing linked with consciousness, which regulate cortico-
cortical communication, among which communication through coherence (CTC) and gating by
inhibition (GBI). The alteration of those mechanisms is presented through results from the
clinical literature. Third, we attempt at linking these findings with concepts that have recently
emerged based on the communication between brain regions based on cross-frequency couplings
between oscillations with specific functional roles. Finally, we suggest possible clinical
implications of this framework in terms of novel neuromodulation protocols in DOC.
5
1. Awareness, wakefulness: a short review of concepts
Conscious perception results from an interplay between two processes interacting with each
other: awareness and wakefulness. Deep sleep switches awareness off, whereas being able of
conscious perception (awareness) of environmental stimuli usually implies a state of
wakefulness, as illustrated in Figure 1. For example, during general anaesthesia, there is both an
absence of conscious perception, awareness and wakefulness. In some peculiar cases, however,
these two components can be unrelated. Unresponsive Wakefulness State (UWS) is an example
of Disorder of Consciousness (DOC) in which wakefulness is present without any detectable
signs of awareness (Laureys and Boly, 2012). Also, during lucid dreaming, there is a form of
awareness in the absence of wakefulness (during sleep) (Voss et al., 2013). Another example is
spatial neglect syndrome, in which patients have no conscious awareness of visual stimuli, while
being awake in the contralateral side of the cortical lesion (Le et al., 2015).
Awareness is supported by attentional, fronto-parietal networks that amplify synaptic
connections within specific cortical pathways (Tallon-Baudry, 2011). This amplification of
relevant stimuli enhances the activated network related to stimulus representation. In parallel, the
concomitant inhibition of irrelevant surrounding networks i) optimizes cortico-cortical routing of
information by constraining the possible propagation of neural activity throughout all possible
cortical "routes", which ii) restricts propagation to a limited number of stimulus-driven
possibilities, and iii) increases the signal-to-noise ratio. Such mechanisms are related to the
concept of functional connectivity, and are detailed further in this review. Importantly, such
mechanisms of active inhibition likely involve cortical inhibition with an active modulation by
thalamocortical inputs (Gabernet et al., 2005), implying that the pattern of thalamo-cortical
activity influences information processing in cortico-cortical networks.
6
Wakefulness depends critically on thalamo-cortical connectivity and neuromodulatory
brainstem inputs to the thalamus (including noradrenaline projections from the locus coeruleus
(Monti, 2011)). For instance, during slow-wave sleep, a low-frequency, synchronized activity
between the cortex and thalamus (the so-called "up-and-down" rhythm (Neske, 2015), prevents
transmission of sub-cortical inputs to the cortex during sleep). This provides an example in
which
thalamo-cortical drastically decreases
information processing by cortico-cortical
networks, resulting in a loss of consciousness. During wakefulness, thalamo-cortical activity is
weakly synchronized (Gent et al., 2018), which is a necessary, but not sufficient condition to
enable consciousness. For example, as aforementioned, wakefulness is present in UWS patients
but cortico-cortical communication is severely impaired (Noirhomme et al., 2010), interfering
with the "awareness" component of consciousness. Another required condition for consciousness
is an efficient large-scale cortico-cortical communication that can support awareness through the
activation of attentional fronto-parietal networks (Luckmann et al., 2014; Ptak et al., 2017).
Therefore, in terms of neuroanatomy, it is possible to link wakefulness with thalamo-cortical,
"vertical" connectivity, while awareness depends on cortico-cortical, "horizontal" connectivity.
This is consistent with the recent view by Naccache (Naccache, 2018) that Minimally Conscious
State (MCS) patients, who are conscious to some degree, exhibit "Cortically Mediated States",
whereas UWS patients do not exhibit such activity, presumably because cortico-cortical
connectivity (critical for awareness) is too severely impaired. More specifically, a "critical mass"
of
information processing requires occurring for
the emergence and maintenance of
consciousness, which is tightly regulated by thalamo-cortical and cortico-cortical connectivity.
Figure 1 presents, in a two-dimensional plane, the continuum of the states of consciousness, as a
function of awareness and wakefulness.
7
Figure 1. Wakefulness and awareness are two essential dimensions of consciousness. In this diagram, several
qualitatively different states of consciousness have been positioned on the 2D-matrix as a function of the associated
axes "content of consciousness" (awareness) and "level of consciousness" (wakefulness). Adapted from (Laureys,
2005).
2. Functional networks, a flexible architecture for conscious processes
The main disadvantage of static network architectures is their limitation in terms of amount
and variety (complexity) of information processing that can take place. The brain takes
advantage of different mechanisms that overcome this limitation, by enabling a dynamic,
transient reconfiguration of brain networks increasing the repertoire of possible responses to
8
inputs (i.e., complexity of input-output functions) (Sporns, 2013). Such transient networks
involving only a few brain regions, coordinated to achieve a specific function limited in duration,
form what is termed functional connectivity. There is a growing interest regarding the functional
networks associated with specific cognitive tasks (Hassan et al., 2015) and novel frameworks
have recently emerged (Avena-Koenigsberger et al., 2017) to explain how brain-scale anatomical
connectivity relates to functional connectivity. Functional networks organize through the
network "means of communication", also termed communication dynamics, that governs
information routing through specific networks, instead of propagating information through the
entire brain network (Avena-Koenigsberger et al., 2017). If that was the case, then information
generated locally would induce distant activity in all connectome-related regions, resulting in a
massively synchronized response with low informational content and complexity.
More specifically, one fundamental question is: what are the mechanisms regulating
communication dynamics and enabling functional networks to emerge in brain-scale networks?
This question is central to understand how the brain optimizes its information processing
capabilities, which are tightly linked with consciousness. We propose that the fundamental
mechanisms underlying communication dynamics are actually cellular-scale mechanisms that 1)
prevent brain-scale neuronal synchronization following a stimulus, and 2) enable the transient
coupling of specific distant brain regions. There has been a considerable amount of interest for
large-scale brain activity patterns linked with consciousness, since those can be measured
through various neuroimaging modalities (e.g., electroencephalography, EEG; functional
magnetic resonance imaging, fMRI). However, mechanisms at the cellular scale have remained
more challenging to address in humans for obvious reasons of invasiveness associated with the
9
required recording techniques in humans. In this section, we review evidence for such
mechanisms that could bridge the micro-circuit and brain-scale levels.
In terms of large-scale neuroanatomical pathways enabling consciousness, long-range
glutamatergic projections between pyramidal neurons through white matter fibers have likely a
critical role (Dehaene and Changeux, 2011) since they enable fast (due to myelin)
communication between distant regions. At the brain-scale level, these white matter fibers are
likely critical to enable conscious access, which involves the transient stabilization of neuronal
activity encoding a specific information pattern, in a network of high-level brain regions
interconnected by long-range connections, with the prefrontal cortex (PFC) acting as a key node
(Dehaene et al., 2006; Berkovitch et al., 2017). On conscious trials, distributed gamma-band
activity reflects a stabilization of local information broadcasted to other areas. Global
broadcasting is thought to make the information accessible to introspection and reportable to
other brain regions (Lamme, 2010). During conscious access to a specific information, other
surrounding global workspace neurons would be inhibited and unavailable for the processing of
other stimuli, therefore remaining preconscious (not reaching consciousness).
At the local scale, a micro-circuit has also been identified as being involved in the
communication between distant brain regions: the projection from pyramidal neurons in a brain
region to VIP-positive (Vasoactive Intestinal Peptide) neurons in another region. By activating
VIP-positive neurons in a distant region, this induces an inhibition of somatostatin-positive
(SST) neurons, which target pyramidal cells dendrites, resulting in a disynaptic disinhibition
(Karnani et al., 2016). Through this disynaptic disinhibition, gamma activity generation can
occur through PV-PV mutual inhibition, and binding between the two involved regions can
10
possibly take place (Munoz et al., 2017), temporarily enabling information transfer and
processing. These cellular-scale mechanisms are summarized in Figure 2.
Figure 2. Basic (schematic) circuits involved in the generation of local and distant oscillations. In a local
cortical network (cortical column), gap-junctional, mutual inhibition of soma-projecting, "fast" GABAergic
interneurons are one of the basic mechanisms of generation for local gamma activity, along with the PING
(Pyramid-InterNeuron Gamma). Conversely, the feedback loop between dendrite-projecting, "slow" GABAergic
interneurons and pyramidal cells can generate low-frequency activity. Importantly, distant communication through
the disynaptic pathway enabling transient generation of gamma oscillations in distant populations. Pyramidal cells in
the source population (left circuit) project on the pyramidal cells of the distant population (right), but also on VIP
interneurons that project on dendrite-projecting SST neurons. Transient activation of VIP neurons from the source
population transiently inhibits SST neurons in the target population, enabling the generation of gamma oscillations
through the PV-PV and PYR-PV circuit. Once the input on distant VIP neurons decreases, SST neurons resume their
inhibitory input, which can terminate gamma oscillations generation.
11
Considering these cellular-scale mechanisms, one crucial question is to understand how they
are involved into the transient emergence and maintenance of functional networks such as those
supporting consciousness. It is established that the state of consciousness is critically dependent
on brain functional cortico-cortical connectivity (Marino et al., 2016; Naro et al., 2018). As a
reminder, functional connectivity refers to statistically significant couplings between temporal
courses of neuronal activity within different regions while anatomical connectivity denotes the
physical connections between brain regions. Functional networks can therefore reflect indirect
connections between brain regions, and are transient depending on which tasks are performed, or
which stimuli are perceived. Even in the absence of any specific task or stimuli, it has been
shown that resting-state networks (e.g., default mode network -DMN-) are also transient
(Kabbara et al., 2017). Experimental evidence supports the idea that functional connectivity can
shed light on the networks involved in various conscious states (Jin and Chung, 2012). For
example, during general anaesthesia-induced loss of consciousness, there is a breakdown in
cortico-cortical functional connectivity (Ferrarelli et al., 2010; Hudetz, 2012; Gomez et al.,
2013), severely impairing the capacity of cortical networks to integrate information and to make
it available at a large scale, as required for conscious perception in IIT or GWT. Similarly, in the
transition from wakefulness to slow-wave sleep, the firing rate in the cortex remains relatively
unchanged during the depolarizing phases of the slow sleep oscillation (Steriade et al., 2001),
while effective brain connectivity is dramatically altered (Tononi and Sporns, 2003; Esser et al.,
2009). Upon falling into NREM sleep, cortical activations also become more local and
stereotypical, impairing effective cortical connectivity (Massimini et al., 2010), as shown using
TMS-evoked EEG responses which remain very close to the stimulation site; while these
responses involve a network of distant brain regions undergoing complex dynamical patterns of
12
successive activation during wakefulness (Guillery and Sherman, 2002; Casali et al., 2013;
Casarotto et al., 2016). These results also emphasize the crucial role of the thalamo-cortical
pathway in cortico-cortical functional connectivity. It is worth noting that the essential role of the
thalamocortical loop as well as so-called "reentrant interactions" were previously considered as
key in the DCH (Tononi and Edelman, 1998).
Consistently with these results obtained during sleep, this breakdown of cortico-cortical
connectivity has also been observed during general anaesthesia and in DOC patients, and even
explored through computational modeling (Esser et al., 2009; Casali et al., 2013). In the GWT
framework,
this explains why consciousness
is
impaired
in such states:
large-scale
communication between distant brain areas is impaired due to thalamo-cortical modulation,
preventing ignition from occurring. In brain-damaged DOC patients, large-scale cortico-cortical
communication can be impaired through the partial destruction of long-range fibers, physically
impeding long-range brain synchrony. In terms of effects at the cellular scale, destruction of
long-range fibers could prevent the synchronization of distant VIP interneurons, which is critical
to induce disynaptic disinhibition and associated gamma activity required for CTC.
Pathological alterations of functional connectivity have been investigated using a variety of
modalities: (1) functional connectivity during ''resting state'' using fMRI or EEG; (2) pulsed
stimulation using transcranial magnetic stimulation (TMS) during EEG recording; and (3) other
perturbation-based approaches investigating brain responses to sensory stimuli (Boly et al.,
2017). The advantage of functional connectivity is that it can be employed to improve the
evaluation and classification of disorders of consciousness (Sanders et al., 2012; Holler et al.,
2014; Rossi Sebastiano et al., 2015; Naro et al., 2018). For example, in mild cognitive
impairment (MCI) patients, it has been shown that impaired consciousness is associated with
13
altered effective connectivity (Varotto et al., 2014; Crone et al., 2015). Failure of large-scale
connectivity, along with a hypersynchrony of local short-range delta and alpha activity were
detected within the DMN and were correlated with the level of awareness in patients with DOC
(Vanhaudenhuyse et al., 2010; Fingelkurts et al., 2013; Maki-Marttunen et al., 2013; Varotto et
al., 2014; Kabbara et al., 2017; Naro et al., 2018). Furthermore, the functional connectivity
pattern of several brain regions, such as the posterior cingulate cortex and precuneus, may even
predict UWS patients' state improvement of consciousness with an accuracy superior to 80%
(Wu et al., 2015).
Beside "passive" investigation of resting-state functional connectivity, the use of TMS-
evoked EEG responses enables the active "probing" of functional connectivity. For example, a
drastic breakdown of functional connectivity has been identified in UWS patients using a
specific TMS protocol triggering, in these patients, a stereotyped, local EEG response similar as
in unconscious sleeping or anaesthetized subjects (Rosanova et al., 2012). Restoring cortical
large-scale effective connectivity with transcranial brain stimulation, such as tACS, in DOCs
could therefore be a useful approach to facilitate partial recovery by enhancing oscillations and
plasticity. One clinical result supporting this idea is the recent demonstration that DLPFC
(dorsolateral prefrontal cortex)-tACS was able to transiently restore the connectivity breakdown
in DOC individuals (Naro et al., 2016).
One fundamental microscopic-scale mechanism involved in information routing in the brain,
and contributing to form functional networks within the anatomical network, is called Gating By
Inhibition (Jensen and Mazaheri, 2010). GBI involves inhibitory processes resulting in the
selective activation of sub-networks and inactivation of other sub-networks. By preventing brain-
scale activation in response to a stimulus, and restricting the number of brain regions engaged in
14
performing tasks, GBI also prevents states of low complexity (e.g., all brain regions displaying
the exact same activity) and therefore inefficient information processing. Thus, GBI processes
suggest that the role of inhibition is more complex than preventing excessive activation of brain
networks, contributing instead to shaping anatomical brain networks into functional networks
(Avena-Koenigsberger et al., 2017). Possible alterations of GBI were reported in studies showing
that EEG alpha power is lower in UWS than in MCS patients (Lehembre et al., 2012; Stefan et
al., 2018), hinting that the neurobiological mechanisms underlying alpha oscillations generation
and associated GBI are profoundly altered in unresponsive patients. Moreover, alpha activity was
highly synchronized and clustered in central and posterior cortical regions in UWS patients
(Lehembre et al., 2012; Stefan et al., 2018), suggesting a possible failure of GBI in the most
severe disorders of consciousness.
3. Information processing in large-scale functional networks through
nested oscillations
One of the most established processes by which distant brain regions engage together in an
activity pattern associated with the performance of a given task is Communication Through
Coherence (CTC) (Fries, 2005, 2009, 2015b; Deco and Kringelbach, 2016; Bonnefond et al.,
2017). CTC involves indeed phase-coupled gamma activity between distant brain regions to
enable information processing. CTC has been suggested to be the substrate of "binding", i.e. the
merging of different features of a stimulus into a single, unified conscious perception (Singer,
2001). More precisely, the excitability fluctuation in a group of neurons provides a specific
signature characterized by a specific frequency band and pattern of discharge (Womelsdorf et al.,
2014), propagating through a large-scale network consisting of anatomically interconnected
15
brain areas and subsequently triggering activity in connected regions. Information processing in
the brain is strongly linked with phase-locked, coordinated-in-time fluctuations of excitability
(Fries, 2005; Fries, 2015a) in networks of distributed neuronal populations. The resulting
oscillations generate a specific neuronal code, and coherence enables the association of
information and communication. Furthermore, CTC involves gamma activity, generated mainly
by GABAergic interneurons (PV-positive basket cells). Taken together, inhibitory processes
appear key for information routing and processing in brain-scale networks involved in
consciousness: GBI shapes brain networks spatially (which brain regions are involved, and
which ones are inhibited), while CTC controls them temporally (information flow). However,
this raises an intriguing question: if gamma rhythms are generated locally by interneuronal -
GABAergic- networks, how can distant brain regions, connected through glutamatergic long-
range fibers, communicate efficiently and achieve CTC? One possibility is that the co-
occurrence of low- and high-frequency neuronal oscillations could provide distant co-activation
(low-frequency, glutamatergic origin) that would then enable binding (high-frequency,
GABAergic origin). This would involve the formation of a functional network of several brain
regions through the low-frequency rhythm, prior to information transfer and processing through
CTC (involving gamma activity).
As a support for this possibility, conscious perception is indeed characterized by an increase
in distributed gamma-band activity (Melloni et al., 2007; Wyart and Tallon-Baudry, 2009).
Interestingly, these fast oscillations are modulated by slow oscillations (Osipova et al., 2008;
Jensen et al., 2014). It has recently been proposed that phase synchronization of low-frequency
oscillations, playing the role of a temporal reference frame for information, carrying high-
16
frequency activity, is a general mechanism of brain communication (Bonnefond et al., 2017).
These nested oscillations might be a key mechanism, not only for cortico-cortical
communication and processing, but also between sub-cortical structures. Emotional memory,
involving both cortical and subcortical structures, indeed engages large network synchronization
through nested theta-gamma oscillations (Bocchio et al., 2017). During in vivo experiments
performed in rodents, a perceived threat (a stimulus announcing a footshock) enhances theta
power and coherence in the amygdala, prefrontal cortex, and hippocampus (Lesting et al., 2011;
Likhtik et al., 2014), while fast gamma bursts are phase-locked to theta oscillations (Stujenske et
al., 2014). Overall, these results support the idea that nested oscillations at theta and gamma
frequencies are a plausible substrate for information channel opening/routing (theta) and
processing/transfer (gamma) within the brain.
Attention is another key element for conscious processing and it is involved in the
synchronization of distant brain regions (Steinmetz et al., 2000; Niebur et al., 2002). The main
underlying brain rhythms involved in attentional processes are alpha and gamma oscillations:
brain regions synchronize gamma oscillations (Womelsdorf et al., 2014), and are modulated by
slow alpha oscillations. Slow oscillations enable inhibition of irrelevant networks, influence local
signal processing, widespread
information exchange, and perception (Sadaghiani and
Kleinschmidt, 2016). Information flow is established by neuronal synchronization at the lower-
frequency bands, namely in the theta (4 -- 7 Hz), alpha (8 -- 13 Hz), and beta (14 -- 25 Hz) bands
(Bonnefond et al., 2017). One possible reason is that low-frequency activity induces a transient
change in excitability in target brain structures, which provides an optimal window for binding
neuronal signals from different regions through high-frequency activity (i.e., gamma) (Canolty et
al., 2006). This provides further support to the idea that low-frequency neural oscillations are
17
mainly involved in establishing transient long-range communication through glutamatergic
projections, while high-frequency neural oscillations are rather involved in information
processing/transfer. It is therefore possible to relate the notion of "integration" with this long-
range, glutamatergic co-activation, enabling brain-scale communication between brain regions;
while "differentiation/segregation" would rather depend on locally generated gamma activity
(and in part on low-frequency activation level, which would result in massively synchronized
activity, and reduced differentiation and complexity). An overview of the aforementioned
mechanisms is proposed in Figure 3.
Figure 3. Schematic overview of transient, selective binding among cortical networks through cellular-scale
mechanisms. A. Schematic diagram of an anatomical network with main projections between regions at the brain
scale. B. Upper panel. Selective binding in subset of cortical regions occurs through the generation of gamma
18
oscillations (mostly through micro-circuits involving basket cells), while the distant disinhibition of specific brain
regions occurs through disynaptic disinhibition (activation of distant VIP neurons inhibits SST neurons, which in
turn decrease their inhibitory projection on pyramidal neurons). This contributes to shape the anatomical network
into a functional network. The alpha rhythm acts as a pulsed inhibition to inhibit "irrelevant" networks, increasing
further the signal-to-noise (SNR) ratio. Lower panel. Decreased integration (e.g., following brain damage) leads to
impaired synchronization of distant brain regions (reflected by decreased low-frequency rhythm on the illustrative
oscillation), and thereby a decrease of binding, which combined to a decrease in GBI efficiency strongly decreases
the overall SNR, leading to dysfunction of the network in terms of integration and binding required for
consciousness. Let us note that, in addition to decreased amplitude of the low-frequency rhythm, the phase
relationship of the nested gamma oscillation could be perturbed (i.e., more random) as compared to the
physiological case. Such potential relationship of nested theta/gamma oscillations remains to be explored in DOC.
Emerging evidence shows that the local versus global information processing balance can be
impaired in neurological disorders. Typically, a recent study investigating functional networks in
Alzheimer's disease patients identified a decrease in brain integration as quantified by the
participation coefficient (reflects communication between distant brain modules), while
segregation as quantified by the clustering coefficient (reflects local communication between
neighbour brain regions) was increased (Kabbara et al., 2018). This is consistent with
neurodegenerative processes, which likely impact the "locking" of specific brain regions or the
inhibition of irrelevant networks, thereby severely impairing large-scale integration of
information. In the context of DOC, recent clinical evidence (Chennu et al., 2017; Rizkallah et
al., 2019) supports this view. In the study by Chennu et al., scalp-level networks were assessed
from DOC patients and pointed at decreased integration within the alpha band. More
specifically, the fronto-parietal network in the alpha band was discriminant between MCS and
UWS patients. In a recent study, Rizkallah et al. (Rizkallah et al., 2019) quantified the level of
local versus global information processing in frequency-dependent functional networks (source
19
level) in DOC patients and controls. Integration in theta band functional networks decreased with
consciousness level, and two anatomical regions were systematically involved between controls
and any patient group: a portion of the left orbitofrontal cortex and the left precuneus. One
possibility is that physical damage to long-range white matter fibers impairs large-scale
integration and the local/global information processing balance. One possible approach to study
such anatomical damage to white matter fibers is diffusion tensor imaging (DTI), as performed
in DOC patients suffering severe brain injury (Fernandez-Espejo et al., 2011; Galanaud et al.,
2012; Luyt et al., 2012), which highlighted widespread disruptions of white matter. Lower
fractional anisotropy was indeed found in the subcortico-cortical and cortico-cortical fiber tracts
of DOC patients as compared to controls (Lant et al., 2016; Weng et al., 2017), suggesting that
major consciousness deficits in DOC patients may be related to altered WM connections between
the basal ganglia, thalamus, and frontal cortex. This is also in line with the effect of lesion of
myelinated fiber tracts, which can result in a failure of communication between distant brain
regions (Adams et al., 2000). Therefore, it seems reasonable that white matter lesions can alter,
modify or prevent both CTC and GBI between large-scale networks. Furthermore, we speculate
that, should the specific phase-locking of gamma oscillations onto theta oscillations be perturbed,
then clinical manifestations associated with DOC might appear (loss of integration and decrease
in the consciousness level).
4. Possible clinical implications
In this review, we have attempted to reconcile the neuroanatomical and neurophysiological
knowledge at the level of micro- and macro-scopic networks, regarding the processes that
underlie the emergence and maintenance of consciousness, and its alterations in DOC patients.
20
The multiplexing of neuronal rhythms through nested oscillations appears as a plausible
mechanism of co-activation in a network of specific distant brain regions (integration), which is
a pre-requisite for a conscious perception. Furthermore, a key mechanism seems to be the subtle
balance between low-frequency activity (associated with "global" processing) and high-
frequency activity (associated with a more "local" processing), which could enable the neuronal
dynamics underlying optimal information routing and processing. Excessive low-frequency
activity (e.g., delta activity) results in massively, synchronized activity resulting in a loss of
complexity in terms of information processing, paralleled in such cases with a loss of
consciousness (e.g., sleep, seizures). Similarly, a lack of fronto-parietal functional coupling
(attentional network) has been recently observed in a recent study, as quantified using high-
resolution EEG in DOC patients (Chennu et al., 2017), suggesting that a sufficient level of
fronto-parietal coupling is required to achieve sufficient neuronal integration and ignition for
conscious perception. More generally, brain dynamics in DOC patients is typically characterized
by a loss of integration at a large-scale (Chennu et al., 2017; Rizkallah et al., 2019), preventing
efficient large-scale coordination of distant brain regions to achieve conscious perception. This
suggests that the low-frequency rhythm required for long-range cortical communication is
decreased, preventing binding in the gamma range and therefore further processing information
and ignition for consciousness access. That being said, what are the possible implications of this
slow/fast activity balance as a candidate mechanism for the complex processing associated with
consciousness?
An interesting perspective, which would also be a form of validation for this mechanism, is
the use of neuromodulation techniques in DOC patients to increase their level of consciousness.
The objective of such neuromodulation techniques could be to "re-balance" local versus global
21
processing, for example through the use of transcranial direct or alternating stimulation
(tDCS/tACS), applied to both a frontal and a parietal site simultaneously, in order to increase
low-frequency synchronization in the theta range. It is the current view that tACS can modulate
endogenous brain rhythms using relatively low-levels electric fields (<1 V/m, as discussed in
(Modolo et al., 2018), which would involve to use a stimulation frequency in the theta range to
increase residual oscillations in this frequency range (i.e., assuming that residual anatomical
connections are still present). Dual-site, fronto-parietal tACS in the theta range could then
provide a non-invasive possibility to increase the level of consciousness in DOC patients,
pending that some residual anatomical connectivity remains in the case of brain-damaged
patients. Interestingly, a recent study (Violante et al., 2017) used dual-site tACS in the theta
range in healthy volunteers reported improved working memory, a function also dependent on
fronto-parietal networks. Another study, using tACS targeting the fronto-temporal network,
reported an increase in working memory performance in seniors to comparable levels than young
participants (Reinhart and Nguyen, 2019). These recent results, obtained in humans, provide
compelling evidence that re-balancing information processing through neuromodulation
protocols could contribute to increase the level of consciousness in some DOC patients (e.g.,
where damage to long-range white matter fibers is not too severe).
Discussion and concluding remarks
The two dimensions of consciousness, awareness and wakefulness, depend on anatomically
distinct pathways: cortico-cortical "horizontal" connectivity and thalamo-cortical "vertical"
connectivity. Excessive "up-and-down-like" thalamo-cortical activity impairs cortico-cortical
connectivity by due to excessive lateral inhibition, thereby preventing CTC of distant brain
22
regions, resulting in drastically altered functional connectivity, in line with the loss of
consciousness in deep sleep or DOC. This control of cortico-cortical communication by thalamo-
cortical activity is fundamental in understanding how attentional processes can emerge by
transiently recruiting efficiently, through nested low-(theta) and high-(gamma) frequency
rhythms, distant brain regions. The evidence reviewed highlights how the balance of nested brain
rhythms with fundamentally different functions can transform an anatomical network into a
transient, successive activation of different sub-networks, i.e. a functional network. This
versatility of reconfiguration of the structural connectome results in an immense and complex
dynamical repertoire of functional networks with specific rhythms and cross-frequency
couplings, probably a key infrastructure enabling consciousness. Among those rhythms, three
appear especially involved in conscious processes: while the function of the theta rhythm appears
linked with "opening" transient channels of communications through distant regions, the alpha
rhythm seems to play the role of pulsed inhibition to increase further the SNR. There is also solid
and converging evidence that gamma oscillations are an excellent candidate for information
processing and transfer.
Importantly, mechanisms identified at the micro-circuit scale between specific types of
interneurons, such as projections from pyramidal cells to distant VIP cells, are critical to provide
a more mechanistic framework for theories of consciousness, notably GWT. Such mechanisms
indeed clarify the conditions under which ignition can occur, while providing links with other
concepts that are not necessarily unified (e.g., CTC, GBI) to enable access to consciousness. In
addition to the recruitment of selective brain regions to have access to the global workspace, it
also appears important to take into account that active inhibitory processes co-occur to improve
the signal-to-noise ratio (e.g., GBI). An overview of those concepts, along with the contribution
23
of the reviewed mechanisms to the increase in neural activity complexity associated with
consciousness, is provided in Figure 4.
Figure 4. Synthesis of the network-level mechanisms underlying the complexity associated with conscious
processes. The structural characteristics of brain circuits at different scales are mentioned with some key oscillatory
rhythms and associated functions.
Taken together, the mechanisms presented in this review suggest that the importance of the
thalamo-cortical pathway, emphasized in the DCH theory, cannot be neglected within the context
of the GW theory: the thalamo-cortical pathway plays actually the role of a "switch", enabling or
24
not efficient integration and communication within cortico-cortical networks through feed-
forward inhibition. Therefore, efficient modulation of the thalamo-cortical pathway is necessary,
but not sufficient, to enable ignition within cortical networks and availability of information at a
large scale. For these reasons, it seems that DCH and GWT are both accurate each from their
perspective, and could be unified to obtain a more integrated vision, through a new framework
accounting for both aspects (ignition, availability of information, and control/routing of
information by thalamo-cortical pathways). In such framework, two balances are critical: the first
one is between vertical (thalamo-cortical) and horizontal (cortico-cortical) connectivity, which
controls the second one between local and distant information processing within cortico-cortical
networks. Consequently, we propose that DCH and GWT could be reconciled through this
balance between horizontal and vertical connectivity, and account for a wider range of
phenomena related to consciousness and its deregulations.
Future directions
- An important step forward would be to investigate further the cross-frequency coupling
between the low-frequency theta rhythm and high-frequency gamma rhythm in healthy controls
as compared to DOC patients during resting state. The identification of such changes could have
implications in terms of diagnostic evaluation, but also regarding novel neuromodulation
protocols that might aim, at least in part, to regulate abnormal cross-frequency couplings.
- Another promising application would be to translate the circuitry presented in this review
into a tractable computational model consisting in a network of brain regions, possibly using the
25
neural mass model approach. Evaluating in silico how the microcircuits are involved into the
generation of nested theta-gamma oscillations, and how TMS-evoked EEG responses at the brain
scale are impacted by synchronized thalamocortical activity, would provide key mechanistic
understanding. Candidate tDCS/tACS protocols could also be tested and evaluated in silico.
- Characterizing further the dynamics of functional networks in DOC patients using EEG, for
example by studying the nature and dynamics of modular states over time (e.g., dwell time) as a
function of the level of consciousness (wake, sleep, DOC such as UWS). Extracting such
dynamical information about functional brain network could have diagnostic implications,
notably to distinguish between MCS and UWS patients.
26
Competing financial interests statement
The authors have no competing financial interests to declare for this work.
Funding
This study is funded by the Future Emerging Technologies (H2020-FETOPEN-2014-2015-
RIA under agreement No. 686764) as part of the European Union's Horizon 2020 research and
training program 2014 -- 2018.
References
Adams JH, Graham DI, Jennett B. The neuropathology of the vegetative state after an acute brain insult. Brain 2000;
123 ( Pt 7): 1327-38.
Avena-Koenigsberger A, Misic B, Sporns O. Communication dynamics in complex brain networks. Nat Rev
Neurosci 2017; 19(1): 17-33.
Baars BJ. A cognitive theory of consciousness. Cambridge England ; New York: Cambridge University Press; 1988.
Berkovitch L, Dehaene S, Gaillard R. Disruption of Conscious Access in Schizophrenia. Trends Cogn Sci 2017;
21(11): 878-92.
Bocchio M, Nabavi S, Capogna M. Synaptic Plasticity, Engrams, and Network Oscillations in Amygdala Circuits
for Storage and Retrieval of Emotional Memories. Neuron 2017; 94(4): 731-43.
Boly M, Massimini M, Tsuchiya N, Postle BR, Koch C, Tononi G. Are the Neural Correlates of Consciousness in
the Front or in the Back of the Cerebral Cortex? Clinical and Neuroimaging Evidence. The Journal of neuroscience :
the official journal of the Society for Neuroscience 2017; 37(40): 9603-13.
Bonnefond M, Kastner S, Jensen O. Communication between Brain Areas Based on Nested Oscillations. eNeuro
2017; 4(2).
Canolty RT, Edwards E, Dalal SS, Soltani M, Nagarajan SS, Kirsch HE, et al. High gamma power is phase-locked
to theta oscillations in human neocortex. Science 2006; 313(5793): 1626-8.
Casali AG, Gosseries O, Rosanova M, Boly M, Sarasso S, Casali KR, et al. A theoretically based index of
consciousness independent of sensory processing and behavior. Sci Transl Med 2013; 5(198): 198ra05.
Casarotto S, Comanducci A, Rosanova M, Sarasso S, Fecchio M, Napolitani M, et al. Stratification of unresponsive
patients by an independently validated index of brain complexity. Annals of neurology 2016; 80(5): 718-29.
Chennu S, Annen J, Wannez S, Thibaut A, Chatelle C, Cassol H, et al. Brain networks predict metabolism,
diagnosis and prognosis at the bedside in disorders of consciousness. Brain 2017; 140(8): 2120-32.
Crone JS, Schurz M, Holler Y, Bergmann J, Monti M, Schmid E, et al. Impaired consciousness is linked to changes
in effective connectivity of the posterior cingulate cortex within the default mode network. Neuroimage 2015; 110:
101-9.
Deco G, Kringelbach M. Metastability and Coherence: Extending the Communication through Coherence
Hypothesis Using a Whole-Brain Computational Perspective. Trends Neurosci 2016; 39(6): 432.
Dehaene S, Changeux JP. Experimental and theoretical approaches to conscious processing. Neuron 2011; 70(2):
200-27.
27
Dehaene S, Changeux JP, Naccache L, Sackur J, Sergent C. Conscious, preconscious, and subliminal processing: a
testable taxonomy. Trends Cogn Sci 2006; 10(5): 204-11.
Dehaene S, Kerszberg M, Changeux JP. A neuronal model of a global workspace in effortful cognitive tasks. Proc
Natl Acad Sci U S A 1998; 95(24): 14529-34.
Dehaene S, Lau H, Kouider S. What is consciousness, and could machines have it? Science 2017; 358(6362): 486-
92.
Dehaene S, Sergent C, Changeux JP. A neuronal network model linking subjective reports and objective
physiological data during conscious perception. Proc Natl Acad Sci U S A 2003; 100(14): 8520-5.
Esser SK, Hill S, Tononi G. Breakdown of effective connectivity during slow wave sleep: investigating the
mechanism underlying a cortical gate using large-scale modeling. Journal of neurophysiology 2009; 102(4): 2096-
111.
Fernandez-Espejo D, Bekinschtein T, Monti MM, Pickard JD, Junque C, Coleman MR, et al. Diffusion weighted
imaging distinguishes the vegetative state from the minimally conscious state. Neuroimage 2011; 54(1): 103-12.
Ferrarelli F, Massimini M, Sarasso S, Casali A, Riedner BA, Angelini G, et al. Breakdown in cortical effective
connectivity during midazolam-induced loss of consciousness. Proc Natl Acad Sci U S A 2010; 107(6): 2681-6.
Fingelkurts AA, Fingelkurts AA, Bagnato S, Boccagni C, Galardi G. Dissociation of vegetative and minimally
conscious patients based on brain operational architectonics: factor of etiology. Clin EEG Neurosci 2013; 44(3):
209-20.
Fries P. A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends Cogn
Sci 2005; 9(10): 474-80.
Fries P. Neuronal gamma-band synchronization as a fundamental process in cortical computation. Annu Rev
Neurosci 2009; 32: 209-24.
Fries P. Rhythms for Cognition: Communication through Coherence. Neuron 2015a; 88(1): 220-35.
Fries P. Rhythms for Cognition: Communication through Coherence. Neuron 2015b; 88(1): 220-35.
Gabernet L, Jadhav SP, Feldman DE, Carandini M, Scanziani M. Somatosensory integration controlled by dynamic
thalamocortical feed-forward inhibition. Neuron 2005; 48(2): 315-27.
Galanaud D, Perlbarg V, Gupta R, Stevens RD, Sanchez P, Tollard E, et al. Assessment of white matter injury and
outcome in severe brain trauma: a prospective multicenter cohort. Anesthesiology 2012; 117(6): 1300-10.
Gent TC, Bandarabadi M, Herrera CG, Adamantidis AR. Thalamic dual control of sleep and wakefulness. Nat
Neurosci 2018.
Gomez F, Phillips C, Soddu A, Boly M, Boveroux P, Vanhaudenhuyse A, et al. Changes in effective connectivity by
propofol sedation. PLoS One 2013; 8(8): e71370.
Guillery RW, Sherman SM. Thalamic relay functions and their role in corticocortical communication:
generalizations from the visual system. Neuron 2002; 33(2): 163-75.
Harnad S. Explaining consciousness: the hard problem. Trends Cogn Sci 1998; 2(6): 234-5.
Hassan M, Benquet P, Biraben A, Berrou C, Dufor O, Wendling F. Dynamic reorganization of functional brain
networks during picture naming. Cortex 2015; 73: 276-88.
Holler Y, Thomschewski A, Bergmann J, Kronbichler M, Crone JS, Schmid EV, et al. Connectivity biomarkers can
differentiate patients with different levels of consciousness. Clin Neurophysiol 2014; 125(8): 1545-55.
Hudetz AG. General anesthesia and human brain connectivity. Brain Connect 2012; 2(6): 291-302.
Jensen O, Gips B, Bergmann TO, Bonnefond M. Temporal coding organized by coupled alpha and gamma
oscillations prioritize visual processing. Trends Neurosci 2014; 37(7): 357-69.
Jensen O, Mazaheri A. Shaping functional architecture by oscillatory alpha activity: gating by inhibition. Front Hum
Neurosci 2010; 4: 186.
Jin SH, Chung CK. Messages from the brain connectivity regarding neural correlates of consciousness. Exp
Neurobiol 2012; 21(3): 113-22.
Kabbara A, Eid H, El Falou W, Khalil M, Wendling F, Hassan M. Reduced integration and improved segregation of
functional brain networks in Alzheimer's disease. J Neural Eng 2018; 15(2): 026023.
Kabbara A, El Falou W, Khalil M, Wendling F, Hassan M. The dynamic functional core network of the human brain
at rest. Sci Rep 2017; 7(1): 2936.
Karnani MM, Jackson J, Ayzenshtat I, Hamzehei Sichani A, Manoocheri K, Kim S, et al. Opening Holes in the
Blanket of Inhibition: Localized Lateral Disinhibition by VIP Interneurons. The Journal of neuroscience : the
official journal of the Society for Neuroscience 2016; 36(12): 3471-80.
King JR, Pescetelli N, Dehaene S. Brain Mechanisms Underlying the Brief Maintenance of Seen and Unseen
Sensory Information. Neuron 2016; 92(5): 1122-34.
Lamme VA. How neuroscience will change our view on consciousness. Cogn Neurosci 2010; 1(3): 204-20.
28
Lant ND, Gonzalez-Lara LE, Owen AM, Fernandez-Espejo D. Relationship between the anterior forebrain
mesocircuit and the default mode network in the structural bases of disorders of consciousness. Neuroimage Clin
2016; 10: 27-35.
Laureys S. The neural correlate of (un)awareness: lessons from the vegetative state. Trends Cogn Sci 2005; 9(12):
556-9.
Laureys S, Boly M. Unresponsive wakefulness syndrome. Arch Ital Biol 2012; 150(2-3): 31-5.
Le A, Stojanoski BB, Khan S, Keough M, Niemeier M. A toggle switch of visual awareness? Cortex 2015; 64: 169-
78.
Lehembre R, Marie-Aurelie B, Vanhaudenhuyse A, Chatelle C, Cologan V, Leclercq Y, et al. Resting-state EEG
study of comatose patients: a connectivity and frequency analysis to find differences between vegetative and
minimally conscious states. Funct Neurol 2012; 27(1): 41-7.
Lesting J, Narayanan RT, Kluge C, Sangha S, Seidenbecher T, Pape HC. Patterns of coupled theta activity in
amygdala-hippocampal-prefrontal cortical circuits during fear extinction. PLoS One 2011; 6(6): e21714.
Likhtik E, Stujenske JM, Topiwala MA, Harris AZ, Gordon JA. Prefrontal entrainment of amygdala activity signals
safety in learned fear and innate anxiety. Nat Neurosci 2014; 17(1): 106-13.
Luckmann HC, Jacobs HI, Sack AT. The cross-functional role of frontoparietal regions in cognition: internal
attention as the overarching mechanism. Prog Neurobiol 2014; 116: 66-86.
Luyt CE, Galanaud D, Perlbarg V, Vanhaudenhuyse A, Stevens RD, Gupta R, et al. Diffusion tensor imaging to
predict long-term outcome after cardiac arrest: a bicentric pilot study. Anesthesiology 2012; 117(6): 1311-21.
Maki-Marttunen V, Diez I, Cortes JM, Chialvo DR, Villarreal M. Disruption of transfer entropy and inter-
hemispheric brain functional connectivity in patients with disorder of consciousness. Front Neuroinform 2013; 7:
24.
Marino S, Bonanno L, Giorgio A. Functional connectivity in disorders of consciousness: methodological aspects and
clinical relevance. Brain Imaging Behav 2016; 10(2): 604-8.
Massimini M, Ferrarelli F, Murphy M, Huber R, Riedner B, Casarotto S, et al. Cortical reactivity and effective
connectivity during REM sleep in humans. Cogn Neurosci 2010; 1(3): 176-83.
Melloni L, Molina C, Pena M, Torres D, Singer W, Rodriguez E. Synchronization of neural activity across cortical
areas correlates with conscious perception. The Journal of neuroscience : the official journal of the Society for
Neuroscience 2007; 27(11): 2858-65.
Melnikoff DE, Bargh JA. The Mythical Number Two. Trends Cogn Sci 2018.
Modolo J, Denoyer Y, Wendling F, Benquet P. Physiological effects of low-magnitude electric fields on brain
activity: advances from in vitro, in vivo and in silico models. Current Opinion in Biomedical Engineering 2018; 8:
38-44.
Modolo J, Thomas AW, Legros A. Neural mass modeling of power-line magnetic fields effects on brain activity.
Front Comput Neurosci 2013; 7: 34.
Monti JM. Serotonin control of sleep-wake behavior. Sleep Med Rev 2011; 15(4): 269-81.
Munoz W, Tremblay R, Levenstein D, Rudy B. Layer-specific modulation of neocortical dendritic inhibition during
active wakefulness. Science 2017; 355(6328): 954-9.
Naccache L. Minimally conscious state or cortically mediated state? Brain 2018; 141(4): 949-60.
Naro A, Bramanti A, Leo A, Cacciola A, Manuli A, Bramanti P, et al. Shedding new light on disorders of
consciousness diagnosis: The dynamic functional connectivity. Cortex 2018; 103: 316-28.
Naro A, Bramanti P, Leo A, Russo M, Calabro RS. Transcranial Alternating Current Stimulation in Patients with
Chronic Disorder of Consciousness: A Possible Way to Cut the Diagnostic Gordian Knot? Brain Topogr 2016;
29(4): 623-44.
Neske GT. The Slow Oscillation in Cortical and Thalamic Networks: Mechanisms and Functions. Front Neural
Circuits 2015; 9: 88.
Niebur E, Hsiao SS, Johnson KO. Synchrony: a neuronal mechanism for attentional selection? Curr Opin Neurobiol
2002; 12(2): 190-4.
Noirhomme Q, Soddu A, Lehembre R, Vanhaudenhuyse A, Boveroux P, Boly M, et al. Brain connectivity in
pathological and pharmacological coma. Front Syst Neurosci 2010; 4: 160.
Osipova D, Hermes D, Jensen O. Gamma power is phase-locked to posterior alpha activity. PLoS One 2008; 3(12):
e3990.
Ptak R, Schnider A, Fellrath J. The Dorsal Frontoparietal Network: A Core System for Emulated Action. Trends
Cogn Sci 2017; 21(8): 589-99.
Reinhart RMG, Nguyen JA. Working memory revived in older adults by synchronizing rhythmic brain circuits. Nat
Neurosci 2019; 22(5): 820-7.
29
Rizkallah J, Annen J, Modolo J, Gosseries O, Benquet P, Mortaheb S, et al. Decreased integration of EEG source-
space networks in disorders of consciousness. Neuroimage Clinical 2019; 23(101841): 1-9.
Rosanova M, Gosseries O, Casarotto S, Boly M, Casali AG, Bruno MA, et al. Recovery of cortical effective
connectivity and recovery of consciousness in vegetative patients. Brain 2012; 135(Pt 4): 1308-20.
Rossi Sebastiano D, Panzica F, Visani E, Rotondi F, Scaioli V, Leonardi M, et al. Significance of multiple
neurophysiological measures in patients with chronic disorders of consciousness. Clin Neurophysiol 2015; 126(3):
558-64.
Sadaghiani S, Kleinschmidt A. Brain Networks and alpha-Oscillations: Structural and Functional Foundations of
Cognitive Control. Trends Cogn Sci 2016; 20(11): 805-17.
Sanders RD, Tononi G, Laureys S, Sleigh JW. Unresponsiveness not equal unconsciousness. Anesthesiology 2012;
116(4): 946-59.
Sergent C, Dehaene S. Is consciousness a gradual phenomenon? Evidence for an all-or-none bifurcation during the
attentional blink. Psychol Sci 2004; 15(11): 720-8.
Singer W. Consciousness and the binding problem. Ann N Y Acad Sci 2001; 929: 123-46.
Sporns O. Structure and function of complex brain networks. Dialogues Clin Neurosci 2013; 15(3): 247-62.
Stefan S, Schorr B, Lopez-Rolon A, Kolassa IT, Shock JP, Rosenfelder M, et al. Consciousness Indexing and
Outcome Prediction with Resting-State EEG in Severe Disorders of Consciousness. Brain Topogr 2018.
Steinmetz PN, Roy A, Fitzgerald PJ, Hsiao SS, Johnson KO, Niebur E. Attention modulates synchronized neuronal
firing in primate somatosensory cortex. Nature 2000; 404(6774): 187-90.
Steriade M, Timofeev I, Grenier F. Natural waking and sleep states: a view from inside neocortical neurons. Journal
of neurophysiology 2001; 85(5): 1969-85.
Stujenske JM, Likhtik E, Topiwala MA, Gordon JA. Fear and safety engage competing patterns of theta-gamma
coupling in the basolateral amygdala. Neuron 2014; 83(4): 919-33.
Tallon-Baudry C. On the neural mechanisms subserving consciousness and attention. Front Psychol 2011; 2: 397.
Tononi G. An information integration theory of consciousness. BMC Neurosci 2004; 5: 42.
Tononi G, Edelman GM. Consciousness and complexity. Science 1998; 282(5395): 1846-51.
Tononi G, Sporns O. Measuring information integration. BMC Neurosci 2003; 4: 31.
Vanhaudenhuyse A, Noirhomme Q, Tshibanda LJ, Bruno MA, Boveroux P, Schnakers C, et al. Default network
connectivity reflects the level of consciousness in non-communicative brain-damaged patients. Brain 2010; 133(Pt
1): 161-71.
Varotto G, Fazio P, Rossi Sebastiano D, Duran D, D'Incerti L, Parati E, et al. Altered resting state effective
connectivity in long-standing vegetative state patients: an EEG study. Clin Neurophysiol 2014; 125(1): 63-8.
Violante IR, Li LM, Carmichael DW, Lorenz R, Leech R, Hampshire A, et al. Externally induced frontoparietal
synchronization modulates network dynamics and enhances working memory performance. Elife 2017; 6.
Voss U, Schermelleh-Engel K, Windt J, Frenzel C, Hobson A. Measuring consciousness in dreams: the lucidity and
consciousness in dreams scale. Conscious Cogn 2013; 22(1): 8-21.
Wendling F, Bartolomei F, Bellanger JJ, Chauvel P. Epileptic fast activity can be explained by a model of impaired
GABAergic dendritic inhibition. The European journal of neuroscience 2002; 15(9): 1499-508.
Weng L, Xie Q, Zhao L, Zhang R, Ma Q, Wang J, et al. Abnormal structural connectivity between the basal ganglia,
thalamus, and frontal cortex in patients with disorders of consciousness. Cortex 2017; 90: 71-87.
Womelsdorf T, Valiante TA, Sahin NT, Miller KJ, Tiesinga P. Dynamic circuit motifs underlying rhythmic gain
control, gating and integration. Nat Neurosci 2014; 17(8): 1031-9.
Wu X, Zou Q, Hu J, Tang W, Mao Y, Gao L, et al. Intrinsic Functional Connectivity Patterns Predict Consciousness
Level and Recovery Outcome in Acquired Brain Injury. The Journal of neuroscience : the official journal of the
Society for Neuroscience 2015; 35(37): 12932-46.
Wyart V, Tallon-Baudry C. How ongoing fluctuations in human visual cortex predict perceptual awareness: baseline
shift versus decision bias. The Journal of neuroscience : the official journal of the Society for Neuroscience 2009;
29(27): 8715-25.
30
|
1805.08239 | 2 | 1805 | 2018-11-26T15:36:21 | The Roles of Supervised Machine Learning in Systems Neuroscience | [
"q-bio.NC",
"cs.LG",
"stat.ML"
] | Over the last several years, the use of machine learning (ML) in neuroscience has been rapidly increasing. Here, we review ML's contributions, both realized and potential, across several areas of systems neuroscience. We describe four primary roles of ML within neuroscience: 1) creating solutions to engineering problems, 2) identifying predictive variables, 3) setting benchmarks for simple models of the brain, and 4) serving itself as a model for the brain. The breadth and ease of its applicability suggests that machine learning should be in the toolbox of most systems neuroscientists. | q-bio.NC | q-bio | The Roles of Supervised Machine Learning
in Systems Neuroscience
Joshua I. Glaser1*, Ari S. Benjamin1*, Roozbeh Farhoodi1*, Konrad P. Kording1,2
1. Department of Bioengineering, University of Pennsylvania
2. Department of Neuroscience, University of Pennsylvania, Member, Canadian Institute For Advanced
Research
*These authors contributed equally to this work.
Contacts: jglaser8, aarrii, roozbeh, kording @seas.upenn.edu
Abstract:
Over the last several years, the use of machine learning (ML) in neuroscience has been rapidly
increasing. Here, we review ML's contributions, both realized and potential, across several
areas of systems neuroscience. We describe four primary roles of ML within neuroscience: 1)
creating solutions to engineering problems, 2)
identifying predictive variables, 3) setting
benchmarks for simple models of the brain, and 4) serving itself as a model for the brain. The
breadth and ease of its applicability suggests that machine learning should be in the toolbox of
most systems neuroscientists.
Introduction:
There is a lot of enthusiasm about machine learning (ML). After all, it has allowed computers to
surpass human-level performance at image classification (He et al. 2015), to beat humans in
complex games such as "Go" (Silver et al. 2016), and to provide high-quality speech to text
(Hannun et al. 2014) in popular mobile phones. Progress in ML is also getting attention in the
scientific community. Writing in the July 2017 issue of Science focusing on "AI Transforms
Science", editor Tim Appenzeller writes, "For scientists, prospects are mostly bright: AI promises
to supercharge the process of discovery" (Appenzeller 2017).
The field of systems neuroscience is no exception. In the last few years there have been many
opinion pieces about the importance of ML in neuroscience (Vu et al. 2018; Barak 2017;
Paninski and Cunningham 2017; Vogt 2018; Hinton 2011). Moreover, when we analyze the
number of journal articles about ML in neuroscience, we find that its use has been continuously
growing over the last 20 years (Fig. 1). Machine learning has been used in many different ways
within this literature. In this review, we will catalog the conceptual applications of ML in systems
neuroscience.
Figure 1: Growth of Machine Learning in Neuroscience.
Here we plot the proportion of neuroscience papers that have used ML over the last two decades. That
is, we calculate the number of papers involving both neuroscience and machine learning, normalized
by the total number of neuroscience papers. Neuroscience papers were identified using a search for
"neuroscience" on Semantic Scholar. Papers involving neuroscience and machine learning were
identified with a search for "machine learning" and "neuroscience" on Semantic Scholar.
On the highest level, ML is typically divided into the subtypes of supervised, unsupervised, and
reinforcement learning. Supervised learning builds a model that predicts outputs from input
data. Unsupervised learning is concerned with finding structure in data, e.g. clustering,
dimensionality reduction, and compression. Reinforcement learning allows a system to learn the
best actions based on the reward that occurs at an end of a sequence of actions. This review
focuses on supervised learning.
Why is creating progressively more accurate regression or classification methods (see Box 1)
worthy of a title like 'The AI Revolution' (Appenzeller 2017)? It is because countless questions
can be framed in this manner. When classifying images, an input picture can be used to predict
the object in the picture. When playing a game, the setup of the board (input) can be used to
predict an optimal move (output). When texting on our smartphones, our current text is used to
create suggestions of the next word. Similarly, science has many instances where we desire to
make predictions from measured data.
In this review, we categorize the ways in which supervised ML promises to assist, or has
already been applied to, problems in systems neuroscience. We believe that applications of
supervised ML in this field can be divided in roughly four categories (Fig. 2). 1) Solving
engineering problems. Machine learning can improve the predictive performance of methods
used by neuroscientists, such as medical diagnostics, brain-computer interfaces, and research
tools. 2) Identifying predictive variables. Machine learning can more accurately determine
whether variables (e.g., those related to the brain and outside world) predict each other. 3)
Benchmarking simple models. We can compare the performance of simple interpretable models
to highly accurate ML models in order to help determine the quality of the simple models. 4)
Serving as a model for the brain. We can argue whether the brain solves problems in a similar
way to ML systems, e.g. deep neural networks. The logic behind each of these applications is
rather distinct.
For the bulk of this review, we will go into further detail about these four roles of ML in
neuroscience. We will provide many examples, both realized and potential, of ML across
several areas of systems neuroscience. More specifically, we will discuss the four roles of ML in
relation to neural function, including neural activity and how it relates to behavior; and neural
structure, i.e., neuroanatomy. We also discuss ML in practice (Box 2), as that is crucial for
useful applications.
Figure 2: Examples of the four roles of supervised machine learning in neuroscience.
1 - ML can solve engineering problems. For example, it can help researchers control a prosthetic limb
using brain activity. 2 - ML can identify predictive variables. For example, by using MRI data, we can
identify which brain regions are most predictive for diagnosing Alzheimer's disease (Lebedev et al.
2014). 3 -
ML can benchmark simple models. For example, we can compare the predictive
performance of
the simple "population vector" model of how neural activity relates to movement
(Georgopoulos, Schwartz, and Kettner 1986) to a ML benchmark (e.g. an RNN). 4 - ML can serve as a
the brain. For example, researchers have studied how neurons in the visual pathway
model of
correspond to units in an artificial network that is trained to classify images (Yamins and DiCarlo 2016).
Role 1: Solving engineering problems
A surprisingly wide array of engineering problems can be cast as prediction problems. Their
common thread is that one would like to estimate some quantity of interest (Y) and can take
measurements (X) that relate to that quantity. The relationship between X and Y, however, is
unknown and might be complicated. We call these 'engineering problems' when the final
quantity, Y, is all that is desired. In these problems, one does not need detailed understanding
of the relationship - the aim is simply to estimate Y as accurately as possible. For example,
email providers wish to filter spam from incoming messages, and only care about how
accurately the emails are sorted.
Traditionally, one would attempt to carefully understand the relationship between X and Y, and
synthesize this into a model. Modern machine learning (ML) is changing this paradigm. Instead
of detailed expert knowledge of a process, a practitioner simply needs a large database of
measurements along with the associated quantity of interest for each measurement. Machine
learning algorithms can then automatically model
their relationship. Once trained, an ML
algorithm can make predictions for new measurements.
This 'engineering' framework is the traditional application of ML and is common in industry. In
neuroscience, there are many problems that resemble this general problem format.
Neural Activity / Function
Many medical applications depend on successfully extracting information about
intention,
sensation, or disease from measurements of neural activity. This is a difficult problem, since the
meaning of neural activity, the 'neural code', is most often not known. Machine learning is now a
ubiquitous tool for this task in situations in which it is possible to obtain large datasets of neural
activity along with the behaviors or diseases of interest. One such application is brain-computer
interfaces (BCIs), which seek to use neural signals to control prosthetic limbs, computer cursors,
or other external objects. Several groups have used modern ML techniques, such as recurrent
neural networks, to improve BCIs using recordings of spikes (Sussillo et al. 2016, 2012), ECoG
(Elango et al. 2017), or EEG (Bashivan et al. 2015). Machine learning can also be used to
predict future neural activity from past neural activity. This application is relevant for epilepsy,
since imminent seizures can be predicted using deep learning (Talathi 2017; Nigam and Graupe
2004) and ensemble methods (Brinkmann et al. 2016).
line of applications,
researchers have used ML to diagnose neurological conditions from activity. Several reviews on
this specific application have recently been published (see (Arbabshirani et al. 2017) for
classification using neuroimaging data, (Rathore et al. 2017) for classification of Alzheimer's
disease, and (Vieira, Pinaya, and Mechelli 2017) for a focus on deep learning approaches). Due
to the large datasets available for neural recordings, ML has improved the standards of
accuracy on these medical applications despite the complexity of the input signals.
In another
An ML approach also promises to assist with the inverse of the above problem: predicting
neural activity from variables in the outside world. Solving this problem is important if we want to
use neural stimulation to induce accurate sensation. A prosthetic eye, for example, could be
built by stimulating retinal ganglion cells in the correct manner according to the output of a
camera (Nirenberg and Pandarinath 2012). The most accurate model of ganglion cell activity is
currently a deep learning model trained to predict activity from naturalistic scenes (McIntosh et
al. 2016). Similarly, prosthetic limbs could provide tactile and proprioceptive sensations if
somatosensory neurons were correctly stimulated (Armenta Salas et al. 2018). Machine
learning models may help to enable these neural prosthetics to induce sensations.
A very similar approach can be used to quantify behavior (D. J. Anderson and Perona 2014),
such as movement, sleeping, and socializing. For example, we may want to quantify movement
of the whole body using cheap video recordings. Recent progress in the field has made video
quantification far more precise. Researchers have used deep learning to estimate human poses
from video (Insafutdinov et al. 2016). Related approaches have recently gotten easier to use
and less data intensive, and have been extended to animal tracking (Mathis et al. 2018; T.
Pereira et al. 2018). Along with estimating poses, we can also directly estimate types of
behavior (e.g., walking vs. stopping vs. jumping) from video (Kabra et al. 2013). Behavior can
also be estimated based on other modalities such as audio recordings (Arthur et al. 2013). This
progress in ML-based behavior tracking is timely, given recent calls to understand neural control
of more naturalistic behavior in more natural environments (Krakauer et al. 2017).
The engineering approach of applying ML is also helping solve the problem of obtaining
accurate estimates of neural activity from raw measurements. Many imaging methods, such as
EEG, MEG, and fMRI, require the solving of an 'inverse problem' -- obtaining the source from
the measurements. For example, researchers estimate the origin of EEG signals within the brain
based on electrode recordings from the scalp. Recently, it has been observed that deep
learning can improve the estimates for imaging (McCann, Jin, and Unser 2017). Neural
networks have improved image denoising (Burger, Schuler, and Harmeling 2012; Xie, Xu, and
Chen 2012) and deconvolution (L. Xu et al. 2014), can provide super-resolution images (Dong
et al. 2016), and can even replace the entire image processing pipeline (Golkov et al. 2016; Vito
et al. 2005). Outside of imaging, the deconvolution of time series data is another common
application. For example, once a researcher has obtained traces of cellular calcium
concentration, there is still the difficult 'inverse problem' of inferring the timing of the underlying
spiking. Competition-style ML on labeled datasets (Theis et al. 2016) provides good solutions
(Berens et al. 2017). In each of these applications, a difficult engineering problem was replaced
by building a large labeled dataset and using ML to learn the desired relationship.
Neuroanatomy / Structure
Just as neural activity can be indicative of disease, so can neuroanatomy. As such, it is often
possible to take anatomical measurements and use machine learning to diagnose disease. For
example, researchers can distinguish between Alzheimer's disease and healthy brains of older
adults using MRI scans (Sarraf, Tofighi, and Others 2016). More generally, neuroanatomical
measurements such as structural MRI and diffusion tensor imaging (DTI) can distinguish healthy
from unhealthy patients across many conditions including schizophrenia, depression, autism
and ADHD (Arbabshirani et al. 2017; Rathore et al. 2017; Vieira, Pinaya, and Mechelli 2017).
Sometimes, ML enables surprising opportunities. For example, using deep convolutional neural
networks, we can surprisingly predict cardiovascular risk factors from retinal fundus photographs
(Poplin et al. 2018). The future will undoubtedly see ongoing efforts to automatically detect
disease from biological data.
Since the majority of research in neuroanatomy is based on imaging techniques, recent
advances in computer vision using ML are becoming important tools for neuroanatomy. Thus,
segmenting and labeling parts of an image, which usually requires manual annotation, is an
especially important task. However, as imaging techniques improve and the volume of data
increases, it will become infeasible to rely on manual annotation. To solve this problem, many
ML techniques have been developed to automatically segment or label new images based on a
techniques in this developing
dataset of previously labeled images. The vast majority of
literature are based, at least in part, on convolutional neural networks (Litjens et al. 2017;
Ronneberger, Fischer, and Brox 2015). This approach has been employed to label medical
images, such as in identifying white matter tracts from MRI scans (Ghafoorian et al. 2017). They
have also been used to understand the connections and morphologies of neurons from electron
microscopy (Helmstaedter et al. 2013; Funke et al. 2017; Turaga et al. 2010). As imaging data
improve in resolution and volume, ML is becoming a crucial and even necessary tool for
reconstructing and mapping neuroanatomy.
Caveats
While ML methods have been able to provide many engineering solutions, ML is not magic
(Wolpert and Macready 1997). Several conditions must be met for an ML method to solve a
problem successfully.
A first consideration is that the selected method must match the structure of the data. For
example, convolutional neural networks make the assumption that images have common local
features (like edges), which allows them to be more successful than standard feedforward
neural networks. In practice, this means that a user of ML must take care to select their method,
or alternatively to preprocess their data, such that the assumptions match how the inputs relate
to outputs. This requires good knowledge of the specifics of the data and also of ML methods. A
workaround to this is to apply automated ML methods, which iterate intelligently through many
possible model configurations and select the best-performing option (Guyon et al. 2015; Elsken,
Metzen, and Hutter 2018; Feurer et al. 2015; Kotthoff et al. 2017). This approach is slow but
often works comparatively well. The general rule, however, is that good ML engineering requires
a good awareness of the data.
is not
Another potential failure mode of ML is "overfitting" to the training data. Ideally, an ML method
trained on. If a method learns to make accurate
should accurately predict data it
predictions on the training data but cannot generalize to new data, it is said to be overfit. To
guard against the worry of overfitting, ML practitioners always report a model's performance on
a test set that the model has not been trained on. The capacity of a method to overfit to data
can be lowered with regularization methods, which penalize the complexity of a model. Still,
overfitting is especially worrisome for small datasets and complex models. While different
methods have different sensitivities to the number of data points, all methods become less
vulnerable to overfitting when datasets are large. Sometimes simpler methods may be better
choices on small datasets, even if a more complex method could better express the underlying
input/output relationship (if there was sufficient data). The risk of overfitting means that all ML
practitioners must be aware of regularization techniques, their dataset size, and the importance
of reporting accuracy on a test set not used for training.
Yet another practical drawback of ML is that it can be slow. For large datasets and complex
models, the time it takes to train the model can be prohibitive without proper hardware, like
GPUs for deep learning. Once a model
is much faster to make
predictions. Still, for applications that require real-time predictions, even the prediction step may
be too slow for some ML methods. For example, predictions for brain machine interfaces often
need to be made in the timescale of tens of milliseconds, which can be a challenge for models
requiring many computations. This tradeoff between complexity and run-time is an important
aspect in choosing a model for many engineers.
is trained, however,
it
Role 2: Identifying predictive variables
Neuroscientists often ask questions of the form, "which variables are related to something of
interest?" For example, which brain regions can predict each other? Which brain regions contain
information related to a subject's decision? Which cell types are affected by a certain disease?
Machine learning (ML) can help to more accurately identify how informative one set of variables
is about another. This is particularly instructive when there is a complex nonlinear relationship
between the variables, which is often the case in neural systems. Answering these types of
questions allows researchers to better understand the relationship between parts of the brain,
stimuli, behavior, and more.
The general strategy resembles that of the engineering applications (Role 1). However, instead
of only searching for maximal predictive accuracy, one examines which input variables lead to
that accuracy. There are many methods to establish so-called 'feature importance' (also known
as 'feature selection'). Two of the simplest are the leave-one-out strategy, in which each
variable is removed and one observes the decrease in accuracy, and the 'best first' strategy, in
which the algorithm is run on each variable alone. Leave-one-out reflects the information in that
variable but not the others, while best-first reflects the total (learnable) task information in each
variable. Both are related but different definitions of what it means for a feature to be important.
The development of feature importance measures is an active field in ML and statistics (Tang,
Alelyani, and Liu 2014). These methods allow us to get insights into which variables are
important for a given problem (with the specific meaning of 'importance' depending on the
measure used).
A more traditional approach for this type of question would be to fit simple models to data, like
linear regression, and to examine the coefficients of the fit. This approach is ubiquitous in
science. Its basic drawback, however, is that one needs to assume a model, which may be
inaccurate. For example, if the model is assumed to be y = mx+b, but the true relationship is y =
cos x, then the value of m (the "interaction between x and y") will be 0 despite there being a
strong relationship between x and y. The ML approach, on the other hand, seeks to maximize
predictive accuracy and in doing so does not need to assume a simple functional form. This has
the advantage that we can evaluate a variable's importance even when the relationship
between inputs and outputs is unknown and possibly nonlinear. Plus, by bootstrapping, we can
even find the confidence interval for their importance values. Machine learning combined with
feature selection approaches can be universally applied to problems regardless of whether we
know the underlying relationship.
Determining the important features can also help us to construct simpler models. Rather than
using many inputs for a model, we can only use the important features as inputs. For example,
determining which morphological features of neurons are most predictive of cell type can lead
us to build more accurate generative models of morphologies (that are based on the most
predictive features). Accurately determining the importance of features within ML algorithms is
thus also beneficial for creating simpler models.
This approach is designed to examine the same variables that serve as raw inputs into the ML
model. Often the "features" we seek are different than the raw inputs. In vision, for example, the
raw input variables may be as simple as single pixels, and the approach of Role 2 would then
simply isolate the most relevant pixels. However, as we review below, many problems in
neuroscience follow the format where the input variables are the variables of interest.
Neural Activity / Function
Neuroscience has a long history of building encoding models, which aim to predict neural
activity (e.g., spikes in an individual neuron, or BOLD signal
in an fMRI voxel) based on
variables in the outside world. This is a common approach to identifying the "role" of a brain
area. The building of encoding models is a regression problem (from external variables to
activity), and its purpose is more akin to feature importance than purely predictive power. This
problem is an open invitation to use ML methods in combination with methods of determining
feature importance.
Machine learning would not be necessary for encoding models if simpler methods were as
accurate at describing neural activity. However, this is usually not the case. For example, we
recently showed that XGBoost and ensemble methods led to significant performance
improvements on datasets from motor cortex, somatosensory cortex, and hippocampus
(Benjamin et al. 2018). These improvements were relative to Generalized Linear Models, which
are ubiquitous in computational neuroscience. Others have also shown predictive improvements
in other areas and modalities using methods such as XGBoost (Viejo, Cortier, and Peyrache
2018) and deep learning (McIntosh et al. 2016; Klindt et al. 2017; Agrawal et al. 2014). These
instances serve as warnings that although simple models may appear interpretable, they may
be missing important aspects of how external variables relate to neural function.
Having improved encoding performance can more generally allow researchers to understand
which covariates are predictive of neural activity. This generalizes the already common
approach of adding additional variables to simple models and observing the increase in
performance (e.g., (Stringer et al. 2018)). For example, research on building encoding models of
(Viejo, Cortier, and Peyrache 2018) looked at the
head-direction neurons using XGBoost
relative contribution of the different covariates (such as the direction of the head) within the
encoding model. This allowed determining how the covariates mattered, without assuming the
form of the relationship.
The reverse problem, "what information can be read-out from activity from this brain area" can
also answer questions about information content and the role of specific brain areas or cell
types. For example, researchers used decoding methods to compare how predictive neural
populations in parietal and prefrontal cortices are about task-relevant variables, over the course
of a decision-making task (Sarma et al. 2016). As another example, we have compared
decoding results from motor cortex from different task conditions to determine how uncertainty
in the brain relates to different behavioral uncertainty (Dekleva et al. 2016; Glaser et al. 2018).
The choice of decoding method has a large impact on performance. We have recently done a
thorough test of different ML methods on datasets from motor cortex, somatosensory cortex,
and hippocampus, and have shown that modern ML methods, such as neural networks and
ensemble methods lead to increased decoding accuracy (Glaser et al. 2017). More accurate
decoding can increase our understanding of how much information is contained in a neural
population about another variable, such as a decision, movement, or location.
Neuroscience researchers often want to determine which variables matter for behavior, so that
they can relate these variables to neural activity. We can apply ML to find what variables are
predictive of behavior, without assuming the form of the relationship. For example, researchers
have aimed to determine which visual features predict where we look next. This is a useful step
in determining the neural basis of gaze control
(Ramkumar et al. 2016). Traditionally,
hand-designed visual features have been used to predict where we look next (Itti and Koch
2001), but recently researchers have more accurately predicted fixation locations using deep
learning (Kümmerer, Theis, and Bethge 2014). As another example, researchers have studied
how features in the environment predict
the songs produced by male Drosophila during
courtship (Coen et al. 2016). Using a generalized linear model with a sparsity prior, this study
found that the distance between the male and female was the strongest predictor. This allowed
the researchers to then investigate the neural pathway that was responsible for distance
instance,
modulating song amplitude. More accurate behavioral models can allow researchers to better
investigate the relationship between neural activity and behavior.
In medicine, it is important to understand the underlying factors that are predictive of disease.
This has be done by finding the importance of neuroimaging features in traditional classification
techniques (e.g., determining which functional connectivity measures are predictive of
Alzheimer's disease in a logistic regression classifier (Challis et al. 2015)). More recently,
studies have used a variety of methods to look inside deep learning classifiers to determine the
important features (e.g., determining fMRI connectivity relationships that are predictive of ADHD
(Deshpande et al. 2015) and schizophrenia (Kim et al. 2016)). This approach has also had
success in animal models. For
in a mouse model of depression, researchers
determined which features (power and coherence) of neural activity in prefrontal cortex and
limbic areas were predictive of pathological behavior (Hultman et al. 2016). They were then able
to use this information to design a neural stimulation paradigm to restore normal behavior. By
using machine learning in each of these applications, the researchers were able to test if their
variables predicted disease without having to assume the form of that relationship.
Neuroanatomy / Structure
Just as with neural activity, machine learning can help researchers better understand how
neuroanatomical features across the brain are predictive of disease. A general approach is to
construct an ML classifier to determine whether a subject has the disease, and then look at the
importance of the features (e.g., brain areas or connections) in that classifier. In one example,
researchers trained an SVM classifier to predict depression based on graph-theory based
features derived from diffusion-weighted imaging, and then looked at the importance of those
features (Sacchet et al. 2015).
In another example, researchers trained a random forest
classifier to predict Alzheimer's disease from structural MRI and then determined which brain
areas were the most predictive features in this classifier (Lebedev et al. 2014). Another general
approach is to compare classification models that are constructed using different features. For
example, the previously mentioned paper (Lebedev et al. 2014) also compared classifiers
constructed with different feature sets, e.g., one using cortical thickness measures and one
using volumetric measures. There are thus multiple ways in which ML can inform us about the
predictive relationship between neuroanatomic features and neurological disease.
Neurons have complicated shapes with varying biological structure that vary widely across brain
regions and across species (Zeng and Sanes 2017). Many approaches have been proposed to
classify neurons: electrophysiology (Teeter et al. 2018), morphology (Vasques et al. 2016),
genetics or transcriptomics (Nelson, Sugino, and Hempel 2006), and synaptic connectivity
(Jonas and Kording 2015). Machine learning can help with this endeavor (Armañanzas and
Ascoli 2015; Vasques et al. 2016). The cell types can be labeled based on |
1604.00091 | 1 | 1604 | 2016-04-01T01:02:30 | Learning warps object representations in the ventral temporal cortex | [
"q-bio.NC"
] | The human ventral temporal cortex (VTC) plays a critical role in object recognition. Although it is well established that visual experience shapes VTC object representations, the impact of semantic and contextual learning is unclear. In this study, we tracked changes in representations of novel visual objects that emerged after learning meaningful information about each object. Over multiple training sessions, participants learned to associate semantic features (e.g. made of wood, floats) and spatial contextual associations (e.g. found in gardens) with novel objects. Functional magnetic resonance imaging was used to examine VTC activity for objects before and after learning. Multivariate pattern similarity analyses revealed that, after learning, VTC activity patterns carried information about the learned contextual associations of the objects, such that objects with contextual associations exhibited higher pattern similarity after learning. Further, these learning-induced increases in pattern information about contextual associations were correlated with reductions in pattern information about the objects visual features. In a second experiment, we validated that these contextual effects translated to real-life objects. Our findings demonstrate that visual object representations in VTC are shaped by the knowledge we have about objects, and show that object representations can flexibly adapt as a consequence of learning with the changes related to the specific kind of newly acquired information. | q-bio.NC | q-bio | Learning warps object representations in the ventral temporal cortex
Alex Clarke1,2, Philip J. Pell1, Charan Ranganath2, Lorraine K. Tyler1
1Centre for Speech, Language and the Brain, Department of Psychology, University of Cambridge, UK. CB2 3EB.
2Center for Neuroscience, University of California Davis, Davis CA, 95618.
Abstract
The human ventral temporal cortex (VTC) plays a critical role in object recognition. Although it is well
established that visual experience shapes VTC object representations, the impact of semantic and
contextual learning is unclear. In this study, we tracked changes in representations of novel visual
objects that emerged after learning meaningful information about each object. Over multiple training
sessions, participants learned to associate semantic features (e.g. "made of wood", "floats") and spatial
contextual associations (e.g. "found in gardens") with novel objects. Functional magnetic resonance
imaging was used to examine VTC activity for objects before and after learning. Multivariate pattern
similarity analyses revealed that, after learning, VTC activity patterns carried information about the
learned contextual associations of the objects, such that objects with contextual associations exhibited
higher pattern similarity after learning. Further, these learning-induced increases in pattern information
about contextual associations were correlated with reductions in pattern information about the object's
visual features. In a second experiment, we validated that these contextual effects translated to real-life
objects. Our findings demonstrate that visual object representations in VTC are shaped by the
knowledge we have about objects, and show that object representations can flexibly adapt as a
consequence of learning with the changes related to the specific kind of newly acquired information.
1
Introduction
The ventral temporal cortex (VTC) is crucial for object recognition (Clarke & Tyler, 2014; Grill-Spector et
al., 1998; Kravitz, Saleem, Baker, Ungerleider, & Mishkin, 2013; Martin, 2007; Ungerleider & Mishkin,
1982) and many studies have demonstrated that visual experience can shape object representations in
human VTC (see Kourtzi & Connor, 2011; Op de Beeck & Baker, 2010). In daily life, objects are not only
processed according to their visual appearance, but also according to their meaning. Little is known,
however, about how visual object representations change as they transition from being meaningless to
meaningful.
In order to understand how learning about meaning changes object representations, it is important to
distinguish between different dimensions of meaning that could be learned. For instance, one can learn
about intrinsic attributes (e.g. has ears, made of metal, floats) that determine the function or
significance of an object, or about spatial contextual associations (e.g. found in the zoo) that enable
objects to be situated in the world. Previous studies, in which participants learned semantic features for
meaningless objects have shown learning-related increases in brain responses (Bellebaum et al., 2013;
James & Gauthier, 2003, 2004; Moore, Cohen, & Ranganath, 2006; Skipper, Ross, & Olson, 2011;
Vuilleumier, Henson, Driver, & Dolan, 2002; Weisberg, van Turennout, & Martin, 2007). These studies
show that learning object meaning can change how objects are represented in VTC, but they do not
address the central issue of how specific aspects of meaning drive changes in the neural representation
of objects.
The aim of the present study was to test if there is a relationship between the specific type of
information people learned and how object representations changed. We used functional magnetic
resonance imaging (fMRI) to examine how representations of pre-experimentally novel objects are
modified as they become meaningful. We examined the effect of learning two aspects of meaning – an
object's semantic category and its contextual association, both linked to regions of the VTC such as the
posterior fusiform gyrus (Chao, Haxby, & Martin, 1999; Clarke & Tyler, 2014; Cox & Savoy, 2003; Huth,
Nishimoto, Vu, & Gallant, 2012), parahipocampal cortex (PHC), and extending to the retrosplenial cortex
(RSC; Aminoff, Gronau, & Bar, 2007; Bar & Aminoff, 2003; Bar, Aminoff, & Schacter, 2008; Stansbury,
Naselaris, & Gallant, 2013). Further, we tracked how learning about different aspects of meaning
impacts on visual shape-based representations.
Participants learned a name and four semantic features for each of 12 novel objects, that created a
semantic category structure (three different categories based on overlap of features), while objects
either did or did not have a contextual feature. We used multivariate representational similarity analysis
(RSA; Kriegeskorte, Mur, & Bandettini, 2008; Nili et al., 2014) to determine how learning these two
aspects of meaning - semantic category and contextual associations- effects neural similarity spaces in
VTC. We predict that if learning meaningful information drives changes in neural similarity spaces, then
objects from the same semantic category, or those associated with a context, will show more similar
activation patterns in VTC than objects not sharing that property. Within the VTC, we predict semantic
category effects to be most prominent in the fusiform, while contextual association effects are predicted
2
to be more widespread including the RSC, PHC, lingual and fusiform. Finally, the visual shape similarity
of the objects is predicted to relate to activation patterns in early visual cortex possibly extending into
the VTC, and allows us to test for a relationship between visual form information and the newly learned
semantic information.
To assess the generalizability of our results, we ran a second experiment to examine how semantic
category and contextual associations influence VTC activity patterns during processing of real-world
objects. Based on previous studies we would expect category and contextual effects for real objects, and
their inclusion here not only demonstrates the generalisability of any learning effects, but importantly
allows for comparisons of the distribution of effects for real and novel objects which can help establish
whether effects of our learning paradigm lead to the kinds of representational changes that mirror the
long-term learned representations that are present for real objects.
Methods
Experiment 1: Novel objects
Overview
An overview of the experimental sessions can be seen in Table 1. The experiment consisted of two
identical fMRI sessions (mean time between scans was 26 days; range 21-28 days) and four behavioural
learning/testing sessions during which the semantic feature information was learned for 12 novel
objects. During the fMRI sessions, participants performed a simple visual task that could be performed
on the objects both before and after learning, and all behavioural sessions were conducted between the
two fMRI sessions. The behavioural sessions were completed within one week, with behavioural session
3 occurring between 48 and 72 hours prior to the second fMRI session, and behavioural session 4
occurring just prior to the second scanning session.
Participants
Twelve healthy participants (6 males, 6 females) completed all MRI and behavioural sessions. All had
normal or corrected to normal vision and were right handed. The average age was 20.7 years (range 19-
23 years). All participants provided informed consent, and the study was approved by the Cambridge
Research Ethics Committee. One participant was excluded due to excessive head movement during both
scanning sessions, and was excluded from all analyses (leaving eleven participants).
Stimuli
A total of 24 non-real, novel objects were used in the study - 12 were assigned to the learning condition
and 12 to the exposure condition. The novel objects were "fribbles" (downloaded from
http://wiki.cnbc.cmu.edu/Novel_Objects), and have a main body and four appendages. Eight fribbles
were selected from three different visual "species", where a species is defined by a common main body.
Within each species, two fribbles were selected from each of four "families", where members of a family
share the main body and the appendages have a variable degree of overlap, and fribbles from the same
3
species but different families share a main body but have no appendages in common. This structure
provides a variable degree of visual similarity across the novel objects based on the main body and
appendages (visual features). Half of the fribbles were used for the learning condition and half for the
exposure condition. In each condition, and for all species, the amount of visual feature overlap was
matched to ensure that effects were due to the semantic, rather than the visual, properties of the
objects.
All fribbles were displayed as grey-scale images in the centre of a white background. Pretesting was used
to ensure that the 24 fribbles did not have a strong resemblance to familiar object categories (animals,
plant life, tools, vehicles, living, nonliving). Eight items from the exposure set were inverted based on
pretests so they no longer showed a resemblance to familiar object categories.
Figure 1. Objects in the learning condition, showing their names and semantic features. Each row shows
a different semantic group, where each member shares 2 or 3 features. Each semantic group is
composed of objects from two different visual groups.
4
The 12 fribbles from the learning condition were each assigned a name and four semantic features (e.g.
made of metal; Figure 1) forming three semantic groups based on semantic feature overlap. Each
semantic group contained four fribbles that were drawn from two different visual species. All items
within a semantic group had two semantic features in common that were not present for items in other
semantic groups (the shared features of that group i.e. Made of metal & Conducts electricity for the four
fribbles on the top row of Figure 1). These shared features provide the basis for semantic similarity
within each group and the basis for category organisation. Further, one semantic feature was shared
between two items in each group (less shared features; e.g. Is durable is present in two fribbles on the
top row of Figure 1) and all other semantic features were unique, with a total of 27 features. Twenty six
of the semantic features were selected from the McRae et al., (2005) feature production norms, and
one feature was chosen by the experimenters.
All semantic features were non-visual (i.e., not describing visual object form) and synonyms were
avoided to ensure features were easily discriminable. Features that signify animacy (e.g., eats) or specify
the function of familiar items (e.g., used for holding liquids) were also avoided to increase the
plausibility of the features for novel objects. The two shared semantic features were a material (e.g.,
made of wood) and a well known property of that material (e.g., burns). Two of the unique semantic
features in each semantic group were contextual features (e.g. Found in...). The remaining semantic
features were sensory (e.g., smells bad), hidden-visual (e.g., yellow on the inside) or general object
properties (e.g. has layers, is lightweight). This selection of the fribbles and assignment of semantic
features allows us to test for high-level visual, semantic feature and contextual similarities in the brain
due to learning.
Procedure
Behavioural sessions
Participants learned to associate semantic information for 12 novel objects over two learning/testing
sessions on separate days (behavioural sessions 1 to 2) and two testing session (behavioural sessions 3
to 4). While learning was only given in the first two sessions, testing was conducted in all behavioural
sessions to track learning rates over time.
Learning sessions
The 12 objects were learned in subsets of 4 items. Initially participants viewed information slides for
each object in the subset. Slides contained an image of the object along with text containing its name
and semantic features (Figure 2a). Participants were instructed they would be tested on their knowledge
of the text and were asked to read each slide carefully. After viewing the information slides, participants
answered a series of questions about the objects and their features (Figure 2b). Questions were
presented in three phases; Phase 1 - associating objects with the shared features. Phase 2 - associating
objects with less shared and unique features. Phase 3 - associating names with the objects/semantic
features. Questions followed the general form of presenting a target and two response options. The
nature of the targets and response options varied across trials and phases but could be an object image,
name, single feature or pair of features. The response options remained on screen until a decision was
made, and feedback was given on each trial to reinforce learning. Information slides for each object
5
were intermixed within the question trials. Each phase was repeated if accuracy was below 80%. In total
there were 252 question trials in behavioural session 1 and 216 question trials in behavioural session 2,
where only phases 2 and 3 were given.
Figure 2. Semantic learning and testing. a) Example information slide presented to participants in order
to learn semantic information. b) Example question trial from phase 2 of learning. c) Example question
from the forced-choice recognition task in the testing phase
Testing
Two methods were used to measure learning performance: a forced-choice recognition task and a
feature-recall task. Forced-choice recognition tests were administered after learning in behavioural
6
sessions 1 and 2, and again during behavioural session 3 (where no learning took place). On each trial,
four semantic features belonging to the same object were presented with three options (Figure 2c), that
were either object images or names. The incorrect options were either from the same or different
semantic groups. There were a total of 48 trials in the testing sessions. Semantic features and response
options remained on screen until a response was made, and no feedback was given. The lowest accuracy
across all three testing sessions was 79% (Table 2), with all participants scoring higher than 85% in the
final testing session and 5/11 participants scoring 100%.
Feature-recall tests were administered at the start of behavioural session 2, 3 and 4. Participants were
given two sheets of paper containing the images of the 12 learned, novel objects, above five empty lines
labelled: "Name", "Feature 1", "Feature 2", "Feature 3", and "Feature 4". Participants were asked to fill
in as much information as possible within a 10 minute time limit. The ordering and location of each
object on the answer sheets was randomised across the three tests, and was not ordered by semantic or
visual groupings. The lowest accuracy in the final testing session (behavioural session 4) was 67%, while
7/11 participants scored 100% (Table 2). Overall, the testing scores indicate that the vast majority of
participants learned the associations between the semantic features and the object images to a high
level of success (9/11 scored at least 98% in the forced-choice testing, and 8/11 scored at least 96% in
the free-recall).
Exposure condition
An additional 12 novel objects were included to act as a control condition for any influence of visual
familiarity and to test for overall effects of learning meaning. These objects were not associated with
semantic features, but participants were exposed to them using a one-back visual matching task that
was performed during behavioural sessions 2 and 3. On each trial an object appeared for 600 ms and
was followed by a 1 s central fixation cross. Participants were instructed to press the right button if the
object on screen did not match the previous item or the left button if it did match the previous item. A
12-trial practice of the one-back task was followed by two 96-trial blocks. Within each block there were
8 repetitions of each item and each item was twice a 'match' target; giving a total of 24 match-trials in
each block. Participants were given feedback on their performance at the end of each block.
fMRI sessions
Procedure
Both pre and post-learning fMRI sessions used the same task and procedure. Participants performed a
visual anomaly detection task where they had to detect when one of the objects' visual features was
'bleached out'. The task ensured participants paid close attention to the images, could perform the task
both prior to and post-learning, and could be equally performed with objects from the learning and
exposure sets. Two modified versions of each of the 24 objects was created for the anomaly detection
task by increasing the brightness and contrast of one of the appendage features by 60% using Adobe
Photoshop CS2. A different feature was modified in each version. One version was used in the first
scanning session and the other in the second scanning session.
7
Each trial consisted of a centrally presented black fixation cross on a white background for 300 ms,
followed by a picture lasting 700ms, then a blank white background between 2 s and 7 s. The
participants' task was to press one button if the object on screen was an unmodified image or another
button if the object was a modified image (where a feature was bleached out; see Fig M4 centre panel
below). There were 15 repetitions of each image; 12 unmodified versions and three modified versions.
Objects were presented in 15 blocks each containing a single presentation of all 24 object with four or
five modified images in each block. The order of items within a block was random, and different for
every block. The position of the four/five modified items in each block was random with the constraint
that no more than two could occur in succession. The presentation and timing of stimuli was controlled
using Eprime version 1.1 (Psychology Software Tools, Pittsburgh, PA).
Scanning
All scanning took place at the MRC Cognition and Brain Sciences Unit, Cambridge, in a Siemens 3-T Tim
Trio MRI scanner (Siemens Medical Solutions, Camberley, UK). Two functional scans were collected in
each scanning session using gradient-echo echoplanar imaging (EPI) sequences collecting 32 slices in
descending order of 3 mm thickness and between slice gap of 0.75 mm and a resolution of 3 x 3 mm.
The field-of-view was 192 x 192 mm, matrix size 64 x 64 with a TE of 30 ms, TR of 2 s and a flip angle of
78⁰. Each functional scan lasted approximately 12 minutes with a short break half-way through. Prior to
functional scanning, a high-resolution structural MRI image was collected using an MPRAGE sequence
with 1 mm isotropic resolution.
fMRI preprocessing
Data from the two scanning sessions were preprocessed independently. Functional images were slice-
time corrected, spatially realigned and smoothed using a 4 mm Gaussian kernel in SPM8 (Wellcome
Institute of Cognitive Neurology, London, UK). These unnormalised images were analysed for each
participant with the general linear model to creating a single beta image for each object based on the 12
repetitions of the unmodified images. In addition to the 24 novel object predictors in the GLM,
predictors were included to capture effects associated with the modified objects, slow trends using 11
regressors for the first block of functional scans and 12 for the second block based on the basis functions
of a discrete cosine transform (minimum frequency = 1/128 Hz), six head motion regressors for each
session along with their first derivatives, and a global mean predictor for each scanning session. We also
included a separate regressor for each fast-motion event, where a fast motion event was defined as
motion greater than 0.7 mm/TR and detected using ArtRepair software (Mazaika, Hoeft, Glover, & Reiss,
2009). The resulting 24 beta images for the unmodified objects were converted to t-images prior to the
RSA analyses.
RSA analysis
We first used a searchlight mapping approach (Kriegeskorte, Goebel, & Bandettini, 2006) to test for
either consistent effects over sessions, or learning-induced changes to how the novel objects were
represented in the pre- and post-learning fMRI sessions. Searchlight analyses were followed by analyses
of anatomically defined ROIs that have been previously linked to semantic category and contextual
effects, and was performed as activation patterns may encode information at the spatial scale of
8
anatomical regions which searchlight analyses may be less sensitive to due to their smaller spatial
neighbourhood (Etzel, Zacks, & Braver, 2013).
Candidate model RDMs
We tested a visual model RDM and three other models capturing different semantic distinctions (Figure
3). The visual features model captures the amount of visual-feature overlap between pairs of objects,
where a feature can be the main body or an appendage. This is a high-level model of visual similarity
that captures visual shape information, rather than simply low-level visual information. Objects from the
same fribble species will always share at least one feature (the main body), while objects from different
fribble species always share no features. The meaning model tests for overall effects of learning
meaning in contrast to the objects in the exposure condition. Here, objects that were learned about are
predicted to cluster together to a greater degree that objects from the exposure condition, which have
no associated meaning. The semantic category model captures the three semantic categories the
objects belonged to, where members of the same semantic category share either 2 or 3 features, and no
features were shared across semantic categories. The contextual model tests where activation patterns
for the learned objects form two clusters according to whether they were associated with a specific
context or have no contextual association (the correlation between the semantic category and
contextual association RDMs was r = 0.19).
Figure 3. Candidate model RDMs tested using representational similarity analysis.
Searchlight
Each candidate model RDM was tested against the observed activation patterns using a searchlight
similarity analysis implemented in the RSA toolbox (Nili et al., 2014) and custom matlab functions,
9
before we tested for significant conjunctions or changes over sessions. At each voxel, object activation
values from grey matter voxels within a spherical searchlight (radius 9 mm) were extracted to calculate
distances between all items (using 1 - Pearson correlation) creating an object dissimilarity matrix based
on that searchlight. This fMRI RDM was correlated with each candidate model RDM (using Spearman's
rank correlation) and the resulting similarity values were Fisher transformed and mapped back to the
voxel at the centre of the searchlight. Similarity maps for each model RDM and each participant were
normalised to the MNI template space and spatially smoothed using a 6 mm FWHM Gaussian kernel.
The similarity maps for each participant were then entered into a group-level random effects (RFX)
analyses using SPM8. We used a paired samples t test design testing for first, common RSA effects
across both sessions using a conjunction analysis (the conjunction null hypothesis) for positive effects,
and second, increased RSA effects in the post-learning session compared to the pre-learning session.
Given our a priori expectation that effects will be in the ventral processing stream, these analyses were
constrained to a ventral stream mask (including the RSC) produced by combining bilateral regions from
the Harvard-Oxford brain atlas (occipital pole, intracalcarine cortex, cuneal cortex, lateral occipital
inferior, fusiform occipital, fusiform temporooccipital, lingual gyrus, parahippocampal posterior) and the
RSC (BA29 & 30). Results are reported using a voxelwise threshold of p < 0.005 and a cluster extent of p
< 0.05 (FWE) correcting for multiple comparisons.
An anatomical ROI analysis was performed to test for effects of the candidate model RDMs in targeted
VTC regions that have been previously implicated in semantic category and contextual processing. This
was done to test for distributed regional effects that may only be present in activation patterns at the
spatial scale of regions, rather than the smaller searchlights. This would be the case if information was
distributed across voxels at a spatial scale larger than the searchlights. A number of ROIs were specified
to cover the VTC, specifically the fusiform (fusiform temporooccipital), lingual (lingual gyrus) and
parahippocampal (parahippocampal posterior). ROIs were defined from the Harvard-Oxford brain atlas
(Desikan et al., 2006), and the retrosplenial cortex was also included and defined as BA29 and BA 30. For
each ROI, object activation values were extracted for all voxels in the ROI to calculate distances between
the 24 items (using 1 - Pearson correlation). The fMRI ROI RDMs were correlated with the candidate
model RDM (using Spearman's rank correlation) and the resulting similarity values were Fisher
transformed. RFX analyses were performed using one-tailed one-sampled t tests against zero when
testing for significant positive effects of each model RDM, and paired-samples t-tests when testing for
significant differences between fMRI sessions or model RDMs.
ROI analysis
Experiment 2: Real objects
Participants
Sixteen participants took part in the study (6 male, 10 female). All were right-handed and were aged
between 19 and 29 (mean 23 years). All participants had normal or corrected to normal vision, and gave
informed consent. The study was approved by the Cambridge Research Ethics Committee.
10
Stimuli
A total of 145 real objects were used, where 131 of these were from one of six object categories (34
animals, 15 fruit, 21 vegetables, 27 tools, 18 vehicles, 16 musical instruments) and 14 additional objects
that did not adhere to a clear category (and were not included in our analyses). Isolated coloured
objects were shown in the centre of a white background, and normalised to a maximum visual angle of
7.5°. All objects were chosen to depict concepts from an anglicised version of the McRae production
norms (McRae et al., 2005; Taylor, Devereux, Acres, Randall, & Tyler, 2012) from which semantic feature
information could be obtained to construct the meaningful model RDMs.
Procedure
Participants performed an overt basic-level naming task. Each trial consisted of a fixation cross lasting
500 ms, before an object for 500 ms followed by a blank screen lasting between 3 and 11 seconds. All
objects were repeated 6 times across 6 different blocks. The object presentation order for each block
was randomised for each participant, although a constant category order was maintained ensuring an
even distribution of object category across the block. The presentation and timing of stimuli were
controlled with Eprime version 1 (Psychology Software Tools, Pittsburgh, PA, USA), and naming accuracy
was recorded by the experimenter during acquisition.
fMRI acquisition
Participants were scanned at the MRC Cognition and Brain Sciences Unit, Cambridge, in a Siemens 3-T
Tim Trio MRI scanner (Siemens Medical Solutions, Camberley, UK). There were 3 functional scanning
sessions using gradient-echo echoplanar imaging (EPI) sequences collecting 32 slices in descending order
of 3 mm thickness and between slice gap of 0.75 mm and a resolution of 3 x 3 mm. The field-of-view was
192 x 192 mm, matrix size 64 x 64 with a TR of 2 seconds, TE of 30 ms and a flip angle of 78⁰. Each
functional session lasted approximately 9-10 minutes, containing two object blocks. Prior to functional
scanning, a high-resolution structural MRI image was collected using an MPRAGE sequence with 1 mm
isotropic resolution.
fMRI preprocessing
Functional images were slice-time corrected, spatially realigned and smoothed using a 4 mm Gaussian
kernel in SPM8 (Wellcome Institute of Cognitive Neurology, London, UK). These unnormalised images
were analysed for each participant with the general linear model to create a single beta image for each
object based on the 6 repetitions. In addition to the 145 object predictors, predictors were included to
capture slow trends using 18 regressors for each session based on the basis functions of a discrete
cosine transform (minimum frequency = 1/128 Hz), six head motion regressors for each session and a
global mean predictor for each scanning session. The resulting 145 beta images were converted to t-
images prior to the RSA analyses. Only objects named correctly on all six repetitions (86%, SE = 1.53%)
were included in further analyses.
RSA analysis
An RSA analysis was performed for the real objects using both searchlight mapping and an anatomical
ROI analysis. The procedures were identical to those used in experiment 1, with the exception that here
11
the searchlight RFX analysis was conducted using a one-sampled t-test testing for positive effects of the
models with no a priori voxel restrictions.
We tested for RSA effects of a contextual association RDM and a semantic category RDM (note that the
high-level visual features and meaning models were not tested, as they can not be defined for our real
objects in the same manner as they are defined for the novel objects). The contextual association model
RDM was defined for real objects in the same manner as we defined it for novel objects. Using the
McRae et al. (2005) production norm data to extract semantic features for our real objects, we can
determine if an object has contextual associations or not. Features that indicated an object has strong
contextual associations took the form; associated with X, found in/near/on X, lives in X, used at/on X,
where X signified a specific location (e.g. deserts, school, Spain). Based on this criterion, 44 out of 131
objects were determined to have spatial contextual associations. The contextual association model RDM
for real objects was constructed in the same way as for novel objects, where objects with contextual
associations will form one cluster, and objects without contextual associations are form a second
cluster. The semantic category RDM captures the category structure of the 131 objects across 6
superordinate categories (animals, fruit, vegetables, tools, vehicles, musical instruments) and indicates
there is more within-category similarity than between-category similarity. The correlation between the
semantic category and contextual association RDMs was r = 0.09.
Results
Experiment 1: Novel objects
The goal of this study was to determine whether learning-induced changes in object representations are
governed by the semantic category and contextual association information people learn, and whether
this impacts on visual form representations. The similarity relations between the objects, as embodied
in the model RDMs, capture the visual and meaningful (learned) similarity spaces that can be tested
against brain activation patterns to track changes in the information represented due to learning. Here,
we used RSA to uncover statistical correspondences between the visual and meaningful (i.e. learned)
similarity structures (Figure 3) and activation pattern similarities across the different objects. We first
explored whether effects of our candidate model RDMs showed significant learning-induced changes, or
consistent effects across sessions, in their relationship to activation patterns. To do this we directly
compared the RSA searchlight maps from the pre- and post-learning scanning sessions, where the pre-
learning session acts as a baseline for how the brain responds to meaningless objects. We then present
an anatomical ROI analysis to look for more spatially distributed effects that extend beyond the size of
our searchlights.
Using searchlight RSA, the visual features model was found to show consistent significant effects across
both sessions in the occipital lobe (peak MNI coordinate: 21, -88, -1; Figure 4a), showing that
searchlights within the visual cortex responded to the same images in a similar manner both before and
after learning. Multidimensional scaling (MDS) further illustrated that the regional activation patterns
12
clustered by visual form similarity, with objects from the same visual groups (shown by different
colours) falling closer together in this 2-dimensional space than objects from different visual groups.
Testing for increased RSA effects after learning compared to before, we found significant effects relating
to the contextual association model. Representational changes were seen in the right VTC with two foci
– one on the fusiform gyrus (change in Spearman's rho = 0.08; peak MNI coordinate: 39, -34, -24) and
one spanning the collateral sulcus (change in Spearman's rho = 0.04; peak MNI coordinate: 24, -52, -16;
Figure 4b). MDS plots further showed how activation patterns in both these areas dissociate along the
contextual association dimension of the stimuli. No effects were seen for the other meaning related
models. These results show that activation patterns in the VTC have been warped through learning,
whereby the same stimuli are represented in a different manner after learning in comparison to before,
while object activation patterns in the early visual regions remained more stable in that they were
significantly related to the visual feature information in both sessions, that were not significantly
different to one another (Note that equivalent results are found by inspecting pre- and post- learning
searchlight maps separately).
Figure 4. RSA searchlight effects for novel objects. a) Searchlights showing common significant effects
across both sessions for the visual features model. MDS plot derived from object pattern similarities
averaged over sessions with each visual group shown in a different colour. b) Searchlights showing
significant increases for the contextual associations model in the post-learning session compared to the
pre-learning session. MDS plots derived from object pattern similarities in the post-learning session for
the two foci, where objects with contextual associations shown in blue and without contextual
associations in red (exposure objects not shown for clarity). Both images are voxelwise p < 0.005, cluster
p < 0.05.
13
While our searchlight analysis tested for effects in local activation patterns (i.e. searchlights),
representational information could be present at the larger scale of anatomical regions which will not
always be captured by smaller searchlights. This would be the case if information was distributed across
voxels at a spatial scale larger than the searchlights. Therefore we also performed an analysis testing the
candidate model RDMs within anatomically defined regions across the posterior ventral temporal lobe,
including the RSC, that have been implicated in semantic category and contextual processing -
specifically bilateral posterior fusiform, lingual, PHC and RSC (Figure 5).
Figure 5. RSA ROI analyses showing pre- and post-learning effects for the 4 model RDMs in the 4
anatomical ROIs. Pre-learning effects shown in light grey, post-learning effects in dark grey. Asterisks
show significant effects of the models (**p<0.01,*p<0.05) and significant changes over sessions shown
by a horizontal bar. PHC = parahippocampal cortex and RSC = retrosplenial cortex. Error bars show
standard errors.
Confirming the results from our searchlight analysis, we observed significant learning-induced increases
in contextual association information in the fusiform in the post-learning session compared to the pre-
learning session (t(10) = 3.37, p = 0.004). Further, the fusiform region showed significant effects of the
contextual association model during the post-learning session (t(10) = 5.10, p = 0.0002), which were
absent in the pre-learning session (t(10) = 0.6). We also observed marginally significant learning-induced
increases for the contextual association model in the RSC (t(10) = 1.67, p = 0.064), which was significant
for the post-learning session (t(10) = 1.96, p = 0.039) but not in the pre-learning session (t(10) = 0.05).
Finally, the lingual region showed significant effects for the contextual association model in the post-
learning session (t(10) = 1.92, p = 0.042), and not in the pre-learning session (t(10) = 0.4), although no
significant change was seen between the two sessions (t(10) = 1.1).
Turning to the visual features model, a significant effect was seen in the fusiform for the pre-learning
activation patterns (t(10) = 1.92, p = 0.042), that was reduced in the post-learning session and no longer
significant (t(10) = 0.6), although cross session comparisons revealed no significant differences between
14
the two sessions (t(10) = 0.85). The visual feature model also showed effects in the lingual region after
learning (t(10) = 3.02, p = 0.006), which showed marginal effects in the pre-learning session (t(10) = 1.62,
p = 0.068) that was not significantly different across the two scanning sessions (t(10) = 0.2) suggesting
the lingual region reflected relatively stable object representations, similar to our searchlight results.
Finally, we note that the other two meaning-related models – the meaning and semantic category
RDMs, both showed significant effects in the post-learning session only (meaning: t(10) = 2.41, p =
0.018, semantic category: t(10) = 2.33, p = 0.021). These effects were absent in both the cross-session
searchlight comparisons and in individual searchlight effects for the post-learning session (not shown)
suggesting that information relating to general meaningfulness, and to the learned semantic categories
the objects belong to, may be coded in more distributed patterns than the searchlight mapping was
sensitive to.
The results from the anatomical ROI analyses confirm what was observed in the our searchlight analyses
- learning about novel objects induced representational changes in VTC, most prominently in the
fusiform gyrus, where patterns reflected the learned contextual associations of objects. Within the
anatomical ROIs we also saw suggestions of a more general meaning-related effect conferred through
the meaning and semantic category models. Moreover, after learning, visual similarity effects in the
fusiform are reduced and learning-induced contextual similarities emerge. Testing the relationship
between representational changes in the fusiform over the sessions showed a significantly greater
change in RSA effects for the contextual model compared to both the visual feature model (t(10) = 2.2, p
= 0.05) and the semantic category model (t(10) = 2.2, p = 0.05). Moreover, the reduction in the visual
feature model effect was significantly correlated with the increased effect for the contextual association
model (R-sq = 0.49, r = -0.7, p = 0.0085; Figure 6) indicating a representational shift in the kind of
information that was represented in the fusiform as a consequence of learning.
Figure 6. Relationship of representational changes between pre- and post-learning sessions for visual
feature and contextual association model RDMs.
Experiment 2: Real objects
In Experiment 1 we found that learning about contextual associations for pre-experimentally novel
objects changed voxel pattern similarity information in VTC. However, in real life, contextual
associations are learned by encountering objects in particular contexts, whereas for our novel objects
associations were learned through reading text. Accordingly, in Experiment 2, we use RSA to test for
15
effects of a contextual association model RDM for real objects in a previously published dataset (Clarke
& Tyler, 2014). This not only allows us test the generalisability of our learning effects, but allows us to
compare the distribution of effects for real and novel objects. Further we test for effects of a semantic
category model, also known to show effects in the VTC, as a means of comparison, using both
searchlight mapping and the anatomical ROIs within the VTC.
Searchlight RSA showed effects for the contextual association model primarily in bilateral posterior VTC
including the fusiform, lingual, and parahippocampal cortices, and also including the calcarine and
inferior occipital gyrus (peak MNI coordinate: -9, -91, 10; Figure 7a). The results echo those previously
reported for the contextual associations of real world objects (e.g. Stansbury et al., 2013), validating the
manner in which we define our contextual model, and partly overlapping with our effects for novel
objects in the collateral sulcus. The semantic category effects were widespread throughout VTC (peak
MNI coordinate: -27, -26, 59), similar to previously reported effects (e.g. Connolly et al., 2012; Huth et
al., 2012) although we would also highlight that there are well known confounds between an object's
semantic category and its visual properties, a factor that we controlled in our novel objects experiment.
Figure 7. RSA effects of the contextual association and semantic category models using real objects. a)
Searchlight results, shown thresholded at voxelwise p < 0.005, cluster p < 0.05. b) RSA effects for the
anatomical ROIs. Dark grey bars show effects for the contextual association model and light grey for the
semantic category model. Asterisks show significant effects of the models (**p<0.01,*p<0.05), with
significant changes over sessions shown by a horizontal bar. PHC = parahippocampal cortex, RSC =
retrosplenial cortex, VTC = ventral temporal cortex. Error bars show standard errors.
Within the anatomical ROIs, we found significant effects of the contextual association model in the
fusiform (t(15) = 4.42, p = 0.0002), lingual (t(15) = 3.34, p = 0.0022) and PHC (t(15) = 2.13, p = 0.025)
confirming our searchlight results are also present at this larger spatial scale (Figure 7b). The semantic
category model showed significant effects in the fusiform (t(15) = 8.61, p < 0.0001), lingual (t(15) = 4.13,
p = 0.0004), PHC (t(15) = 2.44, p = 0.014), and the RSC (t(15) = 2.34, p = 0.017). Further, the ROI similarity
matrix was more similar to the semantic category model than the contextual association model in the
fusiform ROI (t(15) = 5.10, p = 0.0001), showing the inverse relationship to that found for novel objects.
Overall, both searchlight and ROI analyses showed activation patterns to real objects show a significant
16
relationship to the contextual associations model with the strongest effects in the fusiform and lingual
gyri, plus widespread semantic category effects across VTC that also peaked in the fusiform gyrus.
Discussion
The goal of the present study was to determine how learning different kinds of meaningful information
affects the neural representation of objects in VTC. In Experiment 1, we examined the effect of learning
two aspects of meaning: the semantic features of an object, and whether each object had a specific
contextual association (i.e., that an object is found in a particular setting). We found that learning about
the presence of contextual associations induced systematic changes in the similarity structure of
corresponding object representations in VTC. Learning led to the development of two clusters in the
similarity space, one consisting of objects with a strong contextual association (i.e. a 'Found in...'
feature) and another consisting of objects that did not have a strong contextual association. In
Experiment 2, we demonstrated a distinction between images of real-world objects that had prominent
contextual associations and objects that did not elicit strong contextual associations. Collectively, the
results suggest that learned contextual associations exert powerful influences on the neural mechanisms
of object processing. We elaborate on this and related issues below.
Contextual association effects in VTC
Searchlight-based voxel pattern similarity analyses in Experiment 1 revealed that learning contextual
associations affected object similarity spaces in VTC (Figure 4). Within VTC, there was a lateral peak in
the fusiform gyrus and a more medial peak spanning the collateral sulcus where the similarity responses
reflected the presence of learned contextual associations. The peak of the medial effect was posterior to
the parahippocampal gyrus, peaking closely to reported scene selective responses (Aguirre, Zarahn, &
D'Esposito, 1998). Anatomical ROI analyses converged with these results, showing contextual
association effects for both novel and real objects in the fusiform, and changes for novel objects were
also seen in RSC. Together, the findings demonstrate that learning about contextual associations causes
significant representational changes in the VTC.
Large areas of the VTC have been shown to contain information about natural scene categories and their
contextual associations (Stansbury et al., 2013; Walther, Caddigan, Fei-Fei, & Beck, 2009), in addition to
meaningful object information more generally (Huth et al., 2012; Mahon, Anzellotti, Schwarzbach,
Zampini, & Caramazza, 2009; Tyler et al., 2013; Vuilleumier et al., 2002). The emergence of learned
information suggests that representations in the VTC are malleable through short term learning, with
this area strongly implemented in previous learning studies (Gauthier, Tarr, Aanderson, Skudlarski, &
Gore, 1999; James & Gauthier, 2004; Kourtzi, Betts, Sarkheil, & Welchman, 2005; Moore et al., 2006; Op
de Beeck, Baker, DiCarlo, & Kanwisher, 2006; Sigala & Logothetis, 2002; van der Linden, Murre, & van
Turennout, 2008). Our results add to this assertion by showing that learning-induced changes are
specific to the kind of meaning that was learned.
17
Whereas previous fMRI research claims the RSC processes contextual information, the PHC is also
strongly implicated (Aminoff et al., 2007; Bar & Aminoff, 2003; Bar et al., 2008; Ranganath & Ritchey,
2012). In experiment 1, verbal learning about the context in which an object could be found had strong
effects on pattern similarity relationships in RSC, but interestingly, no comparable effects were detected
in PHC. In contrast experiment 2, using real-life objects, did show contextual association effects in the
PHC. This differential pattern in the PHC could be attributed to how contextual information was learned
in the cases of novel and real-life objects. In experiment 1, these contextual associations were learned
verbally, but never visually experienced. This contrasts with our understanding of real-world objects, for
which we predominantly acquire strong contextual associations through repeated visual exposure in
specific contexts. These may suggest that visual experience is key for contextual association effects in
the PHC, while contextual effects in the RSC were seen for novel objects that suggests a more abstracted
role of linking objects to contexts that is not as tied to visual contexts as the PHC (Bar, 2004).
Learning-induced informational warping in the fusiform
Results from Experiment 1 demonstrated that learning meaningful information about novel objects was
associated with changes in visual-form similarity relationships in the fusiform. Our anatomical ROI
analysis showed that the change in the contextual association model effect over sessions was
significantly greater than the change in the visual-features model effect, and critically, the reduction in
the visual feature effect with learning was strongly correlated with the emergence of the contextual
association effect. This result could reflect a dimensional modulation of the representational space
(Folstein, Palmeri, & Gauthier, 2014) where there is a representational shift away from the pre-learning
visually-based object information, to information about object meaning. This result must reflect a
change in how these objects are represented in VTC, rather than a top-down effect, as the same visual
task was used before and after learning that did not require specific access to object meaning. Our
results further highlight how the VTC can flexibly code different kinds of object properties gained
through experience (such as form and meaning) and suggests that meaning plays a critical role in
shaping object representations.
Beyond contextual association effects of meaning
Beyond learned effects of contextual associations in the fusiform, we also found evidence for more
general semantic effects after learning at the spatial scale of anatomical regions. Activation patterns in
the fusiform ROI correlated with the meaning and semantic category RDMs after learning which may
indicate more subtle changes along semantic dimensions, where members of the same semantic
category show more similar activation patterns after learning. A number of explanations could underlie
the more modest semantic category effects compared to the contextual effects for novel objects. One
account would be that participants in Experiment 1 learned to associate objects with labels that did not
correspond to pre-existing object categories (such as animals or tools). Categories defined by their
shared features (e.g. one category would be composed of metal, electrically conducting things), as in
Experiment 1, have little ecological relevance, and participants could not readily leverage pre-existing
category representations to facilitate learning (Op de Beeck & Baker, 2010).
18
This may suggest that learning about object contexts and forming superordinate categories takes place
over different timescales. The significantly stronger effects for the contextual model over the semantic
category model in Experiment 1 in the fusiform could be a consequence of enhanced initial learning for
the contextual associations, which would be beneficial as context provides additional semantic
constraints. Ad hoc testing supports the notion that contextual information aids initial learning, as our
participants showed higher behavioural feature-recall accuracy for objects with a contextual ('Found
in...') feature following the first learning session (session 2: contextual mean feature-recall 66%, non-
contextual mean feature-recall 54%, t(10) = 2.27, p = 0.046) with this advantage disappearing with
further training sessions (session 3: contextual mean feature-recall 91%, non-contextual mean feature-
recall 85%, t(10) = 2.12, p = 0.06; session 4: contextual mean feature-recall 94%, non-contextual mean
feature-recall 92%, t(10) = 1.16, p = 0.27). The ability to form generalisations across items that share
common features could require more extensive exposure to exemplars, especially when objects are
novel and cannot readily fit into existing categories. Consistent with the possibility of different
timescales of context and category learning, the contextual association effect was significantly greater
than the semantic category effect in the fusiform for novel objects, whereas the inverse was found for
real objects. One caveat is that, in the case of real objects, semantic category and visual properties are
highly confounded, whereas this was not the case for novel objects in Experiment 1. Further research is
therefore needed to clarify how contextual and category information interact in the formation of stable
meaningful representations in the brain.
In conclusion, the present results demonstrate that visual object representations change as a
consequence of learning meaningful information for previously meaningless objects. Learning-induced
changes in neural pattern similarity relationships tracked the specific kinds of information that
participants learned about objects, showing how representations in the VTC can flexibly adapt to
represent newly acquired meaningful information. Such changes were specific to the informational
structure that was learned, further emphasising that object representations are shaped by the nature of
the information we learn.
Acknowledgements
This project has received funding to LKT from the European Research Council (ERC) under the European
Union's Horizon 2020 research and innovation programme (grant agreement No 669820), from the
European Research Council under the European Community's Seventh Framework Programme
(FP7/2007-2013)/ ERC Grant agreement n° 249640, and a Guggenheim Fellowship to CR.
References
Aguirre, G. K., Zarahn, E., & D'Esposito, M. (1998). An area within human ventral cortex sensitive to
"building" stimuli: evidence and implications. Neuron, 21(2), 373-383.
19
Aminoff, E., Gronau, N., & Bar, M. (2007). The parahippocampal cortex mediates spatial and nonspatial
associations. Cerebral Cortex, 17(7), 1493-1503. doi: 10.1093/cercor/bhl078
Bar, M. (2004). Visual objects in context. Nat Rev Neurosci, 5(8), 617-629. doi: 10.1038/nrn1476
Bar, M., & Aminoff, E. (2003). Cortical analysis of visual context. Neuron, 38(2), 347-358.
Bar, M., Aminoff, E., & Schacter, D. L. (2008). Scenes unseen: the parahippocampal cortex intrinsically
subserves contextual associations, not scenes or places per se. Journal of Neuroscience, 28(34),
8539-8544. doi: 10.1523/JNEUROSCI.0987-08.2008
Bellebaum, C., Tettamanti, M., Marchetta, E., Della Rosa, P., Rizzo, G., Daum, I., & Cappa, S. F. (2013).
Neural representations of unfamiliar objects are modulated by sensorimotor experience. Cortex,
49(4), 1110-1125. doi: 10.1016/j.cortex.2012.03.023
Chao, L. L., Haxby, J. V., & Martin, A. (1999). Attribute-based neural substrates in temporal cortex for
perceiving and knowing about objects. Nature Neuroscience, 2(10), 913-919.
Clarke, A., & Tyler, L. K. (2014). Object-specific semantic coding in human perirhinal cortex. Journal of
Neuroscience, 34(14), 4766-4775.
Connolly, A. C., Guntupalli, J. S., Gors, J., Hanke, M., Halchenko, Y., Wu, Y., . . . Haxby, J. V. (2012). The
representation of biological classes in the brain. The Journal of Neuroscience, 32(8), 2608-2618.
Cox, D. D., & Savoy, R. L. (2003). Functional magnetic resonance imaging (fMRI) "brain reading":
detecting and classifying distributed patterns of fMRI activity in human visual cortex.
Neuroimage, 19(2 Pt 1), 261-270. doi: S1053811903000491 [pii]
Desikan, R. S., Segonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., . . . Killiany, R. J. (2006).
An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral
based
DOI
10.1016/j.neuroimage.2006.01.021
Neuroimage,
968-980.
interest.
regions
of
31(3),
doi:
Etzel, J. A., Zacks, J. M., & Braver, T. S. (2013). Searchlight analysis: promise, pitfalls, and potential.
Neuroimage, 78, 261-269. doi: 10.1016/j.neuroimage.2013.03.041
Folstein, J. R., Palmeri, T. J., & Gauthier, I. (2014). Perceptual advantage for category-relevant perceptual
Front Psychol, 5, 1394. doi:
and motion.
case of
shape
dimensions:
10.3389/fpsyg.2014.01394
the
Gauthier, I., Tarr, M. J., Aanderson, A., Skudlarski, P., & Gore, J. C. (1999). Activation of the middle
fusiform 'face area' increases with expertise in recognizing novel objects. Nature Neuroscience,
2, 568-573.
Grill-Spector, K., Kushnir, T., Hendler, T., Edelman, S., Itzchak, Y., & Malach, R. (1998). A sequence of
object processing stages revealed by fMRI in the human occipital lobe. Human Brain Mapping, 6,
316-328.
Huth, A. G., Nishimoto, S., Vu, A. T., & Gallant, J. L. (2012). A continuous semantic space describes the
representation of thousands of object and action categories across the human brain. Neuron,
76(6), 1210-1224. doi: 10.1016/j.neuron.2012.10.014
James, T. W., & Gauthier, I. (2003). Auditory and action semantic features activate sensory-specific
perceptual brain regions. Current Biology, 13(20), 1792-1796.
James, T. W., & Gauthier, I. (2004). Brain areas engaged during visual judgments by involuntary access to
novel semantic information. Vision Research, 44(5), 429-439.
20
Kourtzi, Z., Betts, L. R., Sarkheil, P., & Welchman, A. E. (2005). Distributed neural plasticity for shape
learning in the human visual cortex. PLoS Biol, 3(7), e204. doi: 10.1371/journal.pbio.0030204
Kourtzi, Z., & Connor, C. E. (2011). Neural representations for object perception: structure, category, and
adaptive coding. Annual Review of Neuroscience, 34, 45-67. doi: 10.1146/annurev-neuro-
060909-153218
Kravitz, D. J., Saleem, K. S., Baker, C. I., Ungerleider, L. G., & Mishkin, M. (2013). The ventral visual
pathway: an expanded neural framework for the processing of object quality. Trends Cogn Sci,
17(1), 26-49. doi: 10.1016/j.tics.2012.10.011
Kriegeskorte, N., Goebel, R., & Bandettini, P. (2006). Information-based functional brain mapping.
Proceedings of the National Academy of Sciences of the United States of America, 103(10), 3863-
3868.
Kriegeskorte, N., Mur, M., & Bandettini, P. (2008). Representational similarity analysis - connecting the
branches of systems neuroscience. Front Syst Neurosci, 2, 4. doi: 10.3389/neuro.06.004.2008
Mahon, B. Z., Anzellotti, S., Schwarzbach, J., Zampini, M., & Caramazza, A. (2009). Category-specific
organization in the human brain does not require visual experience. Neuron, 63(3), 397-405.
Martin, A. (2007). The representation of object concepts in the brain. Annual Review of Psychology, 58,
25-45. doi: 10.1146/annurev.psych.57.102904.190143
Mazaika, P., Hoeft, F., Glover, G. H., & Reiss, A. L. (2009). Methods and Software for fMRI Analysis for
Clinical Subjects. Paper presented at the Human Brain Mapping.
McRae, K., Cree, G. S., Seidenberg, M. S., & McNorgan, C. (2005). Semantic feature production norms for
a large set of living and nonliving things. Behavior Research Methods, 37, 547-559.
Moore, C. D., Cohen, M. X., & Ranganath, C. (2006). Neural mechanisms of expert skills in visual working
memory. Journal of Neuroscience, 26(43), 11187-11196. doi: 10.1523/JNEUROSCI.1873-06.2006
Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., & Kriegeskorte, N. (2014). A toolbox for
doi:
Comput Biol,
e1003553.
analysis.
10(4),
PLoS
representational
similarity
10.1371/journal.pcbi.1003553
Op de Beeck, H. P., & Baker, C. I. (2010). The neural basis of visual object learning. Trends Cogn Sci,
14(1), 22-30. doi: 10.1016/j.tics.2009.11.002
Op de Beeck, H. P., Baker, C. I., DiCarlo, J. J., & Kanwisher, N. G. (2006). Discrimination training alters
object representations in human extrastriate cortex. Journal of Neuroscience, 26(50), 13025-
13036. doi: 10.1523/JNEUROSCI.2481-06.2006
Ranganath, C., & Ritchey, M. (2012). Two cortical systems for memory-guided behaviour. Nat Rev
Neurosci, 13(10), 713-726. doi: 10.1038/nrn3338
Sigala, N., & Logothetis, N. K. (2002). Visual categorization shapes feature selectivity in the primate
temporal cortex. Nature, 415, 318-320.
Skipper, L. M., Ross, L. A., & Olson, I. R. (2011). Sensory and semantic category subdivisions within the
doi:
Neuropsychologia,
3419-3429.
49(12),
anterior
10.1016/j.neuropsychologia.2011.07.033
temporal
lobes.
21
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(5), 1025-1034. doi:
10.1016/j.neuron.2013.06.034
Taylor, K. I., Devereux, B. J., Acres, K., Randall, B., & Tyler, L. K. (2012). Contrasting effects of feature-
based statistics on the categorisation and identification of visual objects. . Cognition, 122(3),
363-374.
Tyler, L. K., Chiu, S., Zhuang, J., Randall, B., Devereux, B. J., Wright, P., . . . Taylor, K. I. (2013). Objects and
categories: Feature statistics and object processing in the ventral stream. Journal of Cognitive
Neuroscience, 25(10), 1723-1735.
Ungerleider, L., & Mishkin, M. (1982). Two cortical visual systems. In D. J. Ingle, M. A. Goodale, & R. J. W.
Mansfiled (Eds.), Analysis of visual behavior (pp. 549-586). Cambridge, MA: MIT Press.
van der Linden, M., Murre, J. M., & van Turennout, M. (2008). Birds of a feather flock together:
experience-driven formation of visual object categories in human ventral temporal cortex. PLoS
ONE, 3(12), e3995. doi: 10.1371/journal.pone.0003995
Vuilleumier, P., Henson, R. N., Driver, J., & Dolan, R. J. (2002). Multiple levels of visual object constancy
revealed by event-related fMRI of repetition priming. Nature Neuroscience, 5(5), 491-499.
Walther, D. B., Caddigan, E., Fei-Fei, L., & Beck, D. M. (2009). Natural scene categories revealed in
distributed patterns of activity in the human brain. Journal of Neuroscience, 29(34), 10573-
10581. doi: 10.1523/JNEUROSCI.0559-09.2009
Weisberg, J., van Turennout, M., & Martin, A. (2007). A neural system for learning about object function.
Cerebral Cortex, 17(3), 513-521. doi: 10.1093/cercor/bhj176
22
|
1609.01663 | 2 | 1609 | 2016-10-26T16:19:04 | The canonical semantic network supports residual language function in chronic post-stroke aphasia | [
"q-bio.NC"
] | Current theories of language recovery after stroke are limited by a reliance on small studies. Here, we aimed to test predictions of current theory and resolve inconsistencies regarding right hemispheric contributions to long-term recovery. We first defined the canonical semantic network in 43 healthy controls. Then, in a group of 43 patients with chronic post-stroke aphasia, we tested whether activity in this network predicted performance on measures of semantic comprehension, naming, and fluency while controlling for lesion volume effects. Canonical network activation accounted for 22-33% of the variance in language test scores. Whole-brain analyses corroborated these findings, and revealed a core set of regions showing positive relationships to all language measures. We next evaluated the relationship between activation magnitudes in left and right hemispheric portions of the network, and how right hemispheric activation related to the extent of left hemispheric damage. Activation magnitudes in the each hemispheric network were strongly correlated, but four right frontal regions showed heightened activity in patients with large lesions. Activity in two of these regions (inferior frontal gyrus pars opercularis and supplementary motor area) was associated with better language abilities in patients with larger lesions, but poorer language abilities in patients with smaller lesions. Our results indicate that bilateral language networks support language processing after stroke, and that right hemispheric activations related to extensive left hemisphere damage occur outside of the canonical semantic network and differentially contribute to in their relationship to behavior depending on the extent of left hemispheric damage | q-bio.NC | q-bio | Title: The canonical semantic network supports residual language function in chronic
post-stroke aphasia
Authors: Joseph C. Griffis1, Rodolphe Nenert2, Jane B. Allendorfer2, Jennifer Vannest3,
Scott Holland3, Aimee Dietz4, Jerzy P. Szaflarski2
Institutional Affiliations: University of Alabama at Birmingham Department of
Psychology1, University of Alabama at Birmingham Department of Neurology2,
Cincinnati Children's Hospital Medical Center, Cincinnati, OH3, University of Cincinnati
Academic Health Center, Cincinnati, OH4
Corresponding Author Information: Joseph C. Griffis ([email protected])
Department of Neurology and UABEC, University of Alabama at Birmingham, 312
Civitan International Research Center, 1719 6th Avenue South, Birmingham, AL 35294-
0021
Abstract
Current theories of language recovery after stroke are limited by a reliance on small
studies. Here, we aimed to test predictions of current theory and resolve inconsistencies
regarding right hemispheric contributions to long-term recovery. We first defined the
canonical semantic network in 43 healthy controls. Then, in a group of 43 patients with
chronic post-stroke aphasia, we tested whether activity in this network predicted
performance on measures of semantic comprehension, naming, and fluency while
controlling for lesion volume effects. Canonical network activation accounted for 22-
33% of the variance in language test scores. Whole-brain analyses corroborated these
findings, and revealed a core set of regions showing positive relationships to all language
measures. We next evaluated the relationship between activation magnitudes in left and
right hemispheric portions of the network, and characterized how right hemispheric
activation related to the extent of left hemispheric damage. Activation magnitudes in the
each hemispheric network were strongly correlated, but four right frontal regions showed
heightened activity in patients with large lesions. Activity in two of these regions
(inferior frontal gyrus pars opercularis and supplementary motor area) was associated
with better language abilities in patients with larger lesions, but poorer language abilities
in patients with smaller lesions. Our results indicate that bilateral language networks
support language processing after stroke, and that right hemispheric activations related to
extensive left hemisphere damage occur outside of the canonical semantic network and
differentially relate to behavior depending on the extent of left hemispheric damage.
Keywords: fMRI, language recovery, stroke, lesion, right hemisphere
Introduction
Functional neuroimaging of language recovery after stroke
Aphasia commonly occurs in patients with strokes affecting the left middle
cerebral artery (LMCA) territory, and is one of the most debilitating consequences of
stroke since language impairments affect nearly every aspect of daily life [Maas et al.,
2012]. Language recovery after LMCA stroke is variable, with many survivors
experiencing chronic deficits [Charidimou et al., 2014; Lazar et al., 2008; Pedersen,
1995]. Current theories of aphasia recovery are based primarily on evidence obtained
from functional neuroimaging studies [Heiss and Thiel, 2006; Saur and Hartwigsen,
2012; Turkeltaub et al., 2011]. During early recovery, patients have been observed to
show an up-regulation of right hemispheric responses to language tasks [Heiss et al.,
1999; Saur et al., 2006] that is thought to reflect a transient compensatory mechanism
triggered by the acute disruption of left hemispheric function by the ischemic event
[Hartwigsen et al., 2013; Saur et al., 2006; Thiel et al., 2006]. In later stages of recovery,
the gradual reinstatement of left hemispheric capacity for language processing is thought
to decrease the need for right hemispheric compensation, resulting in a return of
relatively balanced hemispheric activation patterns [Hamilton et al., 2011; Heiss and
Thiel, 2006; Saur et al., 2006; Saur and Hartwigsen, 2012; Turkeltaub et al., 2011]. Thus,
the general principle to be inferred from functional neuroimaging studies of language
recovery after stroke might be summarized as follows: good long-term language
outcomes depend primarily on the preservation and/or restoration of close-to-normal
function in canonical brain language networks.
While there is consistent support in the literature for a primary role of the
canonical left hemispheric language networks in supporting language recovery after
stroke, the role of the right hemisphere is less clear, particularly with regard to the
chronic recovery phase. This is likely due, in part, to an oversimplified conceptualization
of right hemispheric contributions to recovery in much of the literature [Turkeltaub et al.,
2012]. Nonetheless, seemingly discrepant findings regarding the role of the right
hemisphere pose an obstacle to developing a deeper understanding of how/when the right
hemisphere contributes to language recovery. For example, there is evidence that
activation of right inferior frontal [Belin et al., 1996; Griffis et al., 2016b; Rosen et al.,
2000; Winhuisen et al., 2005; Winhuisen et al., 2007] and right superior temporal [Heiss
et al., 1999; Karbe et al., 1998; Szaflarski et al., 2013] areas during language tasks are
associated with poorer language function, although there is also evidence that activation
of right inferior frontal [Mattioli et al., 2014; van Oers et al., 2010; Raboyeau et al., 2008;
Saur et al., 2006] and right anterior superior temporal [Crinion and Price, 2005] regions
may support residual language processing and contribute to language recovery.
One explanation for such seemingly discrepant relationships is that when the left
hemisphere is not sufficiently preserved, the initial reliance on right hemispheric
language network homologues persists beyond early recovery, and the lack of left
hemispheric involvement leads to poorer recovery relative to when left hemispheric
function is successfully restored [Heiss et al., 1999; Heiss and Thiel, 2006; Karbe et al.,
1998; Szaflarski et al., 2013; Turkeltaub et al., 2011]. In this case, the direction of the
observed relationship between right hemispheric activation and language task
performance might depend on the characteristics of the patient sample. For example, right
hemispheric activation might show a positive relationship to language abilities in a
sample that primarily consists of severely impaired patients with extensive left
hemispheric damage, whereas a negative relationship might be observed in a sample that
includes patients with various degrees of impairment and varying extents of left
hemispheric damage [Saur and Hartwigsen, 2012].
Sample sizes in functional neuroimaging studies of aphasia recovery
A limitation shared by many of the studies that form the foundation for current
theories of recovery is that they typically utilize small patient samples [Saur and
Hartwigsen, 2012], and this has the potential to (1) increase the risk of detecting spurious
effects that do not generalize beyond the sample being studied, (2) reduce power to detect
real effects, and (3) produce inflated estimates of the magnitudes of detected effects.
Indeed, it has been previously noted in the context of neuroimaging that correlation
estimates based on small samples are also unstable and can be disproportionately
influenced by outlier data points [Poldrack, 2012]. Further, the power to detect between-
subject effects is reduced in small samples [Yarkoni, 2009], and these relationships are
often of primary interest in neuroimaging studies of language recovery after stroke. Even
for real effects, small sample sizes inflate estimates of effect size, such that for a
relationship with a population r of 0.3, the average observed sample r in samples of 20
patients is expected to be around 0.73 [Yarkoni, 2009].
To illustrate the prevalence of small sample sizes in functional neuroimaging
studies (i.e. not including structural/lesion-mapping studies) of post-stroke aphasia, we
performed a literature search using Google Scholar. Using the search terms "aphasia
fmri", "aphasia pet", "aphasia neuroimaging", "aphasia recovery", "aphasia imaging",
and "language recovery after stroke", we identified 84 functional neuroimaging (i.e. not
including structural/lesion-mapping studies) studies published between 1995 and 2016
with accessible texts. Across all 84 studies, the average patient sample size was only
about 10 patients (mean = 10.33, SD = 8.44; Figure 1), and 87% (73/84) had fewer than
20 patients. The single largest sample was reported by a study published this year
[Geranmayeh et al., 2016] that collected data from 53 sub-acute patients (see
Supplementary Material S1 for individual study sample sizes and a full reference list).
While these estimates are based on a relatively coarse literature search, we note that they
are in close agreement with the previous observation by Saur and Hartwigsen [2012] that
neuroimaging studies of therapy-induced aphasia recovery typically feature typically less
than 10 patients.
Despite the limitations of small samples in individual studies, meta-analytic
methods such as activation likelihood estimation (ALE) provide a means for identifying
consistent effects in the literature [Wager et al., 2007]. However, we are aware of only
one such published meta-analysis of the functional neuroimaging of aphasia recovery
literature [Turkeltaub et al., 2011]. Larger studies, in conjunction with meta-analytic
assessments of the published literature, are crucial for resolving discrepancies in the
literature and for building generalizable theories of recovery that are based on robust
empirical evidence.
Statistical control for lesion volume in functional neuroimaging of aphasia
recovery
A second potential limitation shared by most functional neuroimaging studies of
language recovery after stroke is a lack of statistical control for lesion volume effects on
imaging-behavior relationships. This is somewhat surprising given that the use of
statistical controls to reduce confounds related to lesion volume effects is relatively
common in lesion-symptom mapping research [Rorden and Karnath, 2004; Schwartz et
al., 2009; Zhang et al., 2014]. The logic behind using statistical controls to account for
lesion volume effects is that since (1) larger lesions are often associated with more severe
impairments [Allendorfer et al., 2012; Butler et al., 2014; Cheng et al., 2014; Karnath et
al., 2004; Kümmerer et al., 2013; Meltzer et al., 2013; van Oers et al., 2010; Rorden and
Karnath, 2004; Szaflarski et al., 2013; Yarnell et al., 1976], and (2) larger lesions have a
higher probability of including both task-relevant and task-irrelevant voxels [Karnath et
al., 2004; Rorden and Karnath, 2004; Schwartz et al., 2009; Zhang et al., 2014], lesion
volume effects have the potential to introduce bias into measurements of lesion-behavior
relationships at voxels that are primarily damaged in patients with large lesions [Karnath
et al., 2004; Zhang et al., 2014]. In functional neuroimaging, the probability that a voxel
is active during task performance depends on the probability that the tissue at that voxel
is spared, and the probability that the tissue at a given voxel in the left hemisphere is
spared depends on the size of the lesion. Therefore, the potential for lesion volume
confounds is logically extendable to relationships between functional neuroimaging
activation and behavior, and it is possible that a lack of controls for such effects could
introduce bias into functional neuroimaging studies of recovery.
Study aims and hypotheses
In the current study, we first tested what we consider to be the primary prediction
of current models of language recovery after stroke using a relatively large (n=43) sample
of patients with chronic (> 1 year) post-stroke aphasia and a sample (n=43) of
demographically matched healthy controls. Specifically, we tested the prediction that
long-term language outcomes depend on the preservation/restoration of language task-
driven activation in canonical language networks. To test this prediction, we first defined
the canonical semantic network (CSN) as the set of regions that were more strongly
activated by semantic decisions relative to tone decisions in the healthy controls. This
bilateral but predominantly left-lateralized network is well-suited for testing the
prediction that residual language abilities depend on the preservation/recovery of activity
within distributed canonical language networks as it is (1) well-characterized in healthy
individuals, and (2) consists of distributed brain regions that include the left angular
gyrus (AG)/posterior inferior parietal lobule (pIPL), left middle temporal gyrus (MTG),
left parahippocampal gyrus (PHG), left dorsal superior frontal gyrus (dSFG), left inferior
frontal gyrus (IFG), and left posterior cingulate cortex (PCC) [Binder et al., 1997; Binder
et al., 1999; Binder et al., 2008; Binder et al., 2009; Kim et al., 2011; Szaflarski et al.,
2002; Xu et al., 2016]. Thus, we tested whether task-evoked activity in the CSN
identified in controls predicted in-scanner performance on the semantic decision task,
out-of-scanner performance on verbal fluency tasks, and out-of-scanner performance on a
picture naming task in the patient group (with and without controls for lesion volume
effects). We note that while the in-scanner and out-of-scanner tasks differ in their
emphases on speech comprehension and speech production, respectively, previous
studies suggest that activity evoked by speech comprehension tasks may also relate
performance on other measures of language function that include speech production tasks
[e.g. Saur et al., 2006; Szaflarski et al., 2013]. Other previous studies also suggest that
speech production networks (i.e. regions activated during word/verb generation) are also
activated (albeit somewhat less strongly) during sentence comprehension tasks
[Piervincenzi et al., 2016], and that estimates of language lateralization based on
activations measured during both speech production and auditory semantic decision tasks
show similar relationships to estimates based on intracarotid amobarbital measurements
[Szaflarski et al., 2008].
Additionally, we tested the explanation that strong right hemispheric activation in
chronic patients reflects compensation driven by a reduced capacity of the left
hemisphere to perform language-related processing. As noted in the previous paragraph,
the CSN has bilateral components [e.g. Binder et al., 2009], and thus activation of the
right hemispheric network might be expected to increase to compensate for reduced left
hemispheric network function [e.g. Hartwigsen et al., 2013]. While previous studies have
investigated how activity in the right hemisphere relates to the effects of damage to
specific left hemispheric regions [Blank et al., 2003; Sims et al., 2016; Turkeltaub et al.,
2011], we are not aware of any studies that have directly measured the how activation in
the right hemisphere relates to the extent of damage sustained by the left hemisphere.
Further, the notion that language-task evoked activation in the right hemisphere is
most pronounced in patients with extensive left hemispheric lesions is commonly
discussed in the literature [Hamilton et al., 2011; Heiss and Thiel, 2006]. Despite its
intuitive appeal, there appears to be little empirical evidence to support this claim. For
example, a recent review of the mechanisms of language recovery after stroke [Hamilton
et al., 2011] cites only two reports to support the unambiguous assertion that "larger
lesions involving eloquent cortex of the left hemisphere are associated with greater
recruitment of the right hemisphere during language tasks". One is an earlier review
[Heiss and Thiel, 2006] that discusses this idea within the context of a larger hierarchical
model of recovery but does not itself provide empirical support for the veracity of this
claim in stroke patients, and the other [Kertesz et al., 1979] is a computerized
tomography study that shows more extensive left hemispheric damage to be predictive of
more severe aphasia and poorer language recovery after stroke, but that does not relate
lesion size to any measurement of functional activation. Therefore, we sought to
empirically test the hypotheses that right hemispheric language task activation in chronic
stroke patients reflects compensation driven by reduced left hemispheric function, and
that this effect is most pronounced in patients with the most extensive left hemispheric
damage. We expected, based on the literature discussed above, that higher levels of right
CSN activation during semantic decisions would be associated with (1) lower levels of
left CSN activation, and (2) more extensive left hemispheric damage. As noted above, the
CSN is predominantly left-lateralized (i.e. the right and left CSN are not mirror-
symmetric), and so we also assessed these effects in right hemispheric regions mirroring
the left CSN.
Finally, because we are not aware of any studies that have directly measured the
effect of left hemispheric lesion volume on regional activation in the right hemisphere,
we characterized regions in the right hemisphere where task-driven responses correlated
with left hemispheric lesion volume. To address the question of whether activity in these
regions might reflect beneficial compensation that is most pronounced for patients with
extensive left hemispheric damage, we further assessed whether relationships between
activity in these regions and language measures depended on the extent of left
hemispheric damage.
All study procedures were approved by the Institutional Review Boards of the
Materials and Methods
Participants
participating institutions and were performed in accordance with Declaration of Helsinki
ethics principles and principles of informed consent. For the current study, we used
functional MRI data collected from 43 chronic post-stroke aphasia patients and 43
healthy controls. Imaging and behavioral data for the post-stroke aphasia patients were
collected as part of several separate studies performed by our laboratory. Prior to
inclusion, all participants were screened to exclude individuals that had diagnoses of
degenerative/metabolic disorders, had severe depression or other psychiatric disorders,
were pregnant, were not fluent in English, or had any contraindication to MRI/fMRI.
Patients were included in the current study if they had a single left hemispheric stroke
resulting in aphasia at least 1 year prior to data collection. Data for the control
participants were selected from a database of 150 healthy individuals collected as part of
several studies performed by our laboratory, and were selected based on age group (19-
29, 30-39, 40-49, 50-59, 60+), handedness as determined by the Edinburgh Handedness
Inventory [Oldfield, 1971], and sex to minimize differences in demographics with respect
to the patient group. Participant demographics are shown in Table 1. A more detailed
characterization of patient demographics is provided in the Supplementary Material (S2).
Language measures
Prior to MRI scanning, all participants were administered a battery of
neuropsychological language assessments. All participants performed the Boston Naming
Test (BNT) [Kaplan et al., 2001], Semantic Fluency Test (SFT) [Kozora and Cullum,
1995], and Controlled Oral Word Association Test (COWAT) [Lezak et al., 1995]. The
BNT requires patients to name a series of black and white line drawings that contain both
animate and inanimate items that vary in frequency of use (e.g. bed vs. abacus), and the
number of correctly named pictures serves as a measure of naming ability. The SFT
requires patients to generate as many words as they can think of that fit a given category
prompt (animals/fruits and vegetables/things that are hot) within a one-minute time limit,
and the number of words generated serves as a measure of category fluency. The
COWAT requires patients to generate as many words as they can think of that begin with
a particular letter (C/F/L) within a one-minute time limit, and the number of words
generated serves as a measure of phonemic fluency. Performance scores for the COWAT
and SFT were very strongly correlated across patients (r=0.92), so they were averaged
together to define a single combined measure of verbal fluency. Performance scores for
the verbal fluency measure were correlated, albeit less strongly, with performance scores
for the BNT (r=0.76). Individual patient language task data are provided in the
Supplementary Material (S2).
Neuroimaging data collection
MRI data were collected at the University of Alabama at Birmingham using a 3T
head-only Siemens Magnetom Allegra scanner located in the Civitan International
Research Center Functional Imaging Laboratory. These data consisted of 3D high-
resolution T1-weighted anatomical scans (TR/TE = 2.3 s/2.17 ms, FOV =
25.6×25.6×19.2 cm, matrix = 256x256, flip angle = 9 degrees, slice thickness = 1mm),
and two T2*-weighted gradient-echo EPI pulse functional scans (TR/TE = 2.0 s/38.0 ms,
FOV = 24.0x13.6x24.0, matrix = 64x64, flip angle = 70 degrees, slice thickness = 4 mm,
165 volumes per scan). MRI data were also collected at the Cincinnati Children's
Hospital Medical Center on a 3T research-dedicated Phillips MRI system located in the
Imaging Research Center. These data consisted of 3D high-resolution T1-weighted
anatomical scans (TR/TE = 8.1 s/2.17 ms, FOV = 25.0×21.0×18.0 cm, matrix = 252x211,
flip angle = 8 degrees, slice thickness = 1mm) and two T2*-weighted gradient-echo EPI
pulse sequence functional scans (TR/TE = 2.0 s/38.0 ms, FOV = 24.0x13.6x24.0, matrix
= 64x64, flip angle = 70 degrees, slice thickness = 4 mm, 165 volumes per scan).
Functional MRI scans were acquired while participants completed alternating 30
second blocks of a semantic decision/tone decision task. This fMRI task was selected
because it has been previously shown to result in robust activation in canonical areas
involved in semantic language processing [Binder et al., 1997; Kim et al., 2011], and has
been used extensively to evaluate language network activation in healthy and diseased
populations that include patients with chronic post-stroke aphasia [e.g. Eaton et al., 2008;
Szaflarski et al., 2008; Szaflarski et al., 2011].
The active condition (semantic decision -- SD) was performed 5 times during
each scan. Each block of the active condition consisted of 8 trials where participants were
presented with spoken English animal names. On each trial, participants decided if the
presented animal met the criteria "native to the United States" and "commonly used by
humans". If both criteria were satisfied, then the participants responded "1" by using their
non-dominant hand to press a button. If both criteria were not satisfied, then the
participants responded "2" by using their non-dominant hand to press a second button.
The control condition was performed 6 times during each scan. Each block of the control
condition (tone decision – TD) consisted of 8 trials where participants were presented
with brief sequences of four to seven 500- and 750-Hz tones. On each trial, participants
decided if the sequence contained two 750-Hz tones. If the sequence contained two 750
Hz tones, then they pressed the button designated "1" with their non-dominant hand.
Otherwise, they pressed the button designated "2" with their non-dominant hand. Each
scan lasted 7 minutes and 15 seconds. Prior to completing the task in the MRI scanner, all
participants confirmed their understanding of the task by performing a mock run that
included a sequence of five sets of tones followed by a sequence of five nouns
designating different animals. In-scanner task data were not collected for 4 patients due to
hardware issues, and they were excluded from analyses investigating relationships
between fMRI activation and in-scanner performance.
Neuroimaging data processing
All MRI data were processed using Statistical Parametric Mapping (SPM)
[Friston et al., 1995] version 12 running in MATLAB r2014b (The MathWorks, Natick
MA, USA). Functional MRI scans were pre-processed using a standard pre-processing
pipeline consisting of slice-time correction, realignment/reslicing, co-registration of the
fMRI data to the corresponding anatomical scan, unified segmentation with optimized
tissue priors for lesioned brains and normalization of the anatomical scan to MNI space
[Ripollés et al., 2012; Seghier et al., 2008], normalization of the functional scan to MNI
space using the transformation applied to the anatomical scan, and spatial smoothing
using an 8mm full-width half maximum Gaussian kernel. This pipeline enables accurate
template registration and normalization even for patients with structural abnormalities
such as those observed in stroke patients [Ripollés et al., 2012]. To reduce the potential
for motion-related artifacts, functional scans were motion-corrected by replacing volumes
with >0.5mm motion with an interpolated volume created from adjacent volumes
[Mazaika et al., 2005].
Lesion probability maps were created for each patient using a voxel-based naïve
Bayes classification algorithm that was developed by our laboratory and is implemented
in the lesion_gnb toolbox for SPM12 [Griffis et al., 2016a]. While automated
classification with this method compares favorably to manual lesion delineation for large
and small lesions [Griffis et al., 2016a], we opted to manually threshold the resulting
posterior probability maps to ensure that they precisely reflected the extent of the lesion.
Our decision to manually threshold the resulting posterior probability maps was primarily
motivated by the potential for automated methods to introduce false positive voxel
clusters in patients with small lesions [e.g. Griffis et al., 2016a; Wilke et al., 2011], and
while an arbitrary cluster threshold (i.e. 100 voxels) is often sufficient to remove these
clusters, it is not guaranteed. Further, as we note in our report describing the validation of
our method [Griffis et al., 2016a] it is important to inspect the lesion masks as a quality
control step, and by manually thresholding the probability maps produced by the
automated classification procedure, we were able to ensure that the final lesion masks
accurately reflected each patient's lesion. The resulting binary lesion masks were used to
estimate lesion volume and lesion-ROI overlaps, and were used in all further lesion
analyses. Lesion frequencies across all 43 patients are shown in Figure 2A. Individual
patient lesion images are provided in the Supplementary Material (S2).
For each participant, the pre-processed fMRI data were fit to a general linear
model (GLM) where task blocks were modeled as boxcar regressors convolved with a
canonical hemodynamic response function (HRF). In order to account for variability in
the time-to-peak of the HRF, time and dispersion derivatives were included as basis
functions [Meinzer et al., 2013]. Note that since data were collected using a blocked
design, separate modeling of correct vs. incorrect trials was not possible. Linear contrast
estimate maps were then computed to quantify the difference in activation magnitudes
between the active and control conditions. These contrast maps were used for all further
functional MRI analyses.
Statistical Analyses
Group differences on behavioral measures were assessed for descriptive purposes
using two-tailed independent samples t-tests with degrees of freedom adjustment for
unequal variances.
To identify regions that showed significant group-level activation for each group,
the first-level contrast estimate maps from all individuals in each group were entered into
separate second-level t contrasts quantifying the difference in peak HRF magnitude
between the SD and TD conditions. Regional activation was considered significant if it
survived a combined voxel-level intensity (p-value) threshold of 0.01 and a cluster-level
extent threshold of p<0.05 corrected (kcrit=99 voxels) to control the whole-brain family-
wise error rate (FWE) at 0.05 as determined by 1000 Monte Carlo simulations [Slotnick
et al., 2003]. The thresholded map obtained for the control group was then binarized,
creating a CSN region-of-interest (ROI) mask where statistically significant voxels had a
value of 1 and all other voxels had a value of 0. For each patient, the mean contrast
estimate across all voxels within this CSN ROI was then extracted from their first-level
contrast estimate map, quantifying the average magnitude of the task-driven response in
the CSN. These estimates were used in subsequent ROI correlation analyses. Group
differences in activation within the CSN were assessed for descriptive purposes using an
independent samples t-test with degrees of freedom adjustment for unequal variances.
Lesion-ROI overlaps were calculated for each patient using the lesion masks created
during pre-processing. The correlation between total lesion volumes and lesion-ROI
overlaps indicated that lesion-ROI overlap was a close linear transformation of lesion
volume (r=0.93).
To assess the unique relationships between activity in the CSN and language
function in chronic patients, the mean contrast estimate quantifying activation within the
HC-defined CSN was then entered as an independent variable in 3 separate partial
correlation analyses. The dependent variables for each analysis were the performance
scores for the in-scanner SD task (% correct responses) and performance scores for the
out-of-scanner naming and fluency tasks. Each model included total lesion volume as a
covariate to account for variance attributable to lesion volume effects. Analyses were
repeated without lesion volume control. Linear correlation analyses assessing imaging-
behavior relationships were also performed for the control group using the CSN ROI. To
fully characterize relationships between regional language task activation and
performance on the in-scanner and out-of-scanner language tasks, three whole-brain
multiple regressions (one for each language task) were performed that each included
lesion volume as a covariate. Because patient data were collected on different scanners,
all correlational analyses for the patient group were repeated with scanner included as a
covariate, with and without lesion volume control. These analyses are provided in the
Supplementary Material (Supplementary Figure 4).
To test whether the magnitude of activation in right hemispheric portions of the
CSN was increased in patients with lower levels of activation in left hemispheric portions
of the network, the CSN ROI mask was split into left and right hemispheric components.
For each patient, the mean contrast estimate was then extracted from each hemispheric
ROI as described above. Linear correlations were assessed between the mean contrast
estimates obtained from each hemisphere with and without lesion volume control.
Correlations were also assessed between left hemispheric and right hemispheric
activation in controls. To explicitly test for this effect in right hemispheric homologues of
the left hemisphere network and control for differences in extent between left and right
hemispheric portions of the CSN, these analyses were also repeated using a right
hemispheric ROI that was defined by simply mirroring the left-hemispheric ROI to the
right hemisphere. To assess whether out-of-network right hemispheric homologues of the
left hemispheric CSN (i.e. regions that were activated in the left, but not right hemisphere
in controls) differed from the right hemispheric CSN with regard to their relationship to
left hemispheric CSN activation, a right hemispheric out-of-network homologue ROI was
created by subtracting the right hemispheric CSN from the mirrored left hemispheric
CSN (see Figure 4A for illustrations of each ROI). Differences in activation between left
and right hemispheric networks were assessed with dependent samples t-tests for each
group. Between-group differences in activation for left and right hemispheric networks
(and differences in activation between left and right hemispheric networks) were assessed
with unequal variance t-tests. Between-group differences in the strengths of correlations
between left and right CSN ROIs were assessed using z-tests [Fisher, 1921].
Lastly, to assess the effect of lesion size on task-driven activation in the right
hemisphere, we performed an additional whole-brain linear regression analysis that was
restricted to the right hemisphere and included only total lesion volume as a predictor.
Post-hoc moderation analyses were performed for right hemisphere clusters that showed
significant positive correlations between activation magnitudes and lesion volume in
order to assess whether activity in these regions reflected beneficial compensation in
patients with larger lesions. For each identified cluster, three separate multiple regression
analyses (one for each language measure) were performed that each included mean
activation magnitudes for that cluster, total lesion volume (entered as a proportion of the
maximum lesion volume), and the interaction term between mean cluster activation and
lesion volume. For these post-hoc analyses, only the interaction effects were of interest,
since we our hypothesis was that the benefit of recruiting these regions would depend on
the extent of left hemispheric damage.
Statistical tests were two-tailed and significance thresholds were set as follows.
Results for group comparisons and for each set of ROI-based analyses were considered
significant if they survived a Bonferroni-Holm step-down correction procedure to control
the FWE at 0.05 across each set of tests, since this method provides good protection
against Type I errors and better power than standard Bonferroni correction [Aickin and
Gensler, 1996]. T-contrast maps for the effects of interest from each of the whole-brain
multiple regression analysis were thresholded at p<0.01, uncorrected at the voxel level
and multiple comparisons corrected at the cluster level to control the cluster-wise FWE at
0.01 (kcrit = 126 voxels, as determined by 1000 Monte Carlo simulations) for each
contrast. Since power is often lower for analyses of moderation [Jaccard et al., 1990], the
post-hoc moderation analyses were multiple comparisons corrected using a more lenient
False Discovery Rate (FDR) correction threshold of 0.10 [Benjamini and Hochberg,
1995] computed across the interaction term p-values for all moderation models.
We note that a recent report [Eklund et al., 2016] has called into question the
validity of cluster-based correction methods in fMRI research, particularly when they are
used in conjunction with voxel-level thresholds above 0.001. Our choice to use the voxel-
level threshold of 0.01 was primarily motivated by its prevalence in the neuroimaging
literature of post-stroke aphasia, and because detection power is reduced for group
analyses of highly heterogeneous groups such as stroke patients [e.g. Meinzer et al.,
2013]. In addition, we note that for the whole-brain regression analyses, we employed a
more stringent whole-brain cluster-correction threshold of 0.01 (as opposed to the
conventional 0.05 threshold), even when our analyses were restricted to only the right
hemisphere. Thus, we have attempted to balance providing control over false positives
while maintaining sufficient power to detect effects of interest. In addition, we provide
voxel-wise FDR thresholded activation maps and cluster/peak statistics for each group in
Supplementary Figure 4 to (1) demonstrate that similar activation patterns are detected at
stringent voxel-wise correction thresholds, and (2) provide more precise anatomical
information about peak activation locations for each group.
Results
Comparisons of overall CSN activation and task performance between
patients and controls
Patients performed more poorly on language tasks than controls (in-scanner SD %
correct: t74.77=-4.71, p<0.001, corrected; BNT: t43.38=-6.80, p<0.001, corrected; fluency:
t82.28=-13.44, p<0.001, corrected). Summary statistics are shown in Figure 2B. The
semantic networks recruited by each group are shown in Figure 2C. Cluster peak co-
ordinates and statistics for regions activated by each group are provided in Table 2.
Controls showed significantly higher levels of overall activation in the CSN than patients
(t67.72 = 3.98, p<0.001, corrected).
Region of interest behavioral correlations
Total lesion volume showed negative relationships with each of the language
measures (in-scanner SD correct: r=-0.22, p=0.19, corrected; fluency: r =-0.58, p<0.001,
corrected; naming: r=-0.49, p=0.004, corrected), although the relationship was not
significant for the in-scanner SD task. Lesion volume showed a trend-level but non-
significant negative relationship with mean activation in the CSN (r=-0.27, p=0.15,
corrected). The partial correlation analyses revealed significant positive relationships
between activity within the CSN and in-scanner performance on the SD task (partial
r=0.47, p=0.008, corrected), out-of-scanner performance on the fluency measure (partial
r=0.62, p<0.001, corrected), and out-of-scanner performance on the naming measure
(partial r=0.49, p=0.005, corrected) that were linearly independent of the effects of total
lesion volume (Figure 2D). Relationships between CSN activation and in-scanner
performance on the SD task (r=0.51, p=0.005, corrected), out-of-scanner performance on
the fluency measure (r=0.65, p<0.001, corrected), and out-of-scanner performance on the
naming measure (r=0.55, p=0.001, corrected) were stronger without lesion volume
control. These results are shown in Figure 2D.
Additional analyses assessed relationships between mean CSN activation
magnitudes and behavioral measures in controls. Mean CSN activation magnitudes
showed a significant positive relationship to performance on the in-scanner SD task
(r=0.37, p=0.03, corrected), but did not show significant relationships with performance
on the fluency (r=0.09, p=1.0, corrected) or naming (r=0.09, p=1.0, corrected) tasks
(Figure 2D).
Whole-brain behavioral correlations
The results for the whole-brain multiple regressions for each language measure
that included total lesion volume as a covariate are shown in Figure 3A. Cluster peak
locations and statistics for each whole-brain regression are provided in Table 3. Figure
3B shows regions that showed positive relationships between activation and performance
on all three language measures, and Figure 3C shows regions that showed positive
relationships between activation and any of the language measures relative to the
locations of regions associated with the CSN identified in healthy controls. Whole-brain
analyses repeated without including lesion volume as a covariate showed very similar
results (Supplemental Tables 7-9).
Correlations between left and right hemispheric activation magnitudes
Partial correlation analyses (controlling for lesion volume) for the patient group
revealed a strong positive relationship between activation magnitudes in the left CSN and
activation magnitudes in the right CSN (partial r=0.83, p<0.001, corrected). Nearly
identical results were obtained without lesion volume control (r=0.83, p<0.001,
corrected). A strong positive relationship was also observed between activation
magnitudes in the left CSN and activation magnitudes in the mirrored left CSN in the
right hemisphere with (partial r=0.81, p<0.001, corrected) and without (r=0.80, p<0.001,
corrected) lesion volume control. However, correlations were much weaker between
activation magnitudes in the left CSN and activation magnitudes in out-of-network
homologues of the left hemispheric CSN in the right hemisphere with (partial r= 0.66,
p<0.001, corrected) and without (r=0.65, p<0.001, corrected) lesion volume control. Left
and right hemispheric ROIs are shown in Figure 4A. Scatterplots illustrating the
relationships between residual left and right hemispheric activation magnitudes (i.e. after
partialling out lesion volume) are shown in Figure 4B (top row).
Correlation analyses for the control group also revealed a strong positive
relationship between activation magnitudes in the left CSN and activation magnitudes in
the right CSN (r=0.75, p<0.001, corrected). A weaker relationship was observed between
activation magnitudes in the left CSN and activation magnitudes in the mirrored left CSN
in the right hemisphere (r=0.43, p=0.008, corrected). No significant relationship was
observed between activation magnitudes in the left CSN and activation magnitudes in
out-of-network homologues of the left CSN in the right hemisphere (r=-0.04, p=0.79,
corrected). Scatterplots illustrating these relationships are shown in Figure 4B.
The strengths of correlations between left CSN and right CSN activation
magnitudes did not significantly differ between patients and controls (z=0.52, p=0.6,
corrected), but the strengths of correlations between activation magnitudes in the left
CSN and mirrored left CSN in the right hemisphere (z=2.98, p=0.006, corrected) and
between activation magnitudes in the left CSN and out-of-network homologues of the left
CSN in the right hemisphere (z=3.72, p<0.001, corrected) were significantly stronger in
patients than in controls.
Comparisons of left and right hemispheric activation magnitudes
Activation magnitudes for each group are shown in Figure 4C. Neither patients
(t42=0.96, p=0.34, corrected) nor controls (t42=1.72, p=0.19, corrected) showed significant
differences in activation magnitudes between the left CSN and right CSN. Both patients
(t42=3.89, p=0.001, corrected) and controls (t42=9.22, p<0.001, corrected) showed
significantly larger activation magnitudes in the left CSN compared to the mirrored left
CSN in the right hemisphere. Both patients (t42=5.42, p<0.001, corrected) and controls
(t42=12.92, p<0.001, corrected) also showed significantly larger activation magnitudes in
the right CSN compared out-of-network homologues of the left CSN in the right
hemisphere.
Patients had significantly lower activation magnitudes in both the left (t66.37=-
3.60, p<0.001, corrected) and right (t74.28=-4.23, p<0.001, corrected) CSN than controls.
Patients also had significantly lower activation magnitudes in the mirrored left CSN in
the right hemisphere (t67.17=-2.92, p=0.01, corrected). Patient and control activation
magnitudes in out-of-network homologues of the left CSN in the right hemisphere did not
significantly differ (t60.15=0.47, p=1.0, corrected).
The difference between activation in the left CSN and the right CSN was not
significantly larger in controls than in patients (t81.89=0.38, p=1.0, corrected). Controls
showed a significantly larger difference in activation between the left hemispheric CSN
and the mirrored left CSN in the right hemisphere (t83.82=3.99, p<0.001, corrected).
Controls also showed a significantly larger difference between activation magnitudes in
the right CSN and out-of-network homologues of the left CSN in the right hemisphere
(t82.15=6.26, p<0.001, corrected).
Lesion volume effects on activation magnitudes
Activation magnitudes in both the left CSN (r=-0.24, p=0.35, corrected) and right
CSN (r=-0.31, p=0.18, corrected) showed non-significant negative relationships to lesion
volume. A weaker negative relationship was observed between lesion volume and
activation magnitudes in the mirrored left CSN in the right hemisphere (r=-0.12, p=0.90,
corrected). Lesion volume effects showed a near-zero correlation with activation
magnitudes in out-of-network homologues of the left CSN in the right hemisphere (r=-
0.01, p=0.97, corrected). These results are shown in Figure 4D.
The results from the whole-brain regression analysis assessing the effects of
lesion volume on activation in the right hemisphere are shown in Figure 5A. Cluster
statistics and peak locations are displayed in Table 4. All of the regions showing positive
relationships with lesion volume fell outside of the canonical semantic network in right
frontal cortex (Figure 5B). Notably, with the exception of the right middle frontal gyrus
(MFG), these regions did not correspond to out-of-network homologues of the left CSN.
In contrast, several of the regions showing negative relationships between activation and
lesion volume overlapped with the CSN (Figure 5B).
To assess whether there was a benefit of activating out-of-network right-
hemisphere regions that depended on the extent of left hemispheric damage, we
performed additional post-hoc regression analyses to assess whether the lesion size might
moderate the effect of activity in each right hemispheric cluster showing positive
correlations to total lesion volume (hot clusters in Figure 5A). This revealed significant
interactions between lesion volume and right SMA activation for the fluency (t39=2.97,
FDRp = 0.04) and naming (t39=2.82, FDRp = 0.04) measures (Figure 5C), and between
lesion volume and right IFG pars opercularis activation for the fluency (t39=2.50, FDRp =
0.07) measure. For each of the observed interaction effects, higher levels of activation in
the right hemispheric cluster was associated with lower language test scores for patients
with smaller lesions, but with higher language test scores for patients with larger lesions
(Figure 5C). Note that the high and low values shown in Figure 5C are calculated for +1
and -1 standard deviations from the mean of each variable.
Discussion
Functional neuroimaging studies of post-stroke aphasia are foundational to
theories of language recovery after stroke, but this theoretical foundation is primarily
composed of small studies. To address this shortcoming and assess the validity of
predictions based on current theories of recovery, we tested what we consider to be their
primary prediction – that successful long-term language recovery depends on the
preservation and/or restoration of language processing in canonical language networks.
Importantly, we tested this prediction in one of the largest samples of post-stroke aphasia
patients studied to date in the functional neuroimaging literature, and utilized statistical
controls to reduce the potential for confounding effects related to lesion volume on
imaging-behavior relationships. We also attempted to bring clarity to contested issues in
the literature – how language task-driven activity in the unaffected right hemisphere
relates to left hemispheric damage and function, and whether the recruitment of right
hemispheric regions by patients with extensive left hemispheric damage supports residual
language function. We discuss our results and their relation to the broader literature in the
following sections.
Canonical language network contributions to post-stroke language function
Drawing from the broader functional neuroimaging literature of language
recovery after stroke, we expected that activation of the CSN for language processing
would positively predict language functions in chronic patients. The results of both our
ROI-based and whole-brain analyses matched this expectation, and indicate that
activation of the CSN is a moderate-to-strong predictor of language function in chronic
patients (Figures 2 and 3). While activation magnitudes in this network only predicted
performance on the in-scanner task for controls, they predicted performance on all three
language measures for patients (Figure 2D). Further, our whole-brain analyses of the
patient data revealed that the positive relationships between regional task-driven
activation magnitudes and language task performance were primarily localized within or
adjacent to the CSN identified in controls (Figure 3). Thus, our results corroborate the
general implications of previous functional neuroimaging studies of language outcomes
in patients with LMCA stroke [Fridriksson et al., 2010; Fridriksson et al., 2012; Heiss et
al., 1999; Karbe et al., 1998; van Oers et al., 2010; Rosen et al., 2000; Saur et al., 2006;
Szaflarski et al., 2013]. Importantly, this result supports the emphasis of current theory on
the preservation and/or restoration of function in canonical language networks as a key
factor that enables successful long-term language recovery after stroke.
Notably, our results indicate that despite the fact that the auditory semantic
decision task lacks an expressive language component, the level of CSN activation
evoked was a good predictor of expressive language capacity in our patient sample
(Figure 2D/Figure 3). As noted in the introduction, previous studies suggest that
relationships between activation evoked by auditory (i.e. receptive) language tasks and
language function in chronic stroke patients may not be specific to receptive language
functions [e.g. Saur et al., 2006; Szaflarski et al., 2013]. Our results suggest that
activation in the CSN evoked by auditory semantic decisions may provide an index of
general language network preservation/recovery.
This interpretation is consistent with evidence that the abrupt and catastrophic
disruption of neural function by stroke leads to widespread disruptions of communication
and regulation in distributed brain networks [Baldassarre et al., 2016; Geranmayeh et al.,
2016; He et al., 2007; Ovadia-Caro et al., 2013; Siegel et al., 2016], and that the
preservation/restoration of typical function in canonical language networks is indicative
of successful language recovery [Heiss et al., 1999; Saur et al., 2006]. Indeed, post-stroke
deficits in complex cognitive functions [Siegel et al., 2016] that include language
[Geranmayeh et al., 2016] and attention [Baldassarre et al., 2016; He et al., 2007] may be
conceptualized as behavioral manifestations of dysfunction in large-scale functional brain
networks. Along these lines, optimal functional recovery after stroke may depend on the
preservation/restoration of functional dynamics that most strongly resembles those
observed in the pre-stroke brain [Carter et al., 2012].
Thus, the magnitude of task-evoked responses in the distributed CSN may index
the degree to which pre-stroke functional dynamics are preserved/restored in patients
with chronic post-stroke aphasia. While discussions of how distributed networks
contribute to recovery from stroke often focus on measures of intrinsic (i.e. resting state)
network function [e.g. Carter et al., 2012], there is substantial evidence that task-evoked
brain networks are strongly influenced by the intrinsic network architecture [Binder et al.,
1999; Binder, 2012; Cole et al., 2014; Dosenbach et al., 2006; Dosenbach et al., 2007;
Fox et al., 2005; Fox et al., 2006; Muhle-Karbe et al., 2015; Xu et al., 2016], and thus it is
possible that the presence of strong activation in the CSN during auditory semantic
decisions may be indicative of well-preserved and/or successfully restored intrinsic
network dynamics. We must emphasize, however, that this explanation is tentative and
must be confirmed by future studies that are capable of relating task-evoked responses in
canonical language networks to intrinsic network dynamics in patients with chronic post-
stroke aphasia.
Co-activation of left and right hemispheric networks and right hemispheric
effects of lesion volume
Based on current models of language recovery after stroke [Hamilton et al., 2011;
Heiss and Thiel, 2006; Saur et al., 2006; Saur and Hartwigsen, 2012], we expected that
patients with lower levels of activation in the left CSN would show higher levels of
activation in the right CSN. Assuming that this would reflect compensatory up-regulation
of the right CSN following extensive left hemispheric damage, we also expected that
activation in the right CSN would be increased in patients with larger left hemispheric
lesion volumes. However, our results did not bear out this prediction. Rather, we found
that for both patients and controls, activation magnitudes in the right CSN were strongly
positively correlated with activation magnitudes in the left CSN (Figure 4B). Further, in
patients, activation magnitudes in both the left and right CSNs showed negative but non-
significant correlations with lesion volume (Figure 4C). This suggests that the right CSN
and left CSN form a coherent functional unit that is activated by semantic processing.
Indeed, a basic role of certain right hemispheric regions in language processing is
suggested by findings that right hemispheric activation during semantic decision
[Donnelly et al., 2011], semantic comprehension [van Ettinger-Veenstra et al., 2010],
sentence completion [van Ettinger-Veenstra et al., 2012], and word fluency tasks [van
Ettinger-Veenstra et al., 2012] correlates with various measures of language function in
healthy individuals.
Notably, a recent ALE meta-analysis of 12 functional neuroimaging studies of
aphasic patients (total n=105) and healthy controls (total n=129) suggests that specific
regions in both hemispheres are consistently recruited by aphasic patients across different
language task paradigms [Turkeltaub et al., 2011]. Turkeltaub and colleagues [2011] thus
proposed that activity in most right hemispheric regions, with the notable exception of the
IFG pars triangularis, likely reflects either compensatory recruitment or co-activation
with homotopic areas in the left hemisphere. Our finding that activation magnitudes in
the left and right CSN were strongly correlated suggests that at least with regard to the
CSN, activity in right hemispheric portions of canonical networks likely reflects co-
activation rather than compensation for left hemispheric damage. The finding that neither
patients nor controls showed significant differences between the left and right CSN
supports this conclusion.
However, we also found that both patients and controls showed significantly
higher activation in the left CSN than in homologous areas in the right hemisphere, and
this effect was most pronounced for homologous areas that were not associated with the
right CSN (Figure 5C). Further, both patients and controls showed significantly higher
levels of activity in the right CSN compared to these out-of-network left CSN
homologues (Figure 5C). This suggests that co-activations are spatially constrained to the
subset of right hemispheric regions that typically activate as part of the CSN. While the
correlation between activation magnitudes in the left CSN and out-of-network right
hemispheric homologues was substantially weaker than the correlation between
activation magnitudes in the left CSN and right CSN in patients (Figure 5B), controls
showed essentially no relationship between activation magnitudes in the left CSN and
out-of-network right hemispheric homologues (Figure 5B). This suggests that inter-
hemispheric co-activations may be less spatially constrained in stroke patients than in
healthy controls. Speculatively, this may reflect weakened inter-hemispheric inhibition,
as this has been proposed as a potential source of increased activations in right
hemispheric homologues of left hemispheric language areas in patients with post-stroke
aphasia [Hamilton et al., 2011; Heiss and Thiel, 2006].
Nonetheless, our data support a functional distinction between right hemispheric
regions that show robust co-activation with distributed language networks and out-of-
network right hemispheric homologues of left hemispheric language areas, although
conclusions regarding the source of this distinction cannot be drawn from this study.
Future studies using task-based (i.e. effective) connectivity metrics to assess how
activation in these regions relate to interactions between left and right hemispheric
networks are necessary to address such questions.
Right hemispheric compensation in patients with extensive left hemispheric
damage
Recent evidence from structural MRI studies also suggests a compensatory role of
certain right hemispheric regions in supporting language function in chronic patients. For
example, grey matter volume in right dorsal stream temporo-parietal areas is increased in
chronic post-stroke aphasia patients relative to both healthy controls and chronic left
hemispheric stroke patients without aphasia [Xing et al., 2015]. Importantly, higher grey
matter volume in these regions is associated with better residual language functions in
chronic post-stroke aphasia patients [Xing et al., 2015]. Similarly, increased fractional
anisotropy (FA) has been observed in the right IFG pars opercularis for chronic post-
stroke aphasia patients relative to healthy controls, and higher FA in this region is
associated with better speech production abilities for patients [Pani et al., 2016]. Our
findings add to this literature by showing that functional activation in specific right
frontal regions during language task performance is associated with better expressive
language abilities in chronic post-stroke aphasia patients with extensive left hemispheric
damage.
Notably, larger lesions were associated with higher levels of activity in several
right frontal areas, including the right inferior frontal gyrus pars opercularis and
supplementary motor area, both of which were not part of the CSN (Figure 5A/B).
Activation magnitudes in the right IFG pars opercularis and right SMA showed
relationships to out-of-scanner language measures that differed in direction between
patients with larger vs. smaller lesions (Figure 5C). One explanation as to why stronger
activations of these regions were associated with poorer performance for patients with
smaller lesions is that activity in these regions interferes with the functions of task-
relevant areas when canonical regions are intact, but supports residual language function
when canonical regions are damaged. Alternately, stronger activations of these regions
for patients with smaller lesions may be reflective of focal damage to critical left
hemispheric areas that impede the restoration of canonical network function during
recovery, and result in chronic language deficits comparable to those observed in patients
with more widespread left hemispheric damage. Based on evidence that both the right
IFG and right SMA support language processing during early recovery, but show reduced
involvement in later stages when canonical networks are recovered [Saur et al., 2006], we
consider this to be most likely. If this is the case, the identification of such critical regions
by future studies could provide important insights into the factors contributing to poor
long-term language recovery after stroke.
Along these lines, we note that previous studies suggest that activation in the right
IFG pars opercularis is increased in patients with lesions affecting the left IFG [Blank et
al., 2003; Turkeltaub et al., 2011]. In addition, maintained right frontal activation has
been previously reported in patients with left posterior temporal lesions that recovered
less successfully than patients with lesions affecting the left frontal cortices or the left
basal ganglia [Heiss et al., 1999]. Given that (1) our results indicate that larger lesions
correlated with larger activation magnitudes in right frontal cortices, and (2) activation
magnitudes in right IFG/SMA were associated with better expressive language functions
in patients with larger lesions, we speculate that the observed benefit of activating the
right IFG/right SMA in patients with larger lesions may reflect beneficial compensation
that supports expressive language functions (e.g. top-down selection and sequencing).
Based on the findings of Heiss and colleagues [1999], we further speculate that patients
with lesions affecting left posterior temporal regions may also recruit these regions, but
that these patients may show reduced benefit from their involvement, presumably
resulting from an inability to restore function in other distributed portions of canonical
networks. We stress that this must be regarded as speculative, as the current study did not
directly investigate relationships between lesion location and right hemispheric
activation. Rather, the methods used in this study were primarily intended to address the
question of how lesion extent relates to right hemispheric activation. Because the
methodology of this study was not ideal for addressing questions about how lesion
location relates to regional activation, future studies using methodology better suited to
addressing this question are necessary before strong conclusions can be drawn about how
the site of damage relates to right hemispheric activations.
A common network supporting comprehension, naming, and fluency
Activity in a subset of regions (shown in Figure 3B) correlated positively with
performance on all of the language measures, suggesting that this set of regions may
support processes critical for language processing after stroke. This set of regions
included the left dorsal dSFG, the left AG/superior lateral occipital gyrus, bilateral
precuneus and PCC, bilateral PHG and fusiform gyri, and the right temporal pole. These
regions are commonly associated with the default mode network [Dosenbach et al., 2007;
Fox et al., 2005; Raichle et al., 2001; Raichle and Snyder, 2007; Vincent et al., 2008].
Based on observations that this network is more active during both semantic processing
tasks and task-free resting states relative to tasks that do not involve semantic processing
(and generally more active during semantic processing than during rest), this network has
been proposed to form a core "conceptual network" that is involved in both explicit
semantic processing and ongoing manipulations of memory/conceptual representations in
the absence of explicit tasks [Binder et al., 1999; Binder et al., 2009; Leech and Sharp,
2014]. Notably, the PCC, precuneus, and AG are functionally connected to the PHG and
co-activate with parahippocampal areas during memory retrieval [Sestieri et al., 2011].
The AG [Uddin et al., 2010] and PCC [Greicius et al., 2009] also possess direct structural
connections to medial temporal lobe structures such as the hippocampus that play a
critical role in memory. Speculatively, since each of the language measures utilized in
this study involved a memory component (e.g. recalling facts about animals for the in-
scanner semantic decision task, recalling words that begin with a given letter or fit a
given category for the COWAT/SFT, and recalling names of visual objects for the BNT),
it is possible that activity in these regions reflects their role in supporting memory access
and manipulation. In addition, recent intrinsic functional connectivity evidence suggests
that regions such as the left dSFG and left AG may function as hubs that support ongoing
interactions among canonical language modules (i.e. perisylvian language areas),
executive control modules (i.e. fronto-parietal network), and memory/simulation modules
(i.e. default-mode network) [Xu et al., 2016].
Additionally, the right anterior temporal lobe has previously been reported to
support auditory verbal comprehension in patients with damage to left posterior temporal
areas, and may represent an independent module capable of auditory language processing
when left posterior temporal structures are compromised [Crinion and Price, 2005].
When connectivity between left and right anterior temporal areas is preserved, left frontal
access to inputs processed by the right anterior temporal lobe may provide a
compensatory mechanism for achieving top-down modulations (e.g. selection or
integration) of semantic content via inter-hemispheric pathways [Warren et al., 2009].
In addition, medial posterior default mode regions frequently interact with elements of
other functional networks [de Pasquale et al., 2012] and possess diverse structural
connections to language-relevant cortical (i.e. inferior parietal, superior temporal, anterior
cingulate, and supplementary/pre-motor cortices) and subcortical (including the striatum
and multiple thalamic nuclei) areas whose connections may be disrupted by LMCA
stroke [Cavanna and Trimble, 2006; Greicius et al., 2009]. Thus, this set of regions could
potentially act to relay information among surviving language areas when primary
pathways are affected by stroke, providing a potential "back-up" interface among
bilateral frontal, temporal, and parietal areas to support manipulations of memory
representations and/or speech inputs when canonical cortico-cortical pathways (e.g. the
arcuate fasciculus) are no longer viable. While plausible, and consistent with a proposed
role of these regions as cross-network "connectors" [Xu et al., 2016], this should be
regarded as speculation, as delineating the precise roles these regions play in supporting
language processing after stroke is beyond the scope of this study.
Lesion volume effects in functional neuroimaging of aphasia recovery
Because we used statistical controls to account for lesion volume effects, our
results are not likely to be driven by differences in lesion size or overall damage to the
network among patients. Despite the potential for lesion volume differences to introduce
bias, we note that our results were nearly identical (albeit with somewhat larger effects)
when we did not use controls for lesion volume. This may reflect the relatively low
correlation between lesion volume and canonical network activation in this sample.
Nonetheless, while we contend that the use of statistical controls is important to reduce
the potential for biases related to differences in lesion extent, it is possible that lesion
volume is too coarse of a measure to accomplish this goal. An alternative approach may
be to utilize more spatially sensitive measures, such as the percent of damage sustained
by different anatomical areas [e.g. Xing et al., 2015]. Determining the optimal solution to
account for lesion biases in functional neuroimaging studies of stroke patients may
ultimately require dedicated studies using simulated data to understand the severity of
This study has several limitations that must be acknowledged. First, as with many
biases likely to be introduced by varying levels of brain damage and to allow for an
objective comparison of different approaches for mitigating such biases.
Limitations
functional imaging studies, it is not possible to definitively conclude that activation in
any given region is necessary to perform the task. Further, even if this were possible, it
would not be possible to conclude from these data which aspect of the task a hypothetical
"task-critical" region might support. Thus, while our results are interpreted as indicating
that canonical network regions likely support residual language abilities after left
hemispheric stroke, and that specific out-of-network regions in the right hemisphere may
compensate for extensive left hemispheric damage, conclusions about the nature of the
contributions of these regions to task performance cannot be considered definitive.
Second, while the current study suggests that extensive left hemispheric damage is
associated with the recruitment of select right frontal regions to accomplish semantic
processing, it cannot address important questions such as how damage to specific regions
in the left hemisphere influences semantic processing or affects out-of-network responses
in the right hemisphere; future research is necessary to address these questions. Third,
altered vascular dynamics in chronic stroke patients have the potential to lead to changes
in neurovascular coupling and influence the BOLD response, and this may lead to
reductions in measured activation magnitudes for stroke patients relative to controls [e.g.
D'Esposito et al., 2003; Veldsman et al., 2015]. It is necessary to consider the potential
for such effects when interpreting the results of this and other studies using functional
MRI to study clinical populations such as stroke patients. Fourth, this study was not
intended to address the question of how lesion location relates to right hemispheric
activation. We stress that strong conclusions regarding this relationship cannot be drawn
from the results of this study. Finally, the current study does not allow for conclusions
regarding the role of functional interactions among regions associated with the CSN,
between left and right hemispheric portions of the CSN, or between the CSN and other
networks in supporting language recovery after stroke, as it is limited to assessments of
In conclusion, this study aimed to test current theories of language recovery after
co-activation during task performance. Future studies using functional and/or effective
connectivity measures are necessary to enable such conclusions.
Conclusions
stroke and bring clarity to discrepancies in the literature. Our results confirm the
prediction that fMRI activation in the canonical semantic network predicts performance
on multiple measures of language function in chronic patients independently of lesion
volume effects, and reveal a core set of default mode regions that likely play a key role in
supporting basic aspects of residual language functions in the years after stroke. Our
results also suggest that left and right hemispheric portions of the semantic network co-
activate to accomplish language processing after stroke, contrary to the notion that
activity in right hemispheric homologues of the left hemispheric network is up-regulated
to compensate for damage to the left hemispheric network. In contrast, our results suggest
that patients with the most damage to the left hemispheric network recruit specific right
hemispheric areas that have been previously shown to play a primary role in supporting
early recovery. We suspect that these regions may support top-down selection or motor
aspects of residual language function during the chronic stage in these patients. The
findings described here emphasize the importance of canonical language networks for
supporting long-term language recovery after stroke, clarify the contributions of in- and
out- of network right hemispheric areas to language recovery and their relationships to
left hemisphere damage, and provide a more stable foundation for future studies of
language recovery after stroke.
Acknowledgements
Amber Martin
Christi Banks
NIH R01 HD068488
NIH R01 NS048281
References
Aickin M, Gensler H (1996): Adjusting for multiple testing when reporting research
results: The Bonferroni vs Holm methods. Am J Public Health 86:726–728.
Allendorfer J, Kissela B, Holland S, Szaflarski J (2012): Different patterns of language
activation in post-stroke aphasia are detected by overt and covert versions of the
verb generation fMRI task. Med Sci Monit Monit 18.
Benjamini Y, Hochberg Y (1995): Controlling the False Discovery Rate: A Practical and
Powerful Approach to Multiple Testing. J R Stat Soc Ser B 57:289–300.
Baldassarre A, Ramsey L, Rengachary J, Zinn K, Siegel JS, Metcalf N V., Strube MJ,
Snyder AZ, Corbetta M, Shulman GL (2016): Dissociated functional connectivity
profiles for motor and attention deficits in acute right-hemisphere stroke. Brain
139:2024–2038.
Belin P, Van Eeckhout P, Zilbovicius M, Remy P, François C, Guillaume S, Chain F,
Rancurel G, Samson Y (1996): Recovery from nonfluent aphasia after melodic
intonation therapy: a PET study. Neurology 47:1504–1511.
Binder JR, Frost JA, Hammeke TA, Cox RW, Rao SM, Prieto T (1997): Human brain
language areas identified by functional magnetic resonance imaging. J Neurosci
17:353–362.
Binder JR (2012): Task-induced deactivation and the "resting" state. Neuroimage
62:1086–1091.
Binder JR, Desai RH, Graves WW, Conant LL (2009): Where is the semantic system? A
critical review and meta-analysis of 120 functional neuroimaging studies. Cereb
Cortex 19:2767–2796.
Binder JR, Swanson SJ, Hammeke TA, Sabsevitz DS (2008): A comparison of five fMRI
protocols for mapping speech comprehension systems. Epilepsia 49:1980–1997.
Binder J, Frost J, Hammeke T, Bellgowan PSF, Rao SM, Cox RW (1999): Conceptual
processing during the conscious resting state: A functional MRI study. J Cogn
Neurosci 11:80–93.
Blank SC, Bird H, Turkheimer F, Wise RJS (2003): Speech production after stroke: The
role of the right pars opercularis. Ann Neurol 54:310–320.
Butler RA, Lambon Ralph MA, Woollams AM (2014): Capturing multidimensionality in
stroke aphasia: mapping principal behavioural components to neural structures.
Brain:3248–3266.
Carter AR, Shulman GL, Corbetta M (2012): Why use a connectivity-based approach to
study stroke and recovery of function? Neuroimage 62:2271–2280.
Cavanna AE, Trimble MR (2006): The precuneus: a review of its functional anatomy and
behavioural correlates. Brain 129:564–83.
Charidimou A, Kasselimis D, Varkanitsa M, Selai C, Potagas C, Evdokimidis I (2014):
Why is it difficult to predict language impairment and outcome in patients with
aphasia after stroke? J Clin Neurol 10:75–83.
Cheng B, Forkert ND, Zavaglia M, Hilgetag CC, Golsari A, Siemonsen S, Fiehler J,
Pedraza S, Puig J, Cho T-H, Alawneh J, Baron J-C, Ostergaard L, Gerloff C,
Thomalla G (2014): Influence of Stroke Infarct Location on Functional Outcome
Measured by the Modified Rankin Scale. Stroke 45:1695–1702.
Cole MW, Bassett DS, Power JD, Braver TS, Petersen SE (2014): Intrinsic and task-
evoked network architectures of the human brain. Neuron 83:238–251.
D'Esposito M, Deouell LY, Gazzaley A (2003): Alterations in the BOLD fMRI signal
with ageing and disease: a challenge for neuroimaging. Nat Rev Neurosci 4:863–
872.
Donnelly KM, Allendorfer JB, Szaflarski JP (2011): Right hemispheric participation in
semantic decision improves performance. Brain Res 1419:105–116.
Dosenbach NUF, Visscher KM, Palmer ED, Miezin FM, Wenger KK, Kang HC,
Burgund ED, Grimes AL, Schlaggar BL, Petersen SE (2006): A Core System for the
Implementation of Task Sets. Citeulike:JOUR. Neuron 50:799–812.
Dosenbach NU, Fair DA, Miezin FM, Cohen AL, Wenger KK, Dosenbach RAT, Fox
MD, Snyder AZ, Vincent JL, Raichle ME, Schlaggar BL, Petersen SE (2007):
Distinct brain networks for adaptive and stable task control in humans. Proc Natl
Acad Sci U S A 104:11073–8.
Eaton KP, Szaflarski JP, Altaye M, Ball AL, Kissela BM, Banks C, Holland SK (2008):
Reliability of fMRI for studies of language in post-stroke aphasia subjects.
Neuroimage 41:311–22.
Eklund A, Nichols TE, Knutsson H (2016): Cluster failure: Why fMRI inferences for
spatial extent have inflated false-positive rates. Proc Natl Acad Sci 113:201602413.
van Ettinger-Veenstra HM, Ragnehed M, Hällgren M, Karlsson T, Landtblom A-M,
Lundberg P, Engström M (2010): Right-hemispheric brain activation correlates to
language performance. Neuroimage 49:3481–8.
van Ettinger-Veenstra H, Ragnehed M, McAllister A, Lundberg P, Engström M (2012):
Right-hemispheric cortical contributions to language ability in healthy adults. Brain
Lang 120:395–400.
Fisher R (1921): On the probable error of a coefficient of correlation deduced from a
small sample. Metron 1:3–32.
Fox MD, Corbetta M, Snyder AZ, Vincent JL, Raichle ME (2006): Spontaneous neuronal
activity distinguishes human dorsal and ventral attention systems. Proc Natl Acad
Sci U S A 103:10046–51.
Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME (2005): The
human brain is intrinsically organized into dynamic, anticorrelated functional
networks. Proc Natl Acad Sci U S A 102:9673–8.
Fridriksson J, Bonilha L, Baker JM, Moser D, Rorden C (2010): Activity in preserved left
hemisphere regions predicts anomia severity in aphasia. Cereb Cortex 20:1013–9.
Fridriksson J, Richardson JD, Fillmore P, Cai B (2012): Left hemisphere plasticity and
aphasia recovery. Neuroimage 60:854–63.
Friston KJ, Holmes AP, Worsley KJ, Poline J-P, Frith CD, Frackowiak RSJ (1995):
Statistical parametric maps in functional imaging: A general linear approach. Hum
Brain Mapp 2:189–210.
Geranmayeh F, Leech R, Wise RJS (2016): Network dysfunction predicts speech
production after left hemisphere stroke. Neurology 86:1296–1305.
Griffis JC, Allendorfer JB, Szaflarski JP (2016a): Voxel-based Gaussian naïve Bayes
classification of ischemic stroke lesions in individual T1-weighted MRI scans. J
Neurosci Methods 257:97–108.
Griffis JC, Nenert R, Allendorfer JB, Szaflarski JP (2016b): Interhemispheric Plasticity
following Intermittent Theta Burst Stimulation in Chronic Poststroke Aphasia.
Neural Plast 2016:20–23.
Greicius MD, Supekar K, Menon V, Dougherty RF (2009): Resting-state functional
connectivity reflects structural connectivity in the default mode network. Cereb
Cortex 19:72–8.
Hamilton RH, Chrysikou EG, Coslett B (2011): Mechanisms of aphasia recovery after
stroke and the role of noninvasive brain stimulation. Brain Lang 118:40–50.
Hartwigsen G, Saur D, Price CJ, Ulmer S, Baumgaertner A, Siebner HR (2013):
Perturbation of the left inferior frontal gyrus triggers adaptive plasticity in the right
homologous area during speech production. Proc Natl Acad Sci U S A 110:16402–7.
He BJ, Snyder AZ, Vincent JL, Epstein A, Shulman GL, Corbetta M (2007): Breakdown
of functional connectivity in frontoparietal networks underlies behavioral deficits in
spatial neglect. Neuron 53:905–18.
Heiss WD, Kessler J, Thiel A, Ghaemi M, Karbe H (1999): Differential capacity of left
and right hemispheric areas for compensation of poststroke aphasia. Ann Neurol
45:430–438.
Heiss W, Thiel A (2006): A proposed regional hierarchy in recovery of post-stroke
aphasia. Brain Lang 98:118–23.
Jaccard J, Wan CK, Turrisi R (1990): The Detection and Interpretation of Interaction
Effects Between Continuous Variables in Multiple Regression. Multivariate Behav
Res 25:467–478.
Kaplan E, Goodglass H, Weintraub S, Segal O, van Loon-Vervoorn A (2001): Boston
naming test. Pro-ed.
Karbe H, Thiel A, Weber-luxenburger G, Kessler J, Heiss W (1998): Brain Plasticity in
Poststroke Aphasia : What Is the Contribution of the Right Hemisphere? Brain Lang
230:215–230.
Karnath HO, Berger MF, Kuker W, Rorden C (2004): The anatomy of spatial neglect
based on voxelwise statistical analysis: A study of 140 patients. Cereb Cortex
14:1164–1172.
Kertesz A, Harlock W, Coates R (1979): Computer tomographic localization, lesion size,
and prognosis in aphasia and nonverbal impairment. Brain Lang 8:34–50.
Kim KK, Karunanayaka P, Privitera MD, Holland SK, Szaflarski JP (2011): Semantic
association investigated with functional MRI and independent component analysis.
Epilepsy Behav 20:613–622.
Kozora E, Cullum CM (1995): Generative naming in normal aging: Total output and
qualitative changes using phonemic and semantic constraints. Clin Neuropsychol
9:313–320.
Kümmerer D, Hartwigsen G, Kellmeyer P, Glauche V, Mader I, Klöppel S, Suchan J,
Karnath HO, Weiller C, Saur D (2013): Damage to ventral and dorsal language
pathways in acute aphasia. Brain 136:619–629.
Lazar RM, Speizer AE, Festa JR, Krakauer JW, Marshall RS (2008): Variability in
language recovery after first-time stroke. J Neurol Neurosurgery, Psychiatry
79:530–534.
Leech R, Sharp DJ (2014): The role of the posterior cingulate cortex in cognition and
disease. Brain 137:12–32.
Lezak MD, Howieson DB, Loring DW, Hannay JH, Fischer JS (1995):
Neuropsychological assessment (3) Oxford University Press. New York.
Maas MB, Lev MH, Ay H, Singhal AB, Greer DM, Smith WS, Harris GJ, Halpern EF,
Koroshetz WJ, Furie KL (2012): The Prognosis for Aphasia in Stroke. J Stroke
Cerebrovasc Dis 21:350–357.
Mattioli F, Ambrosi C, Mascaro L, Scarpazza C, Pasquali P, Frugoni M, Magoni M,
Biagi L, Gasparotti R (2014): Early aphasia rehabilitation is associated with
functional reactivation of the left inferior frontal gyrus: a pilot study. Stroke 45:545–
52.
Mazaika PK, Whitfield S, Cooper JC (2005): Detection and repair of transient artifacts in
fMRI data. Neuroimage 26:S36.
Meinzer M, Beeson PM, Cappa S, Crinion J, Kiran S, Saur D, Parrish T, Crosson B,
Thompson CK (2013): Neuroimaging in aphasia treatment research: consensus and
practical guidelines for data analysis. Neuroimage 73:215–24.
Meltzer JA, Wagage S, Ryder J, Solomon B, Braun AR (2013): Adaptive significance of
right hemisphere activation in aphasic language comprehension. Neuropsychologia
51:1248–1259.
Muhle-Karbe PS, Derrfuss J, Lynn MT, Neubert FX, Fox PT, Brass M, Eickhoff SB
(2015): Co-Activation-Based Parcellation of the Lateral Prefrontal Cortex
Delineates the Inferior Frontal Junction Area. Cereb Cortex:1–17.
van Oers CMM, Vink M, van Zandvoort MJE, van der Worp HB, de Haan EHF,
Kappelle LJ, Ramsey NF, Dijkhuizen RM (2010): Contribution of the left and right
inferior frontal gyrus in recovery from aphasia. A functional MRI study in stroke
patients with preserved hemodynamic responsiveness. Neuroimage 49:885–93.
Oldfield RC (1971): the Assessment and Analysis of Handedness: the Edinburgh
Inventory. Neuropsychologia 9:97–113.
Ovadia-Caro S, Villringer K, Fiebach J, Jungehulsing GJ, van der Meer E, Margulies DS,
Villringer A (2013): Longitudinal effects of lesions on functional networks after
stroke. J Cereb Blood Flow Metab 33:1279–85.
Pani E, Zheng X, Wang J, Norton A, Schlaug G (2016): Right hemisphere structures
predict poststroke speech fluency. Neurology 86:1574–1581.
de Pasquale F, Della Penna S, Snyder AZ, Marzetti L, Pizzella V, Romani GL, Corbetta
M (2012): A cortical core for dynamic integration of functional networks in the
resting human brain. Neuron 74:753–64.
Pedersen P (1995): Aphasia in acute stroke: incidence, determinants, and recovery. Ann
Neurol 38:659–666.
Piervincenzi C, Petrilli A, Marini A, Caulo M, Committeri G, Sestieri C (2016):
Multimodal assessment of hemispheric lateralization for language and its relevance
for behavior. Neuroimage.
Poldrack RA (2012): The future of fMRI in cognitive neuroscience. Neuroimage
62:1216–20.
Raboyeau G, De Boissezon X, Marie N, Balduyck S, Puel M, Bézy C, Démonet JF,
Cardebat D (2008): Right hemisphere activation in recovery from aphasia: Lesion
effect or function recruitment? Neurology 70:290–298.
Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard D a, Shulman GL (2001):
A default mode of brain function. Proc Natl Acad Sci U S A 98:676–82.
Raichle ME, Snyder AZ (2007): A default mode of brain function: a brief history of an
evolving idea. Neuroimage 37:1083–90; discussion 1097–9.
Ripollés P, Marco-Pallarés J, de Diego-Balaguer R, Miró J, Falip M, Juncadella M,
Rubio F, Rodriguez-Fornells A (2012): Analysis of automated methods for spatial
normalization of lesioned brains. Neuroimage 60:1296–306.
Rorden C, Karnath H-O (2004): Using human brain lesions to infer function: a relic from
a past era in the fMRI age? Nat Rev Neurosci 5:813–819.
Rosen HJ, Petersen SE, Linenweber MR, Snyder AZ, White DA, Chapman L,
Dromerick AW, Fiez JA, Corbetta MD (2000): Neural correlates of recovery from
aphasia after damage to left inferior frontal cortex. Neurology 55:1883–1894.
Saur D, Hartwigsen G (2012): Neurobiology of language recovery after stroke: lessons
from neuroimaging studies. Arch Phys Med Rehabil 93:S15-25.
Saur D, Lange R, Baumgaertner A, Schraknepper V, Willmes K, Rijntjes M, Weiller C
(2006): Dynamics of language reorganization after stroke. Brain 129:1371–84.
Schwartz MF, Kimberg DY, Walker GM, Faseyitan O, Brecher A, Dell GS, Coslett HB
(2009): Anterior temporal involvement in semantic word retrieval: voxel-based
lesion-symptom mapping evidence from aphasia. Brain 132:3411–27.
Seghier ML, Ramlackhansingh A, Crinion J, Leff AP, Price CJ (2008): Lesion
identification using unified segmentation-normalisation models and fuzzy clustering.
Neuroimage 41:1253–1266.
Sestieri C, Corbetta M, Romani GL, Shulman GL (2011): Episodic memory retrieval,
parietal cortex, and the default mode network: functional and topographic analyses.
J Neurosci 31:4407–20.
Siegel JS, Ramsey LE, Snyder AZ, Metcalf N V, Chacko RV, Weinberger K, Baldassarre
A, Hacker C, Shulman GL (2016): Common and specific disruptions of network
connectivity predict impairment in multiple behavioral domains after stroke. PNAS
I:1–10.
Sims JA, Kapse K, Glynn P, Sandberg C, Tripodis Y, Kiran S (2016): The relationships
between the amount of spared tissue, percent signal change, and accuracy in
semantic processing in aphasia. Neuropsychologia 84:113–126.
Slotnick SD, Moo LR, Segal JB, Hart J (2003): Distinct prefrontal cortex activity
associated with item memory and source memory for visual shapes. Cogn Brain Res
17:75–82.
Szaflarski JP, Binder JR, Possing ET, McKiernan KA, Ward BD, Hammeke TA (2002):
Language lateralization in left-handed and ambidextrous people: fMRI data.
Neurology 59:238–244.
Szaflarski JP, Allendorfer JB, Banks C, Vannest J, Holland SK (2013): Recovered vs.
not-recovered from post-stroke aphasia: the contributions from the dominant and
non-dominant hemispheres. Restor Neurol Neurosci 31:347–60.
Szaflarski JP, Holland SK, Jacola LM, Lindsell C, Privitera MD, Szaflarski M (2008):
Comprehensive presurgical functional MRI language evaluation in adult patients
with epilepsy. Epilepsy Behav 12:74–83.
Szaflarski JP, Vannest J, Wu SW, DiFrancesco MW, Banks C, Gilbert DL (2011):
Excitatory repetitive transcranial magnetic stimulation induces improvements in
chronic post-stroke aphasia. Med Sci Monit 17:CR132-9.
Thiel A, Schumacher B, Wienhard K, Gairing S, Kracht LW, Wagner R, Haupt WF,
Heiss W-D (2006): Direct demonstration of transcallosal disinhibition in language
networks. J Cereb Blood Flow Metab 26:1122–7.
Turkeltaub PE, Coslett HB, Thomas AL, Faseyitan O, Benson J, Norise C, Hamilton RH
(2012): The right hemisphere is not unitary in its role in aphasia recovery. Cortex
48:1179–1186.
Turkeltaub PE, Messing S, Norise C, Hamilton RH (2011): Are networks for residual
language function and recovery consistent across aphasic patients? Neurology
76:1726–34.
Uddin LQ, Supekar K, Amin H, Rykhlevskaia E, Nguyen DA, Greicius MD, Menon V
(2010): Dissociable connectivity within human angular gyrus and intraparietal
sulcus: evidence from functional and structural connectivity. Cereb Cortex 20:2636–
46.
Veldsman M, Cumming T, Brodtmann A (2015): Beyond BOLD: Optimizing functional
imaging in stroke populations. Hum Brain Mapp 36:1620–1636.
Vincent JL, Kahn I, Snyder AZ, Raichle ME, Buckner RL (2008): Evidence for a
frontoparietal control system revealed by intrinsic functional connectivity. J
Neurophysiol 100:3328–42.
Wager TD, Lindquist M, Kaplan L (2007): Meta-analysis of functional neuroimaging
data: Current and future directions. Soc Cogn Affect Neurosci 2:150–158.
Warren JE, Crinion JT, Lambon Ralph MA, Wise RJS (2009): Anterior temporal lobe
connectivity correlates with functional outcome after aphasic stroke. Brain
132:3428–42.
Wilke M, de Haan B, Juenger H, Karnath H-O (2011): Manual, semi-automated, and
automated delineation of chronic brain lesions: a comparison of methods.
Neuroimage 56:2038–46.
Winhuisen L, Thiel A, Schumacher B, Kessler J, Rudolf J, Haupt WF, Heiss WD (2005):
Role of the contralateral inferior frontal gyrus in recovery of language function in
poststroke aphasia: a combined repetitive transcranial magnetic stimulation and
positron emission tomography study. Stroke 36:1759–63.
Winhuisen L, Thiel A, Schumacher B, Kessler J, Rudolf J, Haupt WF, Heiss WD (2007):
The right inferior frontal gyrus and poststroke aphasia: a follow-up investigation.
Stroke 38:1286–92.
Xing S, Lacey EH, Skipper-Kallal LM, Jiang X, Harris-Love ML, Zeng J, Turkeltaub PE
(2015): Right hemisphere grey matter structure and language outcomes in chronic
left hemisphere stroke. Brain.
Xu Y, Lin Q, Han Z, He Y, Bi Y (2016): Intrinsic functional network architecture of
human semantic processing: Modules and hubs. Neuroimage 132:542–555.
Yarkoni T (2009): Big Correlations in Little Studies: Inflated fMRI Correlations Reflect
Low Statistical Power-Commentary on Vul et al. (2009). Perspect Psychol Sci
4:294–298.
Yarnell P, Monroe P, Sobel L (1976): Aphasia outcome in stroke: a clinical
neuroradiological correlation. Stroke 7:516–522.
Zhang Y, Kimberg D, Coslett H, Schwartz M, Wang Z (2014): Multivariate lesion-
symptom mapping using support vector regression. Hum Brain Mapp 35:997.
Tables
Table 1. Participant demographics
Group N Age
Patients 43 53 (15) 25 M 0.85 (0.43)
Controls 43 54( 14) 23 M 0.80 (0.41)
*Mean (SD) are shown for age/handedness; M-Male
Table 2. Cluster and peak statistics for control and patient SD activations
Sex Handedness Lesion volume (ml)
105.24 (76.29)
N/A
x
-4
2
Peak Location
y
z
Extent
t-value
46 36
44277 14.3308
-34 32
44277 10.8963
-8
44277 10.6975 -36 30
3981 11.8496 -40 -64 36
-60 32
1051
50
125
-12
66
2
-76 -28
3519
10
3519
-54 -52
8
3519
-66 -48
42
7349
40 56
-2
-48 28
7349
8
7349
16 54 30
253
-52
-8
-18
-44 -64 28
1085
-68 32
254
54
217
0
-32 30
-60 -32 -12
585
111
40
-24 -28
265
4
-46 14
Group
Controls L Superior Medial Gyrus
R Posterior Cingulate
L IFG (p. Orbitalis)
L Angular Gyrus
R Angular Gyrus
R Mid Orbital Gyrus
Patients R Cerebelum (VIII)
R Cerebelum (IX)
R Cerebelum (VII)
L Superior Medial Gyrus
L IFG (p. Triangularis)
R Superior Frontal Gyrus
L Middle Temporal Gyrus
L Middle Temporal Gyrus
R Angular Gyrus
Posterior Cingulate
L Inferior Temporal Gyrus
R Fusiform Gyrus
R Posterior Cingulate
*Note: Cluster peaks are provided for the top 3 peaks/cluster that are separated by
a minimum distance of 30mm.
Table 3. Cluster and peak statistics for correlations between patient SD activation
and language task performance while controlling for lesion volume.
y
8.1312
3.3213
5.1981
4.1282
3.8702
4.9037
4.7172
4.0744
4.409
4.4082
4.0334
3.9907
3.9528
3.7833
2.9714
Peak Location
t-value
x
z
Sign
Positive L Angular Gyrus
L Middle Temporal Gyrus
Cerebellar Vermis (VII)
Posterior Cingulate Gyrus
L Calcarine Gyrus
Dorsal Anterior Pons
L Middle Orbital Gyrus
L Middle Temporal Gyrus
L Inferior Temporal Gyrus
R Angular Gyrus
L Parahippocampal Gyrus
R Parahippocampal Gyrus
L Superior Frontal Gyrus
R IFG (p. Orbitalis)
R Precuneus
Extent
Semantic Decisions
1913
1913
5291
5291
5291
544
961
785
671
547
457
275
175
160
211
-42 -68 48
4.9167
-50 -46
8
2.5892
-2
-78 -18
4.8921
0
-32 32
4.4726
-94 12
4
4.3329
0
-12 -28
4.6492
-42 54
0
4.5543
-2
4.4425
-68 -32
-32
-6
-38
4.3857
-58 40
4.2227 54
-22 -28 -18
4.0258
3.9305 20
-32 -20
3.745
-20 56 32
3.7179 50 38
-18
-62 66
4
3.6912
7588
Positive R Superior Medial Gyrus
7588
L Middle Frontal Gyrus
7588
L Posterior-Medial Frontal
L Brainstem (Inf. Colliculus?) 6302
L Fusiform Gyrus
6302
R Cerebelum (IX)
6302
1227
L Angular Gyrus
L Posterior Cingulate
4805
R Cerebelum (VI)
4805
4805
L Calcarine Gyrus
R Fusiform Gyrus
232
R Medial Temporal Pole
1532
1532
R Middle Temporal Gyrus
1532
R Parahippocampal Gyrus
L IFG (p. Orbitalis)
587
Negative R Middle Temporal Gyrus
155
Naming
2
50 38
6.0189
-34 18 52
5.8247
-4
20 66
5.1085
-2
-30 -20
5.2447
-34 -18 -30
4.9057
8
-56 -38
4.8778
-48 -62 32
5.2081
-2
-46 18
4.6649
-80 -14
4.0471 28
6
-92
3.6055
0
4.616
34
-18 -34
4.5911 46
6
-36
-16 -20
3.7147 68
-30
-2
3.685
16
4.0197
-12
-40 24
-3.3541 48
-48
2
Negative L Postcentral Gyrus
R Rectal Gyrus
R Inferior Temporal Gyrus
R Temporal Pole
135
153
421
258
3.6875
2
24
-30
3.6576 56
-12 -30
3.6249 36 14
-24
-3.4808 -60 -10 30
Fluency
5204
5204
5204
1889
268
474
354
265
265
762
2516
2516
2516
202
139
54 46
2
5.5202
-34 16 52
4.494
14 74
6
3.4022
-46 -60 30
4.2896
48
2
-34
4.207
-40 -34 -14
3.9476
44 24
-14
3.921
-42
3.9092 16
-8
2.6356 16 14
-18
-46 -38
3.8878 14
3.8798
-66 64
2
3.5849
0
-36 32
-28 -56 66
2.4993
-42 26
3.8209
-16
3.6031
-18
-38
6
Positive R Superior Medial Gyrus
L Middle Frontal Gyrus
R Posterior-Medial Frontal
L Angular Gyrus
R Inferior Temporal Gyrus
L Fusiform Gyrus
R IFG (p. Orbitalis)
R Fusiform Gyrus
R Superior Orbital Gyrus
R Cerebelum (IX)
R Precuneus
Posterior Cingulate
L Superior Parietal Lobule
L IFG (p. Orbitalis)
L Fusiform Gyrus
Note: Cluster peaks are provided for the top 3 peaks/cluster that are separated by
a minimum distance of 30mm
Table 4. Cluster and peak statistics for right hemispheric correlations between SD
activation and lesion volume.
Sign
x
Extent
277
200
143
221
399
223
1137
1137
1137
160
Peak Location
Positive R IFG (p. Opercularis)
R Posterior-Medial Frontal
R Middle Frontal Gyrus
R Precentral Gyrus
Negative R Inferior Temporal Gyrus
R Mid Orbital Gyrus
R Cerebelum (Crus I)
R Cerebelum (IX)
R Cerebelum (Crus I)
R Anterior Thalamus/Putamen
Note: Cluster peaks are provided for the top 3 peaks/cluster that are separated by
a minimum distance of 30mm
Figures
z
t-value
y
8
3.9446 56 12
8
56
3.0469
6
3.0383 34 40 20
40
-2
2.9707 54
-4.1944 56 -14 -34
-3.9815 2
46
4
-3.8737 56 -62 -32
-3.3749 14 -56 -48
-3.1767 34 -82 -24
-3.7215 10
-6
-2
Figure 1. Sample sizes in the functional neuroimaging literature of post-stroke
aphasia. Histogram illustrating frequencies of different sample sizes across 84 functional
neuroimaging studies of aphasia published since 1995. Bins correspond to single integer
values. The average sample size was 10.33, with a standard deviation of 8.44. The single
largest sample was 53 patients. 67 (81%) of studies had samples sizes less than or equal
to 15 patients, and 73 (87%) of studies had sample sizes less than or equal to 20 patients.
Individual study sample sizes and neuroimaging modalities are provided in
Supplementary Table 1, and the full reference list is also provided in Supplementary
Material 1.
Figure 2. Data characterization and partial behavioral correlation results. A. Lesion
frequencies across all 43 patients. Minimum colorbar values indicate voxels lesioned in
only 1 patient, and maximum colorbar values indicate voxels lesioned in 32 patients
(maximum lesion overlap). B. Means and standard errors for stroke patients and healthy
controls are shown for each language measure. C. Areas showing significantly more
activity during the SD condition relative to the control condition in the healthy controls
(left) and areas showing significantly more activity during the SD condition relative to
the TD condition in the stroke patients (right). Both maps are intensity thresholded at
p<0.01, uncorrected and cluster-corrected at p<0.05 (99 voxels). Colorbar values indicate
t-statistics. Note that the canonical semantic network (CSN) region of interest (ROI) was
defined based on the results shown in (C, left). D. Scatterplots illustrate the partial
correlations between the residuals for activation in the CSN ROI and the residuals for
performance on each language measure (y-axes) after removing the effects of lesion
volume for the patient group (top). Scatterplots illustrate correlations between activation
in the CSN ROI and performance on each language measure for the control group
(bottom). The line of best fit (red) and 95% confidence intervals (red dashes) are shown
on each plot. Each relationship shown was significant at p<0.05, family-wise error
corrected. Note: All brain renderings are in neurological convention.
Figure 3. Whole-brain performance regression results. A. Positive correlations (hot
colors) and negative correlations (cold colors) identified for the patient group between
task-driven activation and performance on the SD task (top) combined fluency measure
(middle) and Boston Naming Test (bottom). B. The subset of regions showing positive
correlations between SD activation and performance on all language measures in the
patient group are shown in red. C. The canonical network identified in controls is shown
in green, regions within the CSN identified in healthy controls where SD activation was
positively related to performance on any language measure in patients are shown in red,
and regions outside of the CSN where SD activation was positively related to
performance on any language measure in patients are shown in blue. Note that the
overlays shown in (C) are qualitative illustrations of how the quantitatively identified
behavioral relationships in patients relate to the canonical network identified in controls,
and are intended to illustrate how activity supporting residual language task performance
relates to the CSN. Each map is intensity thresholded at p<0.01, uncorrected and cluster-
corrected at p<0.01 (126 voxels). Colorbar values indicate t-statistics. All brain
renderings are in neurological convention.
Figure 4. Relationships between left and right hemispheric activation. A. Brain
renderings show the ROIs corresponding to the left hemispheric CSN (LH CSN), right
hemispheric CSN (RH CSN), mirrored left hemispheric CSN (Mirror LH CSN), and out-
of-network left hemispheric homologues (OON-LH-Hom). B. Scatterplots illustrating the
relationship between residual (with the effects of lesion volume partialled out) SD-driven
activity in the LH CSN ROI (y-axes) and each right hemispheric ROI (x-axes) are shown
for the stroke group (top). Scatterplots illustrating the relationship between SD-driven
activity in the LH CSN ROI (y-axes) and each right hemispheric ROI (x-axes) are also
shown for the control group (bottom). C. Line plots illustrate relative activation
magnitudes in the LH CSN ROI and each right hemispheric ROI for the stroke (left, red)
and control (middle, blue) groups. Means and standard errors of activation magnitudes at
each ROI are also shown for both groups (right). D. Scatterplots illustrate the
relationships between left hemispheric lesion volume and activity in each ROI.
Figure 5. Lesion effects on regional activation in the right hemisphere. A. Positive
correlations (hot colors) and negative correlations (cold colors) between task-driven
activation and lesion volume in right hemispheric areas are shown (top). B. Areas of
overlap (red) and no overlap (blue) between regions where increased activation is
associated with larger left hemispheric lesion volumes and the right hemispheric CSN
ROI are shown on the top row. Areas of overlap (red) and no overlap (blue) between
regions where decreased activation is associated with larger left hemispheric lesion
volumes and the right hemispheric CSN ROI are shown on the middle row. Areas of
overlap (red) and no overlap (blue) between regions where increased activation is
associated with larger left hemispheric lesion volumes and the out-of-network left
hemispheric homologues (OON-LH-Hom -- green) ROI are shown on the bottom row.
Note that the overlays shown in (B) are qualitative illustrations of the quantitatively
identified lesion volume effects on right hemispheric activations, and illustrate how right
hemispheric activations associated with lesion volume effects relate to the RH CSN ROI
(top, middle) and to the OON-LH-Hom ROI (bottom). C. Interaction plots are shown for
the moderating effects of lesion volume on the relationship between right inferior frontal
gyrus (IFG) activation and fluency scores (left), right supplementary motor area (SMA)
activation and fluency scores (middle), and right SMA activation and naming scores
(right). Each of these relationships survived a False Discovery Rate threshold of 0.1.
High and low values shown in the interaction plots correspond to +1 and -1 standard
deviations from the mean of each variable. All maps are intensity thresholded at p<0.01,
uncorrected and cluster-corrected at p<0.01 (126 voxels). Colorbar values indicate t-
statistics. Note: all brain renderings are in neurological convention.
Study
Study
PET
Abutalebi et al., 2009
PET
Breier et al., 2009
PET
Fridriksson et al., 2009
PET
Martin et al., 2009
EEG
Specht et al., 2009
NIRS Warren et al., 2009
PET
PET
fMRI
PET
PET
Supplementary Material
Supplementary Analysis 1. Sample sizes from literature search
Supplementary Table 1. Aphasia patient sample sizes in 84 functional neuroimaging
studies of aphasia between 1995 and 2016.
N Modality
N Modality
Weiller et al., 1995
6
1
fMRI
Belin et al., 1996
7
23 MEG
Cappa et al., 1997
8
fMRI
11
Heiss et al., 1997
6
fMRI
2
Thomas et al., 1997
11
12
PET
Sakatani et al., 1998
10
PET
24
Karbe et al., 1998
Fridriksson et al., 2010a
7
fMRI
15
Heiss et al., 1999
Fridriksson et al., 2010b
23
26
fMRI
Miura et al., 1999
Postman-Caucheteux et al.,
1
fMRI
3
2010
Musso et al., 1999
Rochon et al., 2010
4
4
fMRI
Warburton et al., 1999
6
Saur et al., 2010
fMRI
21
Rosen et al., 2000
6 PET+fMRI Sharp et al., 2010
PET
9
Blasi et al., 2002
8
fMRI
5
Leger et al., 2002
1
6
fMRI
Blank et al., 2003
14
fMRI
13
Cardebat et al., 2003
8
fMRI
14
Perani et al., 2003
5
4
fMRI
Fernandez et al., 2004
1
fMRI
8
Kurland et al., 2004
2
fMRI
14
Naeser et al., 2004
4
10
PET
Peck et al., 2004
3
fMRI
16
Zahn et al., 2004
7
30
fMRI
Crinion et al., 2005
17
25 MEG
Crosson et al., 2005
2
fMRI
27
De Boissezon et al.,
7
fMRI
8
2005
Martin et al., 2005
5
24
PET
Pulvermuller et al.,
9
fMRI
14
2005
Winhuisen et al., 2005
11
16
fMRI
Meinzer et al., 2006
1
fMRI
4
Connor et al., 2006
fMRI
12
8
Crinion et al., 2006
22
6
fMRI
Fridriksson et al., 2006 3
fMRI
12
Saur et al., 2006
14
1
fMRI
fMRI
fMRI
PET
PET
fMRI
fMRI
fMRI
fMRI Weiduschat et al., 2011
Allendorfer et al., 2012
fMRI
Fridriksson et al., 2012
fMRI
fMRI
Meltzer et al., 2013
Szaflarski et al., 2013
fMRI
Thompson et al., 2013
PET
fMRI
Thiel et al., 2013
Abel et al., 2014
EEG
PET
Brownsett et al., 2014
Jarso et al., 2014
fMRI
Mattioli et al., 2014
fMRI
PET
Mohr et al., 2014
Robson et al., 2014
fMRI
fMRI
Seghier et al., 2014
Thompson et al., 2010a
Thompson et al., 2010b
Van Oers et al., 2010
Papoutsi et al., 2011
Szaflarski et al., 2011a
Szaflarski et al., 2011b
Tyler et al., 2011
fMRI
fMRI
fMRI
PET
fMRI
fMRI
fMRI
PET
fMRI
Bonakdarpour et al.,
5
2007
Fridriksson et al., 2007 3
Meinzer et al., 2007
1
Winhuisen et al., 2007
9
Vitali et al., 2007
2
Eaton et al., 2008
4
Meinzer et al., 2008
11
Raboyeau et al., 2008
10
Richter et al., 2008
16
*References for supplementary Table 1 are at the end of the supplemental material.
Supplementary Analysis 2. Additional characterization of patient data
Van Hees et al., 2014
Zhu et al., 2014
Abel et al., 2015
Bonakdarpour et al., 2015
Kiran et al., 2015
Spironelli et al., 2015
Griffis et al., 2016
Sims et al., 2016
Geranmayeh et al., 2016
8
13
14
5
8
17
8
14
53
fMRI
fMRI
fMRI
fMRI
fMRI
EEG
fMRI
fMRI
fMRI
Supplementary Figure 1. Behavioral measures. A. Plots of ordered scores for semantic
decisions (left), averaged fluency (middle), and picture naming (right). Scores are shown
separately for stroke patients (top, red) and controls (bottom, blue). B. Plots of
relationships between patient scores on (1) naming and semantic decisions (left), (2)
fluency and semantic decisions (middle), and (3) fluency and naming (right). For each
plot, patient scores are shown as z-scores based on the control data.
Supplementary Table 2. Patient characteristics
Patient Age Sex EHI TSS BNT COWAT SFT Average Fluency %SD Correct
27
21
4
5
20
11
2
8
15
0
8
6
36
0
24
3
4
19
20
31
9
18
13
0
0
2
8
11
9
2
20
7
10
7
1
1
11
2
5
47
39
9
6
42
27
14
20
30
12
10
20
62
2
60
3
10
35
45
45
21
36
12
0
0
1
15
22
14
3
23
18
19
7
6
16
12
7
19
37
30
6.5
5.5
31
19
8
14
22.5
6
9
13
49
1
42
3
7
27
32.5
38
15
27
12.5
0
0
1.5
11.5
16.5
11.5
2.5
21.5
12.5
14.5
7
3.5
8.5
11.5
4.5
12
1
2
3
4
5
6
7
8
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
39
F
63
F
78
41
F
54 M
46 M
52 M
56 M
53 M
55 M
48 M
63 M
56
F
23 M
50 M
F
48
70
F
68 M
59 M
F
23
24
F
78
F
65 M
F
58
72
F
50 M
57 M
51 M
43 M
24 M
F
67
62
F
44
F
62 M
31 M
61 M
64 M
38
F
53
F
54 M
0.55
1.00
0.50
1.00
0.90
0.58
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.91
0.82
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.83
1.00
-1.00
0.91
1.00
1.00
1.00
-1.00
0.91
1.00
0.92
1.0
4.1
5.8
1.6
1.0
1.0
3.4
5.0
1.2
6.1
1.0
1.0
1.0
1.0
1.0
2.0
3.3
1.0
1.0
1.0
3.4
14.0
13.0
1.5
2.9
2.1
1.1
1.3
2.3
2.2
4.4
2.1
2.6
4.8
9.6
2.7
1.8
9.2
3.3
59
57
9
7
53
60
32
50
58
22
60
33
60
2
60
11
9
53
59
59
58
55
40
0
0
2
37
50
21
6
33
41
54
21
25
51
46
34
33
68
50
38
66
46
48
50
76
72
20
40
NA
76
28
72
0
50
72
76
58
36
68
36
NA
42
48
66
44
2
50
62
30
63
NA
NA
30
72
74
55
1.00
0.71
1.00
1.00
1.3
1.3
3.4
12.4
31
3
50
34
1
0
4
2
46 M
F
90
29
F
67 M
40
41
42
43
*EHI – Edinburgh Handedness Inventory, TSS – time since stroke, BNT – Boston
Naming Test, COWAT – Controlled Oral Word Association Test, SFT – Semantic
Fluency Test, %SD Correct -- % Semantic Decision Correct.
Note: The average fluency score is the average of the COWAT and SFT.
5
0
11.5
9
9
0
19
16
34
0
0
26
Supplementary Figure 2. Patient lesions. Patient lesions are shown on example slices.
Patients are organized in columns, as indicated by the text above each column.
Supplementary Analysis 3. Evaluation of result dependence on use of lesion volume
and scanner covariates
additional analyses were performed to assess the stability of our results with and without
scanner included as a covariate (since patients were collected on 2 different scanners) and
with and without lesion volume as a covariate (where applicable). These analyses were
performed to asses the effects of including scanner and lesion volume covariates on
group results. Patients 1-20 were collected on the Phillips scanner, while patients 21-43
were collected on the Siemens scanner. We note that median lesion volumes significantly
differed between patients collected on each scanner (Mann-Whitney U-test p=0.027), and
mean lesion volumes showed a trend-level difference (t41 = 1.83, p=0.075) between
scanners. Control participants were all collected on the Phillips scanner.
Scanner and lesion volume covariates: Mean canonical semantic network
(CSN) activation still predicted fluency (partial r =0.56, p<0.001), naming (partial r=0.49,
p=0.001), and SD performance (partial r=0.46, p=0.006) when both scanner and lesion
volume were partialled out. Activity in left and right hemispheric portions of the CSN
remained significantly correlated when both lesion volume and scanner were partialled
out (partial r=0.79, p<0.001). The same was true when the mirror ROI was used (partial
r=0.80, p<0.001), and when the out-of-network homologue ROI was used (r=0.70,
p<0.001). Whole-brain regression results for models that included both scanner and
lesion volume as covariates are shown in Supplementary Figure 3A.
0.56, p<0.001), naming (partial r=0.51, p<0.001), and SD performance (partial r=0.46,
p=0.004) when only scanner was partialled out. Activity in the whole CSN remained
negatively but non-significantly correlated with total lesion volume when scanner was
partialled out (r=-0.17, p=0.28). Activity in left and right hemispheric portions of the
CSN remained significantly correlated when only scanner was partialled out (partial
r=0.79, p<0.001). The same was true when the mirror ROI was used (partial r=0.80,
p<0.001), and when the out-of-network homologue ROI was used (r=0.69, p<0.001). Left
CSN (partial r = -0.14, p=0.38) and right CSN (partial r=-0.22, p=0.17) activation
remained negatively, but less significantly, correlated with lesion volume when scanner
was partialled out. Whole-brain regression results for models that included only scanner
as a covariate are shown in Supplementary Figure 3B.
No covariates: Correlational results are provided in the main text. Whole-brain
regression results for models that did not include either scanner or lesion volume as
covariates are shown in Figure 3C.
Scanner covariate only: Mean CSN activation still predicted fluency (partial r =
While our main analyses were performed using lesion volume as a covariate,
We note that results were highly similar for all analyses, indicating that scanner
and lesion volume covariates had little effect on the results. While the minor effects of
scanner are perhaps not surprising due to the use of identical scan parameters, our results
suggest that lesion volume differences had only minor effects on functional MRI results
in this sample.
Supplementary Figure 3. Group analyses with and without different covariates included
in the models. A. Whole-brain regressions with both scanner and lesion volume included
as covariates. B. Whole-brain and right-hemispheric regressions with scanner included as
a covariate. C. Whole-brain regressions with no covariates. Note: RH lesion volume
effects are only shown for (B) because lesion volume is the variable of interest in this
analysis, and thus scanner is the only applicable covariate.
Statistical maps are thesholded at a voxelwise p<0.01 and cluster-corrected at p<0.01
(126 voxels). Importantly, nearly identical results were obtained with and without each
covariate included in the models.
Supplementary Analysis 4. Semantic decision activation peaks for voxel-wise (FDR)
correction thresholds.
Supplementary Figure 4. Voxel-wise false discovery rate (vFDR) thresholded semantic
decision activation maps for controls (A, vFDR<0.0005) and patients (B, vFDR<0.05)
are shown along with tables detailing cluster and peak statistics. These data are intended
to (1) demonstrate that the overall patterns of activation are robust even at stringent
voxel-wise thresholds, and (2) provide more precise information about the peak locations
for each group than the more liberally thresholded maps utilized in the primary analyses.
Supplementary Analysis 5. Assessment of damage to "core conceptual network"
that showed positive correlations between activity and all three language measures.
To assess whether or not differences in activation of the core conceptual network that
showed positive relationships to all three language measures might be driven by frequent
damage to the network, we assessed lesion frequencies for each voxel included in the
network. In addition, we assessed the relationship between mean activation within the
network and total damage to the network. The results of these additional analyses are
shown in Supplementary Figure 5. Only the left angular gyrus was lesioned in greater
than 4 (~10% of) patients, and the voxel with the highest lesion frequency was damaged
in 25/43 patients. Activation of the network was weakly negatively (not significantly)
related to the amount of damage sustained by the network.
Supplementary Figure 5. A. Lesion frequencies for regions where activation showed
significant positive relationships to all language measures (i.e. network shown in Figure
3B of the main text) are shown on a partially inflated brain (top), and on slices of a
template brain with cross-hairs centered on the voxel with maximum in-mask lesion
frequency. Colorbar values indicate the number of patients lesioned at each voxel. Only
the left angular gyrus was lesioned in at least 10% (4/43) of patients, with slightly greater
than half of patients (25/43) having damage to the left cerebral white matter underlying
this region. B. Mean activation within the network showed a negative but non-significant
relationship to the extent of damage to the network. This suggests that low levels of
activity within this network are not likely to reflect widespread damage to these regions,
but rather may reflect the effects of focal damage to regions that were more frequently
damaged (such as the left AG) or to inter-regional connections.
References for results from literature search shown in Supplementary Table 1.
Abel S, Weiller C, Huber W, Willmes K (2014): Neural underpinnings for model-
oriented therapy of aphasic word production. Neuropsychologia 57:154–165.
Abel S, Weiller C, Huber W, Willmes K, Specht K (2015): Therapy-induced brain
reorganization patterns in aphasia. Brain 138:1097–112.
Abutalebi J, Rosa PA Della, Tettamanti M, Green DW, Cappa SF (2009): Bilingual
aphasia and language control: A follow-up fMRI and intrinsic connectivity study.
Brain Lang 109:141–156.
Allendorfer J, Kissela B, Holland S, Szaflarski J (2012): Different patterns of language
activation in post-stroke aphasia are detected by overt and covert versions of the
verb generation fMRI task. Med Sci Monit Monit 18.
Belin P, Van Eeckhout P, Zilbovicius M, Remy P, François C, Guillaume S, Chain F,
Rancurel G, Samson Y (1996): Recovery from nonfluent aphasia after melodic
intonation therapy: a PET study. Neurology 47:1504–1511.
role of the right pars opercularis. Ann Neurol 54:310–320.
retrieval learning modulates right frontal cortex in patients with left frontal damage.
Neuron 36:159–170.
Cardebat D (2005): Subcortical aphasia: A longitudinal PET study. Stroke 36:1467–
1473.
oxygen level dependent (BOLD) signal in patients with stroke-induced and primary
progressive aphasia. NeuroImage Clin 8:87–94.
Bonakdarpour B, Parrish TB, Thompson CK (2007): Hemodynamic response function in
patients with stroke-induced aphasia: implications for fMRI data analysis.
Neuroimage 36:322–31.
Behavioral and Neurophysiologic Response to Therapy for Chronic Aphasia. Arch
Phys Med Rehabil 90:2026–2033.
Brownsett SLE, Warren JE, Geranmayeh F, Woodhead Z, Leech R, Wise RJS (2014):
Blank SC, Bird H, Turkheimer F, Wise RJS (2003): Speech production after stroke: The
Blasi V, Young AC, Tansy AP, Petersen SE, Snyder AZ, Corbetta M (2002): Word
Breier JI, Juranek J, Maher LM, Schmadeke S, Men D, Papanicolaou AC (2009):
De Boissezon X, Démonet JF, Puel M, Marie N, Raboyeau G, Albucher JF, Chollet F,
Bonakdarpour B, Beeson PM, DeMarco AT, Rapcsak SZ (2015): Variability in blood
Eaton KP, Szaflarski JP, Altaye M, Ball AL, Kissela BM, Banks C, Holland SK (2008):
Fernandez B, Cardebat D, Demonet JF, Joseph PA, Mazaux JM, Barat M, Allard M
Cappa SF, Perani D, Grassi F, Bressi F, Alberoni M, Franceschi M, Bettinardi M, Todde
Cognitive control and its impact on recovery from aphasic stroke. Brain 137:242–
54.
M, Fazio M (1997): A PET Follow-up Study of Recovery after Stroke in Acute
Aphasics. Brain Lang 56:55–67.
Cardebat D, Demonet JF, De Boissezon X, Marie N, Mari RM, Lambert J, Baron JC,
Puel M (2003): Behavioral and Neurofunctional Changes over Time in Healthy and
Aphasic Subjects: A PET Language Activation Study. Stroke 34:2900–2906.
Connor LT, DeShazo Braby T, Snyder AZ, Lewis C, Blasi V, Corbetta M (2006):
Cerebellar activity switches hemispheres with cerebral recovery in aphasia.
Neuropsychologia 44:171–7.
Crinion JT, Warburton EA, Lambon-Ralph MA, Howard D, Wise RJS (2006): Listening
to narrative speech after aphasic stroke: The role of the left anterior temporal lobe.
Cereb Cortex 16:1116–1125.
Crinion J, Price CJ (2005): Right anterior superior temporal activation predicts auditory
sentence comprehension following aphasic stroke. Brain 128:2858–71.
Crosson B, Moore AB, Gopinath K, White KD, Wierenga CE, Gaiefsky ME, Fabrizio
KS, Peck KK, Soltysik D, Milsted C, Briggs RW, Conway TW, Gonzalez Rothi LJ
(2005): Role of the right and left hemispheres in recovery of function during
treatment of intention in aphasia. J Cogn Neurosci 17:392–406.
Reliability of fMRI for studies of language in post-stroke aphasia subjects.
Neuroimage 41:311–22.
(2004): Functional MRI follow-up study of language processes in healthy subjects
and during recovery in a case of aphasia. Stroke 35:2171–2176.
Fridriksson J (2010): Preservation and modulation of specific left hemisphere regions is
vital for treated recovery from anomia in stroke. J Neurosci 30:11558–11564.
Fridriksson J, Baker JM, Moser D (2009): Cortical mapping of naming errors in aphasia.
Hum Brain Mapp 30:2487–2498.
Fridriksson J, Bonilha L, Baker JM, Moser D, Rorden C (2010): Activity in preserved left
hemisphere regions predicts anomia severity in aphasia. Cereb Cortex 20:1013–9.
Fridriksson J, Morrow-Odom L, Moser D, Fridriksson A, Baylis G (2006): Neural
recruitment associated with anomia treatment in aphasia. Neuroimage 32:1403–
1412.
GC, Rorden C (2007): Neural correlates of phonological and semantic-based anomia
treatment in aphasia. Neuropsychologia 45:1812–22.
aphasia recovery. Neuroimage 60:854–63.
production after left hemisphere stroke. Neurology 86:1296–1305.
following Intermittent Theta Burst Stimulation in Chronic Poststroke Aphasia.
Neural Plast 2016:20–23.
Fridriksson J, Richardson JD, Fillmore P, Cai B (2012): Left hemisphere plasticity and
Geranmayeh F, Leech R, Wise RJS (2016): Network dysfunction predicts speech
Griffis JC, Nenert R, Allendorfer JB, Szaflarski JP (2016): Interhemispheric Plasticity
Fridriksson J, Moser D, Bonilha L, Morrow-Odom KL, Shaw H, Fridriksson A, Baylis
van Hees S, McMahon K, Angwin A, de Zubicaray G, Read S, Copland DA (2014): A
Kiran S, Meier EL, Kapse KJ, Glynn PA. (2015): Changes in task-based effective
Kurland J, Naeser MA, Baker EH, Doron K, Martin PI, Seekins HE, Bogdan A, Renshaw
Heiss WD, Karbe H, Weber-Luxenburger G, Herholz K, Kessler J, Pietrzyk U, Pawlik G
functional MRI study of the relationship between naming treatment outcomes and
resting state functional connectivity in post-stroke aphasia. Hum Brain Mapp
35:3919–31.
Heiss WD, Kessler J, Thiel A, Ghaemi M, Karbe H (1999): Differential capacity of left
and right hemispheric areas for compensation of poststroke aphasia. Ann Neurol
45:430–438.
(1997): Speech-induced cerebral metabolic activation reflects recovery from
aphasia. J Neurol Sci 145:213–217.
Jarso S, Li M, Faria A, Davis C, Leigh R, Sebastian R, Tsapkini K, Mori S, Hillis AE
(2014): Distinct mechanisms and timing of language recovery after stroke. Cogn
Neuropsychol 0:1–22.
Karbe H, Thiel A, Weber-luxenburger G, Kessler J, Heiss W (1998): Brain Plasticity in
Poststroke Aphasia : What Is the Contribution of the Right Hemisphere? Brain Lang
230:215–230.
connectivity in language networks following rehabilitation in post-stroke patients
with aphasia. Front Hum Neurosci 9:1–20.
P, Yurgelun-Todd D (2004): Test-retest reliability of fMRI during nonverbal
semantic decisions in moderate-severe nonfluent aphasia patients. Behav Neurol
15:87–97.
Léger A, Démonet JF, Ruff S, Aithamon B, Touyeras B, Puel M, Boulanouar K,
Cardebat D (2002): Neural substrates of spoken language rehabilitation in an
aphasic patient: an fMRI study. Neuroimage 17:174–183.
Martin PI, Naeser MA, Doron KW, Bogdan A, Baker EH, Kurland J, Renshaw P,
Yurgelun-Todd D (2005): Overt naming in aphasia studied with a functional MRI
hemodynamic delay design. Neuroimage 28:194–204.
Martin PI, Naeser MA, Ho M, Doron KW, Kurland J, Kaplan J, Wang Y, Nicholas M,
Baker EH, Fregni F, Pascual-Leone A (2009): Overt naming fMRI pre- and post-
TMS: Two nonfluent aphasia patients, with and without improved naming post-
TMS. Brain Lang 111:20–35.
Mattioli F, Ambrosi C, Mascaro L, Scarpazza C, Pasquali P, Frugoni M, Magoni M,
Biagi L, Gasparotti R (2014): Early aphasia rehabilitation is associated with
functional reactivation of the left inferior frontal gyrus: a pilot study. Stroke 45:545–
52.
Meinzer M, Obleser J, Flaisch T, Eulitz C, Rockstroh B (2007): Recovery from aphasia
as a function of language therapy in an early bilingual patient demonstrated by
fMRI. Neuropsychologia 45:1247–1256.
Meinzer M, Flaisch T, Breitenstein C, Wienbruch C, Elbert T, Rockstroh B (2008):
Functional re-recruitment of dysfunctional brain areas predicts language recovery in
chronic aphasia. Neuroimage 39:2038–46.
Meinzer M, Flaisch T, Obleser J, Assadollahi R, Djundja D, Barthel G, Rockstroh B
(2006): Brain regions essential for improved lexical access in an aged aphasic
patient: a case report. BMC Neurol 6:28.
Meltzer JA, Wagage S, Ryder J, Solomon B, Braun AR (2013): Adaptive significance of
Miura K, Nakamura Y, Miura F, Yamada I, Takahashi M, Yoshikawa A, Mizobata T
right hemisphere activation in aphasic language comprehension. Neuropsychologia
51:1248–1259.
(1999): Functional magnetic resonance imaging to word generation task in a patient
with Broca's aphasia. J Neurol 246:939–942.
Mohr B, Difrancesco S, Harrington K, Evans S, Pulvermuller F (2014): Changes of right-
hemispheric activation after constraint-induced, intensive language action therapy in
chronic aphasia: fMRI evidence from auditory semantic processing1. Front Hum
Neurosci 8:1–15.
Musso M, Weiller C, Kiebel S, Müller SP, Bülau P, Rijntjes M (1999): Training-induced
brain plasticity in aphasia. Brain 122 ( Pt 9:1781–90.
Naeser MA, Martin PI, Baker EH, Hodge SM, Sczerzenie SE, Nicholas M, Palumbo CL,
Goodglass H, Wingfield A, Samaraweera R, Harris G, Baird A, Renshaw P,
Yurgelun-Todd D (2004): Overt propositional speech in chronic nonfluent aphasia
studied with the dynamic susceptibility contrast fMRI method. Neuroimage 22:29–
41.
van Oers CA, Vink M, van Zandvoort MJE, van der Worp HB, de Haan EHF, Kappelle
LJ, Ramsey NF, Dijkhuizen RM (2010): Contribution of the left and right inferior
frontal gyrus in recovery from aphasia. A functional MRI study in stroke patients
with preserved hemodynamic responsiveness. Neuroimage 49:885–93.
Papoutsi M, Stamatakis EA, Griffiths J, Marslen-Wilson WD, Tyler LK (2011): Is left
fronto-temporal connectivity essential for syntax? Effective connectivity,
tractography and performance in left-hemisphere damaged patients. Neuroimage
58:656–64.
Peck KK, Moore AB, Crosson BA, Gaiefsky M, Gopinath KS, White K, Briggs RW
(2004): Functional Magnetic Resonance Imaging before and after Aphasia Therapy:
Shifts in Hemodynamic Time to Peak during an Overt Language Task. Stroke
35:554–559.
Perani D, Cappa SF, Tettamanti M, Rosa M, Scifo P, Miozzo A, Basso A, Fazio F
(2003): A fMRI study of word retrieval in aphasia. Brain Lang 85:357–368.
Postman-Caucheteux WA, Birn RM, Pursley RH, Butman JA, Solomon JM, Picchioni D,
McArdle J, Braun AR (2010): Single-trial fMRI shows contralesional activity linked
to overt naming errors in chronic aphasic patients. J Cogn Neurosci 22:1299–1318.
reorganization of language in both hemispheres of patients with chronic aphasia.
Neuroimage 28:481–9.
Raboyeau G, De Boissezon X, Marie N, Balduyck S, Puel M, Bézy C, Démonet JF,
Cardebat D (2008): Right hemisphere activation in recovery from aphasia: Lesion
effect or function recruitment? Neurology 70:290–298.
Richter M, Miltner WHR, Straube T (2008): Association between therapy outcome and
right-hemispheric activation in chronic aphasia. Brain 131:1391–1401.
Robson H, Zahn R, Keidel JL, Binney RJ, Sage K, Lambon Ralph MA (2014): The
anterior temporal lobes support residual comprehension in Wernicke's aphasia.
Brain 137:931–43.
Rochon E, Leonard C, Burianova H, Laird L, Soros P, Graham S, Grady C (2010):
Neural changes after phonological treatment for anomia: An fMRI study. Brain
Pulvermüller F, Hauk O, Zohsel K, Neininger B, Mohr B (2005): Therapy-related
Rosen HJ, Petersen SE, Linenweber MR, Snyder AZ, White DA, Chapman L,
Seghier ML, Bagdasaryan J, Jung DE, Price CJ (2014): The Importance of Premotor
Specht K, Zahn R, Willmes K, Weis S, Holtel C, Krause BJ, Herzog H, Huber W (2009):
Lang 114:164–179.
Dromerick a W, Fiez JA, Corbetta MD (2000): Neural correlates of recovery from
aphasia after damage to left inferior frontal cortex. Neurology 55:1883–1894.
Sakatani K, Xie Y, Wemara L, Li S, Zuo H (1998): Language-Activated Cerebral Blood
Oxygenation and Hemodynamic Changes of the Left Prefrontal Cortex in Poststroke
Aphasic Patients. Stroke 29:12–14.
Saur D, Lange R, Baumgaertner A, Schraknepper V, Willmes K, Rijntjes M, Weiller C
(2006): Dynamics of language reorganization after stroke. Brain 129:1371–84.
Saur D, Ronneberger O, Kümmerer D, Mader I, Weiller C, Klöppel S (2010): Early
functional magnetic resonance imaging activations predict language outcome after
stroke. Brain 133:1252–1264.
Cortex for Supporting Speech Production after Left Capsular-Putaminal Damage. J
Neurosci 34:14338–14348.
Sharp DJ, Turkheimer FE, Bose SK, Scott SK, Wise RJS (2010): Increased frontoparietal
integration after stroke and cognitive recovery. Ann Neurol 68:753–6.
Sims JA, Kapse K, Glynn P, Sandberg C, Tripodis Y, Kiran S (2016): The relationships
between the amount of spared tissue, percent signal change, and accuracy in
semantic processing in aphasia. Neuropsychologia 84:113–126.
Joint independent component analysis of structural and functional images reveals
complex patterns of functional reorganisation in stroke aphasia. Neuroimage
47:2057–2063.
Spironelli C, Angrilli A (2015): Brain plasticity in aphasic patients: intra- and inter-
hemispheric reorganisation of the whole linguistic network probed by N150 and
N350 components. Sci Rep 5:12541.
Szaflarski JP, Eaton K, Ball AL, Banks C, Vannest J, Allendorfer JB, Page S, Holland
SK (2011a): Poststroke aphasia recovery assessed with functional magnetic
resonance imaging and a picture identification task. J Stroke Cerebrovasc Dis
20:336–345.
Szaflarski JP, Allendorfer JB, Banks C, Vannest J, Holland SK (2013): Recovered vs.
not-recovered from post-stroke aphasia: the contributions from the dominant and
non-dominant hemispheres. Restor Neurol Neurosci 31:347–60.
Szaflarski JP, Vannest J, Wu SW, DiFrancesco MW, Banks C, Gilbert DL (2011b):
Excitatory repetitive transcranial magnetic stimulation induces improvements in
chronic post-stroke aphasia. Med Sci Monit 17:CR132-9.
Thiel A, Hartmann A, Rubi-Fessen I, Anglade C, Kracht L, Weiduschat N, Kessler J,
Rommel T, Heiss W-D (2013): Effects of noninvasive brain stimulation on language
networks and recovery in early poststroke aphasia. Stroke 44:2240–6.
processing in aphasia: Changes in lateralization patterns during recovery reflect
cerebral plasticity in adults. Electroencephalogr Clin Neurophysiol 102:86–97.
Thompson CK, Riley EA, den Ouden DB, Meltzer-Asscher A, Lukic S (2013): Training
verb argument structure production in agrammatic aphasia: Behavioral and neural
recovery patterns. Cortex 49:2358–2376.
Thomas C, Altenmuller E, Marckmann G, Kahrs J, Dichgans J (1997): Language
Vitali P, Abutalebi J, Tettamanti M, Danna M, Ansaldo A-I, Perani D, Joanette Y, Cappa
Thompson CK, den Ouden DB, Bonakdarpour B, Garibaldi K, Parrish TB (2010): Neural
Thompson CK, Bonakdarpour B, Fix SF (2010): Neural mechanisms of verb argument
structure processing in agrammatic aphasic and healthy age-matched listeners. J
Cogn Neurosci 22:1993–2011.
plasticity and treatment-induced recovery of sentence processing in agrammatism.
Neuropsychologia 48:3211–3227.
Tyler LK, Marslen-Wilson WD, Randall B, Wright P, Devereux BJ, Zhuang J, Papoutsi
M, Stamatakis EA. (2011): Left inferior frontal cortex and syntax: Function,
structure and behaviour in patients with left hemisphere damage. Brain 134:415–
431.
SF (2007): Training-induced brain remapping in chronic aphasia: a pilot study.
Neurorehabil Neural Repair 21:152–60.
Warburton E, Price CJ, Swinburn K, Wise RJ (1999): Mechanisms of recovery from
aphasia: evidence from positron emission tomography studies. J Neurol Neurosurg
Psychiatry 66:155–61.
Warren JE, Crinion JT, Lambon Ralph MA, Wise RJS (2009): Anterior temporal lobe
connectivity correlates with functional outcome after aphasic stroke. Brain
132:3428–42.
Weiduschat N, Thiel A, Rubi-Fessen I, Hartmann A, Kessler J, Merl P, Kracht L,
Rommel T, Heiss WD (2011): Effects of repetitive transcranial magnetic stimulation
in aphasic stroke: a randomized controlled pilot study. Stroke 42:409–15.
Noth J, Diener HC (1995): Recovery from Wernicke's aphasia: A positron emission
tomographic study. Ann Neurol 37:723–732.
Winhuisen L, Thiel A, Schumacher B, Kessler J, Rudolf J, Haupt WF, Heiss WD (2005):
Role of the contralateral inferior frontal gyrus in recovery of language function in
poststroke aphasia: a combined repetitive transcranial magnetic stimulation and
positron emission tomography study. Stroke 36:1759–63.
The right inferior frontal gyrus and poststroke aphasia: a follow-up investigation.
Stroke 38:1286–92.
Zahn R, Drews E, Specht K, Kemeny S, Reith W, Willmes K, Schwarz M, Huber W
(2004): Recovery of semantic word processing in global aphasia: A functional MRI
study. Cogn Brain Res 18:322–336.
Zhu D, Chang J, Freeman S, Tan Z, Xiao J, Gao Y, Kong J (2014): Changes of functional
connectivity in the left frontoparietal network following aphasic stroke. Front Behav
Neurosci 8:167.
Weiller C, Isensee C, Rijntjes M, Huber W, Müller S, Bier D, Dutschka K, Woods RP,
Winhuisen L, Thiel A, Schumacher B, Kessler J, Rudolf J, Haupt WF, Heiss WD (2007):
|
1702.07227 | 1 | 1702 | 2017-02-20T10:33:03 | The knowledge paradox: why knowing more is knowing less | [
"q-bio.NC"
] | To provide an explanation of the evolution of scientific knowledge, I start from the assumption that knowledge is based on concepts, and propose that each concept about reality is affected by vagueness. This entails a paradox, which I term Knowledge Paradox (KP): i.e. we need concepts to acquire knowledge about the real world but each concept is a step away from reality. The KP provides a unifying context for the sorites and the liar paradoxes. Any concept is viewed as a sorites, i.e. it is impossible to set a boundary between what is, and what is not, the entity to which the concept refers. Hence, any statement about reality can be reduced to a liar, wherefrom the KP follows in its most general form: -If I know, then I do not know-. The KP is self-referential but not contradictory, as it can be referred to two levels of knowledge: -if I know(epistemic), then I do not know(ontic)-, where the ontic level is made unachievable by concept vagueness. Such an interpretation of scientific knowledge provides an understanding of its dynamics. Concept proliferation within theories produces periods of knowledge decay that are episodically reversed by the formulation of new theories based on a smaller number of synthetic concepts. | q-bio.NC | q-bio | The knowledge paradox: why knowing more is knowing less
Bruno Burlando
Dipartimento di Sienze e Innovazione Tecnologica
Università del Piemonte Orientale
Viale Teresa Michel 11
15121 Alessandria
Italy
e-mail : [email protected]
1
Abstract
To provide an explanation of the evolution of scientific knowledge, I start from the assumption
that knowledge is based on concepts, and propose that each concept about reality is affected by
vagueness. This entails a paradox, which I term Knowledge Paradox (KP): i.e. we need concepts to
acquire knowledge about the real world but each concept is a step away from reality. The KP
provides a unifying context for the sorites and the liar paradoxes. Any concept is viewed as a
sorites, i.e. it is impossible to set a boundary between what is, and what is not, the entity to which
the concept refers. Hence, any statement about reality can be reduced to a liar, wherefrom the KP
follows in its most general form: "If I know, then I do not know". The KP is self-referential but not
contradictory, as it can be referred to two levels of knowledge: "if I knowepistemic, then I do not
knowontic", where the ontic level is made unachievable by concept vagueness. Such an interpretation
of scientific knowledge provides an understanding of its dynamics. Concept proliferation within
theories produces periods of knowledge decay that are episodically reversed by the formulation of
new theories based on a smaller number of synthetic concepts.
.
2
Introduction
This article concerns an analysis of the meaning and progress of scientific knowledge. I will not
make attempts at defining the limits of science, i.e. dealing with the long debated problem of
demarcation (Resnik, 2000). Scientific knowledge will be simply defined here as a particularly
structured form of conceptual knowledge, deriving from perceptual or instrumental experience of
the world.
It is generally assumed that the purpose of science is to develop laws and theories in order to
explain, understand and predict phenomena (Chalmers, 1990). Accordingly, it is believed that the
scientific method has been one of the driving forces of human progress, while our knowledge about
the world is usually related to the ability of theories to make predictions about reality (Gauch,
2003). However, progressive accumulation of scientific evidence can reveal shortcomings in extant
paradigms, ultimately leading to theoretical rearrangements. Such a trend has been emphasized by
Kuhn, with his distinction between normal and revolutionary phases in the progress of science
(Kuhn 1962).
Cognitive limits of concepts
I start from the assumption that the failure of scientific theories derives from intrinsic cognitive
limits of concepts. The term 'concept' will be used here in a wide acception, i.e. as an idea or notion
that serves to designate a category of entities, or more in general, to provide a model of reality.
Even though non-conceptual content of thought has been considered in the analysis of human
cognitive processes and awareness (Bermudez, 1995), it is generally believed that conceptual
content of thought is typical of scientific reasoning (Achinstein, 1968). Hence, our mind uses
concepts in order to describe the real world, and such a process is generally considered an essential
step in the advancement of science.
I wish to stress that the shortcomings of scientific paradigms, which may lead to a collapse of
theoretical buildings, would depend on the reduction of the world to a set of concepts, i.e. discrete
3
model objects, thereby yielding an artifactual vision of reality. A vivid reminiscence of these ideas
can be found in the words of the French poet and philosopher Paul Valery: «Tout ce qui est simple
est faux, mais tout ce qui ne l'est pas est inutilisable» (Guégan and Renaud, 2004). On a similar
ground, my analysis holds that concepts can be viewed as an unavoidable simplification of reality,
which inevitably leads to its falsification. These arguments are to some extent related to
instrumentalism (White, 1943), which assumes that concepts and theories are merely useful
instruments, or 'fictions', whose function is not to depict reality but rather to provide descriptions
and predictions about the physical world. However, I am pushing my analysis further on, by
proposing that concepts introduce elements of inconsistency within scientific theories that sooner or
later lead to thorough revisions.
Concept vagueness and the sorites
The failure of concepts to provide a reliable description of the world addresses the problem of
vagueness. This latter is strictly related to the sorites paradox, which entails the possibility that the
borderline between a predicate and its negation be fuzzy (Tye, 1994; Keefe, 2000). In his analysis
of the sorites, James Cargile refers to the impossibility of setting an exact point of transition
between a developing tadpole and an adult frog (Cargile, 1969). It may be assumed that a proper
example of vagueness had to be chosen, whereas in contrast my argument is aimed at showing that
any real entity would have been appropriate, since I assume that vagueness is intrinsic to concepts
and, therefore, any concept is susceptible to the sorites.
Obviously, the vagueness of concepts does not affect the practical use of the real objects to
which they refer. In general, vagueness is ruled out whenever a certain approximation of reality can
be tolerated in order to accomplish a particular project, as commonly occurs in everyday life,
technical fields and professional work. Conversely, the vagueness of concepts heavily affects the
rigorous in-depth analyses that need to be deployed in scientific activities.
4
Fundamental concepts of physics, like matter, time, and space, providing basic reference frames
to scientific theories, are affected by vagueness. After the arising of quantum mechanics, there has
been growing awareness that space and time are vague on a microscopic scale (Dolan, 2005).
Moreover, their vagueness does not disappear on a macroscopic scale, as shown by the problem of
infinite divisibility highlighted by Zeno's paradoxes (Mueller 1969), while their unwarranted use
can lead to the puzzling situations of time travel paradoxes (Nahin, 1999). Also, the concept of
matter has been made uncertain by quantum physics, under which the identity of elementary
particles is suggested to be irreducibly vague (French and Krause, 2003; Chibeni, 2004).
In the field of biology, the formulation of the concept of cell has been a milestone of enormous
theoretical importance (Baker, 1953). However, the identity of cells within organisms remains
vague, due to the occurrence of syncytial tissues, endocytobiosis, incomplete cell divisions, and
exchange of material between a cell and its surroundings. Similar reasoning can be applied to a
number of biological entities.
In summary, science is affected by vagueness at its deepest foundations. The formulation of so-
called exact laws is only possible if an appropriate armamentarium of concepts allows the
edification of theoretical systems within which these laws are meaningful, as occurs e.g. for
gravitation, electromagnetism and thermodynamics. By contrast, the lack of a suitable set of
concepts renders extremely difficult the determination of exact laws for nonlinear systems, such as
fluid turbulence, living beings, and stock markets (Bar-Yam, 1997).
Knowledge Paradox and the progress of science
Assuming that any concept about reality generates a sorites, since it involves some discretization
of the real world, it follows that the formulation of concepts gives rise to vagueness and move us
away from reality. This configures a paradox that I call the Knowledge Paradox (KP): scientific
activities yield discoveries that are implemented within extant theories by means of new, ad hoc
concepts, thus producing increasingly complex conceptual frameworks that progressively put our
5
vision of reality out of focus. Consequently, the accumulation of concepts within theories does not
produce an advancement of knowledge, but rather a progressive decay.
In order to clarify such a counterintuitive idea, I will use an example from the field of molecular
biology. If a new, bona-fide protein is found, say protein A, a new concept is needed to implement
such a discovery within the molecular theory of life. However, it is not clear if the notion of protein
A refers to the protein's native form, to one of its post-translational modifications, or to one of the
interactions established by the protein with other molecules, including the bulk of solvent water
molecules. Now, given that each molecule of protein A is in a particular chemico-physical status as
regards to its interactions with other molecules, a specific concept would be needed for each protein
A molecule. Yet, even if this impracticable conceptualization could be done, it would not solve the
problem, as molecular interactions are subjected to continuous change along time. Hence, the
concept of the protein A is vague, representing at most an approximation of reality, and
consequently it adds vagueness to the molecular theory of life. As stated above, this does not imply
that the new concept cannot be helpful at all, but its use will be only profitable within theoretical
boundaries entailing a simplification of reality. Conversely, if the aim is to achieve a better
understanding of reality, the new concept will eventually become a hindrance.
As shown above, the accumulation of concepts negatively affects the description of reality.
However, from time to time exhaustive revisions of data may succeed in producing synthetic
theories that explain facts by a smaller number of new, revolutionary concepts. The evolution of
scientific knowledge would therefore consist of episodic advancements, characterized by a
reduction in the number of concepts, interspersed among long periods of decay characterized by
concept accumulation (Fig. 1). Such a pattern can be easily recognized. By introducing his theory of
natural selection, Darwin (1859) provided a synthetic explanation for the astonishing set of life
diversity that had been collected through the years. About one century later, the theory of the
genetic code yielded a unifying conceptual basis for biological heredity (Crick, 1967). Wegener's
continental drift (1915), and the ensuing theory of plate tectonics, provided a synthetic explanation
6
for the movement of Earth's plates, earthquakes, volcanoes, oceanic trenches, and mountain range
formation. Also, by introducing the concept of "quark" in the 1960s, Gell-Mann and Zweig offered
a unifying account of the variety of hadron particles described during the first half of the 20th
century (Lichtenberg and Rosen, 1980). Similar features are shared by other major scientific
breakthroughs, such as those carried out by Newton, Maxwell and Freud (Kuhn, 1962).
Knowledge Paradox and the liar
If no theory can describe the real world in a completely affordable way, this must hold true for
the KP itself. In other words, if any concept is vague when applied to reality, then the concept of
'concept' is vague too, and consequently also the KP is affected by vagueness. Hence, the KP
makes a statement that, in the ultimate analysis, concerns its own meaning, and therefore is
equivalent to the liar paradox (Martin, 1978). The liar is endowed with self-referentiality, or in other
terms, it implies and is implied by its negation (Suber, 1990). Hence, it represents a contradiction
built into the structure of language, while on the other hand the KP entails an endless loop within
the structure of knowledge. Similar to the antinomy of the liar: "If this sentence is true, then it is not
true", the KP can be formulated as: "If I know, then I do not know", which seems to make
contradiction unavoidable.
A solution to the liar has been proposed by Tarski (1956) who, in his approach to the semantic
concept of truth, adopted a hierarchy of languages consisting of object-language, to which the
definition of truth is applied, and of meta-language, providing the definition of truth for object-
language. By following a similar approach, knowledge can be analyzed from the standpoint of a
meta-knowledge level, by making a distinction between our ideas about physical entities, viz. the
epistemic level, and the real status of these entities, viz. the ontic level (Atmanspacher & Primas,
2003). Such a distinction allows a Tarskian reformulation of the KP providing two meanings for 'I
know', whereby it should read: "If I knowepistemic, then I do not knowontic".
7
According to this view, the formulation of concepts generates a gap between ontic and epistemic
knowledge that cannot be amended, since any attempt at reducing this gap can only be carried out
by formulating other concepts, and so on. Hence, the KP sets the limits of knowledge, while its self-
referentiality highlights the unamendability of these limits.
A symptom of the ontic/epistemic disjunction of knowledge is the arising of inconsistencies in
scientific theories that prelude the advent of their crisis. In Physics, striking examples have been the
discoveries of the photoelectric effect and the black body radiation, which raised inconsistencies in
classic physics and opened the way to quantum mechanics (Planck, 1901). In Biology, the Central
Dogma of molecular genetics holds that biological information flows from DNA or RNA to
proteins, and not the reverse (Crick, 1958). However, the discovery of prions (Prusiner, 1982), and
the finding in yeast that these proteins can transmit information through generations and modify the
expression of the genome (True & Lindquist, 2000), has posed a serious challenge to the overall
consistency of the Dogma.
Inconsistency of knowledge
Combined analyses of the liar and the sorites have been carried out with the aim of finding a
unifying treatment (Tappenden, 1993). The KP provides a link between the sorites and the liar by
showing that any statement about the world is, or is entailed by, a statement that makes use of a
sorites, and can therefore be reconducted to a liar. In the above example of protein A, the sorites
emerges from the impossibility of setting a boundary between what is protein A and what is not
protein A. Hence, the liar follows: "If this entity is (epistemologically) protein A, then it is not
(ontologically) protein A".
In summary, knowledge is intrinsically inconsistent. So far as we formulate new concepts, then,
the more concepts we have, the more inconsistencies we raise, the more we worsen knowledge
about the world. By contrast, if we manage to reduce the number of concepts, by formulating
synthetic ones, we reduce inconsistencies and eventually obtain an improvement of knowledge. So,
8
contrarily to what it may be intuitively assumed: the more concepts, the less knowledge, and vice
versa.
References
Achinstein, P. 1968. Concepts of Science: A Philosophical Analysis. Baltimore: The Johns Hopkins
University Press.
Atmanspacher, H., Primas, H., 2003. Epistemic and ontic quantum realities. In: Castell L.,
Ischebeck O. (Eds), Time, Quantum and Information, Berlin: Springer, pp 301–321.
Bar-Yam, Y. 1997. Dynamics of Complex Systems. Cambridge, MA: Perseus Books.
Baker, J.R. 1953. The cell-theory: a restatement, history and critique. Quarterly Journal of
Microscopical Science 94: 407-440.
Bermudez, J. L. 1995. Non-conceptual content: From perceptual experience to subpersonal
computational states. Mind and Language 10: 333-369.
Cargile, J. 1969. The sorites paradox. Br. J. Phyl. Sci. 20: 193-202.
Chalmers, A. 1990. Science and its Fabrication. Minneapolis, MN: University of Minnesota Press.
Chibeni, S.S. 2004. Ontic vagueness in microphysics. Sorites 15: 29-41. Retrieved February 14,
2017, from http://www.sorites.org/Issue_15/chibeni.htm
Crick, F.H.C. 1967. The Croonian lecture, 1966. The genetic code. Proc. R. Soc. Lond. Ser. B 167:
331-347.
Crick, F.H.C. 1958. Central dogma of molecular biology. Nature 227: 561- 563.
Darwin, C. 1859. On the Origin of Species. London: John Murray.
Dolan, R.P. 2005. The inflaton spacetime model: making sense of the standard models of particle
physics
and
cosmology.
Retrieved
February
15,
2017
from
http://home.earthlink.net/~dolascetta/ISTModel.pdf
French, S. and Krause, D. 2003. Quantum vagueness. Erkenntnis 59: 97-124.
Gauch, H.G.Jr. 2003. Scientific Method in Practice. Cambridge, UK: Cambridge University Press.
9
Guégan, J.F., Renaud, F. 2004. Vers une écologie de la santé. In Barbault R., Chevassus-au-Louis
B. (Eds.): Biodiversité et Changements Globaux. ADPF/MAE, Paris, France. pp. 100-135.
Keefe, R. 2000. Theories of vagueness. Cambridge, UK: Cambridge University Press.
Kuhn, T.S. 1962. The Structure of Scientific Revolutions. Chicago, IL: University of Chicago Press.
Lichtenberg, D.B., Rosen, S.P. (Eds) 1980. Developments in the Quark Theory of Hadrons, Vol. 1.
1964-1978. Nonantum, MA: Hadronic Press.
Mueller, I. 1969. Zeno's paradoxes and continuity. Mind 78: 129-131.
Martin, R. (Ed.) 1978. The Paradox of the Liar. Atascadero, CA: Ridgeview.
Nahin, P.J. 1999. Time Machines: Time Travel in Physics, Metaphysics and Science Fiction. New
York: Springer-Verlag.
Planck, M. 1901. Über das Gesetz der Energieverteilung im Normalspektrum. Annalen Der Physik
4: 553 ff.
Prusiner, SB. 1982. Novel proteinaceous infectious particles cause scrapie. Science 216 : 136-144.
Resnik, D.B. 2000. A pragmatic approach to the demarcation problem. Studies in History and
Philosophy of Science Part A 31: 249-267.
Suber, P. 1990. The Paradox of Self-Amendment: A Study of Law, Logic, Omnipotence, and
Change. New York: Peter Lang Publishing.
Tappenden, J. 1993. The Liar and sorites paradoxes: Toward a unified treatment. The Journal of
Philosophy 90: 551-577.
Tarski, A. 1956. The Concept of Truth in Formalized Languages', Logic, Semantics and
Metamathematics. Oxford: Oxford University Press.
True, H.L., Lindquist, S.L. 2000. A yeast prion provides a mechanism for genetic variation and
phenotypic diversity. Nature 28: 477-483.
Tye, M. 1994. Sorites paradoxes and the semantics of vagueness. In Tomberlin J. (Ed.),
Philosophical Perspectives: Logic and Language. Atascadero, CA: Ridgeview.
Wegener, A. 1915. Die Entstehung der Kontinente und Ozeane. Vieweg, Braunschweig.
10
White, M. 1943. The Origin of Dewey's Instrumentalism. New York: Columbia University Press.
11
Figures
K
knowledge decay
knowledge decay
more concepts needed
to implement new data
new theory with
synthe,c concepts
more concepts needed
to implement new data
t
Figure 1. A hypothetical knowledge index (K), varying along time (t), is defined as being inversely
proportional to the number of concepts used within a theory. So far as increasing numbers of
concepts are needed to implement newly acquired data, the index scales down, indicating a decay of
knowledge. However, from time to time a thorough revision of data gives rise to a synthetic theory
consisting of a few innovative concepts. This renders many concepts obsolete, thus producing a
sudden improvement of scientific knowledge.
12
|
1111.6493 | 1 | 1111 | 2011-11-22T15:38:02 | Depressive patients are more impulsive and inconsistent in intertemporal choice behavior for monetary gain and loss than healthy subjects- an analysis based on Tsallis' statistics | [
"q-bio.NC"
] | Depression has been associated with impaired neural processing of reward and punishment. However, to date, little is known regarding the relationship between depression and intertemporal choice for gain and loss. We compared impulsivity and inconsistency in intertemporal choice for monetary gain and loss (quantified with parameters in the q-exponential discount function based on Tsallis' statistics) between depressive patients and healthy control subjects. This examination is potentially important for advances in neuroeconomics of intertemporal choice, because depression is associated with reduced serotonergic activities in the brain. We observed that depressive patients were more impulsive and time-inconsistent in intertemporal choice action for gain and loss, in comparison to healthy controls. The usefulness of the q-exponential discount function for assessing the impaired decision-making by depressive patients was demonstrated. Furthermore, biophysical mechanisms underlying the altered intertemporal choice by depressive patients are discussed in relation to impaired serotonergic neural systems.
Keywords: Depression, Discounting, Neuroeconomics, Impulsivity, Inconsistency, Tsallis' statistics | q-bio.NC | q-bio | 1
[ http://node.nel.edu/?node_id=7412 NEL 2008]
Depressive patients are more impulsive and inconsistent in intertemporal
choice behavior for monetary gain and loss than healthy subjects- an analysis
based on Tsallis' statistics.
Taiki Takahashi1, Hidemi Oono2, Takeshi Inoue3, Shuken Boku3, Yuki Kako 3, Yuji
Kitaichi3, Ichiro Kusumi3, Takuya Masui3, Shin Nakagawa3, Katsuji Suzuki3, Teruaki
Tanaka3, Tsukasa Koyama3, and Mark H. B. Radford4
1Direct all correspondence to Taiki Takahashi, Unit of Cognitive and Behavioral Sciences
Department of Life Sciences, School of Arts and Sciences, The University of Tokyo,
Komaba, Meguro-ku, Tokyo, 153-8902, Japan ([email protected]).
2Department of Behavioral Science, Faculty of Letters, Hokkaido University, N.10, W.7,
Kita-ku, Sapporo, 060-0810, Japan
3Department of Psychiatry, Graduate School of Medicine, Hokkaido University, N.15,
W.7, Kita-ku, Sapporo, 060-8638
4Symbiosis Group Limited, P.O. Box 1192, Milton, 4064 Australia; and Department of
Behavioral Science, Hokkaido University, N.10, W.7, Kita-ku, Sapporo, 060-0810,
Japan
Acknowledgements: The research reported in this paper was supported by a grant from
the Grant- in-Aid for Scientific Research (“21st century center of excellence” grant) from the
Ministry of Education, Culture, Sports, Science and Technology of Japan.
2
Abstract
Depression has been associated with impaired neural processing of reward and
punishment. However, to date, little is known regarding the relationship between
depression and intertemporal choice for gain and loss. We compared impulsivity and
inconsistency in intertemporal choice for monetary gain and loss (quantified with
parameters in the q-exponential discount function based on Tsallis' statistics) between
depressive patients and healthy control subjects. This examination is potentially
important for advances in neuroeconomics of intertemporal choice, because depression
is associated with reduced serotonergic activities in the brain. We observed that
depressive patients were more impulsive and time-inconsistent in intertemporal choice
action for gain and loss, in comparison to healthy controls. The usefulness of the
q-exponential discount function for assessing the impaired decision-making by
depressive patients was demonstrated. Furthermore, biophysical mechanisms underlying
the altered intertemporal choice by depressive patients are discussed in relation to
impaired serotonergic neural systems.
Keywords: Depression, Discounting, Neuroeconomics, Impulsivity, Inconsistency,
Tsallis' statistics
3
1. Introduction
1.1 Impulsivity and depression
Depressive patients have neuropsychological impairments including deficits in
sensitivity to reward and punishment [1] and decision-making [2]. It is also known that
altered serotonergic neural systems are associated with depression and related
decision-making impairments [3]. Because depressive disorders are strongly associated
with impulsive and risky behavior such as suicide attempts [4,5], it is of importance to
establish a rigorous behavioral framework for assessing the degrees of impairments in
decision-making by depressive subjects. In this study, we examined two distinct
neuropsychological tendencies; i.e., impulsivity and inconsistency in intertemporal
choice (delay discounting) by clinically diagnosed depressive patients. Briefly,
impulsivity in intertemporal choice refers to the degree of preference for (or aversion to)
smaller sooner rewards (punishments) over later larger ones; while inconsistency in
intertemporal choice refers to time-dependency of the impulsivity in intertemporal
choice (we will illustrate these two tendencies in the next section). Because previous
studies indicate that serotonergic functioning in the brain (known to be reduced in
depressed patients) may be related to the evaluation of future rewards [6], these
investigations are of potential importance for neuroeconomic understandings of
depression and biophysical mechanisms of serotonergic systems underlying impaired
decision-making by mood disorder patients. However, to date, little is known about the
relationships between depression, impulsivity and inconsistency in intertemporal choice,
partly due to a difficulty in operationalizing impulsivity and inconsistency in
intertemporal choice in a distinct manner.
Recent studies in econophysics and neuroeconomics have demonstrated the
usefulness of the q-exponential discount function [7-9] to parametrize both impulsivity
and inconsistency in intertemporal choice. We therefore compared both impulsivity and
inconsistency in intertemporal choice for monetary gains and losses between depressive
patients and healthy control subjects, by utilizing the q-exponential discount function. It
is further to be noted that this study is the first to compare both impulsivity and
time-inconsistency in intertemporal choice between control and neuropsychiatric patient
groups by utilizing the q-exponential discount function based on Tsallis' statistics.
1.2 Impulsivity and inconsistency in intertemporal choice
Impulsivity in intertemporal choice (delay discounting) for gain refers to preference for
smaller but more immediate rewards over larger but more delayed ones. Consider the
following example:
4
(A) Choose between (A.1) One cup of coffee now.
(A.2) Two cups of coffee tomorrow.
(B) Choose between (B.1) One cup of coffee in one year.
(B.2) Two cups of coffee in [one year plus one day].
Most people may prefer a sooner smaller reward in (A) (i.e., (A.1), impulsive choice);
while prefer a later larger reward in (B) (i.e., (B.2), patient choice). These examples
demonstrate that most people are patient in making a plan about intertemporal choice in
the distant future, but impulsive in the near future, resulting in "preference reversal" as
time passes [10-12]. It is to be noted that consistent decision-makers should choose
either [(A1) and (B1)] or [(A2) and (B2)] because time-intervals between sooner smaller
rewards and later larger ones are the same (i.e., one day) in both (A) and (B). In
summary, impulsivity in delay discounting of gain corresponds to the degree to which
the subject discount the delayed reward; while inconsistency in delay discounting
corresponds to the dependency of the intensity of aversion to waiting for one day to
obtain an additional cup of coffee, on time-points (now or 1 year later) in the examples
above. It is also to be noted that example A is about the actual intertemporal choice
action; on the other hand, example B is about the intertemporal choice plan. Note that
people cannot actually take future actions now, and therefore choosing (B2) is a future
plan (not an actual action); while choosing (A1) is an actual action. Neuroeconomic
studies have reported that addiction to drugs of abuse is associated with impulsive
intertemporal choice for gain (e.g. choosing immediate rewards from drug intake at the
cost of later larger rewards such as healthy body in later life) [13-16]; while little is
known about the relationships between neuropsychiatric illnesses including addiction
and inconsistency in intertemporal choice.
In intertemporal choice for loss, discounting of delayed loss corresponds to a
decrease in aversion to loss when the loss is delayed. In other words, subjects who avoid
paying small costs immediately and choose to pay larger later costs are strong delay
discounters of loss. Therefore, strong delay discounting of loss (impulsivity in delay
discounting of loss) represents the marked tendency of procrastination about paying a
cost.
1.3 q-exponential discount function based on Tsallis' statistics
Recent econophysical and neuroeconomic studies [7-9] proposed and examined
the following q-exponential discount function based on Tsallis' statistics:
5
V(D)=A/ expq(kqD)=A/[1+(1-q)kqD] 1/(1-q) (Equation 1)
where expq () is the q-exponential function, D is a delay until receipt of a reward, A is
the value of a reward at D=0, and kq is a parameter of impulsivity at delay D=0. Note
that when q=0, equation 1 is the same as a simple hyperbolic discount function [10-12]:
V(D)=A/(1+khD), (Equation 2),
where kh=k0, while q→1, is the same as an exponential discount function proposed in
classical economics [17]:
V(D)=Aexp(-keD), (Equation 3)
where ke=k1.
In any continuous time-discounting functions, a discount rate is defined as
-(dV(D)/dD)/V(D), independently of functional types of discount models, and larger
discount rates corresponds to more impulsive intertemporal choice. In the q-exponential
discount function, the discount rate (q-exponential discount rate) is:
-V’(D)/V(D)=kq/(1+kq(1-q)D). (Equation 4)
We can see that when q=1, the discount rate is independent of delay D, corresponding to
exponential discounting (consistent intertemporal choice); while for q<1, the discount
rate is a decreasing function of delay D, resulting in preference reversal. This can also
be demonstrated by calculating the time-derivative of the q-exponential discount rate:
(d /dD) [-V’(D)/V(D)]= kq
2(1-q)/(kq(1-q)D+1)2 (Equation 5)
which is negative and positive for q<1 and q>1, respectively. Also, impulsivity at delay
D=0 is equal to kq irrespective of q. We have previously shown that the q-exponential
discount function is capable of continuously quantifying human subjects' inconsistency
in intertemporal choice [8,9]. Namely, human agents with smaller q values are more
inconsitent in intertemporal choice. If q is less than 0, the intertemporal choice behavior
is more inconsistent than hyperbolic discounting.
1.4 Objectives of the present study
The aim of this study was to examine impulsivity (kq in Equation 1) and inconsistency
(Equation 2) in delay discounting of gain and loss among depressed patients, in
comparison to healthy people. Based on previous research that has suggested higher
6
impulsivity among depressed patients, we propose the hypothesis that the depressed
patients are more impulsive in intertemporal choice behavior than healthy normal
subjects. It is important to note that our present study is the first to compare
inconsistency in intertemporal choice between healthy controls and neuropsychiatric
patients. Notably, previous findings regarding the relationship between serotonin and
discounting have been mixed. A rodent animal model reported that a reduction in
serotonergic activities were associated with exaggerated impulsivity in intertemporal
choice over several seconds [18]; while a psychopharmacological study with human
subjects did not observe a significant effect of a reduction in serotonergic activities on
intertemporal choice over a year [19]. How our present study resolves the discrepancy
between these studies is also discussed after presentation of our experimental data.
Furthermore, we also aimed to examine the differences in impulsivity and
inconsistency between discounting delayed gain and loss. The rationale is that no study
to date examined the effect of the sign of delayed outcomes (i.e., gain or loss) on
inconsistency in delay discounting, although it has been reported that delayed gain is
more rapidly discounted than delayed loss (i.e., the sign effect in intertemporal choice
[12]).
2. Methods
2.1 Participants
Participants were 29 depressive patients diagnosed with DSM-IVTR (major depressive
disorder and bipolar disorder, most recent episode: depressed) and 15 healthy control
subjects. The depressive patient participants were patients of Hokkaido University
Hospital, which serves a large urban catchments area. Participants’ ages ranged from 27
to 67, with a mean age of 43.83 years (SD = 2.25), and healthy controls’ ages ranged
from 31 to 71 with a mean age of 47.6 years (SD = 3.53). All participants including
depressive patients and healthy control subjects filled an informed consent form before
starting the experiment. The group of depressive patients consisted of both unipolar (i.e.,
patients with only a major depression phase) and bipolar (i.e., patients with both
depressive and manic phases) disorders. It is important to note that bipolar disorder
patients in a depressed state at the time of their participation (not in a manic state) were
included in the present study, in order to exclude the influences of manic mental states
on intertemporal choice. Consequently, there was no significant difference in
intertemporal choice behavior between unipolar and bipolar disorder patients. Therefore,
we combined both unipolar and bipolar patients (referred to as "depressive patients",
hereafter). Moreover, most patients with depression were under medical treatment with
7
several types of antidepressants/mood stabilizers. We did not, however, find any
significant effect of the types of antidepressants/mood stabilizers on intertemporal
choice. Therefore, we did not divide the patients into sub-groups according to the types
of antidepressants they intake. The characteristics of these subjects are described in
more detail in Table 1. Subjects with past or current illegal drug use, assessed with
Mini-International Neuropsychiatric Interview (MINI, see below), were excluded from
the present study, in order to avoid the influences of substance abuse on intertemporal
choice.
2.2 Intertemporal choice task
As a standard instrument measuring degrees to which participants discount delayed
reward and loss, we conducted a face-to-face task developed by
neuropsychopharmacologists Bickel and colleagues [20]. It is to be noted that we have
also utilized this task in our previous studies [8,9,21].
First, participants were seated individually in a quiet room, and faced the
experimenter across a table. After that, participants received the simple instruction that
monetary rewards (or losses) in this experiment were hypothetical, but the experimenter
wanted them to think as though they were real money. Then the participants were asked
to choose between the card describing money delivered immediately (or paid
immediately, in the loss condition) and the card describing money delivered after certain
delay (or paid after a certain delay, in the loss condition). The left card viewed by
participants indicated the amounts of money that could be received immediately (or that
had to be paid immediately, in the loss condition), and the right card indicated 100,000
yen that could be received after a certain delay (or that had to be paid after a certain
delay, in the loss condition.
For the delay discounting tasks, monetary rewards (or losses) and the delay time were
printed on 3×5 index cards. The 27 monetary amounts were 100,000 yen (about
$1,000), 99,000 yen, 96,000 yen, 92,000 yen, 85,000 yen, 80,000 yen, 75,000 yen,
70,000 yen, 65,000 yen, 60,000 yen, 55,000 yen, 50,000 yen, 45,000 yen, 40,000 yen,
35,000 yen, 30,000 yen, 25,000 yen, 20,000 yen, 15,000 yen, 10,000 yen, 8,000 yen,
6,000 yen, 4,000 yen, 2,000 yen, 1,000 yen, 500 yen and 100 yen. The seven time
delays were 1 week, 2 weeks, 1 month, 6 months, 1 year, 5 years and 25 years.
The experimenter turned the 27 cards sequentially. The card started with 100 yen, up to
100,000 yen in the ascending order condition, or started with 100,000 yen, down to 100
yen, in the descending order condition. For each card, participants chose either the
immediate or the delayed reward (or loss). The experimenter wrote down the first
8
delayed reward chosen in the descending gain condition, first immediate reward chosen
in the ascending gain condition, first immediate loss in the descending loss condition,
and first delayed loss in the ascending loss condition. The average of these results (gain
condition and loss condition were calculated separately), were used as the points or
subjective equality (hereafter called the indifference point, a subjective value of delayed
gain/loss) in the following analyses. This procedure was repeated for each of the seven
delays (for more detail, see Bickel et al., 1999). The four conditions (ascending gain
condition, descending gain condition, ascending loss condition, and descending loss
condition) were conducted randomly for each participant. These conditions of the order
of the delay discounting tasks did not significantly influence the results. Because our
aim was to compare the group difference between healthy controls and depressive
patients, we presented estimated parameters of the q-exponential discounting (see the
next section) for group median data of the indifference point at each delay. It should
also be noted that when the analysis was performed at the individual level, essentially
the same conclusions were obtained.
After the determination of indifference points in the delay discounting tasks,
we estimated parameters for the q-exponential discount function (i.e., kq and q in
Equation 1, corresponding to impulsivity and inconsistency, respectively) for gain and
loss, separately. For estimating the parameters, we conducted nonlinear curve fitting
with the Gauss-Newton algorithm implemented in R statistical language (nonlinear
modeling package).
2.3 Questionnaires
Mini-International Neuropsychiatric Interview (MINI)
To examine the participants’ drug abuse and alcohol abuse, we used MINI [22]. This
scale is also a structured diagnostic scale, administered by well-trained psychiatric
doctors when interviewing participants. Although this scale can divided into several
parts for assessing specific psychiatric disorders, we used only 2 parts – those for
assessing drug abuse and alcohol abuse.
Beck Depression Inventory-II (BDI-II)
To assess both healthy and depressive participant’s degrees of depressive tendency, we
assessed Beck’s Depression Inventory (more specifically, BDI-II) [23,24]. This scale is
a commonly used self-report scale that measures severity of depression ‘over the past
week’. The scale consists of 21 items that describe core depressive symptoms, with
9
score being rated on a 4-point scale.
In this study, we used the Japanese version of the BDI-II [25,26].
Demographic questionnaire
In addition to the above measures, participants completed questionnaires about sex,
age, history of smoking, and suicide attempt histories. These demographic data did not
significantly affect intertemporal choice behavior in the present study. Therefore, we
simply compared parameters in the q-exponential discount function for gain and loss
between healthy controls and depressed patients.
2.4 Data analysis
For statistically testing differences in estimated parameters in the q-exponential
discount function at the individual level, we utilized t-tests. It is also to be noted that we
examined the fitness of the discount models with AICc (Akaike Information Criterion
with small sample correction) and observed that the q-exponential discount function
best fitted among other models (i.e., exponential and hyperbolic functions) in both
healthy controls and depressive patients.
All statistical procedures were conducted with R statistical language
(http://www.r-project.org/). Significant level was set at 5% throughout.
3. Results
3.1 Impulsivity and inconsistency in controls and depressive patients
Demographic characteristics of depressive patients and healthy controls are presented in
Table 1. We observed that BDI-II scores were higher for depressive patients than
controls (p<0.05), verifying that our present populations were appropriate for the
objectives of the present study.
The estimated parameters in the q-exponential discount function for gain and
loss at the group level are summarized in Table 2 (Table 2-1 for gain, Table 2-2 for loss).
Fitted q-exponential curves for group median indifference points were presented in Fig.
1 (Fig. 1A for gain, Fig.2B for loss). We can see that depressive patients less
dramatically discounted delayed outcomes in the distant future, compared to healthy
controls, implying that depressive patients make more patient (less impulsive)
intertemporal choice plans in the distant future. However, the patient plan does not
imply that their actual intertemporal choice action is also patient (less impulsive),
because it is possible for preference reversal to occur due to inconsistency in
10
intertemporal choice (see section 1.2). In order to investigate differences in both
impulsivity at delay D=0, (i.e., impulsivity in intertemporal choice action) and
inconsistency in delay discounting of gain and loss by controls and depressive patients
at the individual level, we conducted t-tests on estimated kq (impulsivity at delay D=0)
and q (time-consistency) between controls and depressive patients. Consequently, we
found that depressive patients had significantly larger kq (impulsivity at delay D=0) and
smaller q (time-consistency) for both gain and loss, in comparison to controls (ps<0.05).
In other words, depressive patients were more impulsive and inconsistent in delay
discounting actions for both gain and loss than healthy control subjects (see Table 2, for
group data). Moreover, as noted earlier, there was no significant effect of demographic
variables other than depressive status (e.g. sex, age, histories of suicide attempts, and
the type of antidepressants) on the estimated parameters in the q-exponential discount
function.
3.2 Gain-loss asymmetries
Next, we plotted the estimated q-exponential discount rate defined in Equation 4 (for
gain and loss) of healthy controls and depressive patients (Fig. 2). As can be seen from
Fig.2, depressive patients were more impulsive (i.e., a larger discount rate) in the near
future, but less impulsive (i.e., a smaller discount rate) in the distant future, in
comparison to healthy controls. This indicates that depressive patients may experience
more exaggerated "preference reversal" in their intertemporal choice between plans and
actions. It is important to note that depressive patients' discount rate for future loss at
the delay of several decades is almost zero, indicating that depressive patients had
hypersensitivity to potential bad outcomes in the extremely distant future (a red curve,
Fig 2B).
In order to examine differences in impulsivity and inconsistency in
intertemporal choice for gain and loss at the individual level, we conducted t-tests on
estimated kq (impulsivity at delay D=0) and q (time-consistency) between gain and loss.
We then observed that kq was significantly smaller for gain; while q was larger for gain
(p<0.05), indicating that intertemporal choice for loss is less impulsive (in line with
previous reports on the sign effect on a discount rate [12]) but more inconsistent (see
Table 2 for group data).
4. Discussion
4.1 Impulsivity and inconsistency in depressive patients' intertemporal choice
As far as we know, this study is the first to utilize the q-exponential discount model
11
inspired by Tsallis' non-extensive thermostatistics, in order to examine impairments in
depressive patients' decision-making. We observed that depressive patients were more
impulsive in intertemporal choice actions and more inconsistent in temporal discounting,
in comparison to healthy subjects. Our results imply that (i) depressive patients may
experience preference reversal more frequently and dramatically, than healthy subjects,
(ii) depressive patients may be more sensitive to potentially harmful events occurring in
the extremely distant future, than healthy subjects, and (iii) although depressive patients'
plans about the distant future may be forward-looking, their intertemporal choice
actions (occurring at delay D=0) may be more myopic than healthy controls. Future
studies should examine whether subjects who tend to make infeasible future plans are
more susceptible to depression. Another possibility is that neuronal changes associated
with depression (e.g., hypoactivation of serotonergic systems) may induce
time-inconsistency in temporal discounting. Concerning this possibility, several studies
have indicated that time-inconsistency in temporal discounting may result from
nonlinear distortion of psychological time [27-29].
With respect to biophysical mechanisms underlying serotonergic modulation of
time-perception in intertemporal choice, it should be noted that (i) biophysical
simulation studies indicate that nonlinear psychophysical effects on sensation is
mediated by electrical coupling between neurons [30] and (ii) serotonin modulates
neuronal coupling [31]. Therefore, future neuroeconomic and biophysical studies should
examine whether a decrease in serotoninergic activities induces both distorted
time-perception and time-inconsistent discounting behavior. As noted in the
introduction, findings on the relationship between serotonin and discounting have been
mixed. Mobini and colleagues' study [18] demonstrated a reduction in serotonergic
activities resulted in a significant increase in a discount rate; while Crean and
colleagues' study [19] reported no significant effect. It should be noted that in Mobini
and colleagues' study, the time-range of the intertemporal choice task was short; while
Crean and colleagues' study employed longer delays , i.e., about 1 year (note that both
Mobin et al.'s and Crean et al's studies did not employ the q-exponential discount
function but a simple hyperbolic function [10-12], and therefore the estimated simple
hyperbolic discount rates in Crean et al's study were under the influences of both
relatively longer and shorter delays). Our present results of the relationship between
depression (associated with a reduction in serotonergic activities) and temporal
discounting may resolve the discrepancy: a decrease in serotonergic activities may
increase a discount rate at short delays; while decrease a discount rate at longer delays
(see Figure 2). This analysis is impossible without utilizing the q-exponential discount
12
function.
A recent study reported that intertemporal choice for other is more inconsistent
than for self [9]. It is important to examine whether intertemporal choice for other by
healthy subjects involves similar neuropsychological processes to intertemporal choice
for self by depressive patients. Neurochemically, a previous study implied that reduced
(nor)adrenergic activities (assessed with salivary alpha-amylase levels) are associated
with impulsivity in intertemporal choice [32]. It should therefore be examined whether
(nor)adrenergic activities are likewise related to consistency in temporal discounting.
4.2 Sign effects on intertemporal choice
We observed that the signs of the outcomes (i.e., gain or loss) markedly affect both
impulsivity and inconsistency in intertemporal choice (i.e., kq and q). Although several
neuroeconomic studies examined the neural correlates of discounting delayed monetary
gains [33], little is known regarding the neural processing underlying discounting of
monetary loss. Future studies should examine which neuro-biophysical processes
mediate the observed sign effects.
4.3 Limitations and future directions
As noted, the depressive patients in the present study were medicated with several types
of antidepressants/mood stabilizers. Therefore, it is not completely evident that the
treatments did not dramatically affect the intertemporal choice behavior. Nevertheless,
our present results may not totally be attributable to the effects of antidepressants,
because (i) we did not observe the effects of the types of antidepressants on the
parameters in the q-exponential discount function, and (ii) BDI-II scores were
significantly higher in the depressive patients than the controls, indicating that the
patients were still depressed enough to elicit depression-induced alteration in
intertemporal choice (we carefully scheduled our experiment so that the patients may
participate during their depressive phase). In order to further resolve these issues, future
studies should examine the effects of antidepressants on intertemporal choice by healthy
human subjects (or animals).
References
1. Must A, Szabó Z, Bódi N, Szász A, Janka Z, Kéri S. (2006) Sensitivity to reward and
punishment and the prefrontal cortex in major depression. J Affect Disord.
90(2-3):209-215.
2. Radford MH, Mann L, Kalucy RS. (1986) Psychiatric disturbance and
13
decision-making. Aust N Z J Psychiatry. 20(2):210-217.
3. Must A, Juhász A, Rimanóczy A, Szabó Z, Kéri S, Janka Z. (2007) Major depressive
disorder, serotonin transporter, and personality traits: why patients use
suboptimal decision-making strategies? J Affect Disord. 103(1-3):273-276
4. Lejoyeux, M., Arbaretaz, M., McLoughlin, M., & Ades, J. (2002). Impulse control
disorders and depression. The Journal of Nervous and Mental Disease, 190,
310-314.
5. Kim, C., Lesage, A., Seguin, M., Lipp, O., Vanier, C., & Turecki, G. (2003). Patterns
of comorbidity in male suicide completers. Psychological Medicine, 33,
1299-1309.
6. Schweighofer N, Tanaka SC, Doya K. (2007) Serotonin and the evaluation of future
rewards: theory, experiments, and possible neural mechanisms. Ann N Y Acad
Sci. 1104:289-300.
7. Cajueiro D.O. A note on the relevance of the q-exponential function in the context of
intertemporal choices. Physica A 364 (2006) 385–388
8. Takahashi T, Oono H, Radford MH. (2007) Empirical estimation of consistency
parameter in intertemporal choice based on Tsallis' statistics. Physica A 381:
338-342
9. Takahashi T (2007) A comparison of intertemporal choices for oneself versus
someone else based on Tsallis' statistics. Physica A 385: 637-644.
10. Ainslie, G (2005) Précis of Breakdown of Will. Behav Brain Sci. 28(5):635-650;
discussion 650-73. Review.
11. Mazur, J. E. (1987). An adjusting procedure for studying delayed reinforcement. In:
Commons, M. L., Mazur, J. E., Nevin, J. A., & Rachlin, H. (Eds), Quantitative
analysis of behavior: the effect of delay and of intervening events on
reinforcement value. Vol. 5, Hillsdale, NJ: Erlbaum, 55-73.
12. Frederick, S., Loewenstein, G., & O' Donoghue, T. (2002) Time discounting and
time preference: a critical review. Journal of Economic Literature. 40, 350-401.
13. Madden, G. J., Petry, N. M., Badger, G. J., & Bickel, W. K. (1997). Impulsive and
self-control choices in opioid-dependent patients and non-drug-using control
participants: drug and monetary rewards. Exoerimental and Clinical
Psychopharmacology, 5, 256-262.
14. Madden, G. J., Bickel, W. K., & Jacobs, E. A. (1999). Discounting of delayed
rewards in opioid-dependent outpatients: exponential or hyperbolic discounting
functions. Experimental and Clinical Psychopharmacology, 7, 284-293.
15. Petry, N. M., Bickel, W. K., & Arnett, M. (1998). Shortened time horizons and
14
insensitivity to future consequences in heroin addicts. Addiction, 93, 729-738.
16. Kirby, K. N., Petry, N. M., & Bickel, W. K. (1999). Heroin addicts have higher
discount rates for delayed rewards than non-drug-using controls. Journal of
Experimental Psychology: General, 128, 78-87.
17. Samuelson, P. A. (1937). A note on measurement of utility. Review of Economic
Studies, 4, 155-161.
18. Mobini S, Chiang TJ, Al-Ruwaitea AS, Ho MY, Bradshaw CM, Szabadi E. (2000)
Effect of central 5-hydroxytryptamine depletion on inter-temporal choice: a
quantitative analysis. Psychopharmacology (Berl). 149(3):313-318.
19. Crean J, Richards JB, de Wit H. (2002) Effect of tryptophan depletion on impulsive
behavior in men with or without a family history of alcoholism. Behavioral Brain
Research. 136(2):349-357.
20. Bickel WK, Odum AL, Madden GJ. (1999) Impulsivity and cigarette smoking:
delay discounting in current, never, and ex-smokers. Psychopharmacology (Berl).
146: 447-454.
21. Ohmura Y, Takahashi T, Kitamura N, Wehr P. (2006) Three-month stability of delay
and probability discounting measures. Experimental & Clinical
Psychopharmacology. 14(3):318-328.
22. Sheehan DV, Lecrubier Y, Sheehan KH, Amorim P, Janavs J, Weiller E, Hergueta T,
Baker R, Dunbar GC.(1998) The Mini-International Neuropsychiatric Interview
(M.I.N.I.): the development and validation of a structured diagnostic psychiatric
interview for DSM-IV and ICD-10. Journal of Clinical Psychiatry. 59 Suppl.
20:22-33
23. Beck AT, Erbaugh J, Ward CH, Mock J, Mendelsohn M (1961) An inventory for
measureing depression. Archives of General Psychiatry 4 : 561
24. Beck AT, Young J (1979) Handbook of studies on depression. Journal of Nervous
and Mental Disease. 167: 719
25. Hiroe T, Kojima M, Yamamoto I, Nojima S, Kinoshita Y, Hashimoto N, Watanabe N,
Maeda T, Furukawa TA. (2005) Gradations of clinical severity and sensitivity to
change assessed with the Beck Depression Inventory-II in Japanese patients with
depression. Psychiatry Research. 135(3):229-235.
26. Matsumoto T, Yamaguchi A, Chiba Y, Asami T, Iseki E, Hirayasu Y. (2005)
Self-burning versus self-cutting: patterns and implications of self-mutilation; a
preliminary study of differences between self-cutting and self-burning in a
Japanese juvenile detention center. Psychiatry Clin Neurosci. 59(1):62-69.
27. Takahashi, T. (2005). Loss of self-control in intertemporal choice may be
15
attributable to logarithmic time-perception. Medical Hypotheses, 65, 691-693.
28. Takahashi, T. (2006). Time-estimation error following Weber-Fechner law may
explain subadditive time-discounting. Medical Hypotheses, 67, 1372-1374.
29. Takahashi, T. (2007). Hyperbolic discounting may be reduced to electrical coupling
in dopaminergic neural circuits. Medical Hypotheses, 69(1):195-198.
30. Copelli, M., Roque, A.C. Oliveira R.F. and Kinouchi, O. (2002) Physics of
psychophysics: Stevens and Weber–Fechner laws are transfer functions of
excitable media, Physical Review E, 65, 060901
31. Burrell BD, Sahley CL, Muller KJ. (2002) Differential effects of serotonin enhance
activity of an electrically coupled neural network. Journal of Neurophysiology
87(6):2889-2895.
32. Takahashi T, Ikeda K, Fukushima H, Hasegawa T. (2007) Salivary alpha-amylase
levels and hyperbolic discounting in male humans. NeuroEndocrinology Letters.
28(1):17-20.
33. Kable JW, Glimcher PW. (2007) The neural correlates of subjective value during
intertemporal choice. Nature Neuroscience. 10(12):1625-133.
16
Figure legends
Fig. 1 Indifference points and fitted q-exponential discount function for gain (A) and
loss (B). Black and red dots are indifference points at delays for healthy controls
and depressive patients, respectively. Black and red curves are best-fit
q-exponential discount functions for healthy controls and depressive patients,
respectively. Note that in (B), the vertical axis is the unsigned (absolute)
subjective value of delayed monetary loss.
Fig. 2 Estimated q-exponential discount rate: kq/(1+kq(1-q)D) for gain (A) and loss (B).
Black and red curves are q-exponential discount rates for healthy controls and
depressive patients, respectively.
17
Table 1. Means and standard deviations for demographic variables for depressed patients and healthy
people
Sex (% men)
Age (years)
BDI-II
Depressive patients
Mean
55.18
43.83
20.52*
SD
2.25
2.42
Healthy controls
Mean
40
47.6
8.93
SD
3.53
1.12
*:Significantly larger than healthy controls (p<0.05). BDI-II: Beck Depression Inventory (high
BDI-II scores indicate severe depression).
18
Table 2-1. Group data of estimated parameters in q-exponential discounting for gain
kq (impulsivity)
q (consistency)
Depressive patients
0.0006099
-2.4252055
>
<
Healthy controls
0.0004064
-0.2132638
Note that larger kq and q indicate more impulsive intertemporal choice at delay D=0 and
more consistent intertemporal choice, respectively.
Table 2-2. Group data of estimated parameters in q-exponential discounting for loss
kq (impulsivity)
q (consistency)
Depressive patients
0.0005576
-72.5
>
<
Healthy controls
0.0002696
-1.7333967
Note that larger kq and q indicate more impulsive intertemporal choice at delay D=0 and
more consistent intertemporal choice, respectively.
Fig.1A
19
20
Fig.1B
21
Fig2.A
22
Fig.2B
|
1912.02331 | 1 | 1912 | 2019-12-05T01:17:30 | Investigation of ephaptic interactions in peripheral nerve of sheep using 6 kHz subthreshold currents | [
"q-bio.NC"
] | The objective of this work was to determine whether application of subthreshold currents to the peripheral nerve increases the excitability of the underlying nerve fibres, and how this increased excitability would alter neural activity as it propagates through the subthreshold currents. Experiments were performed on two Romney cross-breed sheep in vivo, by applying subthreshold currents either at the stimulus site or between the stimulus and recording sites. Neural recordings were obtained from nerve cuff implanted on the peroneal or sciatic nerve branches, while stimulus was applied to either the peroneal nerve or pins placed through the lower hindshank. Results showed that subthreshold currents applied to the same site as stimulus increased excitation of underlying nerve fibres (p < 0.0001). With stimulus and subthreshold currents applied to different sites on the peroneal nerve, the primary CAP in the sciatic displayed a temporal shift of -2.5 to -3 us which agreed with statistically significant changes in the CAP waveform (p<0.02). These findings contribute to the understanding of mechanisms in myelinated fibres of subthreshold current neuromodulation therapies. | q-bio.NC | q-bio | 1
Investigation of ephaptic interactions in peripheral nerve of sheep using 6
kHz subthreshold currents
James Hope1,2, Narrendar Ravi Chandra1, Frederique Vanholsbeeck2,3, Andrew McDaid1
1The Department of Mechanical Engineering, The University of Auckland, Auckland 1010, New Zealand
2The Dodd Walls Centre for Photonic and Quantum Technologies, Auckland 1010, New Zealand
3The Department of Physics, The University of Auckland, Auckland 1010, New Zealand
Abstract
The objective of this work was to determine whether application of subthreshold currents to the peripheral nerve
increases the excitability of the underlying nerve fibres, and how this increased excitability would alter neural
activity as it propagates through the subthreshold currents. Experiments were performed on two Romney cross-
breed sheep in vivo, by applying subthreshold currents either at the stimulus site or between the stimulus and
recording sites. Neural recordings were obtained from nerve cuff implanted on the peroneal or sciatic nerve
branches, while stimulus was applied to either the peroneal nerve or pins placed through the lower hindshank.
Results showed that subthreshold currents applied to the same site as stimulus increased excitation of underlying
nerve fibres (p < 0.0001). With stimulus and subthreshold currents applied to different sites on the peroneal nerve,
the primary CAP in the sciatic displayed a temporal shift of -2.5 to -3 µs which agreed with statistically significant
changes in the CAP waveform (p<0.02). These findings contribute to the understanding of mechanisms in
myelinated fibres of subthreshold current neuromodulation therapies.
Keywords: Neuromodulation, ephaptic interactions, nerve cuff, peripheral nerve
1 Introduction
Neuromodulation therapies provide an alternative treatment modality for several drug-resistant neurological
conditions, including Parkinson's, epilepsy, and depression [1-6]. While neuromodulation therapies generally
administer suprathreshold currents to modulate activity in target neural cells, a smaller number administer
subthreshold currents to either suppress or promote neural activity through partial hyperpolarization or
depolarization, respectively, of cell membranes. For example, chronic subthreshold cortical stimulation for
epilepsy [7, 8]; a high frequency (10 kHz) variant of spinal cord stimulation for pain management [2]; the
conditioning current which precedes the suprathreshold current in transcranial magnetic stimulation for motor
cortex studies [9]; and, transcranial current stimulation for depression [6].
1
2
Researchers have demonstrated this partial polarization effect on the vestibular system by administering
subthreshold current, with a band pass filtered random noise waveform, to the mastoid processes and observing
subject sway responses which were highly coherent with the applied current polarity and magnitude [10-12]. This
coherence was explained by stochastic resonance, wherein subthreshold components of periodic stimulus and
random noise stimulus sum to become suprathreshold [13, 14]. In two further studies [15, 16], when researchers
applied a subthreshold current transcutaneously to the tibial nerve through electrodes placed proximal to the ankle,
again with a band pass filtered random noise waveform, participants reported an increase in sensitivity to vibration
applied to the foot. The authors of these latter two studies did not identify the underlying mechanism causing
increased sensitivity, but postulated in one [16] that the applied current might increase synchrony of the sensory
receptor action potentials in the tibial nerve due to augmentation of ephaptic interactions between active nerve
fibres.
In ephaptic interactions, spatiotemporal variations in the electric field around an active neural cell influences
activity in nearby neural cells by altering their membrane potentials. These interactions can improve synchrony
of activity within populations of neural cells or triggering of neural activity in subthreshold cells, and may be
induced either by naturally occurring physiological effects [17-19] or by artificially increasing excitability using
chemicals [20, 21] and subthreshold electrical current [18, 22, 23]. In myelinated nerve fibres, modelling studies
predict ephaptic interactions alter the propagation velocities of action potentials in neighboring active fibres which
improves synchrony [24-27], and an in vivo study on rat showed increased activity in response to an electrical
stimulus when it was temporally coupled with a compound action potential (CAP) [28]. Increasing synchrony and
localized triggering of new action potentials in peripheral nerves using a subthreshold current presents an exciting
prospect because such a paradigm could improve the signal to noise ratio in peripheral nerve interfaces, aid
physical rehabilitation after spinal cord injury and stroke, and provide insight into mechanisms of subthreshold
current neuromodulation therapies.
In the current study, we investigated augmentation of neural activity in hind limb of sheep in vivo by applying
subthreshold, 6 kHz, sinusoidal currents to the peroneal nerve between stimulus and recording sites. This paradigm
differs significantly from the in vivo rat study in [28] which evaluated changes in excitability induced by
surrounding neural activity, and is more similar to the stochastic resonance experiments in [15, 16], though in the
current study subthreshold currents were administered via nerve cuffs instead of transcutaneous electrodes, and
neural recordings were acquired. Two evoked stimulus sites were employed, one via a nerve cuff implanted on
the peroneal nerve, and the second via pin electrodes placed distal to the hock. We hypothesized that (1)
administering subthreshold currents would increase excitability of the underlying fibres, and (2) this increased
excitability would increase synchrony in the neural activity and alter the propagation velocities.
2 Materials and Methods
2.1 Experiment apparatus
Nerve cuff electrode arrays were fabricated from stainless steel foil and silicon using the method described in
[29]. The 28-channel nerve cuff contained a 2 rings of 14 electrodes, spaced 7 mm apart, with 0.46 x 3 mm active
2
3
area on each electrode, Fig. 1a. Electrodes were coated with poly(3,4-ethylenedioxythiophene):p-toluene
sulfonate (PEDOT-pTS) to reduce the electrode-tissue contact impedance. The 2-channel nerve cuffs contained
2 electrodes, each 9.5 x 1 mm, spaced 6 mm apart, Fig. 1b. The 28 channel and 2 channel electrode arrays were
glued (Smooth-On Sil-Poxy®) into elastomer cuffs with dimensions (L x O.D. x I.D) of 20 x 8 x 4.5 mm and 6 x
6 x 3 mm, respectively, and with a slit along one side to allow implant on the nerve .
The 28-ch nerve cuff used for recording was connected to a headstage (INTAN C3314) via an adaptor (INTAN
C3410), Fig. 1c. Neural recordings were acquired from each of the 14 electrodes on one electrode ring, with the
electrodes of the other ring shorted together and used as the reference. Data were low pass filtered at 5 kHz for
anti-aliasing, sampled at 20 kS/s, software notch filtered at 50 Hz to remove mains noise, then streamed via a USB
interface board (INTAN C3100) to host PC and saved as .rhd files for processing later.
Stimulation was administered either through the pins placed along the cannon bone, or through electrodes in the
most distally implanted nerve cuff on the peroneal nerve, and was generated by a bench-top pulse stimulator (AM-
systems 2100) when digitally triggered (National Instruments CompactRIO® NI9403), Fig. 1c. The digital trigger
was recorded on the USB interface board (INTAN C3100).
Subthreshold currents were generated using a PCB with parallel current-source circuits, developed by The EIT
Research Group at University College London, and available for download from https://github.com/EIT-team.
Currents were switched on and off using solid state relays (IXYS CPC1017N) controlled via digital lines (National
Instruments CompactRIO® NI9403), Fig 1c.
2.2 Tissue preparation and handling
All animal procedures were approved by The University of Auckland Animal Ethics Advisory Committee. In
total, two Romney cross breed sheep, female, and weighing 61 and 67 kg, were used in in-vivo experiments.
Anesthesia was induced using by intravenous injection, and maintained using a mixture of isofluorane, oxygen
and medical air administered via endotracheal tube. At the conclusion of experiments subjects were euthanized.
To implant nerve cuffs, once subjects were anesthetised, the left hind leg was extended and loosely fixed in
position around the hock and fetlock joint. An incision was made down the posterior side of the thigh to expose
underlying muscle, then the Semitendinosus and Biceps Femoris muscles were parsed apart, and the sciatic, tibial,
and peroneal nerves were isolated by cutting away adipose tissue [30]. In subject 1, two 28-channel nerve cuffs
were implanted adjacent to one another on the peroneal nerve, Fig. 1d. In subject 2, a 28-channel nerve cuff was
implanted on the sciatic nerve, and four 2-channel nerve cuffs were implanted adjacent to one another on the
peroneal nerve, Fig 1e. After implantation the muscle cavity was filled with physiological saline, preheated to 38
ºC, to cover the nerve cuffs and exposed nerve. A stainless steel pin was placed in this saline which connected via
a 460 kΩ resistor to ground. In both subjects, two pins were placed 100 mm apart, subdermally, along the cannon
bone to allow stimulation of distal sections of the peroneal nerve.
3
4
Electrodes
through 14
1
Electrodes 15
through 28
Elastomer cuff
(a)
1
15
14
28
Electrodes
Elastomer cuff
(b)
(c)
(d)
(e)
Figure 1: The 28 channel electrode array with two columns of 14 electrodes after PEDOT coating (left) and assembled
into an elastomer cuff (right) (a), and four 2 channel electrode arrays before (left) and after (right) assembly into an
elastomer cuff (b); black scale bars are 5 mm. A schematic of the experiment apparatus with solid lines for current
controlled signals, dashed for digital lines, and dotted for analogue voltage recording (c), where STC = subthreshold
current. Schematic of nerve cuff configuration in subject 1 (d) showing two 28-channel cuffs on the peroneal nerve,
and in subject 2 (e) showing four 2 channel cuffs on the peroneal nerve and a 28-channel cuff on the sciatic nerve.
2.3 Experiment protocol
The subthreshold currents were biphasic, 6 kHz, sinusoidal waveform, and, in all but one protocol, +/- 70 µA
amplitude. A 6 kHz sinusoid was selected to be sufficiently above the frequency components of neural activity to
be filtered out during signal processing, and because transient impedance studies have indicated that 6 kHz results
in more resistive current across the node of Ranvier and less capacitive current across the myelin sheath than in
the neighboring frequencies [31, 32].
The amplitude of the stimulus pulses were selected to be 50 % higher than that which produced onset of twitching
of the lower hindshank and phalanges, respectively, for stimulus administered via nerve cuff on the peroneal nerve
and pins apposing the cannon bone. Twitching onset was characterized by manually triggering stimuli and
increasing amplitude while visually monitoring the hind limb.
2.3.1 Hypothesis 1: Subthreshold current contribution to excitation of fibres
In subject 1, the contribution of subthreshold currents to excitation of fibres was tested by administering a pulse
+/-0.3 mA, 50µs/phase, biphasic, square pulse stimulus every 500 ms to the distal 28-ch nerve cuff, Fig. 1d,
through electrodes E11 and E23, Fig. 1a. In conjunction with this stimulus and on the same nerve cuff, no
subthreshold current was applied for 5 seconds, then one subthreshold current was applied through E1 and E15
4
5
for 5 seconds, then a second subthreshold current -- in phase with the first - was added through E4 and E18 for 5
seconds, Fig. 1a. This protocol was performed for a total of 600 seconds. The in-phase nature of the two currents
produces a sinusoid with a +/- 140 µA amplitude within the cuff.
2.3.2 Hypothesis 2: Pin stimulus
In subject 1, augmentation of neural activity using subthreshold currents was investigated using stimulus applied
through two pins inserted in the hindshank and apposing the cannon bone. Here, a +/-10 mA, 1 ms/phase, biphasic,
square pulse was administered, while subthreshold currents were applied and recordings acquired from the
peroneal nerve using the protocol described above in section 2.3.1. This protocol was performed for a total of 600
seconds with both stimulation pins inserted on the posterior side of the cannon bone, then repeated for a total of
600 seconds with the anodic pin inserted on the anterior side of the cannon bone. Finally, the protocol was repeated
for a total of 100 seconds with the stimulus duration halved to 0.5 ms/phase to distinguish neural activity and
stimulus artefacts in the neural recordings.
2.3.3 Hypothesis 2: Nerve cuff stimulus
In subject 2, augmentation of ephaptic interactions in the presence of three subthreshold currents was investigated
using stimulus applied through a nerve cuff. Here, a +/-0.3 mA, 100 µs/phase, biphasic, square pulse was
administered every 500 ms to the distal most 2-channel nerve cuff on the peroneal nerve, while no subthreshold
current was applied for 5 seconds, then subthreshold currents were added one at a time, for 5 seconds, to each of
the three adjacent 2-channel nerve cuffs, Fig. 1e. This protocol was repeated for a total of 600 seconds. The
adjacent nature of the 2 channel nerve cuffs means the sinusoid amplitude did not exceed +/- 70 µA at any point
along the peroneal nerve. Finally, to verify the influence of subthreshold current amplitude on the CAPs, the above
protocol was repeated for a total of 600 seconds with the subthreshold current amplitude reduced to +/- 30 µA.
2.4 Data processing
Recorded data were processed in MATLAB (R2018b Mathworks). Data were parsed into 500 ms duration
segments beginning with the digital trigger to the pulse stimulator, grouped into sets based on the number of
applied subthreshold currents, and then notch filtered at 6 kHz to remove artefacts from the subthreshold currents.
Signals with significant drift or corruption by artefacts, defined as having a mean value outside the bounds of -
100 to 100 µV in the temporal window 20 to 30 ms, were removed.
Two metrics were used to evaluate changes in neural activity caused by the subthreshold currents: CAP amplitude,
and CAP temporal shift. (i) CAP amplitude, defined as the difference between the maxima of the peak and the
minima of the adjoining trough, was calculated for each of the data segments then compared for each subthreshold
current condition using unpaired t-tests. (ii) CAP temporal shift was analysed by up-sampling the mean recorded
waveforms for each current condition by a factor of 100 using cubic spline interpolation to increase temporal
resolution, then comparing the upsampled waveforms using crosscorrelation. In addition, CAP amplitude and
CAP temporal shift were evaluated by calculating the difference in mean recorded waveforms, and identifying
points outside the +/-3 sigma noise threshold.
5
6
3 Results
3.1 Surgery and nerve cuff implantation
The femoral artery crossed over the sciatic, tibial and peroneal nerve branches, Fig. 2a, which hampered isolation
of nerve branches from surrounding adipose tissue, and implantation of the nerve cuffs. The peroneal nerve was
larger in subject 1, at 4 -- 5 mm diameter, than in subject 2, at 3 mm diameter, which was one reason that compelled
the use of different nerve cuff configurations between the two subjects Fig. 2b-c.
Tibial nerve
Peroneal nerve
28ch cuff
Femoral artery
Fetlock
joint
Stimulus
pins
Sciatic
nerve
Peroneal nerve
Four 2ch cuffs
Phalanges
Lower
hindshank
Cannon
bone
Hock
Figure 2: The sciatic, tibial, and peroneal nerve branches and femoral artery within the muscle cavity prior to nerve
cuff implantation (a). One 28 channel nerve cuff implanted on the peroneal nerve in subject 1 (b), and four 2 channel
nerve cuff implanted on the peroneal nerve in subject 2 (c). The hind limb fixed in place around the fetlock joint
and hock, with stimulus pins visible in the lower hindshank (d).
3.2 Hypothesis 1: Subthreshold current contribution to excitation of fibres
Twitching of the lower hindshank was observed in response to the stimulus pulse, Fig. 2d. With two subthreshold
currents applied, sustained extension of the lower hindshank was observed, indicating prolonged application of a
+/- 140 µA amplitude, 6 kHz sinusoid is sufficient to activate motor fibres in the peroneal nerve which innervate
the gastrocnemius muscle.
In all three current conditions -- none, one, and two subthreshold currents -- a CAP was observed between 0.3 and
0.8 ms, and with a peak at 0.4 to 0.45 ms, after commencement of the stimulus pulse. In electrode 8, which
exhibited the largest CAP amplitudes, amplitudes of (mean +/- standard deviation): 1.58 +/- 0.33, 1.99 +/- 0.49,
and 2.22 +/- 0.56 mV, respectively, were observed for none, one, and two subthreshold current conditions, Fig.
3a-c. Comparing these values with one another using unpaired t-tests (N = 300) produced two-tailed p values of
< 0.0001, which is considered extremely statistically significant. Cross correlation produced lag values of 0 for
both subthreshold current conditions. The difference in mean recorded waveforms were not analysed for these
data.
6
7
Stimulus
artefact
CAP
Stimulus
artefact
CAP
Stimulus
artefact
CAP
(a)
(b)
(c)
Amplitude
V
m
1
0.5 ms
Figure 3: CAPs recorded from the peroneal nerve in response to stimulus applied, 30 mm distally, to the same nerve.
Individual voltage recordings (shaded lines) overlaid with the mean and standard deviation (bold lines with error
bars) with none (a), one (b), and two (b) subthreshold currents applied show that the subthreshold currents contribute
to excitation of nerve fibres when applied to the same region of nerve as the stimulus pulse.
3.3 Hypothesis 2: Pin stimulus
Minor twitching of the phalanges were observed in response to the stimulus pulse. As was the case earlier, with
two subthreshold currents applied sustained extension of the lower hindshank was observed.
The 1 ms/phase, biphasic, square pulse produced a stimulus artefact with 2.5 ms artefact from the pulse, followed
by 1.5 ms of ringing, Fig. 4a-c. The shorter duration, 0.5 ms/phase, pulse produced a stimulus artefact with 1.5
ms from the pulse followed by 1.5 ms of ringing, Fig. 4d. With both stimuli, CAPs were observed from 4.5 ms
onwards, indicating the artefact and CAPs could be distinguished form one another.
In both pin configurations and for all three subthreshold current conditions -- none, one, and two subthreshold
currents -- neural activity was observed between 4.5 and 12 ms after commencement of the stimulus pulse, and
contained multiple peaks and troughs with combined amplitudes in the range of 5 to 20 µV, Fig. 4b-c. CAP
amplitudes were not analysed in the individual segments because they were obscured by noise. In the mean
waveforms, where CAPs were visible, no changes in CAP amplitudes were identified above the 3σ noise threshold
because of the large noise, of σ = 0.8 to 2.2 µV, relative to the CAP amplitudes, Fig. 4e-g. Cross correlation
produced lag values of 0 for both subthreshold current conditions. These results indicate no changes in the CAP
amplitude or temporal shift were identified outside the noise threshold.
7
8
Stimulus artefact
2 ms
V
µ
0
0
5
Stimulus artefact
CAPs
2 ms
V
µ
0
2
(a)
(b)
Stimulus artefact
CAPs
Stimulus artefact
CAPs
Stimulus artefact
CAPs
Stimulus artefact
CAPs
(c)
(d)
Stimulus artefact
CAPs
σ = 0.8 µV
(e)
σ = 1.3 µV
(f)
σ = 1.0 µV
σ = 1.2 µV
σ = 1.8 µV
(g)
σ = 2.2 µV
Figure 4: CAPs recorded from the peroneal nerve in response to stimulus applied approximately 300 mm distally
through two pins apposing the cannon bone; with pins either on the posterior side of the cannon bone (a -- b), or the
posterior and anterior side of the cannon bone (c -- d). The individual (shaded lines) and mean recordings (bold lines)
show a large stimulus artefact between 0 and 4 ms (a). A close up of the mean recordings for none, one, and two
subthreshold currents in red, green, and blue, respectively, show multiple CAPs with 5 to 20 µV amplitudes between
4.5 and 12 ms with a 1 ms/phase stimulus pulse (b-c), or 4.5 to 10.5 ms with a 0.5 ms/phase stimulus pulse (d). No
changes in CAP amplitudes can be seen outside the 3σ noise threshold because of the large noise, of σ = 0.8 to 2.2 µV
(e-g).
3.4 Hypothesis 2: Nerve cuff stimulus
In all four current conditions -- none, one, two, and three subthreshold currents -- a stimulus artefact was visible
between 0 and 0.35 ms, followed by a large CAP with a mean amplitude of 2.13 to 2.14 mV and a peak at 0.7 ms
after commencement of the stimulus pulse, Fig. 5a. Two smaller, secondary CAPs followed, with mean amplitudes
of 13 +/- 1 and 35 +/- 1 µV and peaks at 2.85 +/- 0.05 and 5 +/- 0.05 ms, respectively, Fig. 5b. Lastly, a long
duration CAP between 10 and 200 ms was visible with an amplitude between 59 to 61.5 µV and a peak at 50.9 to
53.3 ms, Fig. 5c. Multiple spikes were visible within the 10 to 400 ms temporal range, each with amplitudes of
10 to 400 µV and durations of 0.1 to 1 ms. The number and temporal location of the spikes varied between
individual data segments, and were suspected to be caused by movement artefacts and evoked action potentials
within the muscle fibres.
The primary CAP amplitude was (mean +/- standard deviation): 2.141 +/- 0.086, 2.136 +/- 0.088, 2.130 +/- 0.096,
and 2.140 +/- 0.089 mV, respectively, for each of the four current conditions, Fig. 5a. Comparison of these values
8
9
using unpaired t-tests (N=300) produced two-tailed p values between 0.13 and 0.83, which are considered to be
not statistically significant. For the one, two, and three subthreshold currents, lag values of -5, -5 and -6 were
calculated from cross correlation of the upsampled waveforms within the temporal window of 0.5 to 2 ms,
corresponding to a temporal shifts of -2.5, -2.5 and -3 µs. These temporal shifts were visible in the difference in
mean recorded waveforms, where the one, two, and three subthreshold current conditions all showed a positive
difference at 0.65 ms followed by a negative difference at 0.75 ms, each with amplitudes of 60 to 80 µV, Fig. 5d.
Using the unpaired t-tests (N=300), both the positive and negative peak difference amplitudes produced
statistically meaningful p-values of 0.02 or less, despite large uncertainty at these points of 280 to 350 µV.
The amplitudes of the secondary CAPs and long duration CAPs were not analysed in the individual segments
because they were obscured by noise. In the mean waveforms, where CAPs were visible, no changes in CAP
amplitudes were identified above the 3σ noise threshold, Fig. 5e-f. Cross correlation of the mean waveforms
produced lag values of 0 for all current conditions.
Stimulus
artefact
Large primary
CAP peak
CAPs
Small secondary
CAP peaks
(a)
Long duration
CAP peak
Spikes
(b)
(c)
V
µ
0
0
0
1
V
µ
0
0
2
V
µ
0
0
2
0
0.5
1
1.5
2
4
6
Time (ms)
8
10
100
200
300
400
(d)
(e)
(f)
V
µ
0
5
V
µ
0
1
V
µ
0
1
0
0.5
1
1.5
2
4
6
Time (ms)
8
10
100
200
300
400
(g)
V
µ
0
5
0
0.5
1
1.5
2
Time (ms)
Figure 5: CAPs recorded from the sciatic nerve in response to stimulus applied approximately 70 mm distally on the
peroneal nerve with none (red), one (green), two (blue), and three (purple) subthreshold currents applied (a -- c). The
individual (shaded lines) and mean recordings (bold lines) show a large stimulus artefact between 0 and 0.4 ms,
followed a large primary CAP in the 0 to 2 ms window (a), two small secondary CAPs in the 2 to 10 ms window (b)
and a long duration CAP coinciding with multiple, short duration spikes in the 10 to 400 ms window (c). Significant
differences between the mean recordings are visible for the primary CAP indicating a temporal shift of the CAP in the
presence of the subthreshold currents (d), whereas any differences in the secondary and long duration CAPs are below
the noise threshold (e-f). The differences between the mean recordings in the primary CAP are significantly reduced
with the subthreshold current is reduced from +/- 70 to +/- 30 µA.
9
10
When the amplitude of the subthreshold currents was reduced to +/- 30 µV, lag values of -1, 1 and 0 were
calculated from cross correlation of the upsampled waveforms within the temporal window of 0.5 to 2 ms,
corresponding to a temporal shifts of -0.5, +0.5 and 0 µs in the negative direction. These temporal shifts were
again visible in the difference in mean recorded waveforms, although with markedly smaller amplitudes than
previously, Fig. 5e.
4 Discussion
The first hypothesis, that the subthreshold currents contribute to excitation of the underlying nerve fibres, was
confirmed through two observations: (i) that two subthreshold currents applied to the same section of nerve
produce sustained extension of the hindshank; and, (ii) that the CAP amplitude increases with application of one
subthreshold current, and then increases further with two subthreshold currents, when applied to the same section
of nerve. This result is expected given the summative effect of currents on neural membrane excitability, and
agrees in principle with stochastic resonance experiments [13, 14], as well as the ephaptic interaction study on rat
in vivo [28] where the subthreshold component was instead produced by a CAP.
The second hypothesis, that increased excitability would increase synchrony in the neural activity and augment
the propagation velocities, was confirmed with low confidence by two observations from the large primary CAP
produced in subject 2: (i) a lag of -2.5 to -3 µs in the presence of subthreshold currents; and, (ii) a biphasic
difference in the mean waveforms of around 140 µV peak-peak amplitude in the presence of subthreshold currents.
While confidence in the lag values is limited by the use of upsampling with spline interpolation and because they
represent an increase in propagation velocity of only 0.43 %, they do agree with the statistically significant
biphasic difference in the mean waveforms characteristic of a negative temporal shift.
It is not clear whether the observed change in CAP velocity agrees with modelling predictions in [24-27], that
ephaptic interactions slow down the CAP, because the models did not consider external current. In the present
study, it is conceivable that the partially excited membranes took less time to depolarize, which would produce
the observed increase in CAP velocity. If this was the case, however, it is not clear why the lag values did not
increase linearly with the number of subthreshold currents applied. Finite element modelling of nerve fibres may
provide insight into the observed change in CAP velocity, as well as the best way to configure multiple
subthreshold currents to augment such an effect.
Limitations of this study are the low number of subjects (n=2), low temporal sampling rate (20 kHz), high noise,
and small range of stimulus and subthreshold current amplitudes and frequencies investigated.
5 Conclusion
In this study, we confirmed that administering subthreshold currents increases excitability of the underlying fibres,
and, with low confidence, that this increased excitability changes the propagation velocity of neural activity in the
underlying fibres. While more work needs to be done in this area to improve confidence in the results, these initial
findings contribute to understanding of possible mechanisms of neuromodulation using subthreshold currents.
10
11
Disclosures
The authors have no relevant financial interests in this article and no potential conflicts of interest to disclose.
Acknowledgements
The authors would like to thank David Holder, Kirill Aristovich and Enrico Ravagli, from the EIT Research Group
at University College London, for fabrication of electrode arrays used in this study; Darren Svirskis from the
School of Pharmacy, The University of Auckland for assistance with PEDOT coating electrodes; and staff at the
Faculty of Medical and Health Sciences, The University of Auckland for their help with experiments.
References
Amon, A. and F. Alesch, Systems for deep brain stimulation: review of technical features. Journal of
Neural Transmission, 2017. 124(9): p. 1083-1091.
Verrills, P., C. Sinclair, and A. Barnard, A review of spinal cord stimulation systems for chronic pain.
Journal of pain research, 2016. 9: p. 481-492.
McClintock, S.M., et al., Consensus recommendations for the clinical application of repetitive
transcranial magnetic stimulation (rTMS) in the treatment of depression. The Journal of clinical
psychiatry, 2018. 79(1).
Markert, M.S. and R.S. Fisher, Neuromodulation - Science and Practice in Epilepsy: Vagus Nerve
Stimulation, Thalamic Deep Brain Stimulation, and Responsive NeuroStimulation. Expert Review of
Neurotherapeutics, 2019. 19(1): p. 17-29.
Cristancho, M.A., et al., Vagus Nerve Stimulation (VNS), in Psychiatric Neurotherapeutics:
Contemporary Surgical and Device-Based Treatments, J.A. Camprodon, et al., Editors. 2016, Springer
New York: New York, NY. p. 99-116.
Philip, N.S., et al., Low-intensity transcranial current stimulation in psychiatry. American Journal of
Psychiatry, 2017. 174(7): p. 628-639.
Lundstrom, B., et al., Trial stimulation and chronic subthreshold cortical stimulation to treat focal
epilepsy. Brain Stimulation: Basic, Translational, and Clinical Research in Neuromodulation, 2019.
12(2): p. 502.
Lundstrom, B.N., et al., Chronic Subthreshold Cortical Stimulation to Treat Focal Epilepsy. JAMA
Neurology, 2016. 73(11): p. 1370-1372.
Valero-Cabré, A., et al., Transcranial magnetic stimulation in basic and clinical neuroscience: A
comprehensive review of fundamental principles and novel insights. Neuroscience & Biobehavioral
Reviews, 2017. 83: p. 381-404.
Collins, J.J., et al., Noise-enhanced human sensorimotor function. IEEE Engineering in Medicine and
Biology Magazine, 2003. 22(2): p. 76-83.
Wuehr, M., J. Decker, and R. Schniepp, Noisy galvanic vestibular stimulation: an emerging treatment
option for bilateral vestibulopathy. Journal of Neurology, 2017. 264(1): p. 81-86.
Serrador, J.M., et al., Enhancing vestibular function in the elderly with imperceptible electrical
stimulation. Scientific reports, 2018. 8(1): p. 336.
Wiesenfeld, K. and F. Moss, Stochastic resonance and the benefits of noise: From ice ages to crayfish
and SQUIDs. Nature, 1995. 373(6509): p. 33-36.
Hänggi, P., Stochastic resonance in biology: How noise can enhance detection of weak signals and help
improve biological information processing. ChemPhysChem, 2002. 3(3): p. 285-290.
Breen, P.P., et al., Peripheral tactile sensory perception of older adults improved using subsensory
electrical noise stimulation. Medical Engineering & Physics, 2016. 38(8): p. 822-825.
11
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
12
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
Breen, P.P., et al., A new paradigm of electrical stimulation to enhance sensory neural function. Medical
Engineering & Physics, 2014. 36(8): p. 1088-1091.
Jefferys, J.G., Nonsynaptic modulation of neuronal activity in the brain: electric currents and
extracellular ions. Physiological Reviews, 1995. 75(4): p. 689-723.
Shivacharan, R.S., et al., Self-propagating, non-synaptic epileptiform activity recruits neurons by
endogenous electric fields. Experimental Neurology, 2019. 317: p. 119-128.
Han, K.-S., et al., Ephaptic Coupling Promotes Synchronous Firing of Cerebellar Purkinje Cells.
Neuron, 2018. 100(3): p. 564-578.e3.
Arvanitaki, A., Effects evoked in axon by the activity of a contiguous one. Journal of Neurophysiology,
1942. 5(2): p. 89-108.
Jasper, H. and A. Monnier, Transmission of excitation between excised non‐myelinated nerves. An
artificial synapse. Journal of Cellular and Comparative Physiology, 1938. 11(2): p. 259-277.
Fröhlich, F. and D.A. McCormick, Endogenous electric fields may guide neocortical network activity.
Neuron, 2010. 67(1): p. 129-143.
Faber, D., Field effects in the CNS play functional roles. Frontiers in Neural Circuits, 2010. 4(15).
Capllonch-Juan, M., F. Kölbl, and F. Sepulveda. Unidirectional ephaptic stimulation between two
myelinated axons. in 2017 39th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC). 2017.
Binczak, S., J. Eilbeck, and A.C. Scott, Ephaptic coupling of myelinated nerve fibres. Physica D:
Nonlinear Phenomena, 2001. 148(1-2): p. 159-174.
Reutskiy, S., E. Rossoni, and B. Tirozzi, Conduction in bundles of demyelinated nerve fibres: computer
simulation. Biological cybernetics, 2003. 89(6): p. 439-448.
Schmidt, H. and T.R. Knösche, Action potential propagation and synchronisation in myelinated axons.
PLOS Computational Biology, 2019. 15(10): p. e1007004.
Bolzoni, F. and E. Jankowska, Ephaptic interactions between myelinated nerve fibres of rodent
peripheral nerves. European Journal of Neuroscience, 2019. 50(7): p. 3101-3107.
Christopher, A.R.C., et al., Electrode fabrication and interface optimization for imaging of evoked
peripheral nervous system activity with electrical impedance tomography (EIT). Journal of Neural
Engineering, 2019. 16(1): p. 016001.
Vasconcelos, B.G., et al., Origin and distribution of the ischiatic nerve in mixed-breed sheep. Brazilian
Journal of Veterinary Research and Animal Science, 2014. 51(2): p. 102-110.
Y Aristovich, K., Donega, M., Blochet, C., Avery, J., Hannan, S., Chew, D., S Holder, D., Imaging fast
neural traffic at fascicular level with electrical impedance tomography: proof of principle in rat sciatic
nerve. Journal of Neural Engineering, 2018.
Hope, J., et al., Increasing signal amplitude in electrical impedance tomography of neural activity using
a parallel resistor inductor capacitor (RLC) circuit. Journal of Neural Engineering, 2019.
12
|
1702.01568 | 1 | 1702 | 2017-02-06T11:13:15 | Interpretation of Correlated Neural Variability from Models of Feed-Forward and Recurrent Circuits | [
"q-bio.NC"
] | The correlated variability in the responses of a neural population to the repeated presentation of a sensory stimulus is a universally observed phenomenon. Such correlations have been studied in much detail, both with respect to their mechanistic origin and to their influence on stimulus discrimination and on the performance of population codes. In particular, recurrent neural network models have been used to understand the origin (or lack) of correlations in neural activity. Here, we apply a model of recurrently connected stochastic neurons to interpret correlations found in a population of neurons recorded from mouse auditory cortex. We study the consequences of recurrent connections on the stimulus dependence of correlations, and we compare them to those from alternative sources of correlated variability, like correlated gain fluctuations and common input in feed-forward architectures. We find that a recurrent network model with random effective connections reproduces observed statistics, like the relation between noise and signal correlations in the data, in a natural way. In the model, we can analyze directly the relation between network parameters, correlations, and how well pairs of stimuli can be discriminated based on population activity. In this way, we can relate circuit parameters to information processing. | q-bio.NC | q-bio |
Interpretation of Correlated Neural Variability from Models
of Feed-Forward and Recurrent Circuits
Volker Pernice 1,2
Rava Azeredo da Silveira 1,2,3,∗
1 Department of Physics, ´Ecole Normale Sup´erieure, PSL Research University,
2 Laboratoire de Physique Statistique, Centre National de la Recherche
Scientifique, Universit´e Pierre et Marie Curie, Universit´e Denis Diderot, 75005
75005 Paris, France
3 Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08544,
Paris, France
USA
∗ corresponding author
February 7, 2017
Abstract
The correlated variability in the responses of a neural population to the repeated presenta-
tion of a sensory stimulus is a universally observed phenomenon. Such correlations have been
studied in much detail, both with respect to their mechanistic origin and to their influence on
stimulus discrimination and on the performance of population codes. In particular, recurrent
neural network models have been used to understand the origin (or lack) of correlations in
neural activity. Here, we apply a model of recurrently connected stochastic neurons to in-
terpret correlations found in a population of neurons recorded from mouse auditory cortex.
We study the consequences of recurrent connections on the stimulus dependence of correla-
tions, and we compare them to those from alternative sources of correlated variability, like
correlated gain fluctuations and common input in feed-forward architectures.
We find that a recurrent network model with random effective connections reproduces
observed statistics, like the relation between noise and signal correlations in the data, in a
natural way. In the model, we can analyze directly the relation between network parameters,
correlations, and how well pairs of stimuli can be discriminated based on population activity.
In this way, we can relate circuit parameters to information processing.
1 Author Summary
The response of neurons to a stimulus is variable across trials. A natural solution for reliable
coding in the face of noise is the averaging across a neural population. The nature of this
averaging depends on the structure of noise correlations in the neural population. In turn,
the correlation structure depends on the way noise and correlations are generated in neural
1
circuits. It is in general difficult, however, to tease apart the origin of correlations from the
observed population activity alone. In this article, we explore different theoretical scenarios
of the way in which correlations can be generated, and we relate these to the architecture of
feed-forward and recurrent neural circuits. Analyzing population recordings of the activity
in mouse auditory cortex in response to sound stimuli, we find that population statistics are
consistent with those generated in a recurrent network model. Using this model, we can then
quantify the effects of network properties on average population responses, noise correlations,
and the representation of sensory information.
2
Introduction
In the search for clues about the function of neural circuits, it has become customary to rely
upon recordings of the responses of large populations of neurons. These measurements exhibit
the concerted activity of neural populations in different conditions, such as presentations of
different stimuli, summarized by stimulus-dependent, high-dimensional statistics. The lowest
moment of these statistics are also the best characterized, namely the mean response of
neurons, the variability of response of single neurons, and their pairwise correlations. With
these statistics in hand, one can ask two questions: How are they generated in the neural
population? What purpose, if any, do they serve? While the first question is mechanistic and
the second is functional, the two are intimately linked. We focus here on these questions and
on their connection.
Along the mechanistic line of research, a number of network models have been proposed
to account for the statistics in measurements as well as to determine the relationship between
anatomical and physiological parameters, on the one hand, and observed dynamics, on the
other hand.
In this spirit, for example models of balanced networks have been proposed
to explain asynchronous and irregular spike trains [1, 2, 3]. Mechanistic explanations for
the origin of pairwise correlation include the influence of recurrent connections [4, 5, 6] and
global fluctuations (e.g., from top-down afferents) [7, 8, 9, 10, 11, 12]. Reading the origin of
correlations from the recorded activity in a population of neurons is, however, a difficult task
[13, 14, 7, 15].
Along the functional line of research, the objective is to relate the statistics of neural
responses to the function of neural circuits; for example, to elucidate the role of the statistics
in the representation of sensory information. The relation between correlated variability in
the population response and the accuracy of stimulus representation, in particular, has been
the object of much study in recent years [16, 17, 18, 19, 20, 21, 22, 23, 24, 25]. One reason for
the focus on correlations is their possible effect in suppressing noise along relevant dimensions
[26, 2, 27, 28]. Recent work has also illuminated the importance of the origin of correlations
for their structure as it relates to the effect on stimulus representation [29, 30, 31].
In the present paper, we follow a three-pronged approach to relate the possible mechanistic
origins of correlation to recorded statistics in cortical populations [32], on the one hand, and
to relate network mechanisms to the representation of information, on the other hand. First,
we study models of neural populations using the framework of Poisson processes [33] and we
identify signatures in the population statistics of different mechanistic motifs. Second, we
examine data on cortical populations in the light of our model results, and we find that the
measured statistics are consistent with those generated by a recurrent network of Poisson
neurons. Third, we use an idealized model of a recurrent cortical population, in which we
can manipulate mechanistic parameters, to evaluate how the latter affect the accuracy of the
representation of information. We find that, if correlations are not too weak, correlations
generated within recurrent networks can be distinguished from those that are due to external
2
signals. The estimated parameters from the data suggest that the interpretation of neural
spike trains as Poisson processes implies a strong amplification together with noise generation
by the network dynamics.
3 Materials and Methods
3.1 Experimental data set
The data set was first published and analyzed in [32]. It consists of the activity of neural
populations (46-99 neurons) recorded using calcium imaging in the auditory cortex of mice.
Animals were isoflurane anesthetized (1 %). Signals were obtained from neurons labeled
with the synthetic calcium indicator OGB1. Fluorescence was measured at 30 Hz sampling
rate, and firing rates inferred from temporal deconvolution of the fluorescence signal. Up
to 6 neural populations in each of 14 animals were recorded. The data points we use are
the average firing rates over a window of 250 ms after presentation of each of 65 different
sound stimuli, each for 15 trials. Responses were measured relative to spontaneous activity.
Hence, negative responses occurred, if the stimulus-evoked firing rates were smaller than
the spontaneous ones. The stimuli consisted of a range of different pure tones recorded for
different sound intensities, as well as a number of natural sounds. We do not distinguish
individual stimulus identities, but instead regard the set of stimuli as a diverse and, to some
degree, generic ensemble.
3.2 Description of response variability
If the vector of population activity in trial T of stimulus s is r(s, T ), the average response
across trials for this stimulus is denoted by r(s) = (cid:104)r(s, T )(cid:105)T . To measure covariability across
trials in a given stimulus condition between neurons i and j, we use the noise covariances,
defined as
Cij(s) = cov(ri(s, T ), rj(s, T ))T = (cid:104)ri(s, T )rj(s, T )(cid:105)T − (cid:104)ri(s, T )(cid:105)T(cid:104)rj(s, t)(cid:105)T .
(1)
A measure of the strength of pairwise noise correlations across stimuli is the average correla-
tion coefficient,
Here, (cid:104).(cid:105)s denotes the average over all stimuli presented. This quantity is to be contrasted
with the signal correlation,
(2)
(cid:42)
(cid:43)
.
s
Cij(s)
(cid:112)Cii(s)Cjj(s)
cov(cid:0)ri(s), rj(s)(cid:1)
(cid:113)
var(cid:0)ri(s)(cid:1)
svar(cid:0)rj(s)(cid:1)
s
cN
ij =
cS
ij =
,
s
(3)
which measures the similarity of average responses across stimuli. As a measure of how the
orientation of the high-dimensional distributions changes across stimuli, we use the variance
of the population activity projected along the direction of the normalized mean response,
¯r(s) = r(s)/r(s), namely,
σ2
µ(s) =
Cij(s)¯rj(s)¯ri(s).
(4)
(cid:88)
ij
3
√
This quantity can be compared to the variance projected along the diagonal direction, ¯d =
(1, . . . , 1)T /
N ,
(cid:88)
ij
σ2
d(s) =
Cij(s) ¯di ¯dj,
µ and σ2
d are normalized by the sum of the variances, σ2
which corresponds to a uniform averaging of the covariances. To compare these quantities
across stimuli, σ2
i Cii(s).
If, for a given stimulus, all neurons are equally active on average, then ¯r = ¯d and σ2
d = σ2
µ.
As we show, different circuit models predict different stimulus dependencies of σ2
µ and σ2
d.
These differences are most apparent for the stimuli for which the average population response
differs strongly from a uniform population response. A deviation from a uniform response
can be measured by the angle between mean response and diagonal, or cos(d, r) = ¯rT · ¯d. For
a graphical illustration of these measures, see Fig. 1.
(5)
all(s) = (cid:80)
Figure 1: Properties of response distributions and network scenarios. A: Examples for
distributions where the variability is to a large degree in the diagonal direction (response distribu-
tion indicated by blue ellipse), resulting in a large value of σd, and where the variability is mainly
in the direction of the average response (red ellipse), resulting in a larger value of σµ. B: Response
distributions for two stimulus ensembles. Either σµ (red set) or σd (blue set) remains large across
stimuli. A dependence as in the red set appears in a model of shared gain fluctuations, while a
dependence as in the blue set may arise in densely connected recurrent or feed-forward networks.
C: Different network layouts that induce correlated activity. Connections (arrows) to and be-
tween neurons (dots) vary in strength. Dashed arrows indicate multiplicative effects on firing rates.
4
3.3 Recurrent model of Poisson neurons with noisy input
(cid:90) ∞
ri(t) = rext,i(t) +
We model an intrinsically noisy process of spike generation in neurons, in which the effect of
presynaptic spikes on neural activity is captured through linear modulations of an underlying
firing rate. Specifically, the spike train of neuron i is a realization of an inhomogeneous
Poisson process with a time-dependent firing rate, ri(t), calculated as
gij(τ )sj(t − τ )dτ,
(cid:88)
where rext,i(t) is the external input, sj(t) =(cid:80)
t(cid:48) δ(t−t(cid:48)) is the recurrent input from presynaptic
spike trains, and the causal coupling kernel, gij(τ ), defines the interactions between neurons
in the network. The external input is an analog quantity which can be viewed as resulting
from a convolution between a linear filter and an incoming spike train from "non-local"
presynaptic neurons. We treat it as a noisy, but stationary, signal, i.e., its mean and higher
moments do not depend on time. While this model of interacting point processes [33] defines
the full temporal dynamics of the system, we are only interested in time-averaged rates or,
equivalently, spike counts. Their expected values and (co-)variances across realizations can
be obtained if the coupling matrix is known.
(6)
0
j
Assuming stationarity of input spiking and firing rates, we can solve for the average rate
vector, r = (r1, . . . rN ), across trial realizations, as
Here, I denotes the identity matrix, G is the steady-state coupling matrix with elements
and rext is the average input vector. To describe the correlation in the activity, we consider
the spike count in a long time bin, ∆, defined as
The total input in the time bin is given by
ni(∆) =
si(t)dt.
(9)
(7)
(8)
(11)
(12)
r = (I − G)−1rext.
Gij =
gij(τ )dτ,
(cid:90) ∞
0
(cid:90) t+∆
(cid:90) t+∆
t
rext(∆) =
(10)
and is a random variable with a normalized variance, Vext ≡ var(rext(∆))/∆ (across choices of
different time bins). From spectral properties of the covariance [33] and the relation between
covariance and count correlations [34, 5], it can be shown that the count covariance normalized
by the time bin duration, C, with elements
rE(t)dt,
t
cov(cid:0)ni(∆), nj(∆)(cid:1)
,
Cij ≡ lim
∆→∞
∆
C = (I − G)−1(cid:0)D[r] + D[Vext](cid:1)(I − GT )−1.
is given by the matrix
The term D[r] denotes a diagonal matrix with elements D[r]ij = riδij, and similarly D[Vext]ij =
δijVext,i. The effect of the recurrent network interactions is reflected in the transfer matrix,
B ≡ (I − G)−1.
5
(13)
In Eq. (12), one component of the covariance arises from the variance of the external input,
Vext. Although the intrinsic noise from the spike-generating Poisson process is dynamical,
its effect is that of an additive contribution to the external variance (see also [35]). The
dependence on the rate vector, in Eq. (12), results from the Poisson character of the noise,
in which the variance equates the mean.
In the experimental data set, firing rates are measured relative to spontaneous activity.
To take this into account, an offset, a, can be added to the rates in Eq. (12). Henceforth,
we examine the model through Eqs. (7) and (12) for the mean neural activities and their
covariances, respectively.
3.3.1 Random effective network and stimulus ensemble
The rates and covariances given in Eqs. (7) and (12) depend on the recurrent coupling
matrix within the network, G, only via effective couplings defined by the transfer matrix,
B = (I − G)−1. In numerical calculations, we choose the entries of G independently from a
normal distribution with positive mean; this affords a specific structure to the matrix B. The
empirical mean and variance of the elements of this matrix will be called (cid:104)B(cid:105) and var(B). In
a number of analytical calculations, for the sake of simplicity, we neglect the specific structure
of B and model it as a random matrix with elements independently drawn from a normal
distribution with corresponding mean and variance. The elements of B are not direct, but ef-
fective connections, because they reflect also the effects of indirect connections in the network,
hence the expression "random effective network." The variability of effective connections can
be quantified by the ratio ρ = var(B)/(cid:104)B(cid:105)2.
The stimulus ensemble is modeled as a set of random vectors, rext(s). We assume that
the elements of these vectors are independent across neurons and stimuli, and normal, and
we call the mean and variance of the corresponding normal distribution (cid:104)rext(cid:105) and var(rext).
The relative variability of the inputs is defined as ρE = var(rext)/(cid:104)rext(cid:105). This variability
is high if both excitatory and inhibitory inputs are present, and, together with the network
parameters, it determines the variability of the average responses across stimuli. In numerical
calculations, we set Vext,i(s) = rext,i(s), as dictated by a Poisson process. We allow for input
signal correlations, i.e., the average inputs to a pair of neurons can come with a non-vanishing
correlation coefficient, cin = (cid:104)rext,i(s)i, rext,j(s)(cid:105)s.
3.4 Models of correlated activity from shared inputs or gain
modulations
Correlations in the activity of neurons can have other origins besides recurrent connections. In
parallel with the recurrent network model, we consider two alternative prototypical models in
which correlations originate from shared inputs or from shared gain fluctuations, respectively.
The three different scenarios are illustrated in Fig. 1C. In Appendix 8.2, we show how,
formally, these models can also be cast as special cases of the recurrent model.
3.4.1 Feed-forward model with shared input
We consider a simple, two-layer, feed-forward network, in which N output neurons receive
inputs from N independent presynaptic neurons with spike trains sj(t). The contributions of
the input neurons to the firing rates of the output neurons are determined by feed-forward
connection kernels fij(τ ), so that the observed firing rates are
6
(cid:90) ∞
(cid:88)
ri(t) =
fij(τ )sj(t − τ )dτ.
(14)
0
j
We define the feed-forward coupling matrix F by its elements, Fij = (cid:82) ∞
As we show in Appendix 8.2, this scenario represents a special case of the framework described
in Section 3.3, so that firing rates and count covariances can be calculated correspondingly.
fij(τ )dτ . With
the time-averaged firing rates of the external input for stimulus s, rext(s), the average spike
counts of the output neurons are given by
0
r(s) = F rext(s).
(15)
If the spike trains of the input neurons are Poisson processes, the input variance is equal to
the rate, Vext = rext, but we will allow for more general inputs. The count covariances are
then given by
C = F D[Vext]F T + D[r].
(16)
We denote again by D[r], D[Vext] diagonal matrices with diagonal elements given by the
vectors r, Vext. The first term describes covariances resulting from the combined shared
inputs. The second term results from the contribution to the count variances of the Poisson
spike generation in output neurons, which is independent across neurons.
If rates are observed only up to a stimulus-dependent offset, a, Eq. (16) has to be replaced
by
C = F D[Vext]F T + D[r + a].
(17)
This scenario can be compared directly to the random effective recurrent network model,
if one uses F = B for an identical ensemble of stimuli (see Section 3.3.1). Comparing Eq.
(17) to Eq. (12), we note that the difference between the two models is that the internally
generated noise in the feed-forward model, D[r + a], is uncorrelated and contributes only
to the variances, while in the recurrent network it is filtered by the network and therefore
correlated.
3.4.2 Model with shared gain fluctuations
In addition to recurrence and shared inputs, shared gain fluctuations have also been proposed
as a source of correlation in neural populations by various authors [7, 36, 24, 10]. In this
model, the firing rate of the neural population is given by the product of a constant, stimulus
dependent vector, r(s), and a scalar fluctuating signal, (t). We set the time-averaged value of
the fluctuations to unity, (cid:104)(t)(cid:105)t = 1, so that the time-averaged firing rate is r. The resulting
count covariance matrix is given by
C = D[r] + rrT Vext,
(18)
where the scalar variance, Vext, reflects the strength of the variation of the fluctuating signal.
In other words, the firing rate of a Poisson neuron is modulated multiplicatively by a fluctu-
ating "gain" signal. One consequence is that the covariance of a neuron pair is proportional
to the product of their firing rates. The general model described in Section 3.3 reduces to
this model if for any given stimulus the activity of the population results from a single input
neuron projecting to the output neurons with corresponding weights (see Appendix 8.2).
Here, the average responses can be chosen freely. When comparing this scenario to the
recurrent and feed-forward scenarios, we use the ensemble of average responses resulting from
the corresponding network scenario. If the stimulus dependence of the firing rates is measured
relative to an offset a, the covariances are given by
C = D[r + a] + (r + a)(r + a)T Vext.
(19)
7
3.5 Stimulus discriminability
(cid:16)
(cid:17)−1(cid:16)
(cid:17)
r(s1) − r(s2)
In order to evaluate the influence of correlation on neural coding, we examine the discrim-
inability of a set of stimuli from the population activity. In the models, a stimulus, s, evokes
activity characterized by an average response vector, r(s), and a covariance matrix, C(s).
We seek a simple measure for evaluating the possibility of attributing a given population re-
sponse to one of two discrete stimuli, s1 and s2, unambiguously. For this, we assume that the
high-dimensional distributions of responses are Gaussian, and project the two distributions
on a single dimension where we calculate the signal-to-noise ratio as our measure.
The best projection follows from Fisher's linear discriminant analysis: the most informa-
tive linear combination of neuron responses, denoted by w · r, is achieved if the vector w
points in the most discriminant direction,
w =
C(s1) + C(s2)
.
(20)
The mean and variance of the projected distributions onto the normalized direction, ¯w, are
¯wT r(si) and σ2
as
si = ¯wT C(si) ¯w, for i ∈ {1, 2}. We can then define the "signal-to-noise ratio,"
(cid:12)(cid:12) ¯wT ·(cid:0)r(s1) − r(s2)(cid:1)(cid:12)(cid:12)
.
(21)
S =
σs1 + σs2
Larger values of S correspond to better discriminability.
To quantify the effect of correlation on discriminability, we compare the quantity S, defined
in Eq. (21), with the quantity Sshuffled obtained from a "shuffled data set", in which responses
are shuffled across trials (in experimental data) or off-diagonal covariance elements, Ci(cid:54)=j(s),
are set to 0 (in models). A ratio Sshuffled/S smaller than unity indicates that correlation is
beneficial to discriminability.
We note that S is an approximation for a measure based on the optimal linear classifier,
for which the threshold separating the two one-dimensional projected distributions has to
be calculated numerically. A measure akin to S used in similar contexts is the linear Fisher
information [16], valid for continuous stimuli. An advantageous property of S is its invari-
ance under linear transformations:
if all responses, r, are fed into another network whose
output, Br, results from a product with an invertible matrix B, S does not change. This
obtains because the mean responses are transformed into Br(s), while the covariances are
transformed into BC(s)BT . Intuitively, a simple matrix multiplication is accommodated for
by a corresponding change in the most discriminative direction w on which we project.
3.6 Two-population model
In order to develop an intuitive understanding of the way in which correlations generated in
a recurrent network influence coding, we examine a highly simplified model. The recurrent
network consists of two excitatory sub-populations, labeled E and E(cid:48), each made up of N
neurons. We ask to what extent their activity discriminates two stimuli when these elicit
preferential responses in the two sub-populations, respectively.
The activity in the network is determined by Eqs. (7) and (12). For the sake of simplicity,
we can set the external variance to zero, such that the input to the network is fully defined
by its mean input, rext. We assume some further simplifications, for the sake of calculational
ease: each neuron in sub-population L projects to a fixed number, nKL, of neurons in sub-
population K; all non-zero coupling weights are identical, Gij = gE. Additionally, all neurons
in the same population receive identical external input. Then, the two-component vectors
Rext(s1) = (1 + ∆, 1)T and Rext(s2) = (1, 1 + ∆)T define a pair of stimuli. Each component
denotes the input to each neuron in the corresponding population.
8
We reduce the dimensionality of the system by considering the population activity on a
macroscopic level. The average response of population K across trials is defined as
RK =
rk,
(22)
and its trial-to-trial variability is described by the population covariance matrix, with elements
defined as
ΣKL =
Ckl.
(23)
It turns out that these macroscopic quantities depend only on the overall number of connec-
tions between populations, not on a specific network realization. For example, the sum of
the rates of neurons in one population depends only on the sum of the inputs to all neurons,
but not on how these inputs are distributed among the neurons. This is a consequence of the
linearity of the dynamics and the assumption that each neuron has a fixed number of output
connections, as we show in detail in Appendix 8.4. This idea was already applied in Refs.
[5, 6].
More precisely, we define the coupling within each population, Γs = nEEgE = nE(cid:48)E(cid:48)gE,
and across populations, Γc = nEE(cid:48)gE = nE(cid:48)EgE, which make up a population coupling matrix
(cid:88)
(cid:88)
k∈K
k∈K,l∈L
(cid:18)Γs Γc
(cid:19)
Γc Γs
Γ =
.
(24)
Applying the definition in Eqs. (22) and (23), we can rewrite the microscopic Eqs. (7) and
(12) as population equations,
R(s) = (I − Γ)−1N Rext(s)
(25)
and
Σ(s) = (I − Γ)−1D[R(s)](I − Γ)−1.
(26)
The population transfer matrix is defined by P = (I−Γ)−1. Up to a factor N , these equations
are equivalent to the microscopic equations in the restricted context of the two-population
model.
4 Results
We discuss the characteristics of the noise statistics that emerge from neural dynamics in
models of recurrent networks, feed-forward networks, and networks with global gain fluctua-
tions. We also look for coarse statistical signatures of each of these three model structures,
which can then be compared with data. In this perspective, we analyze cortical recordings in
mouse. These recordings were made in the auditory cortex, while the animal was stimulated
with musical tones [32]. We then discuss the implications of the neural noise statistics for
stimulus encoding. Finally, we present a highly simplified model of network dynamics, paired
down to include only a few parameters, in order to develop an intuition of the link between
network structure and stimulus encoding.
4.1 Population signatures of noise statistics in a recurrent net-
work model
One population feature which distinguishes the possible mechanisms of generation of noise
correlation is the relation between the population-averaged response, (cid:104)r(cid:105)(s) = (cid:104)ri(s)(cid:105)i =
9
1
N
(cid:80)
Cij =(cid:80)
i ri(s), on the one hand, and the population-averaged variance, (cid:104)Cii(cid:105)i(s) = 1
i Cii(s),
or the noise covariance averaged across pairs (cid:104)Cij(s)(cid:105)i(cid:54)=j =
i(cid:54)=j Cij, on the other
hand. In this section, we analyze these relations for Poisson neurons in a recurrent network
with random effective connections, as described in Section 3.3.1. Using Eq. (12), which reads
N (N−1)
N
1
k BikBjk(rk(s) + a + Vext,k) for the pairwise covariances, we derive the expression
(cid:80)
(cid:80)
(27)
(28)
for the average variance, and the expression
(cid:104)Cii(cid:105)i(s) ≈ N(cid:104)B2(cid:105)(cid:16)(cid:104)r(cid:105)(s) + a + (cid:104)Vext(cid:105)(cid:17)
(cid:104)Cij(cid:105)i(cid:54)=j(s) ≈ N(cid:104)B(cid:105)2(cid:16)(cid:104)r(cid:105)(s) + a + (cid:104)Vext(cid:105)(cid:17)
(cid:80)
(cid:80)
N 2
N 2
ij B2
for the average covariance. (See Appendix 8.3.4 for mathematical details and Fig. 2 for nu-
merical results.) The quantity (cid:104)B(cid:105) = 1
ij Bij denotes the average strength of the effective
connections in the network across neuron pairs and, correspondingly, (cid:104)B2(cid:105) = 1
ij the
average of their square. The variance of the input, averaged across neurons, is (cid:104)Vext(cid:105), and a
denotes a potential constant offset in the observed firing rates.
The variance of a Poisson variable is proportional to its mean. This contribution from
the Poisson spike generation is reflected by the term (cid:104)r(cid:105) in the two equations and is the
reason for the linear relation between average response and average variance/covariance - for
non-Poisson spike generation, these relations would be non-linear. As we see, not only the
external noise, (cid:104)Vext(cid:105), but also the internally generated noise are amplified by the recurrent
connections. The amplification factor, or the slope of the linear relation, N(cid:104)B(cid:105)2 in Eq. (28),
is smaller for the covariances, where it depends only on the mean strength (cid:104)B(cid:105) of the effective
connections. For the variances, expressed in Eq. (27), the averaging in the slope N(cid:104)B2(cid:105) is
carried out on the square of the effective connections. The slope is therefore larger, and
depends both on the average strength of connections and their variability. In the relations
for both average population variance and covariance, the ratio between the intercept and the
slope is (cid:104)Vext(cid:105) + a. Up to a possible offset, a, that can result from a measurement of firing
rates relative to a constant baseline activity, it represents the strength of the external noise.
In summary, in the recurrent network model, both internally generated noise and external
noise are amplified (or attenuated) by the recurrent connections, and the parameters in the
relationship between (co-)variances and population response can be related to network and
input parameters.
10
Figure 2: Variability in the recurrent network model. A: Relation between response and
variance averaged across population for randomly chosen stimulus vectors (each dot one stimulus).
Four networks are generated with normally distributed connection strengths. Larger ρ, calculated
from effective connections in the transfer matrix B, indicate larger variability of effective weights.
Lines indicate analytic predictions, Eqs. (27) and (28). B: Same for covariances, averaged across
all neuron pairs. C: The ratio of slope and intercept of linear fits corresponds to the strength
of the input variance, which is identical for all networks, and is consistent for variances and
covariances. See Appendix 8.1 for further details and the numerical parameters.
4.2 Population signatures of noise statistics in feed-forward
network models
In a feed-forward network, both shared input units and global gain fluctuations can yield noise
correlations [7, 10, 24]. Here, we compare the population signatures of the noise statistics in
the case of these two alternatives with those from a recurrent model. This serves to point out
differences between the outputs of the various models as well as to motivate the data analysis
to which we turn below.
The three scenarios differ in the scaling of the population-averaged covariance, (cid:104)Cij(cid:105)i(cid:54)=j(s),
with respect to the population response, (cid:104)r(cid:105)(s) (Fig. 3). We already demonstrated in the pre-
vious section that both variances and covariances scale linearly with the population response
in the recurrent network. In a feed-forward network, pairwise covariances are given by Eq.
(16), which can be rewritten as
(cid:88)
Cij = δij(ri + a) +
FikFjkVext,k.
(29)
k
Note that, in contrast to the recurrent network, the neural firing rates, ri, only affect diagonal
entries, and therefore the variances. The average covariance can be expressed as
(cid:104)Cij(cid:105)i(cid:54)=j(s) ≈ N(cid:104)F(cid:105)2(cid:104)Vext(cid:105)(s),
(30)
and does not directly depend on the average population response, (cid:104)r(cid:105)(s) (see Appendix 8.3.3).
If there are excitatory as well as inhibitory inputs, also the average input variance, (cid:104)Vext(cid:105),
can be uncorrelated to the population response:
for Poisson external input, in each input
channel i the variance is as large as the mean input, Vext,i = rext,i. Nonetheless, because the
variance is always positive, including for negative input, the average variance (cid:104)Vext(cid:105) can be
11
decorrelated from the average input, (cid:104)rext(cid:105), and thus from the average output, (cid:104)r(cid:105). Intuitively,
because inhibitory channels add variance, but decrease the average input, the combined input
does not have to be proportional to the combined variance. The extent to which this is true
depends on the balance between positive and negative elements in rext, which we measure by
the input variability ρE = var(rext)/(cid:104)rext(cid:105).
In a model network with global gain fluctuations, from Eq. (18), the covariances are given
by
Again averaging only across the off-diagonal elements, one finds that the average covariances
scale quadratically with the population response, as
Cij = δijri + rirjVext.
(31)
(cid:104)Cij(cid:105)i(cid:54)=j = Vext(cid:104)r(cid:105)2.
(32)
This is a direct consequence of the fact that pairwise covariances are proportional to the
product of the firing rates of the pair of neurons, which holds for large N independent of the
distribution of average responses. By contrast, as is apparent from the expression for Cij,
the individual variances and the population averaged variance each have both a quadratic
and a linear contribution in r (and (cid:104)r(cid:105), respectively) (see also e.g. [7]). (Further details are
provided in Appendix 8.3.)
Until now, we examined how the average covariance (cid:104)Cij(cid:105)i(cid:54)=j changes across stimuli. Fur-
ther differences between the different scenarios can be related to the detailed structure of
C, which determines the shape of the response distributions. By shape, we refer to the ge-
ometric orientation and extent of the multi-dimensional ellipsoid cloud that corresponds to
the response distribution in the space spanned by the responses of the individual neurons. A
two-dimensional sketch of such an ellipse, and the geometric interpretation of the quantities
we use is depicted in Fig. 1A (see also Methods 3.2). In words, the way the orientation of
the response distribution changes with the average response varies between the three scenar-
ios: the long axis of the ellipsoid indicating the direction of the largest variance is aligned
along the diagonal for the feed-forward and the recurrent network models, but in the direc-
tion of the mean response for the gain fluctuation model (Fig. 1B). To capture this relation
quantitatively, we consider the projected variances, σ2
d (see Fig. 3 and methods for
details). There, the "diagonal direction" corresponds to a response where all neurons are
equally active.
µ and σ2
Quantitatively, we find that in the recurrent and feed-forward network models with rela-
tively strong correlations, there is always a large fraction of variance in the diagonal direction,
independent of the stimulus. The variance projected along the direction of the average re-
sponse is only large if the latter happens to be similar to the diagonal direction. By contrast,
in the gain fluctuation model, the variance projected along the direction of the average re-
sponse is approximately constant across stimuli. This behavior of the gain fluctuation model
is due to the fact that the variability across trials is multiplicative on the mean response, and
hence variability is large in this direction for all stimuli.
In the network models, strong correlations result from effective interactions between neu-
rons or from shared input. Each neuron in the recurrent network, or each input channel in the
feed-forward network, is effectively connected to a large part of the remaining neurons and
contributes to their correlations. Because our networks are not balanced, these contributions
are mostly positive, even for inhibitory input channels. Thus, across trials, responses of all
neurons are strongly correlated and the trial-to-trial fluctuations tend to be of similar size
for each neuron for a given stimulus. By contrast, the variability for any given neuron varies
more strongly across stimuli.
12
Figure 3: Response-covariance relations depend on origin of correlations. A-C: Average
population response (cid:104)r(cid:105) versus average (co-)variances in recurrent network (top), feed-forward net-
work (middle) and gain-fluctuation model (bottom). Each dot corresponds to a random stimulus,
blue dashed lines indicate analytic results, Eqs. (28)-(32). D-F: Dependence of normalized variabil-
ity projected on mean direction on response direction for the three scenarios. In D,E colors indicate
√
results for different networks, in F size of markers indicates strength of gain fluctuations. Square
cN . G-I: Same for the normalized variability
markers on the left indicate numerical value of
projected on diagonal direction. See Appendix for further details and the numerical parameters.
4.3 Relation between signal correlation and noise correlation
in recurrent network models
In the simple scenario we analyze here, average responses and covariances are related because
they both depend on the network architecture. By expressing signal correlations, cS
ij, and the
noise correlations, cN
ij , in terms of network and input parameters, we can derive a relation
between these two sets of quantities (Fig. 4). For a random transfer matrix, B, all neuron
pairs are statistically equivalent, and the strength of correlations can be characterized by
the average of the pairwise signal and noise correlations. In our recurrent network model,
the network architecture affects noise and signal correlations through an effective quantity,
13
namely the signal-to-noise ratio of the elements of the transfer matrix, ρ = var(B)/(cid:104)B(cid:105)2:
,
cS ≡ (cid:104)cS
ij(cid:105)i(cid:54)=j ≈ 1
1 + ρ
cN ≡ (cid:104)cN
ij(cid:105)i(cid:54)=j ≈ 1 + (N − 1)cin
1 + ρ + (N − 1)cin
(33)
(34)
(see Appendix 8.3.4, Eqs. (70) and (77)). Here, cin denotes the strength of signal correlations
in the input across stimuli. Both signal and noise correlations are larger in networks with more
homogeneous entries (smaller ρ). We already showed in Section 4.1 that noise covariances,
(cid:104)Cij(cid:105)i(cid:54)=j, depend on the average strength of the effective connections, while the variances,
(cid:104)Cii(cid:105)i, depend on var(B) as well. If the average input for pairs of neurons across stimuli is
uncorrelated (no input signal correlations, cin = 0), noise and signal correlations are identical
on average. However, due to a prefactor which grows with system size (N − 1 in Eq. (34)),
even weak input signal correlations are strongly amplified and can yield a large effect. Because
noise in the population response to a given stimulus is produced internally in the process of
spike generation, noise covariances are unaffected by this mechanism, so that strong signal
correlations can coexist with weaker noise correlations.
These relations depend on network parameters and hold for the average correlations
(across pairs in a network). The pairwise noise and signal correlation coefficients, cN
ij and
c2
ij, can vary widely, but numerical calculations indicate that they are correlated in a given
network (across pairs) as well.
Figure 4: Relation between signal and noise correlations in the recurrent network
model. A: Dependence of average signal correlations, cS, (continuous lines) and average noise
correlations, cN , (dashed lines) on input correlation and and network properties, Eqs. (33), (34).
Low variability of the transfer matrix elements, ρ, increases cN and cS. Input signal correlations
affect only cS. B: Same data, average signal versus noise correlation. C: Scatter plot of all pairwise
signal versus pairwise noise correlations (dots) in five network realizations for cin = 0.05. Circles
indicate network average across pairs, orange dotted line corresponds to analytic expressions shown
in B for cN and cS.
4.4 Stimulus discrimination in a recurrent network model
Since the network architecture affects both signal and noise correlation, it is natural to ask how
the discriminability of stimuli is affected in turn, as it depends on both quantities. Indeed,
noise correlations can affect the coding properties of a population appreciably [19]. Noise
14
correlations are referred to as favorable for the discrimination of a pair of stimuli when the
shapes of the two correlated response distributions (corresponding to the two stimuli) is such
that variability occurs predominantly in a direction orthogonal to the informative direction
(Fig. 5). The relevance of correlations can be quantified by comparing the discriminability
of stimulus pairs in the case of the full population output and in the case in which noise
correlations are removed by shuffling trials.
In Fig. 5, we visualize, for a recurrent network model, the interplay of the two effects
which affect the influence of correlations: on the one hand, how likely it is that the influence
of correlations be favorable for a pair of stimuli, and, on the other hand, how strong the effect
of correlations is for a given pair. For the first question, we utilize the fact that the noise is
predominantly along the diagonal direction. Consequently, the effect of correlations is favor-
able, if average responses differ in a direction orthogonal to the diagonal. The distribution
of the angle between these two directions will therefore reflect the likelihood of beneficial
correlations. To confirm this picture, and to examine the second question, we plot how the
effect of shuffling on the signal-to-noise ratio, S, depends on this angle (see Eq. (21) for the
definition of S). The network parameters, in this case the variability ρ of the elements of
the effective connections, affect both the distribution of angles and the effects on S - the
distribution of angles via the distribution of average responses, or the signal correlations, and
S by the strength of the noise correlations. However, it turns out that the effect of shuffling,
averaged across all pairs, is the same in each network. This is due to the fact that signal
correlations are on average as strong as noise correlations in all of these networks. In Section
4.8, we analyze a simplified network model to better understand some of these effects.
15
Figure 5: Covariances affect stimulus discriminability A: Sketch of response distributions for
two neurons and two stimuli (blue ellipses). Neural responses are correlated for both stimuli, so that
main axis of variability is parallel to linear separator (orange line). B: Larger ensemble of stimuli for
correlated neurons (large variance along diagonal directions). The effect of correlations is favorable
for stimulus pairs where difference in means is orthogonal to diagonal (red arrow), or unfavorable
when difference in means tends to be parallel to diagonal (green arrow). C: Distribution of cosines
of angle between diagonal and difference in means across stimulus pairs in recurrent networks with
different parameter ρ. In networks with low ρ, unfavorable angles are more frequent. D: Ratio
between discriminability for correlated and shuffled distributions, dots correspond to stimulus
pairs. Lower values indicate that correlations are more beneficial. Dashed horizontal lines indicate
average across all stimulus pairs.
4.5 Variability and correlations in populations of mouse cor-
tical auditory neurons
We use the theoretical results presented in the previous sections to analyze the responses of
a population of neurons to different stimuli. The data set contains the firing rates, collected
during a certain time interval, in response to the presentation of different sound stimuli in
a number of trials (see Section 3.1 and [32]). We will compare the properties of covariances
(across trials) and average responses with the predictions of our different models. Based
on the models, we can then evaluate the effect of network generated variability on stimulus
discrimination.
The trial-to-trial variability in single-neuron output is large (supra-Poisson, Fig. 6). Neu-
rons with larger average response exhibit also a larger variability in their responses. This
tendency is observable across different neurons responding to a given stimulus, for individual
neurons across stimuli and for the average population response across stimuli. The pairwise
correlations in trial-to-trial variability are generally high and pairs with high noise corre-
lations tend to have strong signal correlations. In other words, because signal correlations
measure the similarity of average responses across different stimuli, neurons with similar tun-
ing properties also have similar trial-to-trial variability. Additionally, in populations with
strong average signal correlations, noise correlations are strong as well. These relations be-
16
tween noise and signal correlations can be reproduced in the recurrent network model (Fig.
4). However, similar results may be possible also in alternative scenarios for the generation
of correlated variability. We compare the different models in the next section.
Figure 6: Variability and correlation in mouse auditory cortex. A: Average response
versus response variance in a single population. Colors correspond to 10 randomly chosen stim-
uli. Dots correspond to single neurons, circles to averaged values across population. Dashed line
indicates identity. Inset: All neurons for all stimuli. B: Scatter plot of signal correlation coeffi-
cient (across stimuli) vs. noise correlation coefficient (averaged across stimuli) for different neural
populations measured in the same animal, marginal histograms at top and bottom. Signal and
noise correlations are correlated across pairs. Correlations are high in general, but the amount of
signal and noise correlations varies strongly across populations. Circles denote average across pairs
in each populations. Inset: Black circles: Average signal and noise correlations in all measured
populations. Grey dots: individual pairs, 5 % randomly chosen from all experiments.
4.6 Analysis of population response variability in mouse au-
ditory cortex
Motivated by our investigation of network models, we first examine the relation between the
shape of the response distributions for each stimulus and the pattern of average responses
for the entire set of stimuli. In Fig. 7, we display the dependence of the normalized standard
deviation projected on the direction of the mean response, σµ/σall, and the normalized stan-
dard deviation projected on the diagonal direction, σd/σall, on the angle between average
response and diagonal, through cos(d, r) (see Section 3.2 for details and definitions). We
observe a strong dependence on cos(d, r) of the variance projected on the direction of the
mean response, but not of the variance projected on the diagonal direction. The dependence
of σµ/σall on the stimulus is much stronger in nearly all the measured populations. This
behavior is consistent with a network model, either feed-forward or recurrent, but not with
a model with shared gain fluctuations, where a large part of the variance is consistently in
the direction of the mean response (see Section 4.2.) Note, however, that in the data firing
rates are measured relative to the spontaneous activity; this is reflected in the presence of
negative values in the stimulus evoked activity. Such an offset could affect our comparison
because we do not know the true value of the mean response. To account for this possibil-
ity, we searched for the best possible offset, assuming the model of gain fluctuations for the
17
stimulus-dependent covariances. We then corrected the firing rates by this offset, and eval-
uated the stimulus dependence of σµ and σd, as before (see Appendix 8.3.2 for details), but
found no qualitative change in the results. However, we did not test for mixed models, and
it is possible that part of the correlated variability can be explained by shared fluctuations.
Based on the results presented above we conclude that, of our three scenarios, feed-forward
and recurrent network models are more consistent with the data. In the following, we fit the
parameters of these models to the data to find which of these two provides a better fit. For this
comparison, we analyze the dependence of the average variances, (cid:104)Cij(cid:105)i(cid:54)=j, and covariances,
(cid:104)Cii(cid:105)i, on the population-averaged response, (cid:104)r(cid:105)(s). In the recurrent network (Eqs. (27) and
(28)), both variances and covariances increase linearly with (cid:104)r(cid:105) across stimuli, while in the
feed-forward network (Eq. (32)), mean covariances are not expected to depend strongly on
(cid:104)r(cid:105).
To examine whether the experimental data is consistent with the feed-forward model, we
fitted parameters to the activity of each experimentally observed population (see Appendix
8.3.5), and generated a surrogate data set of responses and covariance matrices, with a match-
ing number of neurons and stimuli. In particular, we obtained values for the variability of
the input ensemble, ρE = var(rext)/(cid:104)rext(cid:105)2, and of the network elements, ρ = var(F )/(cid:104)F(cid:105)2.
It turns out that ρE, ρ > 1: for both input and network elements the variance is much larger
than the mean (see Fig. 8A-B). The reason is the high variability of average responses across
stimuli. With these parameters, the statistics of the distribution of average responses are well
reproduced (Fig. 8C). As we discussed in Section 4.2, for a high value of ρE the feed-forward
model does not predict a strong relation between (cid:104)Cij(cid:105)i(cid:54)=j and (cid:104)r(cid:105)(s). We quantify this re-
lation by the ratio of slope and intercept from a linear fit to the data, which indicates how
strongly the average covariances increase for stimuli that evoke strong population responses.
In the feed-forward scenario, this behavior does not obtain: the slopes of the linear fits in the
model relative to the intercepts are too low in comparison to what is observed in the experi-
ment (Fig. 8D, E). The feed-forward model thus cannot reproduce both the large variability
of neural responses across stimuli and the increase of the average covariance with the average
response.
In a recurrent model, this increase is expected, because noise generated by the neurons is
propagated through the network: the covariances are proportional to the average response (see
Eq. (28)). Moreover, it predicts that the ratios of intercept and slope in the linear behaviors
of variances and covariances are the same (it indicates the variance of the external input).
Indeed, linear fits to these relationships reveal approximately consistent ratios of intercept
and slope. The estimated parameters are summarized in Fig. 8G and H. The external noise,
estimated from the ratio intercept/slope, turns out to be of the same magnitude as the average
rate. Interpreted in terms of the model, the noise resulting from Poisson spike generation
thus contributes as much to neuron variance as the external input. This combined variability
is propagated through the network, and, on average, multiplied by a factor N(cid:104)B(cid:105)2 > 1,
corresponding to the slope in Eq. (28). Because this factor is larger than one, both average
response and noise are amplified by the recurrent connections.
In summary, we find that a model of recurrently connected neurons with random effec-
tive connections captures the observed activity statistics, in particular the relations between
average response and average covariances as well as the consistent direction of population
fluctuations across stimuli. Both purely feed-forward networks and gain fluctuation models
are not consistent with all of these observations. We note, however, that our conclusions are
based on a relatively small data set. A larger number of trials and measurements of absolute
values of firing rates would be desirable for a more stringent test.
18
Figure 7: Dependence of noise distribution orientation on average response in data. A,
B: Typical example, 6 neural populations recorded in the same animal. Relative variances projected
on mean and diagonal direction, versus cosine of angle between mean response and diagonal. Each
marker corresponds to a different stimulus. Different markers/colors denote different populations.
Squares to the left side indicate(cid:112)(cid:104)Cij(cid:105)i(cid:54)=j/(cid:104)Cii(cid:105)i. Solid lines: linear fit. C: Slopes from linear fits
as in A/B from σµ/σall vs. slope from σd/σall for all measured populations. Circles correspond
to slopes for positive, squares to negative cos(r, d). Colors in all panels indicate value of cN in
populations.
19
Figure 8: Scaling of covariances with average response in experiment and model. A,
B: Distribution of parameters ρ and ρE (variability of network connections and stimulus input)
estimated from all experiments. C: Histograms of cos(d, r(s)) from mean responses across stimuli,
for a selection of populations measured in one animal. Filled histograms: experiment. Solid lines:
Model results for 500 randomly generated stimuli and one network with random effective connec-
tions, parameters inferred from data. D: Population averaged response versus population averaged
variances and covariances for all stimuli (full circles) in one experiment and from corresponding
feed-forward model (empty circles). Dashed lines represent linear fits. E: Scatter plot of ratio
slope/intercept from linear fit in all experiments versus corresponding value in the feed-forward
model. Orange circle indicates population used in panel D. F: The ratio intercept/slope is, across
experiments (dots), consistent for covariances and variances. G: Distribution of estimated slopes
for average variances. H: Distribution of estimated (cid:104)Vext(cid:105) (intercept/slope from fit to variances) in
comparison to average and minimum firing rate across experiments.
4.7
Influence of noise correlations on stimulus discrimination
We quantify the influence of noise correlation on stimulus discrimination by the ratio Sshuffled/Soriginal.
S is calculated for the data set before and after shuffling trials. It is defined in Eq. (21) and
denotes, for a pair of stimuli, the difference in average response divided by the standard de-
viation of the responses, both projected on the most discriminative direction. Larger values
indicate that stimuli are easier to discriminate. If Sshuffled/Soriginal is larger than one, then
20
removing correlations by shuffling trials improves stimulus discrimination. Across pairs of
stimuli, this signal-to-noise ratio varies strongly (Fig. 9). On average, stimuli are slightly
easier to discriminate in the shuffled data, so that noise correlations are weakly unfavorable
to the encoding of the stimulus set used in the experiments.
A closer analysis of the response distributions reveals that, to a large degree, the effect of
shuffling can be explained by the relative locations of the two mean responses and the diagonal,
as measured by cos(d, r(s1)−r(s2)) (Fig. 5). This is because, due to the noise correlations, the
main direction of variability is along the diagonal. The overall effect of shuffling on stimulus
discrimination depends on the relation between noise and signal correlations: stronger signal
correlations lead to a stronger dominance of angles that are unfavorable if noise is aligned
along the diagonal direction, and in this case shuffling correlations away will benefit stimulus
discrimination.
Figure 9: Effect of correlations on stimulus discrimination. A: Signal-to-noise ratio S,
dots correspond to pairs of stimuli. Value from observed covariance matrix versus the one based
on shuffled trials, for different populations in the same animal. B: Same data, distribution of
S(shuffled)/S(original). Broad distribution, but mean value greater one for all populations (short
solid lines) indicates that shuffling increases information on average. C: Distribution of cosine of
angle between diagonal and average response difference across stimulus pairs for neural popula-
tions. Large cosines/small angles are most frequent. D: Scatter plot of effect of correlations on
discrimination. Small dots correspond to stimulus pairs. Connected large dots indicate average
value in a bin centered at the corresponding location on the x-axes. E: The distribution of mean
values across all measured populations shows that correlations are on average unfavorable for al-
most all populations. F: Average effect of correlations is stronger for increasing signal correlations.
4.8
Intuition from a two-population model
The relation between network properties and the statistics of variability, on the one hand,
and the relation between the statistics of variability and the encoding of stimuli, on the
other hand, have been studied widely. Above, we describe the structures of correlations that
emerge from different network architecture. We also analyze the structure of correlations in
data from mouse cortex, and we relate them to stimulus encoding. Here, we study a highly
idealized recurrent cortical network in order to develop an understanding of the relation
between network characteristics, response statistics, and stimulus encoding.
21
To understand the effect of recurrent connections on the discriminability of a pair of
stimuli, we consider two populations of excitatory neurons. The neurons in each of these
populations prefer one of the stimuli, and we want to understand the effect of connections
between these differently tuned neurons. The connectivity is regular, in the sense that each
neuron projects to a fixed number of neurons in its own population and to a fixed number of
neurons in the other population. Neurons in each population also receive the same external
input. We start by considering two stimuli, s1,2, where for each stimulus one of the populations
receives a stronger input. To discriminate between these stimuli, it is sufficient to consider
the network response at the population level, that is, the sum of the response of neurons
in each of the two populations. This follows from the regularity of the connections. The
distribution of responses across trials can thus be described by a two-dimensional vector, R,
of the average population response and a covariance matrix, Σ, of the variability. The sum
of the couplings within a population, Γs, and across populations, Γc, determine the response
via a symmetric 2 × 2 coupling matrix, Γ. Here, we will manipulate the entries of Γ, rather
than the effective population coupling matrix, P = (I − Γ)−1 (see Fig. 10 and Methods for
further details of the model).
We start by examining how cross-coupling between the populations can induce beneficial
correlations for stimulus discrimination, by considering the effect of connections between the
populations on the covariances.
If we introduce more connections across populations, the
cross-coupling Γc increases and the response distributions evoked by the two stimuli change
(Fig. 10). The responses of the two populations are amplified, but also more similar - the
difference in population response, R(s1) − R(s2), decreases. Simultaneously, the correlation
between the population responses increases, rendering the distributions more elongated. To
evaluate the effect of shuffling trials, we need to consider the covariances between individual
neuron pairs. As we show in detail in Appendix 8.4, covariances reduce the overlap of the
distributions if Γc > Γs. Then, covariances between neurons belonging to different populations
are larger, on average, than for neurons of the same sub-population, and the signal-to-noise
ratio, S, becomes smaller if covariances are shuffled away (Fig. 10C, D). While it thus appears
that strong coupling between differently tuned neurons induces beneficial pairwise correlations
this is not always the case.
To see this, we compare networks with different parameters, and disentangle the effects of
cross-connections on average responses, variances, and covariances. Increasing the coupling
between populations decreases S both in the shuffled and in the original cases (Fig. 10E),
but this suppression is stronger when covariances are shuffled. The suppression can be traced
back to the amount of noise produced through spike generation. Recurrent amplification
increases firing rates, and the variance of Poisson neurons increases with rate. If the amount
of noise remained constant, S would not depend on Γc:
larger coupling induces stronger
correlations which reshape noise along a direction irrelevant for discrimination, but at the
same time average responses approach each other. These two effects precisely cancel each
other, ultimately because a linear transformation of responses by the network cannot reduce
noise.
Above, we only considered a pair of stimuli for which beneficial correlations are realized in
our network. We now evaluate the effect of cross-connections on a more general stimulus en-
semble. We define it via the input vectors for the two populations, (1+∆ cos(φ), 1+∆ sin(φ))T
for φ ∈ (0, 2π) and a constant ∆ (see Fig. 10G and Appendix 8.4.1). For this ensemble we
see that inevitably correlations will be unfavorable for certain stimulus pairs, depending on
the stimulus dimension in which their inputs differ. The averaged effect of shuffling across
this two-dimensional stimulus ensemble depends not only on the noise correlations, but also
on the distribution of the input, which affects the distribution of the average responses, i.e.
the signal correlations.
22
If excitatory recurrent networks amplify internally generated noise, are feed-forward ar-
chitectures generally more efficient? In the following, we compare the performances in the
two scenarios. As we noted before, the difference between the two scenarios is that the noise
produced by spike generation in the recurrent network is fed back into the network, and
therefore correlated, while it is independent in the feed-forward network. The measure of
performance we used so far is the discriminability of pairs of stimuli. We want to compare
the effect on information of a recurrent network, with a transfer matrix P defined by the con-
nectivity parameters, to a feed-forward network with identical transfer matrix, for changes
of a high-dimensional input in arbitrary directions. A quantity that measures the informa-
tion of a network response statistics in the case of a continuous stimulus is the linear Fisher
information matrix. In the two-population model, it is given by
in the recurrent case and by
IR = (D[R] + Σext)−1
IF = PT (PΣextPT + D[R])−1P
(35)
(36)
in the feed-forward case. Here, Σext denotes the variance of the external input to the two
populations (see Appendix 8.4.3 for details). As figures of merit for the coding of high-
dimensional stimuli, we use the traces of IR and IF, for the recurrent and feed-forward cases
respectively. In the recurrent network, this quantity depends only on the firing rates. These,
in turn, depend on the average strength of the effective connections, (cid:104)P(cid:105) (Fig. 10H): the larger
(cid:104)P(cid:105), and therefore the response, the more noise is generated, and the more information is
lost. In a corresponding feed-forward network, information depends also on the correlations.
However, because the network-generated noise is uncorrelated, increasing firing rates effec-
tively improves the signal-to-noise ratio, and therefore increases the information. Whether a
feed-forward or a recurrent architecture is preferable thus depends on the output firing rate.
From this model, it is not clear how firing rates should be chosen, but additional constraints,
for example imposed by energy consumption are conceivable.
A simple model with two coupled, uniform populations illustrates how variations in the
network parameters can affect stimulus discriminability, through modulations in the average
responses and variability. Broadly speaking, stimulus discriminability depends on the inter-
play between noise creation, average response, and noise correlation, and the relative behavior
of these quantities depends on the network architecture.
23
Figure 10: Model of two coupled excitatory populations. A: Circuit model. Neurons
are coupled with excitatory connections within and across two populations. For each stimulus,
one population receives stronger input. Population responses are described by macroscopic vari-
ables. B: Responses to two stimuli (colored dots on dotted red line) are amplified for stronger
cross-coupling (brighter colors). Simultaneously, response fluctuations are more correlated. If in-
ternally generated noise does not increase with rate, variability is smaller (dashed ellipses). C
and D: Distribution of pairwise covariances within (dark bars) and across populations (light bars),
vertical lines indicate average. If cross-coupling is weak (C), average covariance is larger within
populations than across, if it is strong (D), covariances between neurons of different populations
are larger. E: Dependence of stimulus discriminability S on Γc. Strong couplings implies smaller
discriminability and even more so for shuffled trials. F: Increasing cross-coupling reduces difference
in mean response (black line). Effect on variance in the relevant direction differs for correlated
populations (solid orange line) and model with independent neurons/shuffled correlations (dash-
dotted orange line). G: Red dots: average response for stimulus ensemble. Ellipses: Response
distributions for four stimuli for the recurrent network (blue) and a feed-forward network with
identical average responses (red). Dotted/dashed lines: stimulus pairs where correlations have
positive/negative influence.H: Linear Fisher information, dependence on average of population
transfer matrix elements (horizontal axes) for recurrent and feed-forward scenario. Darker lines
indicate larger variance of transfer matrix elements. Lines overlap in the recurrent scenario.
24
5 Discussion
Despite large amounts of work on the topic, the interpretation of the observed response
fluctuations in population activity, in the context of computation by neural circuits and
information coding, is not elucidated [11].
In the present study, we seek to understand
the influence of correlated variability on stimulus representation, based on the exploration of
model networks of stochastic neurons together with data analysis. We modeled neural activity
using a set of coupled Poisson processes in order to study the relations between covariances
and average responses in a neural population, across different stimuli. We found that these
relations differ depending on how covariances are generated, and that correlations observed in
mouse auditory cortex are consistent with those generated in a recurrent network model. This
model also helps us interpret the effect of the observed correlations on the discriminability of
stimuli and the nature of the propagation of information through the network.
5.1 Origins of the structure of correlations in neural circuits
In the examination of the impact of correlated response variability on stimulus represen-
tation, it is useful to be aware of the possible range of its mechanistic origin. Spike train
variability has been attributed to the dynamics in recurrent networks [1, 37]. Along with this
source of variability, in stable states of asynchronous activity, recurrent network models can
also produce transitions between different states of global activity [3, 38, 39]. Alternatively,
external signals fed to sub-populations of neurons, such as gain fluctuations induced by up-
stream areas, were pointed to as causing high variability and correlations [7, 15, 9, 8, 24].
Population-wide gain fluctuations can also reproduce dependencies between noise and signal
correlations [10].
In general, it is difficult to differentiate shared input or gain fluctuations, and recurrent
connections, as the origin of correlated variability, based on measured activity. Here, instead
of focusing on temporal dynamics [40, 13, 41], we approach the problem by considering the
relation between the network architecture and the structure of population activity statistics,
when a neural population is presented with an ensemble of stimuli. Our analysis makes the
assumption that noise resulting from spike generation in neurons can be modeled as a Poisson
process, and, thus, depends on the firing rate. If the noise generated within the network is
purely additive and independent of firing rates, it is indistinguishable from external noise
that is simply filtered by the network [35]. In the Poisson model, however, the interplay of
internally generated noise and variability due to external signals imprints its signature on the
dependence of covariances on firing rates.
5.2 Influence of correlations on stimulus discrimination in neu-
ral populations
We use a recurrent network with random effective connections to estimate the influence of
the variability generated by the network dynamics on stimulus representation. Based on
the paradigm that response variability amounts to noise that downstream neurons have to
cope with, a number of authors have argued that noise correlations can benefit information
coding [19, 21, 23, 22, 24, 25]. Furthermore, recent studies have examined specifically the
impact of correlations generated by recurrent networks. Reference [30] stresses that it is
the structure of the noise, more than its magnitude, which determines the role it plays for
information coding. Harmful structures of noise correlation, causing information to saturate
in large neural populations, are the consequence of the amount of irreducible external noise
and potentially sub-optimal processing [31], but these structures are difficult to pinpoint
25
in detail in measured activity [30]. By contrast, it was demonstrated that noise internally
generated in a network via spike generation can be averaged out in feed-forward or recurrent
networks, depending on the connectivity in the network and the redundancy of the population
code [27].
In this work, the goal was to start from the observed properties of the activity, and to
interpret mean activity and noise correlations based on outcomes of model network dynamics.
Variations of network properties, such as the architecture of connections, change both the
mean activity and noise correlations, both of which influence the accuracy of the neural code.
Consequently, it is useful to consider both simultaneously to understand network effects. The
effect of recurrent dynamics on input information depends mostly on the amount of internally
generated noise. Our toy model, as well as Ref. [27], argue that recurrent amplification
suppresses information. The reason is essentially that, for Poisson neurons, larger firing
rates imply a higher variance. The stronger the recurrent connections, the more the average
response is amplified, and the more noise is produced internally. Because this noise is fed
back into the network and thus amplified furthermore, the variance is increased to the point
that the signal-to-noise ratio decreases. We also show that under simplifying assumptions,
including a random effective network architecture and a large, random stimulus ensemble, the
amplification factor of the network and the amount of noise can be estimated from measured
responses.
Limitations of our analysis result, in particular, from the assumption of linearly interacting
Poisson processes. Nonlinear transfer functions can affect the relation between rates and
correlation in individual pairs [42, 11]. Using Poisson neurons also implies that the generation
of noise in the network is attributed to individual neurons. The tractability of the model,
however, allows us to obtain relations between observable statistics, like noise and signal
correlations, which can be used as a starting point to interpret correlated variability in terms
of its mechanistic origins.
5.3 Variability in mouse auditory cortex
We tested the consistency of different scenarios involving interacting Poisson neurons with
measured neural population responses, and we interpreted the observed variability in terms
of a recurrent network model. The measured population activity is given as spike counts in
200 ms bins, so that correlations are defined over a relatively slow timescale and, therefore, do
not take into account fast temporal structure (as noted in, e.g., Ref. [43]). Both experimental
variability and correlations are high [44], so that the effects of correlations are potentially
strong and amenable to an analysis such as ours. Although variability and correlations are
often high in anesthetized animals [9], noise and signal correlations were somewhat weaker
in comparable experiments with anesthetized animals [45]. It is conceivable that part of the
correlations are due to experimental artifacts from the calcium imaging technique used in the
experiments (like scattering of fluorescent light by the neuropil), but intracellular recordings
of rates were consistent with calcium recordings [32].
We model the set of stimuli as a generic, high-dimensional ensemble, and the network
connections at the level of a transfer matrix. When analyzing the discriminability of pairs of
stimuli, this enables a broader analysis than one relying on a single stimulus dimension or av-
erage correlation. The recurrent network model explains not only the scaling of variances and
covariances with firing rates, but also the relation between signal and noise correlations, and
the effect of shuffling away covariances on stimulus discrimination, in a natural way. Differ-
ences among populations of neurons can be attributed to changes in the effective connectivity
within populations.
Recent studies have explained correlated variability in terms of "external" gain fluctu-
26
ations acting on a (sub-)population [7, 15, 9, 8, 24]. We showed that recurrent networks
of strongly connected neurons can exhibit population-wide fluctuations as well. Estimated
model parameters indicate an amplification of input signals by excitatory recurrent con-
nections. This suggests that the noise that can be traced back to spike generation in the
population - noise harmful to coding - is comparable in magnitude to the external input
noise.
Our data analysis suffers from an unknown offset in the firing rates, a relatively small
number of trials, and the low temporal resolution. We therefore only compared different pro-
totypical scenarios without trying to evaluate the outcome of a mixed model. In experimental
populations of neurons, the statistics of activity can obviously result from a combination of
both feed-forward and recurrent processes.
6 Acknowledgments
We thank Brice Bathellier for providing the experimental data set as well as commenting on
it, and for discussions.
This work was supported by the CNRS through UMR 8559 and the SNSF Sinergia Project
CRSII3 141801.
7 Author Contributions
Conceived and designed the experiments: VP RS. Performed the experiments: VP. Analyzed
the data: VP. Contributed reagents/materials/analysis tools: VP RS. Wrote the paper: VP
RS.
27
8 Appendix
8.1 Numerical simulations: details and parameters
of a rate offset a = 4, so that C = (I − G)−1(cid:0)D[r] + a + D[Vext](cid:1)(I − GT )−1. The stimulus
In Fig. 2, covariances were calculated for recurrent networks from Eq. (12), with the addition
ensemble consisted of 200 input vectors rext of size N = 60, with entries independently chosen
from a normal distribution with mean 0.2 and standard deviation 1. Rates r were calculated
as in Eq. (7), and we assumed Poisson input with Vext = rext. For the networks, four
connectivity matrices G of size N = 60 were generated. Their entries were chosen from a
normal distribution with standard deviation 1.5/N and mean 0.8/N − 0.9/N , respectively.
For the analytic predictions from Eqs. (28)-(32), the mean and variance of the transfer matrix
B = (I − G)−1 were calculated numerically.
For Fig. 3, for the input ensemble for the recurrent network, elements of 200 random input
vectors rext of length N = 60 were chosen from a normal distribution with mean 0.4 and
variance 2. Connectivity in the recurrent network was as in Fig. 2. For rates and covariances
of the feed-forward scenario, we use Eqs. (15) and (16). We set F = 10 · B, and scaled the
input vectors by a factor 1/10 to obtain rates and correlations of comparable size as in the
recurrent network. Negative output rates were set to zero. For the gain fluctuation model,
responses of the network with ρ = 1.4 were used as the ensemble of stimulus responses. From
these, covariances were calculated, see Eq. (18), for which the variance of external fluctuations
Vext was varied. For this figure, no rate offset was used, a = 0.
The networks in Fig. 4C were identical to the ones used above. 50 stimuli rext with random
entries were generated with mean 0.2 and standard deviation 1.5. Inputs for different neurons
were correlated with correlation coefficient cin = 0.05 across stimuli. Rate offset was 4.
In Fig. 5, networks and stimulus parameters were the same as in Fig. 4, but there were
no input signal correlations, cin = 0.
In Fig. 10 the network connects two populations of 100 neurons each. For the pair of
stimuli, panel B, neurons in the population that was more strongly excited received an input
of 1 + ∆, with ∆ = 0.2, and the others input 1. More general inputs (panel G) were
defined by setting the inputs Rext = (1 + cos(φ)∆, 1 + sin(φ)∆) for φ ∈ (0, 2π). Each neuron
had a fixed number of nEE postsynaptic partners within the same populations. Coupling
strength between connected neurons was set to g = 0.01. We chose a population coupling
within populations of Γs = 0.2, and the number of connections between neurons within
populations such that Γs = nEEg. Connections across populations were chosen accordingly
to realize across coupling values Γc = 0, . . . , 0.4. To visualize the effect of the increase in
internally generated noise due to increased rates (panel B), we compared the covariances
Σ = BpD[R](Bp)T for increasing Γc to the ones where in D[R] the rates of a network with
Γc = 0 were used.
In panels E and F, we calculated Soriginal by inserting the population covariances and
responses in Eq. (21) for the two stimuli, for the correlated case. For the shuffled case, the
full matrix covariance matrix C was generated and the off-diagonal elements set to 0, such
that
(cid:18)(cid:80)
Σshuf f led =
.
(37)
(cid:19)
(cid:80)
i∈E Cii
0
0
i∈E(cid:48) Cii
For panels E-H, population level parameters in the matrices Γ or P were varied directly.
We set the external noise Σext = N D[Rext]. In panel G, Γc = 0.4 and Γs = 0.2. In panel
H, we fixed for the population transfer matrix P the ratio ρ = (cid:104)P(cid:105)2/var(P ) to the values
ρ = 0.5, 1, 2. For each of these values, (cid:104)P(cid:105) was varied between 0.1 . . . 5, so that var(P ) was
fixed. From (cid:104)P(cid:105) and var(P ), we calculated Ps and Pc, see Appendix 8.4.1.
28
8.2 Common framework for the generation of correlations
Here we formulate the three model scenarios described in Sections 3.3 and 3.4 as special cases
of the interacting point processes framework defined in Eq. (6). The derivation is similar as
in [33, 27]. We consider an extended network of two populations which receive only constant
input rfull(t) ≡ rfull. Neurons are divided into an observed population O and an unobserved
external population U . The external population projects to the observed population, but
does not receive feedback. The coupling matrix of integrated kernels of the full system is a
block matrix of the shape
Gfull =
.
(38)
(cid:18)E 0
(cid:19)
F G
The coupling between nodes within the external network is described by the matrix E. If
E = 0, external inputs are independent Poisson processes. The feed-forward weights to the
observed network are defined by F , and by G the recurrent connections within the observed
population. The input vector to the system is rfull = (r0, 0). The components of the vector
r0 together with E determine the firing rates of the input nodes. The 0 represents a vector
of zeros, so that there is no constant input directly to the observed population.
The time dependent firing rates of neurons k ∈ U in the external input population are
determined by
rk(t) =
ekl(τ )sl(t − τ )dτ + r0,k
and the rates of neurons i ∈ O in the observed population by
(cid:90) ∞
0
(cid:88)
l∈U
(cid:90) ∞
(cid:88)
j∈O
0
(cid:90) ∞
(cid:88)
k∈U
0
ri(t) =
gij(τ )sj(t − τ )dτ +
fik(τ )sk(t − τ )dτ.
(39)
(40)
(41)
(42)
The transfer matrix of the full system is
(I − E)−1
Bfull = (I − Gfull)−1 =
(I − G)−1F (I − E)−1
(cid:18)
(cid:19)
≡
(cid:18) BE
0
BF BE B.
(cid:19)
0
(I − G)−1
The average rates of the system are, from applying Eq. (7) to the full system,
(I − Gfull)−1rfull = Bfullrfull = (BEr0, BF BEr0)T ≡ (rext, r)T ,
where rext = BEr0 are the rates of the input neurons and r = BF BEr0 = BF rext corre-
spondingly, the rates of the observed neurons. From Eq. (12), the covariance matrix is given
by
Cfull =
BF BED[rext]BT
E B
BED[rext]BT
EF T BT
D[r] + F BED[rext]BT
.
BT
(43)
(cid:32) BED[rext]BT
E
(cid:16)
(cid:33)
EF T(cid:17)
The block CE ≡ BED[rext]BT
population. The block at the lower right,
E at the upper left describes the covariances of the external
C ≡ B(D[r] + F CEF T )BT
(44)
describes how covariances in the observed network depend on the properties of the unobserved
input neurons.
The recurrent network scenario is obtained for an input population and an output popu-
lation of equal size N , with F = I and any diagonal E. In this case, CE is diagonal, and we
can identify the elements on the diagonal of CE with Vext.
29
For the feed-forward scenario set G = 0 and again identify the diagonal elements of CE
with Vext.
The gain fluctuation model is obtained for G = 0, and if the input population consists
of a single neuron. The matrix F then corresponds to a single column vector, and C =
D[F rext] + F CEF T . With r = F rext and setting Vext = CE/r2
ext (rext is a number in this
case), C = D[r] + rrT Vext.
8.3 Further details on statistics in different model scenarios
8.3.1 Gain fluctuation model
We derive relations between covariances and average response in the gain fluctuation model,
in particular the scaling of average covariances with average rates and the orientation of the
response distribution measured by the projections of the variances in different directions.
From Eq. (18), pairwise covariances are directly related to average responses,
Cij = δijri + rirjVext.
(45)
The relation between average covariance and the population averaged response, Eq. (32),
follows from
(cid:104)Cij(cid:105)i(cid:54)=j =
1
N (N − 1)
rirjVext ≈ Vext
N 2
= Vext(cid:104)r(cid:105)2,
(46)
rj
(cid:16)(cid:88)
(cid:17)(cid:16)(cid:88)
ri
i
j
(cid:17)
(cid:88)
i(cid:54)=j
where the terms r2
i can be neglected if N is large.
To measure changes of the response distribution across stimuli, we use the variance pro-
d, respectively. In
µ does not strongly depend on the
jected in the direction of mean response and diagonal direction, σ2
the following, we give an argument that in this model σ2
stimulus, while σ2
µ and σ2
d does.
Both quantities are normalized by the sum of the variances
σ2
all =
Cii =
i Vext + ri) = r2Vext + N(cid:104)r(cid:105).
(r2
√
The variance projected in the diagonal direction ¯d = (1, . . . , 1)T /
i
N is
(cid:17)
(cid:88)
(cid:88)
(cid:16)
(cid:88)
i
(cid:88)
(47)
(48)
(49)
σ2
d =
Cij(s) ¯di ¯dj =
rirjVext + δijri
/N = N Vext(cid:104)r(cid:105)2 + (cid:104)r(cid:105).
ij
ij
To see that there is a strong dependence of σ2
cos(d, r) = ¯d¯rT =
N(cid:88)
i=1
d/σ2
all on stimulus direction, note that
rir =
√
N(cid:104)r(cid:105)/r.
1√
N
For non-vanishing r, it follows from cos(d, r) = 0 that (cid:104)r(cid:105) = 0 and in this case σ2
If cos(d, r) = 1, r ∝ d and r2 = N(cid:104)r(cid:105)2, and so σ2
N Vext(cid:104)r(cid:105)2+N(cid:104)r(cid:105) . This value
is of the order of one, if the ratio of Vext(cid:104)r(cid:105)2 and (cid:104)r(cid:105), which is of the order of the average
noise correlation coefficient, is not very small (in comparison to one). Consequently, if noise
correlations are not too small, the normalized variance projected on the diagonal direction
strongly depends on the direction of the response vector.
all = N Vext(cid:104)r(cid:105)2+(cid:104)r(cid:105)
d/σ2
d/σ2
all = 0.
The variance projected on the mean direction is
σ2
µ =
¯ri¯rjCij =
¯ri¯rj(rirjVext + δijri) > Vext
j /r2 = Vextr2
i r2
r2
(50)
(cid:88)
(cid:88)
(cid:88)
ij
ij
ij
30
Hence, σ2
r2Vext+N(cid:104)r(cid:105) , which depends only weakly on (cid:104)r(cid:105), if N(cid:104)r(cid:105) is not much
bigger than Vextr2 ≥ VextN(cid:104)r(cid:105)2, that is once again if noise correlations are not too small.
µ/σ2
all >
Vextr2
8.3.2 Additional test for a common origin of shared variability
If rates are only measured up to an unknown offset a, the variance σ2
all, which is obtained
from a projection on the apparent average response r − a may not be constant across stimuli.
In this case, the variation of this quantity across stimuli is not a reliable indicator to exclude
a gain fluctuation model as the source of observed correlations. However, if a large part of
the covariances can be explained by a single component, it should be possible to reconstruct
the common origin from the stimulus dependent covariance matrices. From
µ/σ2
C(s) =(cid:0)r(s) + a(cid:1)T(cid:0)r(s) + a(cid:1)Vext + D(cid:0)r(s) + a(cid:1)
largest eigenvalue of C, by neglecting the contribution of D(cid:0)r(s) + a(cid:1). This approximation
the vectors v(s) = r(s) + a can be obtained approximately by finding the eigenvector with the
(51)
can also be avoided by applying a factor analysis with a single latent component. Because a is
constant across stimuli, up to measurement errors of the estimated covariances, the straight
lines through the mean responses r + x¯v, for x ∈ IR and normalized directions ¯v, will intersect
in the point −a. The best intersection point of multiple lines in the least square sense is given
by
(cid:16)(cid:88)
I − ¯v(s)¯v(s)T(cid:17)−1(cid:16)(cid:88)
− a =
(I − ¯v(s)¯v(s)T )r(s)
(cid:17)
(52)
s
s
and can be used to correct the average responses to r(cid:48)(s) = r(s) + a. We found that
the analysis of σ2
d based on the corrected responses lead to no qualitative change in
the stimulus dependence, indicating that a potential shared component is too weak to be
identified based on the available data.
µ and σ2
8.3.3 Feed-forward model
Here we derive characteristic relations for the response statistics resulting in a feed-forward
network. They illustrate qualitative differences between the predictions of different models
and will be used to extract model parameters from the data.
In the feed-forward model
the mean responses and covariances are, from Eqs. (15) and (16) and allowing an offset in
observed rates,
r = F rext,
FikFjkVext,k.
(53)
(54)
Cij = δij(ri + a) +
(cid:88)
k
In contrast to the recurrent network model (see below), the average covariance is not
correlated to the population averaged rate (cid:104)r(cid:105) across stimuli, Eq. (30). This follows after
taking the average across neurons:
(cid:104)Cij(cid:105)i(cid:54)=j =
1
N (N − 1)
FikFjkVext,k ≈ N(cid:104)Fik(cid:105)ik(cid:104)Fjk(cid:105)jk(cid:104)Vext,k(cid:105)k = N(cid:104)F(cid:105)2(cid:104)Vext(cid:105).
(55)
(cid:88)
i(cid:54)=j,k
Input variances Vext are assumed to be independent of the network structure and Fik indepen-
dent of Fjk, which means that the strengths of connection of an external neuron to different
internal neurons are independent.
If (cid:104)Cij(cid:105)i(cid:54)=j is to be uncorrelated to (cid:104)r(cid:105) across stimuli, the average input variance (cid:104)Vext(cid:105)
needs to be approximately uncorrelated to (cid:104)r(cid:105). This holds in our model even though rk =
31
(cid:80)
0, or the distribution of column sums(cid:80)
i Fikrext,k and in each input channel the variance equals the strength of the input, Vext,k =
rext,k, if either the distribution of inputs across neurons is approximately symmetric around
i Fik of the feed-forward matrix (across columns) is
symmetric around 0. The first case can be realized in a feed-forward network with a similar
number of excitatory and inhibitory input channels. Intuitively, inhibitory inputs decrease
the average output, but contribute positively to the average variance, and thus decorrelate
the two quantities.
Formally, one compares the random variable(cid:80)
the variable(cid:80)
i Fik to
k rext,k. A random variable x is uncorrelated to its absolute value
x if its distribution is symmetric around 0. Here, the relevant variables are the elements
of the input vector rext, and their distribution is approximately symmetric around 0, if the
variance var(rext) is much larger than their mean (cid:104)rext(cid:105), that is if ρE = var(rext)/(cid:104)rext(cid:105) (cid:29) 1.
Consequently, if the variance of external inputs across stimuli is high, the population averaged
covariances in a feed-forward network are uncorrelated to the population response.
ik Fikrext,k =(cid:80)
k Vext,k =(cid:80)
k rk =(cid:80)
(cid:80)
k rext,k
In the following, we motivate that the stimulus dependence of the variances projected
along different directions in the feed-forward model is different from the one in the gain
fluctuation model. The sum of the variances is
ikVext,k ≈ N(cid:0)(cid:104)r(cid:105) + a(cid:1) + N 2(cid:104)Vext(cid:105)(cid:104)F 2(cid:105).
F 2
all = N ((cid:104)r(cid:105) + a) +
σ2
(56)
Due to the Poisson spike generation, there is a linear contribution in (cid:104)r(cid:105) to the average
variance, (cid:104)Cii(cid:105)i = σ2
all/N . This variability does not contribute to covariances, because spikes
are generated independently across neurons, in contrast to the recurrent model, where the
spiking or not spiking of a neuron directly influences post-synaptic firing rates.
The variance projected onto the diagonal direction is
d = (cid:104)r(cid:105) + a +
σ2
1
N
FikFjkVext,k ≈ (cid:104)r(cid:105) + a + N 2(cid:104)Vext(cid:105)(cid:104)F(cid:105)2.
(57)
(cid:88)
i,k
(cid:88)
ijk
d/σ2
all depends only weakly on (cid:104)r(cid:105) if the term N(cid:104)Vext(cid:105)(cid:104)F(cid:105)2 is large
Consequently, the ratio σ2
against (cid:104)r(cid:105)+a, that is if the sum of covariances is larger than the sum of variances. This is the
case, if correlation coefficients are larger than of the order O(1/N ). The variance projected
(cid:88)
onto the direction of the mean response is
(cid:88)
(cid:80)
FikFjkVext,k ¯ri¯rj ≈ a +
i r3
i
r2 +
(cid:104)r(cid:105)2
r2 N 3(cid:104)F(cid:105)2(cid:104)Vext(cid:105).
σ2
µ =
(ri + a)¯r2
i +
i
ijk
(58)
µ/σ2
all, assume again that the sum across variances in the first term a +
To see that the ratio σ2
all strongly depends on the population response, in contrast to
σ2
d/σ2
is not too large
against the sum across covariances in the second term. The second term depends strongly on
(cid:104)r(cid:105): it is 0 for cos(r, d) = 0. For cos(r, d) = 1 it becomes N 2(cid:104)F(cid:105)2(cid:104)Vext(cid:105). In this case, σ2
µ/σ2
is not too small, if ((cid:104)r(cid:105) + a)/(N Vext) is not much larger than 1 (assuming that (cid:104)F 2(cid:105)/(cid:104)F(cid:105)2 is
all
of order one).
i r3
r2
i
Finally, we calculate the average noise correlation coefficient (across stimuli and neuron
(cid:80)
pairs)
cN = (cid:104)
(cid:112)Cii(s)Cjj(s)
Cij(s)
(cid:105)s,i(cid:54)=j.
(59)
For a given stimulus, we assume that the Cij are approximately independent, so that we can
write
.
(60)
(cid:112)Cii(s)Cjj(s)
Cij(s)
(cid:104)
(cid:105)i(cid:54)=j ≈ (cid:104)Cij(s)(cid:105)i(cid:54)=j
(cid:104)Cii(s)(cid:105)i
32
Then, from Eqs. (55) and (56),
cN =
N(cid:104)F(cid:105)2(cid:104)Vext(cid:105)
(cid:104)r(cid:105) + a + N(cid:104)Vext(cid:105)(cid:104)F 2(cid:105) .
(61)
Signal correlations can be calculated analogously as for the recurrent model, see below.
8.3.4 Recurrent network model
The response distributions in the recurrent network are characterized by the average and
covariances of responses given in Eqs. (7) and (12),
and
r = Brext
C = BD[reff ]B,
(62)
(63)
with reff = r+a+Vext, including an offset a. To compare this model to the alternative scenarios
described in the previous sections, we calculate the relation between average (co)variances
and the projected variances and r as well as the signal and noise correlations.
We first derive Eqs. (27) and (28). The average covariance is
(cid:104)Cij(cid:105)i(cid:54)=j =
1
N (N − 1)
BikBjkreff,k ≈ N(cid:104)BikBjk(cid:105)ijk(cid:104)reff,k(cid:105)k.
(64)
(cid:104)Bij(cid:105)(cid:104)(cid:80)
(cid:80)
We assume that N is sufficiently large and that the external input characterized by Vext, rext
In that case reff,k is approximately uncor-
does not depend on the recurrent network.
related to Bik (including the contribution of rk, because (cid:104)Bijrk(cid:105) = (cid:104)Bij
l Bklrext,l(cid:105) ≈
l Bklrext,l(cid:105)). The term (cid:104)reff(cid:105) = (cid:104)r(cid:105) + a + (cid:104)Vext(cid:105) is linear in (cid:104)r(cid:105), with an uncorre-
lated contribution (cid:104)Vext(cid:105), see Appendix 8.3.3. If the elements of B are pairwise independent,
(cid:104)Cij(cid:105)i(cid:54)=j = N(cid:104)B(cid:105)2((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105)).
Under the same assumptions, the sum of variances is
(cid:88)
i(cid:54)=j,k
(cid:88)
σ2
all =
ikreff,k =
B2
ik(Vext,k + rk + a)
ik
(Vext,k + rk + a) = N 2(cid:104)B2(cid:105)((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105)).
Using the same argument as in the feed-forward scenario following Eq. (55), the term N 2(cid:104)B2(cid:105)(cid:104)Vext(cid:105)
only contributes with an offset to the relation between (cid:104)r(cid:105) and (cid:104)Cii(cid:105)i = σall/N .
(cid:104)r(cid:105) as the ones in the feed-forward model. The projected variance on the diagonal is
Next, we show that the projected variances in this scenario have a similar dependence on
σ2
d =
1
N
BikBjkreff,k ≈ N 2(cid:104)B(cid:105)2((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105)).
Consequently, in the ratio σ2
all the dependence on (cid:104)r(cid:105) is canceled out.
d/σ2
The projection along the mean response direction,
(cid:88)
ij,k
σ2
µ =
BikBjk
rk + a + Vext,k
(cid:17)
¯ri¯rj ≈ N 3(cid:104)r(cid:105)2
r2 (cid:104)B(cid:105)2((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105)),
33
(cid:88)
≈ N(cid:104)B2(cid:105)(cid:88)
B2
i,k
k
(cid:88)
ij,k
(cid:16)
(65)
(66)
(67)
(68)
depends strongly on (cid:104)r(cid:105). Because cos(d, r) = 0 implies (cid:104)r(cid:105) = 0, σ2
for cos(d, r) = 1, with r2 = N(cid:104)r(cid:105)2 one gets σ2
all = (cid:104)B(cid:105)2/(cid:104)B2(cid:105).
Apart from the raw covariances, we are interested the average noise correlation coefficient
(across stimuli and neuron pairs). As for the feed-forward model, we approximate cN ≈
(cid:104)Cij(cid:105)s,i(cid:54)=j
(cid:104)Cii(s)(cid:105)s,i
all = 0 in this case, and
. From Eqs. (65) and (66)
µ/σ2
µ/σ2
cN =
N(cid:104)B(cid:105)2((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105))
N(cid:104)B2(cid:105)((cid:104)r(cid:105) + a + (cid:104)Vext(cid:105))
(cid:104)B(cid:105)2
(cid:104)B2(cid:105) ,
=
and because (cid:104)B2(cid:105) = var(B) + (cid:104)B(cid:105)2
cN =
1
1 + var(B)/(cid:104)B(cid:105)2 =
1
1 + ρ
.
(69)
(70)
We see that the correlation coefficient depends on the relative variability of the network
elements. The variances, resulting from the variances of the input channels, are determined
by the mean of the square elements of the network. By contrast, the covariances depend
on the effective weights of inputs to the neuron pairs, and hence on the square of the mean
weight. Correspondingly, we consider the signal covariances CS
ij(s) = cov(ri(s), rj(s))s,
CS
ij = cov(
Bikrext,k(s), Bjlrext,l(s))s =
cov(Bikrext,k(s), Bjlrext,l(s))s.
(71)
BikBjkcov(rext,k(s), rext,l(s))s
BikBjkvar(rext,k(s))s +
BikBjkcov(rext,k(s), rext,l(s))s.
kl
(cid:88)
(cid:88)
(cid:88)
kl
For i (cid:54)= j, one gets
CS
ij =
=
k
Averaged across neurons
(cid:88)
kl
(cid:88)
k(cid:54)=l
(cid:88)
(cid:88)
k(cid:54)=l
k(cid:54)=l
(72)
(73)
(74)
(75)
(76)
(cid:104)CS
ij(cid:105)i(cid:54)=j = N(cid:104)B(cid:105)2var(rext) + N (N − 1)(cid:104)B(cid:105)2cinvar(rext).
For the signal variances, i = j, which correspond to the variance of the rates across stimuli,
(cid:104)CS
ii (s)(cid:105)s =
=
(cid:88)
(cid:88)
k
k
var(Bikrext,k) +
B2
ikvar(rext,k) +
cov(Bikrext,k, Bilrext,l)
cov(rext,k, rext,l)BikBil.
Averaged across neurons, with (cid:104)B2(cid:105) = (cid:104)B(cid:105)2 + var(B):
(cid:104)CS
ii (s)(cid:105)s,i = N var(rext)((cid:104)B(cid:105)2 + var(B)) + N (N − 1)(cid:104)B(cid:105)2var(rext)cin.
Their ratio is
(cid:104)Cij(cid:105)
(cid:104)Cii(cid:105) =
=
N(cid:104)B(cid:105)2var(rext)(1 + (N − 1)cin)
N var(rext)((cid:104)B(cid:105)2 + var(B)) + N (N − 1)(cid:104)B(cid:105)2var(rext)cin
1 + var(B)/(cid:104)B(cid:105)2 + (N − 1)cin
1 + (N − 1)cin
.
This results in an approximate expression for the average signal correlation coefficient
ij(cid:113)
cS = (cid:104) CS
CS
ii CS
jj
ij(cid:105)
(cid:105)i,j ≈ (cid:104)CS
(cid:104)CS
ii(cid:105) =
1 + (N − 1)cin
1 + var(B)/(cid:104)B(cid:105)2 + (N − 1)cin
.
(77)
34
8.3.5 Estimation of network model parameters
We extract parameters of the recurrent and feed-forward network models from the data, both
to test the consistency of the models with the data and to interpret the observed variability.
Because rates and covariance matrices were measured for many different stimuli and thus
provide a large number of constraints, one approach would be to infer as much information
as possible about the full connectivity matrices B or F from the data. However, due to
the relatively small number of trials for each stimulus, we use a model with few parameters.
The set of parameters consists of the network parameters (cid:104)B(cid:105) and var(B) ((cid:104)F(cid:105) and var(F ),
respectively) as well as the parameters of the input ensemble, (cid:104)rext(cid:105), var(rext) and cin.
In particular, we want to infer the ratios ρE = var(rext)/(cid:104)rext(cid:105)2 and ρ = var(B)/(cid:104)B(cid:105)2 (and
correspondingly for F ). Based on these, we can generate surrogate data to test if the observed
scaling of average covariances with average rates is more consistent with the recurrent or the
feed-forward model, Eq. (28) or (30). The experimental data provides constraints in the
form of the population and stimulus averaged rates, their variances and the noise and signal
correlation coefficients. In the models, rates are given by Eqs. (7) and (15), respectively. The
population averaged mean response thus is
(cid:104)r(cid:105) = N(cid:104)B(cid:105)(cid:104)rext(cid:105) or (cid:104)r(cid:105) = N(cid:104)F(cid:105)(cid:104)rext(cid:105).
The variance of rates across stimuli, see Eq. (75), is
var(r) = N var(rext)((cid:104)B(cid:105)2 + var(B) + (N − 1)(cid:104)B(cid:105)2cin)
such that the relative variance of rates
var(r)
(cid:104)r(cid:105)2 =
var(rext)
(cid:104)rext(cid:105)2
1 + ρ + (N − 1)cin
N
(78)
(79)
(80)
depends on the input signal to noise ratio ρE = var(rext)
elements, ρ = var(B)
respectively.
(cid:104)rext(cid:105)2 and the variability of the network
(cid:104)B(cid:105)2 , or ρ = var(F )
(cid:104)F(cid:105)2
The estimates for the remaining parameters ρ and cin are obtained from the measured
values of the ratio of covariances to variances, which correspond approximately to the average
coefficients of noise and signal correlations,
and
(cid:104)(cid:104)Cij(cid:105)i(cid:54)=j
(cid:104)Cii(cid:105)i
(cid:105)s ≈ cN ,
(cid:104)cov(ri, rj)(cid:105)i(cid:54)=j
(cid:104)var(ri)(cid:105)i
≈ cS,
(81)
(82)
using Eqs. (70) and (77). Together with Eq. (80), these relations provide the necessary
constraints for the network models. Strictly speaking, these equations are valid only for the
recurrent network, while for the feed-forward model, Eq. (61) is relevant. In this case ρ is
overestimated, which results in a lower bound for ρE, and thus a conservative estimate of the
input variability.
Under additional assumptions, we can also choose the absolute values of the parame-
ters, for example (cid:104)rext(cid:105) and (cid:104)F(cid:105), such that mean rates and mean covariances correspond to
experimental ones. The mean rates (78) constrain the product of the two parameters
(cid:104)r(cid:105) = N(cid:104)F(cid:105)(cid:104)rext(cid:105).
(83)
35
From Eq. (55) it follows that
(cid:104)Cij(cid:105)i(cid:54)=j = N(cid:104)F(cid:105)2(cid:104)Vext(cid:105),
(84)
and we assume that Vext = rext to relate mean input and input variance. The distribution
of Vext thus is a folded normal distribution, with
(cid:104)rext(cid:105) =(cid:112)var(rext)(cid:112)2/πe−1/2ρx + (cid:104)rext(cid:105)(cid:16)
(cid:17)
1 − 2Φ(−(cid:112)1/ρx)
(with the cumulative normal Φ ), and from this one finds
(cid:104)Cij(cid:105) = (cid:104)F(cid:105)(cid:104)r(cid:105)(cid:104)√
(cid:105)
ρxe−1/2ρx + 1 − 2Φ(−(cid:112)1/ρx)
.
(85)
(86)
To set the absolute values of input versus network strength we made assumptions regarding
the relation between input strength and input variance across trials.
If we measure the
strength of the dependence between mean response and mean covariance, Eq. (55) by the
ratio intercept/slope of a linear fit, however, the result is independent of the absolute values
of (cid:104)F(cid:105) and (cid:104)rext(cid:105), as well as a potential linear factor relating Vext and rext.
8.4 Covariances and rates on a population level in regular
networks
Here we derive the macroscopic Eqs. (25) and (26) for the responses of the two populations:
i∈I ri, and the covariances between responses
k∈K,l∈L Ckl, in the regular networks defined in Section 3.6
depend only on the coupling between populations, see below.
The two populations are of size N , and weights between connected neurons are of equal
strength. By definition, the regularity of the network connectivity implies that for any neuron
k∈K Gkl(cid:48) is identical, and we can define the population
the average response of population I, RI ≡(cid:80)
of populations L, K, ΣKL ≡ (cid:80)
l(cid:48) ∈ from population L, the sum (cid:80)
ΓKL ≡ (cid:88)
coupling matrix, Γ, by
Gkl(cid:48) =
1
N
Gkl = nKLg,
(87)
(cid:88)
k∈K
k∈K,l∈L
where g is the weight of the connections. We want to express R and Σ in terms of Γ. On the
neuron level, rates, r, and covariances, C, depend on the transfer matrix, B = (I− G)−1. We
will define an analogous population transfer matrix P that depends only on Γ and show that
R and Σ can be written in terms of P. We define
P ≡ (I − Γ)−1 = I + Γ + Γ2 + ··· =
Γm.
To relate P to the microscopic quantities r and C, we will use that
(cid:88)
k∈K,l∈L
PKL =
1
N
(cid:88)
k∈K
Bkl =
Bkl(cid:48)
∞(cid:88)
m=0
for any l(cid:48) ∈ L, in complete analogy to Γ. To see this, we note the corresponding relation for
all of the individual terms m
(88)
(89)
(90)
[Γm]KL =
1
N
[Gm]kl.
k∈K,l∈L
(cid:88)
36
(cid:88)
I
As an example consider m = 2. Using the regularity, we see that
[G2]kl =
GkiGil =
ΓKI Gil = N
k∈K,l∈L
l∈L,k∈K,i∈I,I
l∈L,i∈I,I
For the rate of population I, RI , we then find with (89) that
(cid:88)
RI ≡(cid:88)
i∈I
(cid:88)
(cid:88)
(cid:88)
(cid:88)
ri =
Bikrext,k =
Bikrext,k =
i∈I
k∈K
K
i∈I,k∈K
K
because rext,k = Rext,K for any k ∈ K, and thus
R = PN Rext = (I − Γ)−1N Rext,
ΓKI ΓIL = [Γ2]KLN.
(91)
(cid:88)
K
PIKN Rext,K,
(92)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
I
which is Eq. (25). Similarly, for the covariances of the population responses, Σ:
Ckl =
BkiBliri =
PKI Bliri
k∈K,l∈L
k∈K,l∈L,i∈I
(cid:88)
(cid:88)
(cid:88)
l∈L
I
i∈I
ΣKL ≡ (cid:88)
(cid:88)
(cid:88)
I
I,i∈I
=
which corresponds to Eq. (26),
PKI PLI ri =
PKI PLI RI ,
Σ = (I − Γ)−1D[R](I − Γ)−1,
(with the diagonal matrix D[R] with D[R]IJ = δIJ RI ).
8.4.1 Effect of cross-coupling on correlations on a population level
In this section, we evaluate how noise and signal correlations in the population activity arise
from direct and effective coupling. The strength of the cross-coupling between populations,
Γc, determines via the effective cross-coupling, Pc, of the population transfer matrix,
(cid:18)1 − Γs
Γc
(cid:19)
(cid:18)Ps Pc
(cid:19)
Pc Ps
Γc
1 − Γs
≡
,
(97)
P = (I − Γ)−1 =
1
(1 − Γs)2 − Γ2
c
the population rates and covariances. The network is unstable if the eigenvalues of Γ are
larger than one, which corresponds to the constraint Γs + Γc < 1, and therefore Pc < Ps.
On the population level, the effects of cross-coupling between differently tuned neurons
on correlations can be expressed in terms of the variance and the mean of the elements of P,
similar as for the random effective network model. To see this, we consider an ensemble of
(two-component) input vectors Rext(s) =(cid:0)rext(s), rext(cid:48)(s)(cid:1)T with external input components
rext, rext(cid:48) chosen with mean and variance µrext, σ2
rext, independently across stimuli s.
The signal covariance matrix of mean responses R = PRext across the stimulus ensemble
ij = cov(Ri, Rj) which is
is ΣS
(cid:18)P 2
(cid:19)
ΣS = σ2
rext
s + P 2
c
2PsPc
2PsPc
P 2
s + P 2
c
The average signal correlation for the population activity is
(93)
(94)
(95)
(96)
(98)
(cid:104)cS(cid:105) =
1
4
(cid:88)
ij
ij(cid:113)
ΣS
ΣS
(cid:104)P(cid:105)2
iiΣS
jj
=
(cid:104)P(cid:105)2 + var(P )
= 1 +
2PsPc
s + P 2
P 2
c
=
(Ps + Pc)2
(Ps + Pc)2 + (Ps − Pc)2
(cid:104)P(cid:105)2
(cid:104)P 2(cid:105) =
=
1
1 + var(P )/(cid:104)P(cid:105)2
,
(99)
37
using (cid:104)P(cid:105) = (Ps + Pc)/2 and 2var(P ) = (Ps − (Ps + Pc)/2)2 + (Pc − (Ps + Pc)/2)2 = (Ps −
Pc)2/2. The normalized variance of the elements of P thus determines the strength of signal
correlations, and one can interpret ρ = var(P )/(cid:104)P(cid:105)2 as a measure for coupling strength.
A similar relation holds for the average noise correlations, if we average first across stimuli,
then across neurons: the average of the noise covariance matrix across stimuli is (cid:104)Σ(s)(cid:105)s =
(cid:104)P R(s)P T(cid:105) = P(cid:104)R(cid:105)P T . Because (cid:104)Rext(cid:105) = (cid:104)R(cid:48)
ext(cid:105), the average rates of both populations are
identical, and the average across population pairs can be evaluated just as for the signal
correlations.
8.4.2 Condition for favorable correlations
Here, we show that strong connections between populations can induce correlations that are
beneficial for stimulus discrimination, when compared to shuffled trials. Shuffling correlations
is favorable for stimulus discrimination if the covariances within each populations are stronger
than the ones across populations. To see this, assume that the responses of the excitatory
populations for two stimuli are R(s1) = (R0 + ∆R, R0) and R(s2) = (R0, R0 + ∆R). We
also assume that the population covariance matrix is stimulus independent, Σ(s1) = Σ(s2) =
(cid:18) ΣEE ΣE(cid:48)E
(cid:19)
(cid:88)
i∈E
w = ΣEE − ΣE(cid:48)E =
σ2
and that ΣEE = ΣE(cid:48)E(cid:48). Calculating the most discriminative direction w
ΣE(cid:48)E ΣE(cid:48)E(cid:48)
(using Eq. (20) on the macroscopic variables) yields an eigenvector of Σ , (1,−1), and the
variance in that direction corresponds to the smaller eigenvalue of Σ,
(cid:88)
the case if the sum of covariances across populations, (cid:80)
within a population, (cid:80)
Covariances can be considered helpful, if this eigenvalue is smaller when compared to a
shuffled version where all covariances are set to 0, such that σ2
i∈E Cii. This is exactly
i∈E,k∈E(cid:48) Cik, is larger than the one
i(cid:54)=j∈E Cij. Although ΣEE ≥ ΣE(cid:48)E, it is possible that (cid:104)Cij(cid:105)i(cid:54)=j∈E <
Cij − (cid:88)
w(cid:48) =(cid:80)
i∈E,k∈E(cid:48)
(cid:104)Cij(cid:105)i∈E,j∈E(cid:48), because of the contribution of the variances Cii to ΣEE.
noise in the relevant direction. For simplicity, let us assume that Rext = Rext(cid:48) ≡ R. Then
We want to find a condition for which such a removal of covariances by shuffling decreases
i(cid:54)=j∈E
Cii +
Cik.
(cid:18)P 2
Σ = R
s + P 2
c
2PsPc
2PsPc
P 2
s + P 2
c
(cid:19)
(100)
(101)
(102)
The average pairwise across-covariance is Cacross = ΣE(cid:48)E/N 2. Approximately, the average
neuron variance differs from the average within-covariance only by the contribution of the
rate: (cid:104)Cii(cid:105) ≈ R/N + Cwithin. From this we get for the population variance ΣEE = R(P 2
s +
c ) = N(cid:104)Cii(cid:105) + N (N − 1)Cwithin = R + N 2Cwithin. For favorable correlations, we need
P 2
Cacross > Cwithin,
2PsPcR
N 2 >
R(P 2
s + P 2
N 2
c ) − R
and since Ps > Pc
1 > (Ps − Pc)2 ⇔ 1 > Ps − Pc
Using Eq. (97) for the entries of P , this means that
1 − Γs − Γc
(1 − Γs)2 − Γ2
⇔ Γs(1 − Γs) > Γc(1 − Γc).
1 >
(103)
The function Γ(1 − Γ) is a parabola through the points (0,0) and (0,1), with its minimum
at (1/2,-1/4). We have Γs,Γc < 1, so both sides are < 0 and Γc < 1 − Γs. Thus, the
condition can be fulfilled if Γs < 1/2 and within-coupling is smaller than cross-coupling,
Γs < Γc < 1 − Γs.
c
38
8.4.3 Linear Fisher information
We use the signal-to-noise ratio, S, to measure how well discrete pairs of stimuli in a given
ensemble can be discriminated, Section 3.5.
In a model network, we have access to all
possible input dimensions, and we can use an alternative measure that combines these possible
stimulus dimensions. We will use this measure to compare the effects of different network
scenarios on the representation of general stimuli.
If a high-dimensional stimulus varies
continuously, the information in the response distribution about the stimulus can be measured
by the Fisher information matrix [16]. An approximate value, the 'linear Fisher information',
is obtained by neglecting the information in the stimulus dependence of the covariance matrix,
and the entries of the linear Fisher Information matrix are defined by
Imn = ∂mRT Σ−1∂nR,
(104)
for the population covariance matrix Σ and population response R. Here, ∂mR is the deriva-
tive of the response vector with respect to stimulus coordinate m. Our stimulus dimensions
are the coordinates of the input vector Rext and output in a network model is R = PN Rext.
It is easy to see that ∂mR is given by the mth column of P, and hence
I = PT Σ−1P.
(105)
Generally, if the input depends linearly on a one-dimensional stimulus s, Rext(s) = Rext,0 +
s∆Rext, the information about changes in this direction is ∆RT
extI∆Rext. As a measure
for the information about stimulus changes in all possible directions, we use the trace of I,
Tr(I) =(cid:80)
i Iii.
Covariances in a recurrent network with external noise are Σ = P(D[R] + Σext)PT , where
Σext is the diagonal 2 × 2 matrix describing the variance of external input to the two popu-
lations. This results in a recurrent linear Fisher Information matrix
I r = (D[R] + Σext)−1.
(106)
This can be compared to a feed-forward scenario with identical transfer matrix, F = P, where
Σ = PΣextPT + D[R], so that
I f = PT (PΣextPT + D[R])−1P.
(107)
References
[1] van Vreeswijk C, Sompolinsky H. Chaotic balanced state in a model of cortical circuits.
Neural Comput. 1998;10(6):1321–71.
[2] Shadlen MN, Newsome WT. The variable discharge of cortical neurons: implications for
connectivity, computation, and information coding. J Neurosci. 1998;18(10):3870–96.
[3] Litwin-Kumar A, Doiron B. Slow dynamics and high variability in balanced cor-
Nat Neurosci. 2012;15(11):1498–505.
tical networks with clustered connections.
doi:10.1038/nn.3220.
[4] Renart A, la Rocha JD, Bartho P. The asynchronous state in cortical circuits. Science.
2010;327(5965):587–590. doi:10.1126/science.1179850.
[5] Pernice V, Staude B, Cardanobile S, Rotter S.
Correlations
doi:10.1371/journal.pcbi.1002059.
in Neuronal Networks.
How Structure Determines
PLoS Comput Biol. 2011;7(5):e1002059.
[6] Trousdale J, Hu Y, Shea-Brown E, Josi´c K.
Impact of network structure and cel-
lular response on spike time correlations. PLoS Comput Biol. 2012;8(3):e1002408.
doi:10.1371/journal.pcbi.1002408.
39
[7] Goris RLT, Movshon JA, Simoncelli EP. Partitioning neuronal variability. Nat Neurosci.
2014;17(6):858–65. doi:10.1038/nn.3711.
[8] Scholvinck ML, Saleem AB, Benucci A, Harris KD, Carandini M. Cortical State Deter-
mines Global Variability and Correlations in Visual Cortex. J Neurosci. 2015;35(1):170–
178. doi:10.1523/JNEUROSCI.4994-13.2015.
[9] Lin IC, Okun M, Carandini M, Harris KD. The Nature of Shared Cortical Variability.
Neuron. 2015;87(3):644–656. doi:10.1016/j.neuron.2015.06.035.
[10] Ecker AS, Denfield GH, Bethge M, Tolias AS. On the Structure of Neuronal Population
Activity under Fluctuations in Attentional State. J Neurosci. 2016;36(5):1775–1789.
doi:10.1523/JNEUROSCI.2044-15.2016.
[11] Doiron B, Litwin-Kumar A, Rosenbaum R, Ocker GK, Josi´c K. The mechanics of state-
dependent neural correlations. Nat Neurosci. 2016;19(3):383–393. doi:10.1038/nn.4242.
[12] Kohn A, Coen-cagli R, Kanitscheider I, Pouget A. Correlations and Neuronal Population
Information. Annu Rev Neurosci. 2016;39:237–256. doi:10.1146/annurev-neuro-070815-
013851.
[13] Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, et al. Spatio-
temporal correlations and visual signalling in a complete neuronal population. Nature.
2008;454(7207):995–9. doi:10.1038/nature07140.
[14] Vidne M, Ahmadian Y, Shlens J, Pillow JW, Kulkarni J, Litke AM, et al. Modeling
the impact of common noise inputs on the network activity of retinal ganglion cells. J
Comput Neurosci. 2012;33(1):97–121. doi:10.1007/s10827-011-0376-2.
[15] Ecker AS, Berens P, Cotton RJ, Subramaniyan M, Denfield GH, Cadwell CR, et al.
State dependence of noise correlations in macaque primary visual cortex. Neuron.
2014;82(1):235–48. doi:10.1016/j.neuron.2014.02.006.
[16] Abbott LF, Dayan P. The effect of correlated variability on the accuracy of a population
code. Neural Comput. 1999;11(1):91–101.
[17] Wilke SD, Eurich CW. Representational accuracy of stochastic neural populations.
Neural Comput. 2002;14(1):155–189. doi:10.1162/089976602753284482.
[18] Shamir M, Sompolinsky H.
Nonlinear population codes.
Neural Comput.
2004;16(6):1105–1136. doi:10.1162/089976604773717559.
[19] Averbeck BB. Neural correlations, population coding and computation. Nat Rev Neu-
rosci. 2006;7(5):358–366. doi:10.1038/nrn1888.
[20] Tkacik G, Prentice JS, Balasubramanian V, Schneidman E. Optimal population
coding by noisy spiking neurons. P Natl Acad Sci USA. 2010;107(32):14419–24.
doi:10.1073/pnas.1004906107.
[21] Ecker AS, Berens P, Tolias AS, Bethge M.
in populations of diversely tuned neurons.
doi:10.1523/JNEUROSCI.2539-11.2011.
The effect of noise correlations
J Neurosci. 2011;31(40):14272–83.
[22] da Silveira RA, Berry MJ. High-Fidelity Coding with Correlated Neurons. PLoS Comput
Biol. 2014;10(11):e1003970. doi:10.1371/journal.pcbi.1003970.
[23] Hu Y, Zylberberg J, Shea-Brown E. The sign rule and beyond: boundary effects,
flexibility, and noise correlations in neural population codes. PLoS Comput Biol.
2014;10(2):e1003469. doi:10.1371/journal.pcbi.1003469.
[24] Franke F, Fiscella M, Sevelev M, Roska B, Hierlemann A, da Silveira RA. Struc-
tures of Neural Correlation and How They Favor Coding. Neuron. 2016;89(2):409–422.
doi:10.1016/j.neuron.2015.12.037.
[25] Zylberberg J, Cafaro J, Turner MH, Shea-Brown E, Rieke F. Direction-Selective Cir-
cuits Shape Noise to Ensure a Precise Population Code. Neuron. 2016;89(2):369–383.
doi:10.1016/j.neuron.2015.11.019.
40
[26] Zohary E, Shadlen MN, Newsome WT. Correlated neuronal discharge rate and
Nature. 1994;370(6485):140–3.
its implications for psychophysical performance.
doi:10.1038/370140a0.
[27] Beck J, Bejjanki VR, Pouget A.
Insights from a simple expression for linear fisher
information in a recurrently connected population of spiking neurons. Neural Comput.
2011;23(6):1484–1502. doi:10.1162/NECO a 00125.
[28] Renart A, van Rossum MCW. Transmission of population-coded information. Neural
Comput. 2012;24(2):391–407. doi:10.1162/NECO a 00227.
[29] Seri`es P, Latham PE, Pouget A. Tuning curve sharpening for orientation selectivity:
coding efficiency and the impact of correlations. Nat Neurosci. 2004;7(10):1129–35.
doi:10.1038/nn1321.
[30] Moreno-Bote R, Beck J, Kanitscheider I, Pitkow X, Latham P, Pouget A. Information-
limiting correlations. Nat Neurosci. 2014;17(10):1410–1417. doi:10.1038/nn.3807.
[31] Kanitscheider I, Coen-Cagli R, Pouget A. Origin of information-limiting noise correla-
tions. P Natl Acad Sci USA. 2015;112(50):E6973–E6982. doi:10.1073/pnas.1508738112.
[32] Bathellier B, Ushakova L, Rumpel S. Discrete neocortical dynamics predict behavioral
categorization of sounds. Neuron. 2012;76(2):435–49. doi:10.1016/j.neuron.2012.07.008.
[33] Hawkes AG. Point spectra of some mutually exciting point processes. J R Stat Soc B.
1971;33(3):438–443.
[34] Brody CD. Correlations without synchrony. Neural Comput. 1999;11(7):1537–1551.
[35] Grytskyy D, Tetzlaff T, Diesmann M, Helias M.
A unified view on weakly
Front Comput Neurosci. 2013;7(October):131.
correlated recurrent networks.
doi:10.3389/fncom.2013.00131.
[36] Okun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Barth´o P, et al. Diverse cou-
pling of neurons to populations in sensory cortex. Nature. 2015;521(7553):511–515.
doi:10.1038/nature14273.
[37] Brunel N. Dynamics of sparsely connected networks of excitatory and inhibitory spiking
neurons. J Comput Neurosci. 2000;8(3):183–208.
[38] Ponce-Alvarez A, Thiele A, Albright TD, Stoner GR, Deco G. Stimulus-dependent
variability and noise correlations in cortical MT neurons. P Natl Acad Sci USA.
2013;110(32):13162–7. doi:10.1073/pnas.1300098110.
[39] Ostojic S. Two types of asynchronous activity in networks of excitatory and inhibitory
spiking neurons. Nat Neurosci. 2014;17(4):594–600. doi:10.1038/nn.3658.
[40] Paninski L. Maximum likelihood estimation of cascade point-process neural encoding
models. Network - Comp Neural. 2004;15(4):243–262. doi:10.1088/0954-898X/15/4/002.
[41] Macke JH, Buesing L, Cunningham JP, Yu BM, Shenoy KV, Sahani M. Empiri-
cal models of spiking in neuronal populations. Adv Neur In. 2011;24:1350--1358.
doi:10.1.1.230.7630.
[42] Rocha J, Doiron B, Reyes A, Shea-brown E, Josi´c K. Correlation between neural spike
trains increases with firing rate. Nature. 2007;448(7155):802–6. doi:10.1038/nature06028.
[43] Luczak A, Bartho P, Harris KD. Gating of sensory input by spontaneous cortical activity.
J Neurosci. 2013;33(4):1684–95. doi:10.1523/JNEUROSCI.2928-12.2013.
[44] Cohen MR, Kohn A. Measuring and interpreting neuronal correlations. Nat Neurosci.
2011;14(7):811–9. doi:10.1038/nn.2842.
[45] Rothschild G, Nelken I, Mizrahi A. Functional organization and population dy-
namics in the mouse primary auditory cortex. Nat Neurosci. 2010;13(3):353–60.
doi:10.1038/nn.2484.
41
|
1606.02627 | 1 | 1606 | 2016-06-08T16:33:41 | Brains on Beats | [
"q-bio.NC"
] | We developed task-optimized deep neural networks (DNNs) that achieved state-of-the-art performance in different evaluation scenarios for automatic music tagging. These DNNs were subsequently used to probe the neural representations of music. Representational similarity analysis revealed the existence of a representational gradient across the superior temporal gyrus (STG). Anterior STG was shown to be more sensitive to low-level stimulus features encoded in shallow DNN layers whereas posterior STG was shown to be more sensitive to high-level stimulus features encoded in deep DNN layers. | q-bio.NC | q-bio |
Brains on Beats
Umut Gu¸clu1, Jordy Thielen1, Michael Hanke2,3, and Marcel A. J.
van Gerven1
1Radboud University, Donders Institute for Brain, Cognition and
Behaviour, Nijmegen, the Netherlands
2Psychoinformatics lab, Institute of Psychology, Otto-von-Guericke
University, Magdeburg, Germany
3Center for Behavioral Brain Sciences, Magdeburg, Germany
Abstract
We developed task-optimized deep neural networks (DNNs) that achieved
state-of-the-art performance in different evaluation scenarios for auto-
matic music tagging. These DNNs were subsequently used to probe the
neural representations of music. Representational similarity analysis re-
vealed the existence of a representational gradient across the superior
temporal gyrus (STG). Anterior STG was shown to be more sensitive to
low-level stimulus features encoded in shallow DNN layers whereas poste-
rior STG was shown to be more sensitive to high-level stimulus features
encoded in deep DNN layers.
1
Introduction
The human sensory system is devoted to the processing of sensory information
to create our perception of the environment [1]. Sensory cortices are thought to
encode a hierarchy of ever more invariant representations of the environment [2].
An important question thus is what sensory information is processed as one
traverses the sensory pathways from the primary sensory areas to higher sensory
areas. This question is the subject matter of sensory neuroscience.
The majority of the work on auditory cortical representations has been lim-
ited to hand-designed low-level representations such as spectro-temporal mod-
els [3], spectro-location models [4], timbre, rhythm, tonality [5, 6, 7] and pitch [8]
or high-level representations such as music genre [9] and sound categories [10].
For example, Santoro et al. [3] found that a joint frequency-specific modulation
transfer function predicted observed fMRI activity best compared to frequency-
nonspecific and independent models. They showed specificity to fine spectral
modulations along Heschl's gyrus (HG) and anterior superior temporal gyrus
(STG), whereas coarse spectral modulations were mostly located posterior-
laterally to HG, on the planum temporale (PT), and STG. Preference for slow
temporal modulations was found along HG and STG, whereas fast temporal
modulations were observed on PT, and posterior and medially adjacent to HG.
Also, it has been shown that activity in STG, somatosensory cortex, the default
1
mode network, and cerebellum are sensitive to timbre, while amygdala, hip-
pocampus and insula are more sensitive to rhythmic and tonality features [5, 7].
However these efforts have not yet provided a complete algorithmic account of
sensory processing in the auditory system.
Since their resurgence, deep neural networks (DNNs) coupled with functional
magnetic resonance imaging (fMRI) have provided a powerful approach to form
and test alternative hypotheses about what sensory information is processed in
different brain regions. On one hand, a task-optimized DNN model learns a hi-
erarchy of nonlinear transformations in a supervised manner with the objective
of solving a particular task. On the other hand, fMRI measures local changes in
blood-oxygen-level dependent hemodynamic responses to sensory stimulation.
Subsequently, any subset of the DNN representations that emerge from this hier-
archy of nonlinear transformations can be used to probe neural representations
by comparing DNN and fMRI responses to the same sensory stimuli. Con-
sidering that the sensory systems are biological neural networks that routinely
perform the same tasks as their artificial counterparts, it is not inconceivable
that DNN representations are suitable for probing neural representations.
Indeed, this approach has been shown to be extremely successful in visual
neuroscience. To date, several task-optimized DNN models were used to accu-
rately model visual areas on the dorsal and ventral streams [11, 12, 13, 14, 15, 16,
17, 18], revealing representational gradients where deeper neural network layers
map to more downstream areas along the visual pathways [19, 20]. Recently,
[21] has shown that deep neural networks trained to map speech excerpts to
word labels could be used to predict brain responses to natural sounds. Here,
deeper neural network layers were shown to map to auditory brain regions that
were more distant from primary auditory cortex.
In the present work we expand on this line of research where our aim is to
model how the human brain responds to music. We achieve this by probing neu-
ral representations of music features across the superior temporal gyrus using a
deep neural network optimized for music tag prediction. We use the represen-
tations that emerged after training a DNN to predict tags of musical excerpts
as candidate representations for different areas of STG in representational sim-
ilarity analysis. We show that different DNN layers correspond to different
locations along STG such that anterior STG is shown to be more sensitive to
low-level stimulus features encoded in shallow DNN layers whereas posterior
STG is shown to be more sensitive to high-level stimulus features encoded in
deep DNN layers.
2 Materials and Methods
2.1 MagnaTagATune Dataset
We used the MagnaTagATune dataset [22] for DNN estimation. The dataset
contains 25.863 music clips. Each clip is a 29-seconds-long excerpt belonging to
one of the 5223 songs, 445 albums and 230 artists. The clips span a broad range
of genres like Classical, New Age, Electronica, Rock, Pop, World, Jazz, Blues,
Metal, Punk, and more. Each audio clip is supplied with a vector of binary
annotations of 188 tags. These annotations are obtained by humans playing
the two-player online TagATune game. In this game, the two players are either
2
presented with the same or a different audio clip. Subsequently, they are asked
to come up with tags for their specific audio clip. Afterward, players view each
other's tags and are asked to decide whether they were presented the same
audio clip. Tags are only assigned when more than two players agreed. The
annotations include tags like 'singer', 'no singer', 'violin', 'drums', 'classical',
'jazz', et cetera. We restricted our analysis on this dataset to the top 50 most
popular tags to ensure that there is enough training data for each tag. Parts
1-12 were used for training, part 13 was used for validation and parts 14-16 were
used for test.
2.2 Studyforrest Dataset
We used the extended studyforrest dataset [23] for representational similarity
analysis. The dataset contains fMRI data on the perception of musical genres.
Twenty participants (age 21-38 years, mean age 26.6 years), with normal hearing
and no known history of neurological disorders, listened to 25 six-second-long,
44.1 kHz music clips. The stimulus set comprised five clips per each of the five
following genres: Ambient, Roots Country, Heavy Metal, 50s Rock 'n Roll, and
Symphonic. Stimuli were selected according to [9]. The Ambient and Symphonic
genres can be considered as non-vocal and the others as vocal. Participants
completed eight runs, each with all 25 clips.
Ultra-high-field (7 Tesla) fMRI images were collected using a Siemens MAG-
NETOM scanner, T2*-weighted echo-planar images (gradient-echo, repetition
time (TR) = 2000 ms, echo time (TE) = 22 ms, 0.78 ms echo spacing, 1488
Hz/Px bandwidth, generalized auto-calibrating partially parallel acquisition
(GRAPPA), acceleration factor 3, 24 Hz/Px bandwidth in phase encoding di-
rection), and a 32 channel brain receiver coil. Thirty-six axial slices were ac-
quired (thickness = 1.4 mm, 1.4 × 1.4 mm in-plane resolution, 224 mm field-of-
view (FOV) centered on the approximate location of Heschl's gyrus, anterior-
to-posterior phase encoding direction, 10% inter-slice gap). Along with the
functional data, cardiac and respiratory traces, and a structural MRI were col-
lected. In our analyses, we only used the data from the 12 subjects (subjects 1,
3, 4, 6, 7, 9, 12, 14, 15, 16, 17, 18) with no known data anomalies as reported
in [23].
The anatomical and functional scans were preprocessed as follows: Func-
tional scans were realigned to the first scan of the first run and next to the
mean scan. Anatomical scans were coregistered to the mean functional scan.
Realigned functional scans were slice-time corrected to correct for the differ-
ences in image acquisition times between the slices. Realigned and slice-time
corrected functional scans were normalized to MNI space. Finally, a general
linear model was used to remove noise regressors derived from voxels unrelated
to the experimental paradigm and estimate BOLD response amplitudes [24].
We restricted our analyses to the superior temporal gyrus (STG).
2.3 Deep Neural Networks
We developed three task-optimized DNN models for tag prediction. Two of the
models comprised five convolutional layers followed by three fully-connected
layers (DNN-T model and DNN-F model). The inputs to the models were
96000-dimensional time (DNN-T model) and frequency (DNN-F model) domain
3
representations of six second-long audio signals, respectively. One of the models
comprised two streams of five convolutional layers followed by three fully con-
nected layers (DNN-TF model). The inputs to the streams were given by the
time and frequency representations. The outputs of the convolutional streams
were merged and fed into first fully-connected layer.
The layers were similar to the AlexNet layers [25] except for the following
modifications:
• The number of convolutional kernels were halved.
• The (convolutional and pooling) kernels and strides were flattened. That
is, an n × n kernel was changed to an n2 × 1 kernel and an m × m stride
was changed to an m2 × 1 stride.
• Local response normalization was replaced with batch normalization [26].
• Rectified linear units were replaced with parametric softplus units with
initial α = 0.2 and initial β = 0.5 [27].
• Softmax units were replaced with sigmoid units.
We used Adam with parameters α = 0.0002, β1 = 0.5, β2 = 0.999, =
1e−8 and a mini batch size of 36 [28] to train the models by minimizing the
binary cross-entropy loss function. Initial model parameters were drawn from
a uniform distribution as described in [29]. Songs in each training mini-batch
were randomly cropped to six seconds (96000 samples). The epoch in which the
validation performance was the highest was taken as the final model (53, 12 and
12 for T, F and TF models, respectively). The DNN models were implemented
in Keras [30].
Once trained, we first tested the tag prediction performance of the models
and identified the model with the highest performance. To predict the tags of
a 29-second-long song excerpt in the test split of the MagnaTagaTune dataset,
we first predicted the tags of 24 six-second-long overlapping segments separated
by a second and averaged the predictions.
We then used the model with the highest performance for nonlinearly trans-
forming the stimuli to eight layers of hierarchical representations for subsequent
representational similarity analysis [31]. Note that the artificial neurons in the
convolutional layers locally filtered their inputs (1D convolution), nonlinearly
transformed them and returned temporal representations per stimulus. These
representations were further processed by averaging them over time. In contrast,
the artificial neurons in the fully-connected layers globally filtered their inputs
(dot product), non-linearly transformed them and returned scalar representa-
tions per stimulus. These representations were not further processed. These
transformations resulted in n matrices of size m × pi where n is the number
of layers (8), m is the number of stimuli (25) and p is the number of artificial
neurons in the ith layer (48 or 96, 128 or 256, 192 or 384, 192 or 384, 128 or
256, 4096, 4096 and 50 for i = 1, . . . , 8, respectively).
2.4 Representational Similarity Analysis
We used Representational Similarity Analysis (RSA) [31] to investigate how well
the representational structures of DNN model layers match with that of the re-
sponse patterns in STG. In RSA, models and brain regions are characterized
4
by n × n representational dissimilarity matrices (RDMs), whose elements rep-
resent the dissimilarity between the neural or model representations of a pair of
stimuli. In turn, computing the overlap between the model and neural RDMs
provides evidence about how well a particular model explains the response pat-
terns in a particular brain region. Specifically, we performed a region of interest
analysis as well as a searchlight analysis by first constructing the RDMs of STG
(target RDM) and the model layers (candidate RDM). In the ROI analysis, this
resulted in one target RDM per subject and eight candidate RDMs. For each
subject, we correlated the upper triangular parts of the target RDM with the
candidate RDMs (Spearman correlation). We quantified the similarity of STG
representations with the model representations as the mean correlation. For the
searchlight analysis, this resulted in 27277 target RDMs (each derived from a
spherical neighborhood of 100 voxels) and 8 candidate RDMs. For each subject
and target RDM, we correlated the upper triangular parts of the target RDM
with the candidate RDMs (Spearman correlation). Then, the layers which re-
sulted in the highest correlation were assigned to the voxels at the center of
the corresponding neighborhoods. Finally, the layer assignments were averaged
over the subjects and the result was taken as the final layer assignment of the
voxels.
2.5 Control Models
To evaluate the importance of task optimization for modeling STG representa-
tions, we compared the representational similarities of the entire STG region and
the task-optimized DNN-TF model layers with the representational similarities
of the entire STG region and two sets of control models.
The first set of control models transformed the stimuli to the following 48-
dimensional model representations1:
• Mel-frequency spectrum (mfs) representing a mel-scaled short-term power
spectrum inspired by human auditory perception where frequencies or-
ganized by equidistant pitch locations. These representations were com-
puted by applying (i) a short-time Fourier transform and (ii) a mel-scaled
frequency-domain filterbank.
• Mel-frequency cepstral coefficients (mfccs) representing both broad-spectrum
information (timbre) and fine-scale spectral structure (pitch). These rep-
resentations were computed by (i) mapping the mfs to a decibel amplitude
scale and (ii) multiplying them by the discrete cosine transform matrix.
• Low-quefrency mel-frequency spectrum (lq mfs) representing timbre. These
representations were computed by (i) zeroing the high-quefrency mfccs, (ii)
multiplying them by the inverse of discrete cosine transform matrix and
(iii) mapping them back from the decibel amplitude scale.
• High-quefrency mel-frequency spectrum (hq mfs) representing pitch. These
representations were computed by (i) zeroing the low-quefrency mfccs, (ii)
multiplying them by the inverse of discrete cosine transform matrix and
(iii) mapping them back from the decibel amplitude scale.
1These are provided as part of the extended studyforrest dataset [23].
5
The second set of control models were 10 random DNN models with the
same architecture as the DNN-TF model, but with parameters drawn from a
zero mean and unit variance multivariate Gaussian distribution.
3 Results
In the first set of experiments, we analyzed the task-optimized DNN models.
The tag prediction performance of the models for the individual tags was defined
as the area under the receiver operator characteristics (ROC) curve (AUC).
We first compared the mean performance of the models over all tags (Fig-
ure 1). The performance of all models was significantly above chance level
(p < 0.05, Z-test). The highest performance was achieved by the DNN-TF
model (0.8939), followed by the DNN-F model (0.8905) and the DNN-T model
(0.8852). To the best of our knowledge, this is the highest tag prediction
performance of an end-to-end model evaluated on the same split of the same
dataset [32]. The performance was further improved by averaging the predic-
tions of the DNN-T and DNN-F models (0.8982) as well as those of the DNN-T,
DNN-F and DNN-TF models (0.9007). To the best of our knowledge, this is the
highest tag prediction performance of any model (ensemble) evaluated on the
same split of the same dataset [33, 32, 34]. For the remainder of the analyses, we
considered only the DNN-TF model since it achieved the highest single-model
performance.
Figure 1: Tag prediction performance of the task-optimized DNN mod-
els. Bars show AUCs over all tags for the corresponding task-optimized DNN
models. Error bars show ± SE.
We then compared the performance of the DNN-TF model for the individ-
ual tags (Figure 2). Visual inspection did not reveal a prominent pattern in
the performance distribution over tags. The performance was not significantly
correlated with tag category or popularity (p > 0.05, Student's t-test). The
only exception was that the performance for the positive tags were significantly
higher than that for the negative tags (p < 0.05, Z-test).
In the second set of experiments, we analyzed how closely the represen-
tational geometry of STG is related to the representational geometries of the
task-optimized DNN-TF model layers.
First, we constructed the candidate RDMs of the layers (Figure 3). Vi-
sual inspection revealed similarity structure patterns that became increasingly
6
Figure 2: Tag prediction performance of the task-optimized DNN-TF
model. Bars show AUCs for the corresponding tags. Red band shows the mean
± SEM over all tags.
prominent with increasing layer depth. The most prominent pattern was the
non-vocal and vocal subdivision.
Figure 3: RDMs of the task-optimized DNN-TF model layers. Matrix
elements show the dissimilarity (1 - Spearman's r) between the model layer
representations of the corresponding trials. Matrix rows and columns are sorted
according to the genres of the corresponding trials.
Second, we performed a region of interest analysis by comparing the reference
RDM of the entire STG region with the candidate RDMs (Figure 4). While
none of the correlations between the reference RDM and the candidate RDMs
reached the noise ceiling (expected correlation between the reference RDM and
the RDM of the true model given the noise in the analyzed data [31]), they
were all significantly above chance level (p < 0.05, signed-rank test with subject
RFX, FDR correction). The highest correlation was found for Layer 1 (0.6811),
whereas the lowest correlation was found for Layer 8 (0.4429).
Third, we performed a searchlight analysis [35] by comparing the reference
RDMs of multiple STG voxel neighborhoods with the candidate RDMs (Fig-
ure 5). Each neighborhood center was assigned a layer such that the corre-
sponding target RDM and the candidate RDM were maximally correlated. This
analysis revealed a systematic change in the mean layer assignments over sub-
jects along the STG. They increased from anterior STG to posterior STG such
that most voxels in the region of the transverse temporal gyrus were assigned
to the shallower layers and most voxels in the region of the angular gyrus were
assigned to the deeper layers. The corresponding mean correlations between the
target and the candidate RDMs decreased from anterior to posterior STG.
7
Figure 4: Representational similarities of the entire STG region and
the task-optimized DNN-TF model layers. Bars show the mean similarity
(Spearman's r) of the target RDM and the corresponding candidate RDMs over
all subjects. Error bars show ± SE. Red band shows the expected representa-
tional similarity of the STG and the true model given the noise in the analyzed
data (noise ceiling).
Figure 5: Representational similarities of the spherical STG voxel clus-
ters and the task-optimized DNN-TF model layers. (A) Mean represen-
tational similarities over subjects. (B) Mean layers assignments over subjects.
These results show that the representations of increasingly posterior STG
voxel neighborhoods can be modeled with the representations of increasingly
deeper layers of DNNs optimized for music tag prediction. This observation
is in line with the visual neuroscience literature where it was shown that the
increasingly deeper layers of DNNs optimized for visual object and action recog-
nition can be used to model increasingly downstream visual areas on the ventral
and dorsal streams [19, 20]. It also agrees with previous work which used a DNN
trained for speech-to-word mapping to show a representational gradient in au-
ditory cortex [21]. It would be of particular interest to compare the respective
gradients and use the music and speech DNNs as each other's control model
such as to disentangle speech- and music-specific representations in auditory
cortex.
In the last set of experiments, we analyzed the control models. We first
constructed the RDMs of the control models (Figure 6). Visual inspection
revealed considerable differences between the RDMs of the task-optimized DNN-
TF model and those of the control models.
We then compared the similarities of the task-optimized candidate RDMs
and the target RDM versus the similarities of the control RDMs and the target
8
Figure 6: RDMs of the random DNN model layers (top row) and the
baseline models (bottom row). Matrix elements show the dissimilarity (1
- Spearman's r) between the model layer representations of the correspond-
ing trials. Matrix rows and columns are sorted according to the genres of the
corresponding trials.
RDM (Figure 7). The layers of the task-optimized DNN model significantly
outperformed the corresponding layers of the random DNN model (∆r = 0.21,
p < 0.05, signed-rank test with subject RFX, FDR correction) and the four
baseline models (∆r = 0.42 for mfs, ∆r = 0.21 for mfcc, ∆r = 0.44 for lq mfs
and ∆r = 0.34 for hq mfs, signed-rank test with subject RFX, FDR correction).
Figure 7: Representational similarities of the entire STG region and
the task-optimized DNN-TF model versus the representational sim-
ilarities of the entire STG region and the control models. Different
colors show different control models: Random DNN model, mfs model, mfcc
model, lq mfs model and hq mfs model. Bars show mean similarity differences
over subjects. Error bars show ± SE.
These results show the importance of task optimization for modeling STG
representations. This observation also is line with visual neuroscience literature
where similar analyses showed the importance of task optimization for modeling
ventral stream representations [19, 17].
9
4 Conclusion
We showed that task-optimized DNNs that use time and/or frequency domain
representations of music achieved state-of-the-art performance in various eval-
uation scenarios for automatic music tagging. Comparison of DNN and STG
representations revealed a representational gradient in STG with anterior STG
being more sensitive to low-level stimulus features (shallow DNN layers) and
posterior STG being more sensitive to high-level stimulus features (deep DNN
layers). These results, in conjunction with previous results on the visual and
auditory cortical representations, suggest the existence of multiple representa-
tional gradients that process increasingly complex conceptual information as we
traverse sensory pathways of the human brain.
References
[1] B. L. Schwartz and J. H. Krantz, Sensation and Perception. SAGE Publi-
cations, 2015.
[2] J. M. Fuster, Cortex and Mind: Unifying Cognition. Oxford University
Press, 2003.
[3] R. Santoro, M. Moerel, F. D. Martino, R. Goebel, K. Ugurbil, E. Yacoub,
and E. Formisano, "Encoding of natural sounds at multiple spectral and
temporal resolutions in the human auditory cortex," PLOS Computational
Biology, vol. 10, p. e1003412, jan 2014.
[4] M. Moerel, F. D. Martino, K. Ugurbil, E. Yacoub, and E. Formisano,
"Processing of frequency and location in human subcortical auditory struc-
tures," Scientific Reports, vol. 5, p. 17048, nov 2015.
[5] V. Alluri, P. Toiviainen, I. P. Jaaskelainen, E. Glerean, M. Sams, and
E. Brattico, "Large-scale brain networks emerge from dynamic processing
of musical timbre, key and rhythm," NeuroImage, vol. 59, pp. 3677–3689,
feb 2012.
[6] V. Alluri, P. Toiviainen, T. E. Lund, M. Wallentin, P. Vuust, A. K. Nandi,
T. Ristaniemi, and E. Brattico, "From vivaldi to beatles and back: Predict-
ing lateralized brain responses to music," NeuroImage, vol. 83, pp. 627–636,
dec 2013.
[7] P. Toiviainen, V. Alluri, E. Brattico, M. Wallentin, and P. Vuust, "Cap-
turing the musical brain with lasso: Dynamic decoding of musical features
from fMRI data," NeuroImage, vol. 88, pp. 170–180, mar 2014.
[8] R. D. Patterson, S. Uppenkamp, I. S. Johnsrude, and T. D. Griffiths, "The
processing of temporal pitch and melody information in auditory cortex,"
Neuron, vol. 36, pp. 767–776, nov 2002.
[9] M. Casey, J. Thompson, O. Kang, R. Raizada, and T. Wheatley, "Popula-
tion codes representing musical timbre for high-level fMRI categorization
of music genres," in MLINI, 2011.
10
[10] N. Staeren, H. Renvall, F. D. Martino, R. Goebel, and E. Formisano,
"Sound categories are represented as distributed patterns in the human
auditory cortex," Current Biology, vol. 19, pp. 498–502, mar 2009.
[11] D. L. K. 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, pp. 8619–8624, may 2014.
[12] P. Agrawal, D. Stansbury, J. Malik, and J. L. Gallant, "Pixels to vox-
els: Modeling visual representation in the human brain," arXiv:1407.5104,
2014.
[13] S.-M. Khaligh-Razavi and N. Kriegeskorte, "Deep supervised, but not un-
supervised, models may explain IT cortical representation," PLOS Com-
putational Biology, vol. 10, p. e1003915, nov 2014.
[14] 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 Computational Biology, vol. 10, p. e1003963, dec 2014.
[15] T. Horikawa and Y. Kamitani, "Generic decoding of seen and imagined
objects using hierarchical visual features," arXiv:1510.06479, 2015.
[16] 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," arXiv:1601.02970, 2016.
[17] D. Seibert, D. L. Yamins, D. Ardila, H. Hong, J. J. DiCarlo, and J. L.
Gardner, "A performance-optimized model of neural responses across the
ventral visual stream," bioRxiv, 2016.
[18] R. M. Cichy, A. Khosla, D. Pantazis, and A. Oliva, "Dynamics of scene
representations in the human brain revealed by magnetoencephalography
and deep neural networks," NeuroImage, apr 2016.
[19] U. Gu¸clu and M. A. J. van Gerven, "Deep neural networks reveal a gradi-
ent in the complexity of neural representations across the ventral stream,"
Journal of Neuroscience, vol. 35, pp. 10005–10014, jul 2015.
[20] U. Gu¸clu and M. A. J. van Gerven, "Increasingly complex representations
of natural movies across the dorsal stream are shared between subjects,"
NeuroImage, dec 2015.
[21] A. Kell, D. Yamins, S. Norman-Haignere, and J. McDermott, "Speech-
trained neural networks behave like human listeners and reveal a hierarchy
in auditory cortex," in COSYNE, 2016.
[22] E. Law, K. West, M. Mandel, M. Bay, and J. S. Downie, "Evaluation of
algorithms using games: The case of music tagging," in ISMIR, 2009.
11
[23] M. Hanke, R. Dinga, C. Hausler, J. S. Guntupalli, M. Casey, F. R. Kaule,
and J. Stadler, "High-resolution 7-tesla fMRI data on the perception of
musical genres – an extension to the studyforrest dataset," F1000Research,
jun 2015.
[24] K. N. Kay, A. Rokem, J. Winawer, R. F. Dougherty, and B. A. Wandell,
"GLMdenoise: a fast, automated technique for denoising task-based fMRI
data," Frontiers in Neuroscience, vol. 7, 2013.
[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet classification
with deep convolutional neural networks," in NIPS, 2012.
[26] S. Ioffe and C. Szegedy, "Batch normalization: Accelerating deep network
training by reducing internal covariate shift," arXiv:1502.03167, 2015.
[27] J. M. McFarland, Y. Cui, and D. A. Butts, "Inferring nonlinear neuronal
computation based on physiologically plausible inputs," PLOS Computa-
tional Biology, vol. 9, p. e1003143, jul 2013.
[28] D. Kingma and J. Ba, "Adam: A method for stochastic optimization,"
arXiv:1412.6980, 2014.
[29] X. Glorot and Y. Bengio, "Understanding the difficulty of training deep
feedforward neural networks," in AISTATS, 2010.
[30] F. Chollet, "Keras." https://github.com/fchollet/keras, 2015.
[31] N. Kriegeskorte, "Representational similarity analysis – connecting the
branches of systems neuroscience," Frontiers in Systems Neuroscience,
2008.
[32] S. Dieleman and B. Schrauwen, "End-to-end learning for music audio," in
ICASSP, 2014.
[33] S. Dieleman and B. Schrauwen, "Multiscale approaches to music audio
feature learning," in ISMIR, 2013.
[34] A. van den Oord, S. Dieleman, and B. Schrauwen, "Transfer learning by
supervised pre-training for audio-based music classification," in ISMIR,
2014.
[35] N. Kriegeskorte, R. Goebel, and P. Bandettini, "Information-based func-
tional brain mapping," Proceedings of the National Academy of Sciences,
vol. 103, pp. 3863–3868, feb 2006.
12
|
1601.01580 | 1 | 1601 | 2016-01-07T16:02:05 | Topological analysis of the connectome of digital reconstructions of neural microcircuits | [
"q-bio.NC",
"math.AT"
] | A recent publication provides the network graph for a neocortical microcircuit comprising 8 million connections between 31,000 neurons (H. Markram, et al., Reconstruction and simulation of neocortical microcircuitry, Cell, 163 (2015) no. 2, 456-492). Since traditional graph-theoretical methods may not be sufficient to understand the immense complexity of such a biological network, we explored whether methods from algebraic topology could provide a new perspective on its structural and functional organization. Structural topological analysis revealed that directed graphs representing connectivity among neurons in the microcircuit deviated significantly from different varieties of randomized graph. In particular, the directed graphs contained in the order of $10^7$ simplices {\DH} groups of neurons with all-to-all directed connectivity. Some of these simplices contained up to 8 neurons, making them the most extreme neuronal clustering motif ever reported. Functional topological analysis of simulated neuronal activity in the microcircuit revealed novel spatio-temporal metrics that provide an effective classification of functional responses to qualitatively different stimuli. This study represents the first algebraic topological analysis of structural connectomics and connectomics-based spatio-temporal activity in a biologically realistic neural microcircuit. The methods used in the study show promise for more general applications in network science. | q-bio.NC | q-bio |
TOPOLOGICAL ANALYSIS OF THE CONNECTOME OF
DIGITAL RECONSTRUCTIONS OF NEURAL MICROCIRCUITS
PAWE(cid:32)L D(cid:32)LOTKO∗,1, KATHRYN HESS∗, RAN LEVI∗, MAX NOLTE∗, MICHAEL
REIMANN, MARTINA SCOLAMIERO, KATHARINE TURNER, EILIF MULLER,
AND HENRY MARKRAM
Abstract. A recent publication provides the network graph for a neocor-
tical microcircuit comprising 8 million connections between 31,000 neurons
[7]. Since traditional graph-theoretical methods may not be sufficient to un-
derstand the immense complexity of such a biological network, we explored
whether methods from algebraic topology could provide a new perspective
on its structural and functional organization. Structural topological analy-
sis revealed that directed graphs representing connectivity among neurons in
the microcircuit deviated significantly from different varieties of randomized
graph. In particular, the directed graphs contained in the order of 107 sim-
plices groups of neurons with all-to-all directed connectivity. Some of these
simplices contained up to 8 neurons, making them the most extreme neuronal
clustering motif ever reported. Functional topological analysis of simulated
neuronal activity in the microcircuit revealed novel spatio-temporal metrics
that provide an effective classification of functional responses to qualitatively
different stimuli. This study represents the first algebraic topological analysis
of structural connectomics and connectomics-based spatio-temporal activity
in a biologically realistic neural microcircuit. The methods used in the study
show promise for more general applications in network science.
The Blue Brain Project (BBP) has recently generated the first draft digital
reconstruction and simulation of a microcircuit of neurons in the neocortex of a
two-week-old rat (Figure 1A) [7]. This reconstruction is made available through
the Neocortical Microcircuit Portal (https://bbpnmc.epfl.ch) [11]. Based on sparse
anatomical and physiological data for neurons and synapses and on a variety of bi-
ologically motivated organizing principles, the complete connectivity between neu-
rons belonging to a neocortical microcircuit was digitally reconstructed – a micro-
connectome. The structural properties of the reconstruction have been extensively
validated against independent data, and simulations of the reconstruction repro-
duced multiple in vitro and in vivo experiments without adjusting any parameter,
further validating its biological accuracy.
In this article we apply methods from topology to the analysis of 42 variants of
the digital reconstruction, grouped in six sets of seven microciruits each. The first
five sets of microcircuits take into account biological variability in layer heights,
proportions of cell types, and cell densities from five individual rats, while the sixth
set is based on the average reconstruction across the five individuals. To form each
neocortical microcircuit.
Key words and phrases. Topology, directed flag complex, Betti number, Euler characteristic,
∗co-first author and corresponding author.
(1) Partial support provided by the Advanced Grant of the European Research Council GUDHI
(Geometric Understanding in Higher Dimensions).
1
2
D(cid:32)LOTKO ET AL.
set of microcircuits, seven statistically varying instantiations of the microcircuit
were reconstructed [12]. The 42 microcircuits are therefore all distinct, though
the degree of resemblance within each set is higher than that between sets. The
structural connectivity of each reconstructed microcircuit can be represented as a
directed graph with approximately 3 × 104 vertices and 8 × 106 edges, while its
functional connectivity can be represented as a time series of subgraphs formed by
functionally effective connections.
Our topological analysis of the detailed structural and functional connectivity of
these 42 neural microcircuits led to a number of surprising observations. Firstly, we
found that the distribution of directed cliques (directed all-to-all connected subsets)
of neurons by size is highly significantly different from both that in Erdos-R´enyi
random graphs with the same number of vertices and the same average connection
probability and that in more sophisticated random graphs, constructed either by
taking into account distance-dependent probabilities varying within and between
cortical layers or morphological types of neurons, or according to Peters’ Rule
[9], [10] (Figure 1D). In particular, we found that directed cliques of up to eight
neurons are highly prominent motifs in the reconstructed microcircuits: the average
microcircuit incorporates approximately 108 3-cliques and 4-cliques, approximately
107 5-cliques, approximately 105 6-cliques, and approximately 103 7-cliques. Taking
the alternating sum of the numbers of directed cliques of various sizes, we computed
the Euler characteristic (EC) [5] of the 42 reconstructed microcircuits, obtaining in
each case a value on the order of 107, indicating a preponderence of directed cliques
consisting of an odd number of neurons (Figure 2).
Another topological metric that we considered in this analysis are the Betti
numbers (SI, Supplementary Text, ST1.3) associated to a graph via its directed
flag complex (Figure 1C). These are a sequence of natural numbers β0, β1, β2, ...
that measure the higher-order organizational complexity of the network, detecting
“cyclic” chains of intersecting directed cliques. For each graph considered here we
determined its homological dimension, i.e., the maximum n such that βn (cid:54)= 0. We
showed that the reconstructed microcircuits have homological dimension 5 (Figure
2D), whereas the random graphs considered have homological dimension at most 4,
strongly indicating that the microcircuits possess a higher degree of organizational
complexity than the random graphs.
Topological methods also enabled us to distinguish functional responses to differ-
ent input patterns fed into the microcircuit through thalamo-cortical connections.
We ran simulations of neural activity in one of the reconstructed microcircuits dur-
ing one second, over the course of which a given stimulus was applied every 50 ms
(Figure 3). We then binned the output of the simulations by 5 ms timesteps and
associated to each timestep a transmission-response graph, the vertices of which are
all of the neurons in the microcircuit and the edges of which encode connections in
the microcircuit whose activity in that time step leads to firing of the postsynaptic
neuron (Figure 4). The size of the time bins and the precise rule for formation
of the transmission-response graph for each time bin are biologically motivated, as
explained in more detail in the Supplementary Methods section (SI, Supplementary
Methods, SM1).
From the time series of transmission-response graphs for each of 20 trials of two
different stimuli (called Circle and Point for geometric reasons (Figure 4A), we de-
rived time series of two non-topological metrics (mean firing rate and number of
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
3
Figure 1. (A) A sparse visualization of the microcircuit (soma
and dendrites only). Morphological types are color-coded, with
m-types in the same layer having similar colors. (B) Examples of
simplices in dimensions 0 through 3. (C) An example of a directed
graph and its associated flag complex, in which there is one n-
simplex for every directed (n + 1)-clique in the graph. (D) A graph
depicting the average number of simplices in each dimension for
the flag complexes associated to the reconstructed microcircuit (N-
complexes) and the four types of random graphs considered, each
with the same number of vertices as the reconstructed microcircuit,
where shading indicates standard deviation, which was very small
for all except the N-complexes.
edges in the transmission-response graph) and five topological metrics (the num-
ber of 3-cliques, EC, β0, β1, and β2) and applied a Gaussian Bayes classifier (SI,
Supplementary Methods, SM2) to determine how successfully each of the metrics
classified the 40 trials in the time bins corresponding to the first two stimulations
and in the time bins immediately following those stimulations (Figure 5). In each
of those crucial time bins, the metrics that were most successful at classification
the number of 3-cliques (denoted 2D in the figure), β2, and, in one case, the Euler
characteristic (Figure 2A).
We expect the methods applied here will prove useful for studying networks in
general.
ABCDSimplex count0 simplex1 simplex3 simplex2 simplexoriented graphoriented flag complexN-complexesER-complexesP-complexesM-complexesL-complexes 4
D(cid:32)LOTKO ET AL.
Figure 2. (A) An oriented simplicial complex consisting of eight
2-simplices glued together along their 1-dimensional faces, together
with a table of its Betti numbers and numbers of simplices in
dimensions 0,1, and 2 and a computation illustrating that the
Euler characteristic can be computed as the alternating sum of
the Betti numbers or the simplex counts. (B) Graph depicting
the average Euler characteristic of the reconstructed microcircuit
(N-complexes) and of each of the types of random graph consid-
ered, where the whisker indicates standard deviation, which was
very small, except for N-complexes and P-complexes. (C) Box-
and-whisker plots depicting the Euler characteristics of 35 recon-
structed microcircuits, seven for each individual rat. (D) Box-and-
whisker plots depicting the 5th Betti number of 35 reconstructed
microcircuits, seven for each individual rat.
1. Structural topology
We computed the binary adjacency matrices of all 42 digitally reconstructed
microcircuits and then generated the associated directed flag complexes (SI, Sup-
plementary Text, ST1.2), which are oriented simplicial complexes encoding the
connectivity of all orders of the underlying directed graph: to each directed n-
clique (SI, Supplementary Text, ST1.2) in the underlying graph corresponds to an
oriented (n − 1)-simplex in the flag complex, and the faces of a simplex correspond
ABDCEC=1-0+1=6-12+8=20120-simplices1-simplices2-simplices1016128 TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
5
to the directed subcliques of its associated directed clique (Figure 1 B and C). For
each neuron in the microcircuit, there is a vertex in the underlying directed graph
that is labelled with the unique global identification number (GID) of the neuron.
The (j, k)-coefficient of the structural adjacency matrix is 1 if and only if there is a
directed connection in the microcircuit from the neuron with GID j to the neuron
with GID k. We refer to this adjacency matrix as the structural matrix of the mi-
crocircuit and to its associated directed flag complex as a neocortical microcircuit
complex or N-complex .
Having computed each of the 42 N-complexes, we counted the simplices in each
dimension. For comparison with non-biological matrices, we generated five Erdos-
R´enyi random graphs [4] of a comparable size (31,000 vertices) and connection
probability 0.8%, the same as the average arising from the structural matrices of
the microcircuits (SI, Supplementary Methods SM3.1). We refer to the associated
directed flag complexes as ER-complexes.
To have a more biological control, we also generated 20 adjacency matrices,
given by partly randomizing the structural matrix of one of the average microcir-
cuits, taking into account its biologically meaningful division into six layers in 10
cases and into 55 morphological neuron types (m-types) [7] in 10 cases. The ran-
domization was carried out so that the distance-dependent connection probability
for all pairs of layers (respectively, pairs of m-types) was identical to that of the
original matrix, i.e., for each pair of layers (respectively, m-types) the number of
connections between them was the same as that of the original and for each 25 µm
distance bin the number of connections was identical. The matrices are completely
random otherwise (SI, Supplementary Methods SM3.2, SM3.3). We call the asso-
ciated directed flag complexes L-complexes (respectively, M-complexes). Note that
since each m-type is restricted to a fixed layer, the M-complex should retain more
of the structure of the original N-complex than the L-complex. Our final and most
biological control consisted in the generation of 10 connectivity matrices for 31,000
neurons according to Peters’ Rule [9], [10] for which the associated directed flag
complexes are called P-complexes (SI, Supplementary Methods SM3.4). Having
carried out the computations for 10 control matrices out of each randomized set of
20, the very small variance in the results convinced us that no further computations
should be needed.
The resulting distribution of simplices displayed highly consistent behavior among
the N-complexes, all of which we computed, with a small variation among the sam-
ples arising from different rats. Note that Figure 1 represents the analysis only of the
seven N-complexes arising from the average reconstruction because the randomiza-
tions are based on those microcircuits. The ER-complexes showed almost identical
behavior among the different instances, as did the L-complexes, M-complexes, and
P-complexes. On the other hand, the N-complexes exhibited remarkably different
distributions from the various random complexes (Figure 1 D), with much greater
numbers of simplices and simplices of significantly higher dimension. We computed
the Euler characteristic of all N-complexes, as well as that of the various random
complexes, obtaining large positive values in all cases, due to the predominance of
even-dimensional (particularly 2-dimensional) simplices.
The Betti numbers (SI, Supplementary Text, ST1.2) of a simplicial complex pro-
vide a much finer and more sophisticated measure of its organizational complexity
than the dimension-wise simplex count or the Euler characteristic. The n-th Betti
6
D(cid:32)LOTKO ET AL.
Figure 3.
(A) Average firing rate (top-down projection) in the
stimulated microcircuit, plotted during the first 35 ms after the
first stimulation at t=0 ms in the Point vs. Circle experiment. (B)
Raster plots of the same 500 neurons randomly picked from layer
4, for two trials of the circle stimulus. (C) Population PSTH of all
neurons in the microcircuit for three trials of the Circle stimulus.
(D) Mean firing rate of the Circle and Point stimuli, between t and
t + 5 ms, where light shading indicates the standard deviation and
dark shading the error of the mean.
number, βn, counts the number of chains of simplices intersecting along faces to
create an “n-dimensional hole” in the complex, which requires a certain degree of
organization among the simplices. On the other hand, computation of the Betti
numbers is much more expensive than that of the directed flag complex of a di-
rected graph or its Euler characteristic. In fact, the sheer size of the complexes we
considered here made it practically impossible to do so on a computer with 256 GB
of RAM. We succeeded in computing the highest nonzero Betti numbers of the N-
complexes, however, by restricting our attention to the 5-th and 6-th coskeleta (SI,
Supplementary Text, ST1.2). The top Betti number in all N-complexes appeared
in dimension 5, with β5 varying between 1 and 80 (Figure 2D). By contrast, βn = 0
for all n > 3 for all ER-complexes and P-complexes considered, while βn = 0 for
all n > 4 for all L-complexes and M-complexes. Moreover β4 varies between 0 and
6 for all L-complexes and M-complexes, so that these Betti numbers are almost
negligible.
2. Functional topology
We tested our methods on active microcircuits as well. In an experiment that we
call the Point vs. Circle test, we activated in a simulation the incoming thalamo-
cortical fibers of one of the average that the stimulated fibers formed first a point
shape, then a circle shape [7]. The size of the point shape was chosen such that the
average firing rate of the neurons was essentially the same as for the circle shape,
and in both cases the fibers were activated regularly and synchronously with a
frequency of 20 Hz for one second, similar to the whisker deflection approximation
in [7, Figure 17A]. We performed 20 trials of each stimulus (Figure 3). The trials of
-5-0 ms0-5 ms5-10 ms10-15 ms15-20 ms20-25 ms25-30 ms30-35 ms0 Hz> 60 HzCirclePointA050100150t (ms)051015202530354045Mean FR (Hz)CirclePointDuuuuuStimulationB606060000Population FR (Hz)150100500CCircle trials - example population PSTHsCircle trials - L4 example raster plots15010050015001500Neurons TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
7
each stimulus exhibit biological trial-to-trial variability in the neural response, due
to the stochasticity of the synapse models and of some of the ion channel models.
The aim of this experiment was to determine whether our topological methods were
able to classify the two different stimuli, the point and the circle better than the
firing rate, which is largely overlapping for the first two stimulations (see Figure
3D).
After a systematic analysis to determine the appropriate time bin size and con-
ditions for probable spike transmission from one neuron to another (SI, Supplemen-
tary Methods, SM1.4), we divided the activity of the microcircuit into 5 ms time
bins for 1 second after the initial stimulation and recorded for each 0 ≤ n < 200 a
functional connectivity matrix A(n) for the times between 5n ms and 5(n + 1) ms.
The (j, k)-coefficient of the binary matrix A(n) is 1 if and only if the following three
conditions are satisfied, where sj
i denotes the time of the i-th spike of neuron j.
(1) The (j, k)-coefficient of the structural matrix is 1, i.e., there is a structural
(2) There is some i such that 5n ms ≤ sj
connection from the neuron with GID j to the neuron with GID k.
i < 5(n + 1) ms, i.e., the neuron with
l − sj
i < 7.5 ms, i.e., the neuron with
GID j spikes in the n-th time bin.
(3) There is some l such that 0 ms < sk
GID k spikes after the neuron with GID j, within a 7.5 ms interval.
We call the matrices A(n) transmission-response matrices, as it is reasonable to
assume that the spiking of neuron k is influenced by the spiking of neuron j under
conditions (1)–(3) above.
The goal of the Point vs. Circle test was to determine whether topological met-
rics, such as simplex counts, Betti numbers and Euler characteristic, could classify
correctly two groups of stimuli of a similar nature and whether these metrics con-
tain more information than the mean firing rate. In Figure 4C we provide plots of
the time series of the average zeroth, first, and second Betti numbers, of the aver-
age numbers of 1- and 2-simplices, and of the average Euler characteristic for 20
trials of each stimulus. We applied a Gaussian Bayes classifier (SI, Supplementary
Methods, SM2) to each metric in each time bin, to determine their success rate
at classifying the various trials of the stimuli. To compare, we also classified the
stimuli according to the mean firing rates. To allow for a fair comparison, we used
three mean firing rates (between t to t + 5, t + 5 to t + 10, and t + 10 to t + 15 ms)
for the classification at each time step t, since the transmission-response edges for
time step t are based on information from up to t + 12.5 ms.
As illustrated by Figure 5 A, none of the metrics considered, topological or
otherwise, succeeded very well at classifying the stimuli for times between 10 ms
and 50 ms after the initial stimulation, which is not surprising given the strong
similarity between the spatial propagation of activity of the two stimuli during this
period (Figure 3). On the other hand, in the very first time bin, immediately after
the initial stimulation, the 1- and 2-dimensional simplex counts and β1 and β2 all
classify very well. In the second time bin the 2-dimensional simplex count and β2
continue to classify very well, and the Euler characteristic classifies even better.
Immediately after the second stimulation, from 50 ms to 55 ms after the initial
stimulation, none of the metrics performs very well, but the 2-dimensional simplex
count and β2 still have the highest success rate.
In the next time bin, from 55
ms to 60 ms after the initial stimulation, the 2-dimensional simplex count and β2
8
D(cid:32)LOTKO ET AL.
Figure 4.
(A) Schematic representation of the transmission-
response paradigm: there will be an edge from j to k in the graph
associated to particular time bin if and only if there is a physical
connection from neuron j to neuron k, neuron j fires in the time
bin, and neuron k fires at most 7.5 ms after the firing of neuron
j. Here, shading indicates the error of the mean.(B) Schematic
representation of those firing patterns involving a presynaptic and
a postsynaptic neuron that lead to an edge in the transmission-
response graph, with a red block indicating successful transmission
and a white block indicating lack of transmission. (C) Time series
plots of the average value of the metrics 1D (number of 1-simplices),
2D (number of 2-simplices), β0 (the zeroth Betti number, i.e., the
number of connected components), β1 (the first Betti number), β2
(the second Betti number), and EC (the Euler characteristic) for
the Circle and Point stimuli. Here, shading indicates the error of
the mean.
again classify very well and are the only metrics to do so. In all of these cases, the
topological metrics far outperform the metric based on firing rate.
3. Discussion
We have introduced topological analysis of directed graphs encoding structural
or functional connectivity of digital reconstructions of neural microcircuits. We
showed in particular that these directed graphs differed significantly from random
graphs of both Erdos-R´enyi-type and types taking into account biologically con-
strained, distance-dependent connection probabilities. The topological analysis re-
vealed not only the existence of high-dimensional simplices representing the most
extreme form of circuit “motifs” - all-to-all connectivity within a set of neurons -
that have so far been been reported for brain tissue, but also that there are a surpris-
ingly huge number of these structures. We established moreover that topological
methods effectively distinguish functional responses to distinct thalamic stimuli,
introducing a new measure of the spatio-temporal activity responses generated by
neural tissue. The results of our topological analysis of biologically realistic digital
1DStructural connectionTransmission, no responseTransmission-responsePrePostStep∆t2 = 7.5 ms∆t1 = 5 ms1234ABCt (ms)01234561e5t (ms)01234567891e5CirclePointt (ms)1.82.02.22.42.62.83.03.21e4t (ms)0.00.20.40.60.81.01.21e5020406080100120140t (ms)0.00.51.01.52.02.53.01e5020406080100120140t (ms)−0.50.00.51.01.52.02.51e52Dβ1ECβ0β2 TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
9
Figure 5. (A) Times series plot for the first 80 ms of the 40 trials
of the percentage of correct classifications performed by a Gauss-
ian Bayes classifier based on each of the metrics FR (sequences of
mean firing rates over three consecutive time bins), 1D (number
of 1-simplices), 2D (number of 2-simplices), β0 (the zeroth Betti
number, i.e., the number of connected components), β1 (the first
Betti number), β2 (the second Betti number), and EC (the Eu-
ler characteristic). (B) Graphs depicting the percentage of correct
classifications performed by a Gaussian Bayes classifier based on
each of the metrics in four particularly important time bins: from
0 to 5 ms (immediately after the initial stimulation), from 5 to 10
ms, from 50 to 55 ms (the time bin immediately after the second
stimulation), and from 55 to 60 ms.
reconstructions provide a convincing argument for considering topology as a useful
mathematical tool for analyzing the structural and functional connectome of neural
circuits.
Our results lead naturally to many new questions, most notably concerning the
biological significance of the high-dimensional simplices and homology classes we
have discovered in the digitally reconstructed neocortical microcircuits. We intend
to explore these questions in future studies. In particular we hypothesize that the
time series of different topological metrics could reveal an evolving spatio-temporal
code that goes beyond either rate or timing information to one that incorporates the
structural organization. Such metrics could yield a deeper understanding of how
70%80%90%100%Correct classi(cid:31)cations0-5 ms70%80%90%100%5-10 ms70%80%90%100%Correct classi(cid:31)cations70%80%90%100%FR1D2Dβ0β1β2ECFR1D2Dβ0β1β2ECFR1D2Dβ0β1β2ECFR1D2Dβ0β1β2EC50-55 ms55-60 msABFR1D2D012ECβββ01020304050607050%60%70%80%90%100% 10
D(cid:32)LOTKO ET AL.
the structural organization constrains emergent functional states. Age-dependent
changes in such digital reconstructions may help reveal even more complex topo-
logical structures with development, and changes introduced by synaptic plasticity
may reveal structures associated with learning and memory.
We expect the topological approach to studying directed graphs that we im-
plement here will also prove useful in applications of network science outside of
neuroscience, in the study of networks exhibiting intricate directed connectivity
patterns, such as gene and protein networks, VLSI circuits, and electrical grids.
The obvious utility of the directed flag complex in these applications may also
encourage theorists to establish results analogous to those established by Kahle
concerning Betti numbers of undirected flag complexes of random graphs [6].
4. Materials and methods
4.1. Computation of flag complexes and their Betti numbers. We represent
the directed flag complex of a directed graph by a reference-based data structure,
using vectors to store the references to the simplices in the simplicial complex. The
required storage space grows linearly with the number of vertices and with the
number of edges. A publicly available C++ implementation of the code will be
available on http://neurotop.gforge.inria.fr/. All homology computations carried
out for this paper were made with F2 coefficients, using the boundary matrix re-
duced by an algorithm from the PHAT [2] library. For further details, please see
(SI, Supplementary Text, ST2).
4.2. The Point vs. Circle experiment. The stimulated reconstructed microcir-
cuit is innervated by 310 VPM fibers, whose horizontal centers of innervation are
evenly distributed over the microcircuit (one fiber per mini-column). It is therefore
possible to activate the microcircuit with topographically different stimuli by se-
lecting only a subset of these 310 fibers. Here we used two different stimuli, a point
and a circle, which were calibrated by adjusting the respective number of fibers to
evoke an overall similar mean firing rate (i.e., close enough to prevent clearly distin-
guishing between the two stimuli simply by the mean population firing rate). The
microcircuit was stimulated by synchronous spikes, similar to the whisker deflection
experiment described by Markram et al. (2015). The point stimulus consisted of
synchronous spikes in the 46 neighboring fibers of the center of the microcircuit,
whereas the circle stimulus involved 56 fibers near the periphery of the microcircuit.
The stimulation was repeated every 50 ms, but only the firing rates after the first
two stimulations (at 0 and 50 ms) are overlapping.
We used a Gaussian naıve Bayes classifier [8], where we performed 500 classifica-
tion trials, randomly choosing 15 trials of each stimulus to be part of the training
data, and five trials of each stimulus to be part of the test data. We then obtained
the mean ratio of successfully classified test data points using 500 different training
and test sets. The classification of the firing rate used the firing rates of three
consecutive time bins, to make it a fairer comparison, since the edges may contain
firing rate information of more than two time bins, over a range of 12.5 ms.
4.3. Computation of transmission-response matrices. Transmission-response
matrices were calculated according to the specifications mentioned above, using a
custom-written program in the Python programming language. It combined the
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
11
matrix of synaptic connections (structural matrix), constructed as part of the stan-
dard reconstruction process of the BBP, with the spiking output of a simulation run
and user-defined values for time steps ∆t1 and ∆t2 (5 and 7.5 ms in our analyses).
For further details, please see (SI, Supplementary Methods ,SM1).
4.4. Gaussian Bayes classifiers. The Gaussian Bayes classifier minimises the
probability of misclassification under the assumption that the distributions are
Gaussian. We randomly split the data into training and testing sets. Using the
training set we model the distributions of the dot and circle classes by Gaussians
N (µdot, σ2
circle) respectively. Assuming a uniform prior and
Gaussian distributions, Bayes’ theorem provides a classifier
dot) and N (µcircle, σ2
Class(x) = argmax
c∈{dot,circle}
1(cid:112)2πσ2
c
exp
(cid:18)−(x − µc)2
(cid:19)
.
2σ2
c
5. Acknowledgments
This work was supported by funding from the ETH Domain for the Blue Brain
Project (BBP). The BlueBrain IV IBM BlueGene/Q system is financed by ETH
Board Funding to the Blue Brain Project and hosted at the Swiss National Super-
computing Center (CSCS). We thank Ahmet Bilgili for providing the visualization
of the microcircuit in Figure 1. Partial support for P.D. was provided by the GUDHI
project, supported by an Advanced Investigator Grant of the European Research
Council and hosted by INRIA. M.S. was supported by the NCCR Synapsy of the
Swiss National Science Foundation.
References
[1] U. Bauer, M. Kerber, J. Reininghaus, Clear and Compress: Computing Persistent Homology
in Chunks, TopoInVis 2013, in press.
[2] U. Bauer, M. Kerber, J. Reininghaus, Phat library, https://code.google.com/p/phat/.
[3] T. H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein, Introduction to Algorithms, The
MIT Press, 2001.
[4] Erdos, P.; R´enyi, A. On random graphs, I. Publ. Math. Debrecen 6 1959 290–297.
[5] A. Hatcher, Algebraic Topology, Cambridge University Press (Available Online), 2002.
[6] M. Kahle, Sharp vanishing thresholds for cohomology of random flag complexes, Ann. of
Math. (2), 179 (2014), 1085–1107.
[7] H. Markram, et al., Reconstruction and simulation of neocortical microcircuitry, Cell, 163
(2015) no. 2, 456-492.
[8] Pedregosa, Fabian, et al., Scikit-learn: Machine learning in Python, The Journal of Machine
Learning Research 12 (2011): 2825–2830.
[9] Peters, A., and Feldman, M.L. The projection of the lateral geniculate nucleus to area 17 of
the rat cerebral cortex. I. General description, J. Neurocytol., 5 (1976), 6384.
[10] Peters, A., Proskauer, C.C., Feldman, M.L., and Kimerer, L. The projection of the lateral
geniculate nucleus to area 17 of the rat cerebral cortex. V. Degenerating axon terminals
synapsing with Golgi impregnated neurons, J. Neurocytol., 8 (1979), 331357.
[11] S. Ramaswamy,
rat
et al., The neocortical microcircuit collaboration portal:
somatosensory
in Neural Circuits,
cortex , Frontiers
a re-
9 (2015),
source
http://dx.doi.org/10.3389/fncir.2015.00044.
for
[12] M. Reimann, J. King, E. Muller, S. Ramaswamy, and H. Markram, An algorithm to predict
the connectome of neural microcircuits, Frontiers in Computational Neuroscience (2015)
120, doi:10.3389/fncom.2015.00120.
[13] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon, Network Motifs:
Simple Building Blocks of Complex Networks, Science, 25 October 2002: 298 (5594), 824–827.
[DOI:10.1126/science.298.5594.824]
12
D(cid:32)LOTKO ET AL.
SI Appendix
To accompany “Topological analysis of the connectome of digital reconstructions
Pawe(cid:32)l D(cid:32)lotko, Kathryn Hess, Ran Levi, Max Nolte, Michael Reimann, Martina
Scolamiero, Katharine Turner, Eilif Muller, Henry Markram
of neural microcircuits.”
Contents
(1) Supplementary Methods (3 pages)
(2) Supplementary Text (4 pages)
(3) Supplementary Figures (5 pages)
Supplementary Methods
SM1. Optimization of the parameters for the transmission-response
matrices
For fixed values of ∆t1 and ∆t2, the corresponding sequence {A(n)}N
The transmission-response matrices that allow us to analyze activity in an exper-
iment (cf. the section on Functional Topology in the main body of the article) form
a sequence depending on two parameters, ∆t1 and ∆t2. The number of matrices in
the sequence is the duration of the experiment divided by ∆t1. In other words for
a given experiment of duration T and fixed ∆ti, we obtain a sequence of matrices
S(∆t1, ∆t2) = {A(n) = A(n, ∆t1, ∆t2)}N
n=1, where N is the integer value of T /∆t1.
n=1 is ob-
tained as follows. The spiking output of the simulation is first converted into lists
of spike times, one for each neuron. Standard histogram methods, binning by ∆t1,
are applied to each list to determine in which time steps a presynaptic neuron fired.
For each time bin in which a particular neuron fired, the exact timing of its first
spike in that bin is then compared to the full list of spike times of each neuron it
innervates, to ascertain which of them had spiked at most ∆t2 ms after the presy-
naptic neuron. (Spiking of a pair of neurons within ∆t2 ms is ignored if they are not
structurally connected.) For all pre-postsynaptic pairs satisfying this constraint on
spike timing, the corresponding entry in the transmission-response matrix for that
time step is set to 1 and all others to 0. More precisely, the (j, k)-coefficient of the
binary transmission-response matrix A(n) corresponding to the n-th time bin is 1
if and only if the following three conditions are satisfied, where sj
i denotes the time
of the i-th spike of neuron j.
(1) The (j, k)-coefficient of the structural matrix is 1, i.e., there is a structural
connection from the neuron with GID j to the neuron with GID k, so that
they form a pre-post synaptic pair.
GID k spikes after the neuron with GID j, within a ∆t2 ms interval.
Starting with firing data from spontaneous activity in the reconstructed mi-
crocircuit, we generated sequences of 20 transmission-response matrices for ∆ti ∈
{1, 2, 5, 10, 20, 50, 100} ms, thus creating 49 such sequences corresponding to every
possible choice of the pair (∆t1, ∆t2). We refer to each of these sequences as the
true transmission-response matrices corresponding to the pair (∆t1, ∆t2).
(2) There is some i such that n∆t1 ms ≤ sj
with GID j spikes in the n-th time bin.
l − sj
(3) There is some l such that 0 ms < sk
i < (n + 1)∆t1 ms, i.e., the neuron
i < ∆t2 ms, i.e., the neuron with
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
13
In the rest of this section, we describe the procedure for optimizing the choice of
the time intervals ∆t1 and ∆t2 so that the associated true transmission-response
matrices best reflect the actual successful transmission of signals between neurons
in the microcircuit.
SM1.1. Properties of the transmission-response matrix. The nonzero co-
efficients in a transmission-response matrix are a subset of those in the structural
matrix. Due to the partly stochastic behavior of the in silico microcircuit, the sub-
set will vary even for subsequent applications of the same stimulus. In fact, even an
exact repetition of the same conditions will lead to different transmission-response
matrices, if the random number generator is seeded differently. It follows that the
generation of the transmission-response matrices for a given stimulus should be
considered as a stochastic process. With the correct choice of the parameters ∆ti,
the matrices should reflect how the microcircuit processes a stimulus and thus take
into account parameters of neural processing, such as pre-post synaptic interaction.
To find parameters ∆t1 and ∆t2 that maximize the degree to which neural
processing is captured by the transmission-response matrices, we first develop a
stochastic model for synaptic firing that takes into account neural processing and
that depends on ∆t1 and ∆t2. For the purpose of this analysis, we assume that the
true transmission-response matrices are compatible with this model.
Based upon our model for synaptic firing, we formulate a simplified model that
ignores neural processing. For this simplified model and for any choice of parameters
∆t1 and ∆t2, we explain how to obtain transmission-response matrices from actual
firing data, by shuffling the firing data appropriately, then applying the algorithm
for generating a transmission-response matrix of the previous section. Finally, for
each choice of the parameters ∆t1 and ∆t2, we compare the true transmission-
response matrices for spontaneous activity in the reconstructed microcircuit to
those obtained by the simplified generation process. The parameters that we work
with in the main body of the paper are the ∆t1 and ∆t2 that maximize the difference
(measured by the ratio of the numbers of ones in the matrices) between the actual
transmission-response matrices and those resulting from the simplified model.
SM1.2. Stochastic model with neural processing. Fix time intervals ∆t1 and
∆t2. Let A = (aij) denote the structural matrix of a reconstructed microcircuit,
and let A(n) = (an
ij) denote the transmission-response matrix of the n-th time bin,
based on firing data from a trial of simulated activity in the microcircuit, for the
given intervals ∆t1 and ∆t2. By Condition (1) above, if an
ij = 1 for any n, then
aij = 1. It is reasonable to consider A to be static, at least over the time periods
considered here.
We want to compute the probability that an
ij = 1, given that aij = 1, so we
need to determine on which parameters and properties this probability depends.
According to the definition of transmission-response matrices, a presynaptic and a
postsynaptic spike are required for an
ij to be 1. To simplify the analysis somewhat,
we assume that each neuron ni has a time-dependent, instantaneous firing rate
F i(t) that determines spiking probability at time t, i.e., spiking can be described as
an inhomogeneous Poisson process. Under this assumption, the expected number
mi
(t0) of spikes of neuron ni in the interval [t0, t0 + ∆t1] can be computed as
∆t1
(cid:90) t0+∆t1
mi
∆t1
(t0) =
t0
F i(u)du.
14
D(cid:32)LOTKO ET AL.
∆t1
∆t1
∆t1
K i
−mi
(t0) denotes the probability that neuron ni spikes at least once in the interval
If K i
[t0, t0 + ∆t1], then
(t0) = 1 − P(cid:0)mi
the interval [t0, t0 + ∆t1], then P(cid:0)mi
(t0)(cid:1) = 1 − e
(t0)(cid:1) is the probability that X = 0.) If the
where P(λ) is the Poisson probability mass function with parameter λ at 0. (Recall
that if X is a random variable that counts the number of spikes of neuron ni in
(t) ≈ F i(t) · ∆t1. Moreover,
change in F i(t) is slow compared to ∆t1, then mi
1 − P(λ) ≈ λ for small values of λ. For small enough ∆t1, the expected number
(t0) of spikes of neuron ni will certainly be small, and change in F i(t) will be
mi
slow in compared to ∆t1, so that we may assume that
(t0) ≈ F i(t0) · ∆t1.
(t0),
K i
∆t1
∆t1
∆t1
∆t1
∆t1
For the postsynaptic spike the situation is more complicated. As there is a
causal relation between presynaptic and postsynaptic firing, mediated by synaptic
transmission, we need to estimate the conditional probability of at least one post-
synaptic spike, given that at least one presynaptic spike occured. Let ni and nj
denote neurons such that aij = 1. Let s0 ∈ [t0, t0 + ∆t1] denote the time of the first
presynaptic spike in this interval. Let X j
(s0) denote the random variable whose
value is the number of times neuron nj spiked in the time window [s0, s0 + ∆t2].
Let Y i
(t0) denote the random variable whose value is the number of times neuron
ni spiked in the time interval [t0, t0 + ∆t1]. We need to estimate the conditional
probability
∆t2
∆t1
P(cid:0)X j
(s0) > 0 Y i
∆t1
∆t2
(t0) > 0(cid:1).
The nature of this interaction is very intricate and depends on the identities of the
presynaptic and postsynaptic neurons, the spiking history of the presynaptic neuron
before s0, and all other synaptic input the postsynaptic neuron received. It can
be described as governed by some function Gij modulating the spiking probability
of the postsynaptic neuron nj. This function takes as parameters the expected
number of spikes of neuron nj in the interval [s0, s0 + ∆t2], the time t0, and the
“spiking history” of the presynaptic neuron ni until s0, which we write as a function
si∗(t) evaluated at s0, giving rise to the expression
P(cid:0)X j
(s0) > 0 Y i
∆t1
∆t2
(t0) > 0(cid:1) = 1 − e
−Gij (mj
∆t2
(s0),t0,si∗(s0)).
Summarizing the analysis above, the following formula provides a good estimate
of the probability that an
ij = 1 if aij = 1, for small enough ∆t1 and ∆t2, where
s0 denotes the time of the first presynaptic spike in the interval [n∆t1, (n + 1)∆t1]
and t0 = n∆t1.
(cid:16)
≈ F i(t0) · ∆t1 · Gij(cid:0)F j(s0) · ∆t2, t0, si∗(s0)(cid:1).
(t0)(cid:17) ·(cid:16)
(s0),t0,si∗(s0))(cid:17)
−Gij (mj
1 − e
1 − e
P(cid:0)an
ij = 1aij = 1(cid:1) =
−mi
(1)
∆t2
∆t1
This conditional probability encodes not only the distinctive features of the struc-
tural connectivity (via aij) but also the potentially stimulus-dependant neuron-
specific firing rates (via F i and F j) and their co-variation. Most crucially, it
captures the stimulus-dependent functional modulation of postsynaptic firing by
a presynaptic spike as well. We assume that the true transmission-reponse matri-
ces capture the actual transmission of spikes according to the model of synaptic
firing described by this formula.
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
15
SM.1.3. Null hypotheses: no neural processing. We introduce here a sim-
plified model of synaptic spiking that is based upon formula [1] but that ignores
pre-post synaptic interaction. We then explain how to obtain transmission-response
matrices that correspond to this simplified model from firing data arising from sim-
ulated activity.
We begin by setting each Gij to be the projection onto the first component,
ignoring the pre-post synaptic interaction. After this simplification, the approxi-
mation obtained in the previous section now reads
P (an
ij = 1aij = 1) ≈ F i(t0) · F j(s0) · ∆t1 · ∆t2,
where s0 denotes the time of the first presynaptic spike in the interval [n∆t1, (n + 1)∆t1]
and t0 = n∆t1, as before. Since this drastic simplification neglects the central as-
pect of neural computation - pre-post synaptic interaction - it gives rise to control
cases for each pair of parameters (∆t1, ∆t2) and each choice of firing rate func-
tions F i(t). Comparison of the true transmission-response matrices for each pair of
parameters to the corresponding control matrices for the same pair and a specific
choice of the functions F i(t) will allow us to determine values for ∆t1 and ∆t2 for
which the true transmission-response matrix optimally reflects neural processing.
We assume moreover that the individual firing rates consist of a neuron-dependent
frequency that is up- or down-regulated by a global time series, i.e., that F i(t) =
f (i) · F (t), for some function F (t) and some constant f (i) for each neuron ni.
Transmission-response matrices corresponding to this simplified model for fixed ∆t1
and ∆t2, which we call simplified transmission-response matrices, can be generated
by first shuffling all recorded spikes from simulated activity in the reconstructed
microcircuit, while preserving both the number of spikes per neuron and per time
bin, then applying the usual transmission-response matrix generation method.
SM.1.4. Optimization of parameters. The difference between the true transmission-
response matrices and the control case described above is a consequence of the pre-
post synaptic interaction. Comparison with the control case enables us therefore
to measure how well that interaction is captured in the true transmission-response
matrices. In particular, it is reasonable to optimize the parameters ∆t1 and ∆t2
so that the difference between the true transmission response matrices arising from
actual simulation data and those arising in the control cases is maximized, as a
maximal difference indicates that the effect of the pre-post synaptic interaction is
captured optimally by the true transmission-response matrices.
The comparison between the true transmission-response matrices and the control
cases was carried out by first producing 20 true transmission-response matrices and
20 simplified transmission-response matrices based on firing data obtained from
spontaneous activity in the reconstructed microcircuit for every pair (∆t1, ∆t2),
where ∆ti ∈ {1, 2, 5, 10, 20, 50, 100} ms for i = 1, 2. The number of ones in each
matrix was then computed and the average taken over each set of 20 matrices. Since
no stimulus was applied to the microcircuit, the averages computed are meaningful,
since the firing data should be fairly homogeneous across the time bins.
The average number of ones in the transmission-response matrix arising from
simulated actitivity in the reconstructed microciruit, as a function of ∆t1 and ∆t2,
is illustrated in Figure S1. Figure S2 shows the ratio of the average number of
ones in the true transmission-response matrices to the average number of ones in
the simplified transmission-response matrices, for various values of ∆t1 and ∆t2.
16
D(cid:32)LOTKO ET AL.
In all cases we find that the maximum lies between ∆t2 = 5 ms and ∆t2 = 10 ms,
leading us to choose to work with ∆t2 = 7.5 ms. For ∆t1 we find a maximum at
50 ms, but we use ∆t1 = 5 ms (for which the maximum ratio is only slightly lower
than for ∆t1 = 50 ms) instead to avoid more than one spike per neuron per bin.
SM2. Gaussian Bayes classifiers
Suppose there is a distribution ρ over R × {c1, c2, . . . , ck}, where {c1, c2, . . . , ck}
is a set of class labels. We can project ρ onto each of the coordinates to construct
a real-valued random variable X and a class-label-valued random variable Y . We
wish to build a classifier C : R → {c1, c2, . . . , ck} which will, for any real number,
choose the most likely class to which it might belong. That is,
C(x) = argmax
c∈{c1,c2,...,ck}
P (Y = cX = x),
where P (AB) is the probability of A conditional on B and argmaxa∈A f (a) denotes
the element a ∈ A such that f (a) is maximal. This element of A will in practice
always be unique.
Bayes’ theorem states that
P (Y = cX = x)P (X = x) = P (X = xY = c)P (Y = c).
A Bayesian classifier picks the class with the highest conditional probability, which
using Bayes’ theorem is
C(x) = argmax
c∈{c1,c2,...ck}
P (X = xY = c)P (Y = c)
.
P (X = x)
Usually ρ itself is unknown and must be infered from sample data. We then also
assume some model distribution to estimate ρ from these samples. The Gaussian
Bayes classifier is the Bayes’ classifier under the assumption that the distribution
of each separate class is Gaussian.
After calculating the means and variances of the sample data within each of
the classes separately, we model their respective distributions by the Gaussians
N (µci, σ2
ci
). If p(A) denotes the probability density function of A, then
P (X = xY = c)
P
(Y = c)P (X = x) =
=
p(X = xY = c)
P
1(cid:112)2πσ2
c
exp
(Y = c)p(X = x)
(cid:18)−(x − µc)2
(cid:19) P (Y = c)
2σ2
c
p(X = x)
A common situation, such as in our analysis, is a uniform prior. A uniform prior
over {c1, c2, . . . , ck} means P (Y = ci) = 1/k for all i. If we assume a uniform prior,
then the factor P (Y =c)
p(X=x) is common to all classes and thus does not affect which class
achieves the maximum. Thus we get the formula
1(cid:112)2πσ2
c
exp
(cid:18)−(x − µc)2
(cid:19)
.
2σ2
c
C(x) = argmax
c∈{c1,c2,...ck}
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
17
SM3. Randomization of connection matrices and other control cases
We created four types of random matrices of sizes and connection probabilities
similar to the connectivity matrices of the BBP reconstruction.
SM3.1. Generation of Erdos-R´enyi random matrices. For this basic control
we first computed the overall connection probability in the reconstruction and found
it to be 0.8%. We then generated random, binary square matrices of size 3.1× 104,
where 1’s were placed at random off-diagonal in the matrix with probability 0.8%.
SM3.2. Randomization preserving the distance-dependent connectivity
between layers. Input for this randomization method were the structural matrix
and the matrix of pairwise soma distances, both generated as part of the standard
BBP reconstruction process. The rows and columns of both matrices were first
grouped into N = 6 groups according to the layer of the neuron they correspond
to. This effectively partitioned both matrices into N ∗ N = 36 submatrices each.
For each pair of submatrices, the soma distances were grouped into bins of size
25µm. Next, in the submatrix corresponding to each distance bin, we first replaced
all 1’s by 0’s and then replaced randomly chosen 0’s by 1’s, so that the total number
of 1’s was preserved. Creation of autapses, i.e., a connection from a neuron to itself,
was avoided by creating a separate bin for distances of 0µm.
The result was a connection matrix with the same number of connections between
each pair of layers and the same distance-dependent connection probability between
pairs of layers, to within 25µm, as the original matrix.
SM3.3. Randomization preserving the distance-dependent connectivity
between m-types. This randomization method was identical to the preceding
randomization, preserving connectivity between layers, except that the neurons
were partitioned initially into N = 55 groups of morphological types instead of
only six layers.
SM3.4. Generation of connection matrices according to Peters’ Rule.
For this control case, we started with a connection matrix that placed a connection
not just where a synaptic connection was found in the reconstructed microcircuit,
but between each pair of neurons whose arbors came within close proximity (closer
than 3µm). The resulting matrix had approximately 16 times more connections
than the structural matrix. These connections were then pruned randomly with
a uniform probability until the same number of connections as in the structural
matrix was attained.
18
D(cid:32)LOTKO ET AL.
Supplementary Text
Supplementary Text
To accompany “Topological analysis of the connectome of digital reconstructions
Pawe l D(cid:32)lotko, Kathryn Hess, Ran Levi, Max Nolte, Michael Reimann, Martina
Scolamiero, Katharine Turner, Eilif Muller, Henry Markram
of neural microcircuits.”
ST1. The topological toolbox
Most of the mathematical methods we describe here are part of the basic toolbox
of algebraic topology, though perhaps not as well known in the directed variants
presented here. We give a brief account of these concepts for the benefit of the
non-expert, and refer to literature for the reader interested in further details.
We explain first how to associate to any directed graph a simplicial complex
known as its directed flag complex , then recall two types of important invariants of
simplicial complexes, which turn out to be very useful for analyzing the digitally
reconstructed microcircuits: the Euler characteristic and Betti numbers. We then
describe the data structures and algorithms that we implemented in order to con-
struct the flag complexes of the directed graphs representing the microcircuits and
to compute their Euler characteristics and Betti numbers.
ST1.1. Directed graphs. A directed graph G consists of a pair of finite sets
(V, E) and a function τ : E → V × V . The elements of the set V are the vertices
of G, the elements of E are the edges of G, and the function τ associates with each
edge an ordered pair of vertices. The direction of an edge e with τ (e) = (v1, v2)
is taken to be from τ1(e) = v1, the source vertex , to τ2(v) = v2, the target vertex .
The function τ is required to satisfy the following two conditions.
(1) For each e ∈ E, if τ (e) = (v1, v2), then v1 (cid:54)= v2, i.e., there are no loops in
the graph.
(2) The function τ is injective, i.e., for any pair of vertices (v1, v2), there is at
most one edge directed from v1 to v2.
A vertex v ∈ G is said to be a sink if τ1(e) (cid:54)= v for all e ∈ E, and a source is if
τ2(e) (cid:54)= v for all e ∈ E.
To compare two graphs, we require the following notion. A morphism of directed
graphs from a directed graph G = (V, E, τ ) to a directed graph G(cid:48) = (V (cid:48), E(cid:48), τ(cid:48))
consists of a pair of set maps α : V → V (cid:48) and β : E → E(cid:48) such that β takes an edge
in G with source v1 and target v2 to an edge in G(cid:48) with source α(v1) and target
α(v2), i.e., τ(cid:48) ◦ β = (α, α) ◦ τ . Two graphs G and G(cid:48) are isomorphic if there is
morphism of graphs (α, β) : G → G(cid:48) such that both α and β are bijections, which
we call an isomorphism of directed graphs (Figure S3).
A path in a directed graph G consists of a sequence of edges (e1, ..., en) such that
for all 1 ≤ k < n, the target of ek is the source of ek+1, i.e., τ2(ek) = τ1(ek+1).
The length of the path (e1, ..., en) is n, the number of edges of which the path is
composed. If, in addition, target of en is the source of e1, i.e., τ2(en) = τ1(e1), then
(e1, ..., en) is an oriented cycle.
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
19
ST1.2. Simplicial complexes. An abstract oriented simplicial complex is a col-
lection S of finite, ordered sets with the property that if σ ∈ S, then every subset
τ of σ is also a member of S. A subcomplex of an abstract oriented simplicial
complex is a sub-collection S(cid:48) ⊆ S that is itself an abstract oriented simplicial com-
plex. Henceforth, we simplify terminology and usually refer to abstract oriented
simplicial complexes merely as simplicial complexes.
The elements of a simplicial complex S are called its simplices. A simplicial
If σ ∈ S, we
complex is said to be finite if it has only finitely many simplices.
define the dimension of σ, denoted dim(σ), to be σ − 1, the cardinality of the set
σ minus one. If σ is a simplex of dimension n, then we refer to σ as an n-simplex of
S. The set of all n-simplices of S is denoted Sn. A simplex τ is said to be a face of
σ if τ is a subset of σ of a strictly smaller cardinality. A front face of an n-simplex
σ = (v0, ..., vn) is a face τ = (v0, ..., vm) for some m < n. Similarly, a back face of
σ is a face τ(cid:48) = (vi, . . . , vn) for some 0 < i < n. If σ = (v0, . . . , vn) ∈ Sn, then the
ith face of σ is the (n − 1)-simplex σi obtained from σ by removing the vertex vi.
A simplicial complex gives rise to a topological space by means of the construc-
tion known as geometric realization. In brief, one associates a point (a standard
geometric 0-simplex) with each 0-simplex, a line segment (a standard geometric 1-
simplex) with each 1-simplex, a filled-in triangle (a standard geometric 2-simplex)
with each 2-simplex, etc., glued together along common faces. The intersection of
two simplices in S, neither of which is a face of the other, is a proper subset, and
hence a face, of both of them. In the geometric realization this means that the geo-
metric simplices that realize the abstract simplices intersect on common faces, and
hence give rise to a well-defined geometric object. A geometric n-simplex is nothing
but a (n + 1)-clique, canonically realized as a geometric object. An n-simplex is
In this case the
said to be oriented if there is a linear ordering on its vertices.
corresponding (n + 1)-clique is said to be a directed (n + 1)-clique.
If S is a simplicial complex, then the union S (n) = Sn ∪ ··· ∪ S0, which is called
the n-skeleton of S, is a subcomplex of S. We say that S is n-dimensional
if
S = S (n), and n is minimal with this property. If S is n-dimensional, and k ≤ n,
then the collection Sk ∪ . . . ∪ Sn is not a subcomplex of S because it is not closed
under taking subsets. However if one adds to that collection all the faces of all
simplices in Sk ∪ . . . ∪ Sn, one obtains a subcomplex of S called the k-coskeleton
of S, which we will denote by S(k). The computational usefulness of coskeleta will
become clear when we discuss homology computation (ST1.3).
Directed graphs give rise to abstract oriented simplicial complexes in a natural
way. Let G = (V, E, τ ) be a directed graph. The directed flag complex associated
to G is the abstract simplicial complex S = S(G), with S0 = V and whose n-
simplices Sn for n ≥ 1 are (n + 1)-tuples (v0, . . . , vn), of vertices such that for each
0 ≤ i < j ≤ n, there is an edge in G from vi to vj. Notice that because of the
assumptions on τ , an n-simplex in S is characterised by the (ordered) sequence
(v0, . . . , vn), but not by the underlying set of vertices. For instance (v1, v2, v3) and
(v2, v1, v3) are distinct 2-simplices with the same set of vertices.
ST1.3. Homology, Betti numbers, and Euler characteristic. We now recall
certain well known invariants of simplicial complexes arising in algebraic topol-
ogy, which are preserved under a class of morphisms that is relevant in algebraic
topology and that includes isomorphisms. These invariants serve to measure the
20
D(cid:32)LOTKO ET AL.
“complexity” of simplicial complexes, from various topological perspectives, leading
us to refer to them as metrics.
Homology is an important algebraic invariant of topological spaces.
In this
paper we use only mod-2 simplicial homology, computationally the simplest variant
of homology, which is why we choose to work with it in applications, though other
types of simplicial homology may provide deeper information. We do not give a
complete account of homology here, but rather an elementary description of what
it is and its basic properties.
Let F2 denote the field of two elements, which we denote by 0 and 1. Let S be
a finite simplicial complex. Define the chain complex C∗(S, F2) to be the sequence
{Cn = Cn(S, F2)}n≥0, such that Cn is the F2-vector space whose basis elements
are the n-simplices σ ∈ Sn, for each n ≥ 0. In other words, the elements of Cn
are formal linear combinations of n-simplices in S with coefficients in F2. For each
n ≥ 0, there is a linear transformation called a differential
∂n : Cn+1 → Cn
defined by ∂n(σ) = σ0 +σ1 +···+σn for every n-simplex σ, where σi is the i-th face
of σ, as defined above. Having defined ∂n on the basis, one extends the definition
linearly to the entire vector space Cn.
The n-th Betti number βn(S) of a simplicial complex S is the F2-vector space
dimension of its n-th mod 2 homology group, which is defined by
Hn(S, F2) = Ker(∂n−1)/Im(∂n).
∂n as a (cid:0)Sn × Sn+1(cid:1)-matrix Dn with coefficients in F2, then one can easily
Computing the Betti numbers is conceptually very easy. Let Sn denote the
number of n-simplices in the simplicial complex S. If one encodes the differential
compute its nullity, null(∂n), and its rank , rk(∂n), which are the F2-dimensions of
the null-space and the column space of Dn, respectively. The Betti numbers of S
are then a sequence of natural numbers defined by
β0(S) = dimF2(C0) − rk(∂0),
and βn(S) = null(∂n−1) − rk(∂n).
The n-the Betti number βn counts the number of “n-dimensional holes” in the
geometric realization of S. When S = S(G) is the directed flag complex of a directed
graph G, both the simplices of S and these “n-dimensional holes” can be regarded
as particularly important “metamotifs” [13] in the graph G.
It is easy to show that the n-th Betti number of a simplicial complex S is equal
to that of its (n − 1)-st coskeleton S(n−1), i.e., βn(S) = βn(S(n−1)), for all n. This
observation turns out to be computationally very useful, since there is no need
to store the simplices of dimension less than n − 1 that are not faces of higher
dimensional simplices in order to compute βn(S).
In this paper it was exactly
this trick that allowed us to compute the top dimensional homology of the 42 N-
complexes we worked with.
Homology actually encodes far more information than what is intimated here,
which can potentially be used for analyzing networks, but for the purposes of this
article the description above will suffice.
If S is a simplicial complex and Sn denotes the cardinality of the set of n-
simplices in S, then the Euler characteristic of S is defined to be
(cid:88)
n≥0
χ(S) =
(−1)nSn.
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
21
There is a well known, close relationship between Euler characterstic and Betti
If {βn}n≥0 is the sequence of Betti
numbers [5], which is expressed as follows.
numbers for S, then
(cid:88)
n≥0
χ(S) =
(−1)nβn(S).
See Figure 2A for a specific example.
ST1.4. Hasse Diagrams. A Hasse diagram, otherwise known as a directed
acyclic graph, is a directed graph H = (V, E, τ ) with no oriented cycles. Hasse
diagrams can be used to encode various combinatorial, geometric, and topologi-
cal structures, such as posets and cubical complexes. Below we explain in detail
how Hasse diagrams encode simplicial complexes. We include this discussion here
because our computational algorithm (Algorithm 1) is based on this idea.
A Hasse diagram H is said to be stratified if for each v ∈ V , every path from v
to any sink has the same length. Thus in a stratified Hasse diagram the vertices are
naturally partitioned into disjoint strata, where every directed path from a vertex
in the k-th stratum Vk to any sink is of length k. In particular, the 0-th stratrum
V0 is the set of sinks of H. Moreover, for all e ∈ E, there exists k > 0 such that
τ1(e) ∈ Vk and τ2(e) ∈ Vk−1. Note that if H and H(cid:48) are isomorphic Hasse diagrams,
and H is stratified, then so is H(cid:48).
An orientation ς on a Hasse diagram H consists of a linear ordering <ς,v of the
set Ev of edges with source v, for every vertex v of H. If H = (V, E, τ ) and H(cid:48) =
(V (cid:48), E(cid:48), τ(cid:48)) are Hasse digrams equipped with orientations ς and ς(cid:48), respectively,
then a morphism of oriented Hasse diagrams from (H, ς) to (H(cid:48), ς(cid:48)) is a morphism
of directed graphs (α, β) : H → H(cid:48) such that for every v ∈ V , the restriction of
β to a set map Ev → Eα(v) preserves the orientation, i.e, if e <ς,v e(cid:48) for some
e, e(cid:48) ∈ Ev, then β(e) <ς(cid:48),α(v) β(e(cid:48)). A morphism (α, β) of oriented Hasse diagrams
is an isomorphism if α and β are bijections. A stratified Hasse diagram equipped
with an orientation is called admissible.
Vertices in the k-th stratum of a stratified Hasse diagram H are said to be of level
k. If k < n, and v, u are vertices of levels k and n respectively, then we say that v
is a face of u if there is a path in H from u to v. If H is also oriented and therefore
admissible, and there is a path (e1, ..., en−k) from u to v such that ei = min Eτ1(ei)
for all 1 ≤ i ≤ n − k, we say that v is a front face of u. Similarly, v is a back face
of u if there is a path (e1, ..., en−k) from u to v such that ei = max Eτ1(ei) for all
1 ≤ i ≤ n − k. We let Face(u) denote the set of all faces of u and Face(v)k the set
of those that are of level k, while Front(u) and Back(u) denote its sets of front and
back faces, respectively. See Figure S4 for an illustration of the concepts introduced
above.
Example 1. If G = (V, E, τ ) is a directed graph, then G can be equivalently repre-
sented by an admissible Hasse diagram with level 0 vertices V , level 1 vertices E,
and directed edges from each e ∈ E to its source and target. The ordering on the
edges in the Hasse diagram is determined by the orientation of each edge e in G.
Every simplicial complex S gives rise to an admissible Hasse diagram HS as
follows. The level d vertices of HS are the d-simplices of S. There is a directed
edge from each d-simplex to each of its (d − 1)-faces. The stratification on HS is
thus given by dimension, and the orientation is given by the natural ordering of the
faces of a simplex from front to back. See Figure S5.
22
D(cid:32)LOTKO ET AL.
The Euler characteristic of a stratified Hasse diagram H = (V, E, τ ) is defined
to be the integer
χ(H) =
(−1)kVk.
(cid:88)
k≥0
It is easy to see that isomorphic stratified Hasse diagrams have the same Euler
characteristic. It is also straight forward to show that if H is a stratified Hasse
diagram associated to a simplicial complex S, then the Euler characteristic of H
coincides with that of S.
ST2. Data structures and algorithms
In this section we describe our basic data structures and provide a detailed
overview of the algorithm that constructs the directed flag complex associated to
a directed graph. We also indicate briefly how our homology computations were
performed. A publicly available C++ implementation of the code will be available
on http://neurotop.gforge.inria.fr/.
ST2.1. Data structures. We represent an admissible Hasse diagram H cor-
responding to the directed flag complex of a directed graph G = (V, E, τ ) by a
reference-based data structure, using vectors to store the references to the vertices
of the diagram. Each vertex v ∈ H stores the following information.
(1) Ver(v): A vector of the vertices of G determining the simplex of the flag
complex to which v corresponds.
(2) Tar(v): A vector of references to the vertices that are targets of edges with
source v.
(3) Src(v): A vector of references to the vertices that are sources of edges with
target v.
The admissible Hasse diagram H is thus represented by an ordered set of d vectors,
where d is the maximal level in H, and where the i-th vector contains the references
to all level i vertices.
Let Sint denote the size of integer data types, and for a given graph G = (V, E, τ ),
let V and E denote the cardinalities of the corresponding sets. Each edge of the
Hasse diagram is stored in two vertices of the diagram. If each reference requires
Sint storage, then we require O(E · Sint) space to store all references. In addition,
each vertex stores the vector of vertices in V of the simplex in the flag complex of G
to which it corresponds, which requires an additional O(Sint·d) of space per vertex.
The total size of a Hasse diagram is thus bounded by O((Sint · d) · V + E · Sint).
In particular, the required storage space grows linearly with the number of vertices
and with the number of edges. For our complexity analysis below we assume that
accessing any vertex, using Tar or Src, takes O(1) time.
ST2.2. Creation of the directed flag complex associated to a directed
graph. We describe our algorithm that creates a directed simplicial complex given
a directed graph G. The output is a Hasse diagram H, stored as the data structure
described above. The identifier Ver(v) of a vertex v in H, corresponding to a simplex
σ in the directed flag complex, is the vector of vertices in G that represents σ.
For every level n ≥ 1 vertex v in H such that Ver(v) = [v0, . . . , vn], the algorithm
additionally records a vector Uv of references to level 0 vertices u satisfying the
following properties:
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
23
Algorithm 1 Directed flag complex generation.
Input: A directed graph G = (V, E, τ ).
Output: A Hasse diagram H representing the directed flag complex associated to G.
1: Convert G to level 0 and level 1 vertices of H (cf. Example 1).
2: for every level 1 vertex e ∈ H do
3:
exist e1, e2 such that τ1(e1) = τ1(e), τ1(e2) = τ2(e) and τ2(e1) = τ2(e2) = u
if
then
Add u to Ue;
next level nodes – empty vector of references to nodes;
for top–level vertex e ∈ H do
for Every u ∈ Ue do
4:
5: dim = 2;
6: repeat
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24: until next level nodes = ∅
25: Return H;
Create a node t of a Hasse diagram;
Ver(t) = [Ver(e), u];
Ut = Ue;
Add e to Tar(t);
Add t to Src(e);
for Every bd ∈ Tar(e) do
if u ∈ Ver(cbd) then
Add cbd to Tar(t);
Add t to Src(cbd);
Ut = Ut ∩ Ucbd;
Add next level nodes to H;
dim = dim + 1;
for Every cbd ∈ Src(bd) do
Add t to next level nodes;
(1) u (cid:54)= vi for all 0 ≤ i ≤ n, and
(2) for every u ∈ Uv and every 0 ≤ i ≤ n, there exists an edge in G from vi to
u.
Finally, we assume that the graph G itself is given as an admissible Hasse dia-
gram, as described in Example 1. Under these assumptions Algorithm 1 below is
used to create the directed flag complex associated to G.
ST2.3. Discussion of Algorithm 1. At the start of the algorithm (Line 1) only
levels 0 and 1 of the Hasse diagram H, which are the same as those of the Hasse
diagram representation of G itself, have been created (cf. Example 1). The for
loop in the line 2 initialises the creation of the vectors Uv for level 1 vertices. For
every level 1 vertex e, the vector Ue stores the references to all the 0-simplices that,
together with e, will form a level 2 vertex t. The construction of level 2 vertices in
H is performed during the first iteration of the repeat-until loop starting in Line 6.
We analyze the generation of level 2 vertices as a generic case, since the ar-
guments may clearly be generalised to higher levels. The if condition in Line 3
ensures that the vertex u will be the terminal vertex of the 2-dimensional simplex
corresponding to the level 2 vertex t, created in the first iteration of the repeat-until
loop (Line 6). Moreover the level 1 vertex e will correspond to a front face of the
24
D(cid:32)LOTKO ET AL.
2-simplex associated to t. Therefore, the ordering of Ver(e) can be extended to or-
dering of Ver(t), as in Line 11. Thus all level 2 vertices corresponding to 2-simplices
in the directed flag complex of G will be created by the algorithm. Also, since every
simplex has a unique 1-dimensional front face, every 2-simplex will be created only
once by this process.
Notice also that the if condition in Line 3 ensures that only triangles in G
consisting of three edges oriented as (v1, v2), (v2, v3), and (v1, v3) will give rise
to level 2 vertices in H. It follows by induction that the analogous condition on
orientations is then automatically satisfied for simplices of dimension greater than
2. To see this, fix n ≥ 2, and suppose that all simplices of dimension less than or
equal to n have the desired property. Fix an n-simplex S = [v0, . . . , vn] and u ∈ US.
By definition of the set US, there is an edge from vi to u for every i ∈ {0, . . . , n}.
Note that u ∈ US(cid:48) for any S(cid:48) ∈ Tar(S). The previous iteration of the repeat-until
loop (Line 6) created an oriented simplex from S(cid:48) together with u, of which u is the
last vertex. Since the ordering of elements in S(cid:48) is a restriction of the ordering of
elements in S, the ordering of a n + 1 dimensional simplex [v0, . . . , vn, u] restricted
to any face yields the orientation of that face. It follows that Algorithm 1 does
indeed construct a directed flag complex.
We now discuss the termination of Algorithm 1. If a level n vertex v is a face of a
level (n + 1) vertex w, then the last vertex u in Ver(w) is not present in Ver(v), but
is listed in Uv. From Lines 12 and 21 of the algorithm it is clear that Uw ⊂ Uv and
moreover that u (cid:54)∈ Uv. The cardinalities of the vectors U(−) are therefore decreasing
for the newly created vertices. More precisely, for a vertex t and its faces si, there
exist i such that Ut (cid:12) Usi. Level n + 1 vertices are created only if there exist a
level n vertex t such that Ut (cid:54)= ∅. Since the cardinality of the U(−) decreases with
each iteration of the repeat loop, the algorithm will terminate.
We remark finally that the size of the directed flag complex corresponding to
a given directed graph G may be exponential in the size of G. In that case, the
process of creation of a complex is usually stopped at some fixed dimension n. The
time complexity of Algorithm 1 is proportional to the size of the output complex
H, multiplied by maximal level of a vertex in H (due to the target-source search
performed in Line 15) of the algorithm.
ST2.4. Homology and Betti numbers. All homology computations carried out
for this paper were made with F2 coefficients, using the boundary matrix reduced
by an algorithm from the PHAT [2] library.
TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
25
6. Supplementary Figures
Figure S1. Average number of ones in the true transmission-
response matrices for different pairs of parameters (∆t1, ∆t2) in a
simulation of spontaneous, in-vivo-like activity (Ca 1.2)
1251020501001251020501001.21.62.02.42.83.23.64.04.44.8Delta t2 (ms)Delta t1 (ms)log10 number of ones 26
D(cid:32)LOTKO ET AL.
randomized
and
S2. Comparing
Figure
non-randomized
transmission-response matrices: average number of ones in a true
transmission-response (t-r) matrix divided by the average number
of ones obtained when the recorded spikes were randomized
before calculating the t-r matrix. Matrices were calculated from
simulated spontaneous, ongoing activity with different values for
∆t1 (in different colors) and ∆t2 (along the x-axis). For each
pair (∆t1, ∆t2), matrices for 20 time steps were calculated, and
the mean ratio is shown. Spikes were randomized by shuffling the
identities of the firing neurons, thus conserving the number of
spikes in any given time step and the total number of spikes fired
by each neuron.
Figure S3. (A-C) Examples of directed graphs. Graphs (A) and
(B) are isomorphic, where the isomorphism is given by the map
sending vertex a to 1, b to 2, c to 3, and d to 4. Graphs (A) and
(B) are not isomorphic to graph (C). Vertex b in graph (A) is a
sink, vertex a in the same graph is a source. Graph (C) has no
sources or sinks, which explains the lack of isomorphism to graphs
(A) and (B).
0204060801000.80.91.0∆t1 = 1 ms∆t1 = 2 ms∆t1 = 5 ms∆t1 = 10 ms∆t1 = 20 ms∆t1 = 50 ms∆t1 = 100 ms∆t2 (ms):seno fo rebmun fo oitaRdelffuhs / lautca1.11.21.31.41.5shuffled = non-shuffledacdb12341234(A)(B)(C) TOPOLOGICAL ANALYSIS OF NEURAL MICROCIRCUITS
27
Figure S4. (A) A Hasse diagram that is not stratified, due to the
edge from the vertex 1 to 5. (B) A stratified Hasse diagram, where
vertices 5, 6, and 7 are the vertices of level 0, vertices 2, 3, and 4
are of level 1, and vertex 1 is of level 2. This is also an admissible
Hasse diagram, where the outgoing edges are ordered from left to
right. Vertex 2 is a front face of vertex 1, while vertex 3 is neither
a front nor a back face of a vertex 1, and vertex 4 is back face of
a vertex 1.
Figure S5. Top: The geometric realization of a simplicial com-
plex consisting of seven 0-simplices (labeled 1,...,7), ten 1-simplices,
and four 2-simplices. The orientation on the edges is denoted by
arrows, i.e., the tail of an arrow is its source vertex, while the head
of an arrow is its target. Bottom: The Hasse diagram correspond-
ing to the simplicial complex above. Level k vertices correspond
to k-simplices of the complex and are labeled by the ordered sets
of vertices that constitute the corresponding simplex. Note that,
e.g., vertex 23 is a back face of a vertex 123 and a front face of a
vertex 234.
123455672341(A)(B)1234512132324343545123234345135426767565767567 28
D(cid:32)LOTKO ET AL.
Geometrica, Inria, Saclay, France
Laboratory for Topology and Neuroscience, ´Ecole Polytechnique F´ed´erale de Lau-
sanne, Lausanne, Switzerland
Institute of Mathematics, University of Aberdeen, Aberdeen, UK
Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne, Lausanne, Switzer-
land
Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne, Lausanne, Switzer-
land
Laboratory for Topology and Neuroscience, ´Ecole Polytechnique F´ed´erale de Lau-
sanne, Lausanne, Switzerland
Laboratory for Topology and Neuroscience, ´Ecole Polytechnique F´ed´erale de Lau-
sanne, Lausanne, Switzerland
Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne, Lausanne, Switzer-
land
Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne, Lausanne, Switzer-
land
|
1605.06544 | 1 | 1605 | 2016-05-20T21:38:32 | Inference by Reparameterization in Neural Population Codes | [
"q-bio.NC"
] | Behavioral experiments on humans and animals suggest that the brain performs probabilistic inference to interpret its environment. Here we present a new general-purpose, biologically-plausible neural implementation of approximate inference. The neural network represents uncertainty using Probabilistic Population Codes (PPCs), which are distributed neural representations that naturally encode probability distributions, and support marginalization and evidence integration in a biologically-plausible manner. By connecting multiple PPCs together as a probabilistic graphical model, we represent multivariate probability distributions. Approximate inference in graphical models can be accomplished by message-passing algorithms that disseminate local information throughout the graph. An attractive and often accurate example of such an algorithm is Loopy Belief Propagation (LBP), which uses local marginalization and evidence integration operations to perform approximate inference efficiently even for complex models. Unfortunately, a subtle feature of LBP renders it neurally implausible. However, LBP can be elegantly reformulated as a sequence of Tree-based Reparameterizations (TRP) of the graphical model. We re-express the TRP updates as a nonlinear dynamical system with both fast and slow timescales, and show that this produces a neurally plausible solution. By combining all of these ideas, we show that a network of PPCs can represent multivariate probability distributions and implement the TRP updates to perform probabilistic inference. Simulations with Gaussian graphical models demonstrate that the neural network inference quality is comparable to the direct evaluation of LBP and robust to noise, and thus provides a promising mechanism for general probabilistic inference in the population codes of the brain. | q-bio.NC | q-bio |
Inference by Reparameterization in Neural
Population Codes
Rajkumar V. Raju
Department of ECE
Rice University
Houston, Tx 77005
[email protected]
Dept. of Neuroscience, Dept. of ECE
Baylor College of Medicine, Rice University
Xaq Pitkow
Houston, Tx 77005
[email protected]
Abstract
Behavioral experiments on humans and animals suggest that the brain performs
probabilistic inference to interpret its environment. Here we present a new general-
purpose, biologically-plausible neural implementation of approximate inference.
The neural network represents uncertainty using Probabilistic Population Codes
(PPCs), which are distributed neural representations that naturally encode prob-
ability distributions, and support marginalization and evidence integration in a
biologically-plausible manner. By connecting multiple PPCs together as a proba-
bilistic graphical model, we represent multivariate probability distributions. Ap-
proximate inference in graphical models can be accomplished by message-passing
algorithms that disseminate local information throughout the graph. An attractive
and often accurate example of such an algorithm is Loopy Belief Propagation
(LBP), which uses local marginalization and evidence integration operations to
perform approximate inference efficiently even for complex models. Unfortunately,
a subtle feature of LBP renders it neurally implausible. However, LBP can be
elegantly reformulated as a sequence of Tree-based Reparameterizations (TRP)
of the graphical model. We re-express the TRP updates as a nonlinear dynamical
system with both fast and slow timescales, and show that this produces a neurally
plausible solution. By combining all of these ideas, we show that a network of
PPCs can represent multivariate probability distributions and implement the TRP
updates to perform probabilistic inference. Simulations with Gaussian graphical
models demonstrate that the neural network inference quality is comparable to
the direct evaluation of LBP and robust to noise, and thus provides a promising
mechanism for general probabilistic inference in the population codes of the brain.
1
Introduction
In everyday life we constantly face tasks we must perform in the presence of sensory uncertainty. A
natural and efficient strategy is then to use probabilistic computation. Behavioral experiments have
established that humans and animals do in fact use probabilistic rules in sensory, motor and cognitive
domains [1, 2, 3]. However, the implementation of such computations at the level of neural circuits is
not well understood.
In this work, we ask how distributed neural computations can consolidate incoming sensory in-
formation and reformat it so it is accessible for many tasks. More precisely, how can the brain
simultaneously infer marginal probabilities in a probabilistic model of the world? Previous efforts
to model marginalization in neural networks using distributed codes invoked limiting assumptions,
either treating only a small number of variables [4], allowing only binary variables [5, 6, 7], or
29th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
restricting interactions [8, 9]. Real-life tasks are more complicated and involve a large number of
variables that need to be marginalized out, requiring a more general inference architecture.
Here we present a distributed, nonlinear, recurrent network of neurons that performs inference
about many interacting variables. There are two crucial parts to this model: the representation and
the inference algorithm. We assume that brains represent probabilities over individual variables
using Probabilistic Population Codes (PPCs) [10], which were derived from using Bayes' Rule
on experimentally measured neural responses to sensory stimuli. Here for the first time we link
multiple PPCs together to construct a large-scale graphical model. For the inference algorithm, many
researchers have considered Loopy Belief Propagation (LBP) to be a simple and efficient candidate
algorithm for the brain [11, 12, 13, 14, 8, 5, 7, 6]. However, we will discuss one particular feature of
LBP that makes it neurally implausible. Instead, we propose that an alternative formulation of LBP
known as Tree-based Reparameterization (TRP) [15], with some modifications for continuous-time
operation at two timescales, is well-suited for neural implementation in population codes.
We describe this network mathematically below, but the main conceptual ideas are fairly straight-
forward: multiplexed patterns of activity encode statistical information about subsets of variables,
and neural interactions disseminate these statistics to all other encoded variables for which they are
relevant.
In Section 2 we review key properties of our model of how neurons can represent probabilistic
information through Probabilistic Population Codes. Section 3 reviews graphical models, Loopy
Belief Propagation, and Tree-based Reparameterization. In Section 4, we merge these ingredients
to model how populations of neurons can represent and perform inference on large multivariate
distributions. Section 5 describes experiments to test the performance of network. We summarize
and discuss our results in Section 6.
2 Probabilistic Population Codes
Neural responses r vary from trial to trial, even to repeated presentations of the same stimulus x.
This variability can be expressed as the likelihood function p(rx). Experimental data from several
brain areas responding to simple stimuli suggests that this variability often belongs to the exponential
family of distributions with linear sufficient statistics [10, 16, 17, 4, 18]:
p(rx) = φ(r) exp(h(x) · r),
example, if the posterior distribution is Gaussian (Figure 1), then p(xr) ∝ exp(cid:0)− 1
(1)
where h(x) depends on the stimulus-dependent mean and fluctuations of the neuronal response and
φ(r) is independent of the stimulus. For a conjugate prior p(x), the posterior distribution will also
have this general form, p(xr) ∝ exp(h(x) · r). This neural code is known as a linear PPC: it is
a Probabilistic Population Code because the population activity collectively encodes the stimulus
probability, and it is linear because the log-likelihood is linear in r. In this paper, we assume responses
are drawn from this family, although incorporation of more general PPCs with nonlinear sufficient
statistics T(r) is possible: p(rx) ∝ exp(h(x) · T(r)).
An important property of linear PPCs, central to this work, is that different projections of the
population activity encode the natural parameters of the underlying posterior distribution. For
with a · r and b · r encoding the linear and quadratic natural parameters of the posterior. These
projections are related to the expectation parameters, the mean and variance, by µ = b·r
a·r and σ2 = 1
a·r.
A second important property of linear PPCs is that the variance of the encoded distribution is inversely
proportional to the overall amplitude of the neural activity. Intuitively, this means that more spikes
means more certainty (Figure 1).
The most fundamental probabilistic operations are the product rule and the sum rule. Linear PPCs
can perform both of these operations while maintaining a consistent representation [4], a useful
feature for constructing a model of canonical computation. For a log-linear probability code like
linear PPCs, the product rule corresponds to weighted summation of neural activities: p(xr1, r2) ∝
p(xr1)p(xr2) ⇐⇒ r3 = A1r1 + A2r2. In contrast, to use the sum rule to marginalize out variables,
linear PPCs require nonlinear transformations of population activity. Specifically, a quadratic
nonlinearity with divisive normalization performs near-optimal marginalization in linear PPCs [4].
Quadratic interactions arise naturally through coincidence detection, and divisive normalization is a
nonlinear inhibitory effect widely observed in neural circuits [19, 20, 21]. Alternatively, near-optimal
2 x2a · r + xb · r(cid:1),
2
marginalizations on PPCs can also be performed by more general nonlinear transformations [22]. In
sum, PPCs provide a biologically compatible representation of probabilistic information.
Figure 1: Key properties of linear PPCs. (A) Two single trial population responses for a particular
stimulus, with low and high amplitudes (blue and red). The two projections a · r and b · r encode the
natural parameters of the posterior. (B) Corresponding posteriors over stimulus variables determined
by the responses in panel A. The gain or overall amplitude of the population code is inversely
proportional to the variance of the posterior distribution.
3
Inference by Tree-based Reparameterization
3.1 Graphical Models
(cid:81)
To generalize PPCs, we need to represent the joint probability distribution of many variables. A
natural way to represent multivariate distributions is with probabilistic graphical models. In this work,
we use the formalism of factor graphs, a type of bipartite graph in which nodes representing variables
are connected to other nodes called factors representing interactions between 'cliques' or sets of
variables (Figure 2A). The joint probability over all variables can then be represented as a product
over cliques, p(x) = 1
c∈C ψc(xc), where ψc(xc) are nonnegative compatibility functions on the
set of variables xc = {xcc ∈ C} in the clique, and Z is a normalization constant. The distribution of
Z
interest will be a posterior distribution p(xr) that depends on neural responses r. Since the inference
algorithm we present is unchanged with this conditioning, for notational convenience we suppress
this dependence on r.
In this paper, we focus on pairwise interactions, although our main framework generalizes naturally
to richer, higher-order interactions. In a pairwise model, we allow singleton factors ψs for variable
nodes s in a set of vertices V , and pairwise interaction factors ψst for pairs (s, t) in the set of edges
E that connect those vertices. The joint distribution is then
(cid:89)
s∈V
p(x) =
1
Z
(cid:89)
(s,t)∈E
ψs(xs)
ψst(xs, xt)
(2)
3.2 Belief Propagation and its neural plausibility
variable, ps(xs) =(cid:82)
x\xs
The inference problem of interest in this work is to compute the marginal distribution for each
p(x) d(x\xs). This task is generally intractable. However, the factorization
structure of the distribution can be used to perform inference efficiently, either exactly in the case of
tree graphs, or approximately for graphs with cycles. One such algorithm is called Belief Propagation
(BP) [11]. BP iteratively passes information along the graph in the form of messages mst(xt) from
node s to t, using only local computations that summarize the relevant aspects of other messages
upstream in the graph:
mn+1
st
(xt) =
dxs ψs(xs)ψst(xs, xt)
mn
us(xs)
bs(xs) ∝ ψs
mus(xs)
(3)
(cid:89)
u∈N (s)
(cid:90)
xs
(cid:89)
u∈N (s)\t
where n is the time or iteration number, and N (s) is the set of neighbors of node s on the graph. The
estimated marginal, called the 'belief' bs(xs) at a node s, is proportional to the local evidence at
that node ψs(xs) and all the messages coming into node s. Similarly, the messages themselves are
3
ria.rb.rb.r =a.rBANeuronindex ip(xr)xNeuralresponsePosterior1σ=a.r determined self-consistently by combining incoming messages -- except for the previous message
from the target node t.
This message exclusion is critical because it prevents evidence previously passed by the target node
from being counted as if it were new evidence. This exclusion only prevents overcounting on a tree
graph, and is unable to prevent overcounting of evidence passed around loops. For this reason, BP is
exact for trees, but only approximate for general, loopy graphs. If we use this algorithm anyway, it is
called 'Loopy' Belief Propagation (LBP), and it often has quite good performance [12].
Multiple researchers have been intrigued by the possibility that the brain may perform LBP [13,
14, 5, 8, 7, 6], since it gives "a principled framework for propagating, in parallel, information and
uncertainty between nodes in a network" [12]. Despite the conceptual appeal of LBP, it is important
to get certain details correct: in an inference algorithm described by nonlinear dynamics, deviations
from ideal behavior could in principle lead to very different outcomes.
One critically important detail is that each node must send different messages to different targets to
prevent overcounting. This exclusion can render LBP neurally implausible, because neurons cannot
readily send different output signals to many different target neurons. Some past work simply ignores
the problem [5, 7]; the resultant overcounting destroys much of the inferential power of LBP, often
performing worse than more naïve algorithms like mean-field inference. One better option is to
use different readouts of population activity for different targets [6], but this approach is inefficient
because it requires many readout populations for messages that differ only slightly, and requires
separate optimization for each possible target. Other efforts have avoided the problem entirely by
performing only unidirectional inference of low-dimensional variables that evolve over time [14].
Appealingly, one can circumvent all of these difficulties by using an alternative formulation of LBP
known as Tree-based Reparameterization (TRP).
3.3 Tree-based Reparameterization
Insightful work by Wainwright, Jakkola, and Willsky [15] revealed that belief propagation can
be understood as a convenient way of refactorizing a joint probability distribution, according to
approximations of local marginal probabilities. For pairwise interactions, this can be written as
(cid:89)
s∈V
(cid:89)
(s,t)∈E
p(x) =
1
Z
ψs(xs)
ψst(xs, xt) =
(cid:89)
(cid:89)
s∈V
Ts(xs)
(s,t)∈E
Tst(xs, xt)
Ts(xs)Tt(xt)
(4)
where Ts(xs) is a so-called 'pseudomarginal' distribution of xs and Tst(xs, xt) is a joint pseu-
domarginal over xs and xt (Figure 2A -- B), where Ts and Tst are the outcome of Loopy Belief
Propagation. The name pseudomarginal comes from the fact that these quantities are always locally
consistent with being marginal distributions, but they are only globally consistent with the true
marginals when the graphical model is tree-structured.
These pseudomarginals can be constructed iteratively as the true marginals of a different joint
distribution pτ (x) on an isolated tree-structured subgraph τ. Compatibility functions ψ from factors
remaining outside of the subgraph are collected in a residual term rτ (x). This regrouping leaves the
joint distribution unchanged:
(5)
The factors of pτ are then rearranged by computing the true marginals on its subgraph τ, again
preserving the joint distribution. In subsequent updates, we iteratively refactorize using the marginals
of pτ along different tree subgraphs τ (Figure 2C).
p(x) = pτ (x)rτ (x)
Figure 2: Visualization of tree reparameterization. (A) A probability distribution is specified by
factors {ψs, ψst} on a tree graph. (B) An alternative parameterization of the same distribution in
terms of the marginals {Ts, Tst}. (C) Two TRP updates for a 3× 3 nearest-neighbor grid of variables.
4
p(x)=pi(x)ri(x)p(x)=pj(x)rj(x)x1x2x3x1x2x3AOriginalIteration iIteration jTree reparameterizedBC Typical LBP can be interpreted as a sequence of local reparameterizations over just two neighboring
nodes and their corresponding edge [15]. Pseudomarginals are initialized at time n = 0 using the
s (xs) ∝ ψs(xs) and T 0
st(xs, xt) ∝ ψs(xs)ψt(xt)ψst(xs, xt). At iteration n + 1,
original factors: T 0
the node and edge pseudomarginals are computed by exactly marginalizing the distribution built from
(cid:89)
previous pseudomarginals at iteration n:
(cid:1) T n+1
Notice that, unlike the original form of LBP, operations on graph neighborhoods(cid:81)
(cid:1)(cid:0)(cid:82) T n
(cid:0)(cid:82) T n
st ∝
T n+1
s
u∈N (s)
T n
su dxu
∝ T n
1
T n
s
T n+1
t
T n+1
s
st dxs
st dxt
(cid:90)
T n
st
(6)
s
u∈N (s) do not
differentiate between targets.
4 Neural implementation of TRP updates
4.1 Updating natural parameters
TRP's operation only requires updating pseudomarginals, in place, using local information. These are
appealing properties for a candidate brain algorithm. This representation is also nicely compatible
with the structure of PPCs: different projections of the neural activity encode the natural parameters
of an exponential family distribution. It is thus useful to express the pseudomarginals and the TRP
inference algorithm using vectors of sufficient statistics φc(xc) and natural parameters θn
c for each
c · φc(xc)). For a model with at most pairwise interactions, the TRP
(cid:88)
clique: T n
updates (6) can be expressed in terms of these natural parameters as
c (xc) = exp (θn
st = θn
θn+1
st + Qsθn+1
s + Qtθn+1
t + gE(θn
st)
(7)
s = (1 − ds)θn
θn+1
s +
gV (θn
su)
u∈N (s)
where ds is the number of neighbors of s, and Qs, gV and gE are matrices and nonlinear functions
(for vertices V and edges E) that are determined by the particular graphical model (see below). Since
the natural parameters reflect log-probabilities, the product rule for probabilities becomes a linear
sum in θ, while the sum rule for probabilities must be implemented by nonlinear operations g on θ.
In the concrete case of a Gaussian graphical model, the joint distribution is given by p(x) ∝
exp (− 1
2 x(cid:62)Ax + b(cid:62)x), where A and b are the natural parameters, and the linear and quadratic
functions x and xx(cid:62) are the sufficient statistics. When we reparameterize this distribution by
pseudomarginals, we again have linear and quadratic sufficient statistics: two for each node, φs =
(− 1
t , xs, xt)(cid:62). Each of these vectors
of sufficient statistics has its own vector of natural parameters, θs and θst.
To approximate the marginal probabilities, the TRP algorithm initializes the pseudomarginals to
(cid:62). To update θ, we must extract the matrices Q
θ0
s = (Ass, bs)
and nonlinear functions g that recover the univariate marginal distribution of a bivariate gaussian Tst.
s, xs)(cid:62), and five for each edge, φst = (− 1
st = (Ass, Ast, Att, bs, bt)
s, xsxt, − 1
(cid:62) and θ0
2 x2
2 x2
2 x2
For Tst(xs, xt) ∝ exp(cid:0)− 1
(cid:90)
Ts(xs) =
s − θ2xsxt − 1
2 θ1x2
dxt Tst(xs, xt) ∝ exp
2 θ3x2
(cid:1), this marginal is
(cid:19)
θ4θ3 − θ2θ5
xs
+
Using this, we can determine the form of the weight matrices and the nonlinear functions in the TRP
updates (7).
(8)
(9)
(10)
2
t + θ4xs + θ5xt
x2
s
2
(cid:18)
− θ1θ3 − θ2
(cid:18)0
θ3
Qt =
θ3
0
0
0
1
(cid:19)(cid:62)
(cid:33)(cid:62)
(cid:33)(cid:62)
0
0
1
0
0
3;su − θn
θn
4;suθn
2;suθn
5;su
,
,
θn
3;su
θ2;stθ5;st
θ3;st
,
θ2;stθ4;st
θ1;st
0 0
0 0
θn
1;suθn
0
(cid:18)1
(cid:32)
(cid:32)(cid:0)θn
0
0
0
1
(cid:19)(cid:62)
3;su −(cid:0)θn
(cid:0)θn
(cid:1)2
θn
3;su
2;su
(cid:1)2
(cid:1)2
2;st
θ3;st
, 0,
2;st
θ1;st
Qs =
gV (θn
su) =
gE(θn
st) =
where θi;st is the ith element of θst. Notice that these nonlinear functions are all quadratic functions
with a linear divisive normalization.
5
4.2 Separation of Time Scales for TRP Updates
An important feature of the TRP updates is that they circumvent the 'message exclusion' problem
of LBP. The TRP update for the singleton terms, (6) and (7), includes contributions from all the
neighbors of a given node. There is no free lunch, however, and the price is that the updates at time
n + 1 depend on previous pseudomarginals at two different times, n and n + 1. The latter update is
therefore instantaneous information transmission, which is not biologically feasible.
To overcome this limitation, we observe that the brain can use fast and slow timescales τfast (cid:28) τslow
instead of instant and delayed signals. We convert the update equations to continuous time, and
θ =
introduce auxiliary variables θ which are lowpass-filtered versions of θ on a slow timescale: τslow
−θ + θ. The nonlinear dynamics of (7) are then updated on a faster timescale τfast according to
(cid:88)
u∈N (s)
θs = −ds
τfast
θs +
gV (θsu)
τfast
θst = Qsθs + Qtθt + gE(θst)
(11)
where the nonlinear terms g depend only on the slower, delayed activity θ. By concatenating these
two sets of parameters, Θ = (θ, θ), we obtain a coupled multidimensional dynamical system which
represents the approximation to the TRP iterations:
(12)
Here the weight matrix W and the nonlinear function G inherit their structure from the discrete-time
updates and the lowpass filtering at the fast and slow timescales.
Θ = W Θ + G(Θ)
4.3 Network Architecture
To complete our neural inference network, we now embed the nonlinear dynamics (12) into the
population activity r. Since different projections of the neural activity in a linear PPC encode natural
parameters of the underlying distribution, we map neural activity onto Θ by
r = U Θ
(13)
where U is a rectangular Nr × NΘ embedding matrix that projects the natural parameters and their
low-pass versions into the neural response space. These parameters can be decoded from the neural
activity as Θ = U +r, where U + is the pseudoinverse of U.
Applying this basis transformation to (12), we have r = U Θ = U (W Θ + G(Θ)) = U W U +r +
UG(U+r). We then obtain the general form of the updates for the neural activity
r = WLr + GN L(r)
(14)
where WLr = U W U +r and GN L(r) = UG(U +r) correspond to the linear and nonlinear computa-
tional components that integrate and marginalize evidence, respectively. The nonlinear function on r
inherits the structure needed for the natural parameters, such as a quadratic polynomial with a divisive
normalization used in low-dimensional Gaussian marginalization problems [4], but now expanded to
high-dimensional graphical models. Figure 3 depicts the network architecture for the simple graphical
model from Figure 2A, both when there are distinct neural subpopulations for each factor (Figure 3A),
and when the variables are fully multiplexed across the entire neural population (Figure 3B). These
simple, biologically-plausible neural dynamics (14) represent a powerful, nonlinear, fully-recurrent
network of PPCs which implements the TRP update equations on an underlying graphical model.
5 Experiments
We evaluate the performance of our neural network on a set of small Gaussian graphical models
with up to 400 interacting variables. The networks time constants were set to have a ratio of
τslow/τfast = 20. Figure 4 shows the neural population dynamics as the network performs inference,
along with the temporal evolution of the corresponding node and pairwise means and covariances.
The neural activity exhibits a complicated timecourse, and reflects a combination of many natural
parameters changing simultaneously during inference. This type of behavior is seen in neural activity
recorded from behaving animals [23, 24, 25].
6
Figure 3: Distributed, nonlinear, recurrent network of neurons that performs probabilistic inference
on a graphical model. (A) This simple case uses distinct subpopulations of neurons to represent
different factors in the example model in Figure 2A. (B) A cartoon shows how the same distribution
can be represented as distinct projections of the distributed neural activity, instead of as distinct
populations. In both cases, since the neural activities encode log-probabilities, linear connections are
responsible for integrating evidence while nonlinear connections perform marginalization.
Figure 4: Dynamics of neural population activity (left) and the expectation parameters of the posterior
distribution that the population encodes (right) for one trial of the tree model in Figure 2A.
Figure 5 shows that our recurrent neural network accurately infers the marginal probabilities, and
reaches almost the same conclusions as loopy belief propagation. The data points are obtained from
multiple simulations with different graph topologies, including graphs with many loops. Figure 6
verifies that the network is robust to noise even when there are few neurons per inferred parameter;
adding more neurons improves performance since the noise can be averaged away.
Figure 5: Inference performance of our neural network (blue) and standard loopy belief propa-
gation (red) for a variety of graph topologies, including square grids up to 20 × 20 and densely
connected graphs with up to 25 variables. The expectation parameters (means, covariances) of the
pseudomarginals closely match the corresponding parameters for the true marginals.
6 Conclusion
We have shown how a biologically-plausible nonlinear recurrent network of neurons can repre-
sent a multivariate probability distribution using population codes, and can perform inference by
reparameterizing the joint distribution to obtain approximate marginal probabilities.
7
linearconnectionssingletonprojectionspairwiseprojectionsnonlinearconnectionslinearconnectionssingletonpopulationspairwisepopulationsnonlinearconnectionsr12r23r1r2r3ABminmaxTimeNeural activity rTimeTimeInferred expectationparametersMeansCovariancesTrue meanInferred meanNeural networkLBPTrue varianceInferred varianceTrue covarianceInferred covariance Figure 6: Network performance is robust to noise, and improves with more neurons. (A) Neural
activity performing inference on a 5 × 5 square grid, in the presence of independent spatiotemporal
Gaussian noise of standard deviation 0.1 times the standard deviation of each signal. (B) Expectation
parameters (means, variances) of the node pseudomarginals closely match the corresponding parame-
ters for the true marginals, despite the noise. Results are shown for one or five neurons per parameter
in the graphical model, and for no noise (i.e. infinitely many neurons).
Our network model has desirable properties beyond those lauded features of belief propagation. First,
it allows for a thoroughly distributed population code, with many neurons encoding each variable and
many variables encoded by each neuron. This is consistent with neural recordings in which many
task-relevant features are multiplexed across a neural population [23, 24, 25].
Second, the network performs inference in place, without using a distinct neural representation for
messages, and avoids the biological implausibility associated with sending different messages about
every variable to different targets. This virtue comes from exchanging multiple messages for multiple
timescales. It is noteworthy that allowing two timescales prevents overcounting of evidence on loops
of length two (target to source to target). This suggests a novel role of memory in static inference
problems: a longer memory could be used to discount past information sent at more distant times,
thus avoiding the overcounting of evidence that arises from loops of length three and greater. It may
therefore be possible to develop reparameterization algorithms with all the convenient properties of
LBP but with improved performance on loopy graphs.
Previous results show that the quadratic nonlinearity with divisive normalization is convenient and
biologically plausible interpretable, but this precise form is not necessary: other pointwise neuronal
nonlinearities are capable of producing the same high-quality marginalizations in PPCs [22]. In a
distributed code, the precise nonlinear form at the neuronal scale is not important as long as the effect
on the parameters is the same.
More generally, however, different nonlinear computations on the parameters implement different
approximate inference algorithms. The distinct behaviors of such algorithms as mean-field inference,
generalized belief propagation, and others arise from differences in their nonlinear transformations.
Even Gibbs sampling can be described as a noisy nonlinear message-passing algorithm. Although
LBP and its generalizations have strong appeal, we doubt the brain will use this algorithm exactly.
The real nonlinear functions in the brain may implement even smarter algorithms.
To identify the brain's algorithm, it may be more revealing to measure how information is represented
and transformed in a low-dimensional latent space embedded in the high-dimensional neural responses
than to examine each neuronal nonlinearity in isolation. The present work is directed toward this
challenge of understanding computation in this latent space. It provides a concrete example showing
how distributed nonlinear computation can be distinct from localized neural computations. Learning
this computation from data will be a key challenge for neuroscience. In future work we aim to
recover the latent computations of our network from artificial neural recordings generated by the
model. Successful model recovery would encourage us to apply these methods to large-scale neural
recordings to uncover key properties of the brain's distributed nonlinear computations.
Author contributions
XP conceived the study. RR and XP derived the equations. RR implemented the computer simulations.
RR and XP analyzed the results and wrote the paper.
8
minmaxTrue parametersNneuronsBAno noiseNeural activity rInferred parametersMeanVarianceTimeNparams= 1NneuronsNparams= 5 Acknowledgments: XP and RR were supported by a grant from the McNair Foundation and by
the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior
Business Center (DoI/IBC) contract number D16PC00003.1
References
[1] Knill DC, Richards W (1996) Perception as Bayesian inference. Cambridge University Press.
[2] Doya K (2007) Bayesian brain: Probabilistic approaches to neural coding. MIT press.
[3] Pouget A, Beck JM, Ma WJ, Latham PE (2013) Probabilistic brains: knowns and unknowns. Nature
neuroscience 16: 1170 -- 1178.
[4] Beck JM, Latham PE, Pouget A (2011) Marginalization in neural circuits with divisive normalization. The
Journal of neuroscience 31: 15310 -- 15319.
[5] Ott T, Stoop R (2006) The neurodynamics of belief propagation on binary markov random fields. In:
Advances in neural information processing systems. pp. 1057 -- 1064.
[6] Steimer A, Maass W, Douglas R (2009) Belief propagation in networks of spiking neurons. Neural
Computation 21: 2502 -- 2523.
[7] Litvak S, Ullman S (2009) Cortical circuitry implementing graphical models. Neural computation 21:
3010 -- 3056.
[8] George D, Hawkins J (2009) Towards a mathematical theory of cortical micro-circuits. PLoS Comput Biol
5: e1000532.
[9] Grabska-Barwinska A, Beck J, Pouget A, Latham P (2013) Demixing odors-fast inference in olfaction. In:
Advances in Neural Information Processing Systems. pp. 1968 -- 1976.
[10] Ma WJ, Beck JM, Latham PE, Pouget A (2006) Bayesian inference with probabilistic population codes.
Nature neuroscience 9: 1432 -- 1438.
[11] Pearl J (1988) Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan
Kaufmann.
[12] Yedidia JS, Freeman WT, Weiss Y (2003) Understanding belief propagation and its generalizations.
Exploring artificial intelligence in the new millennium 8: 236 -- 239.
[13] Lee TS, Mumford D (2003) Hierarchical bayesian inference in the visual cortex. JOSA A 20: 1434 -- 1448.
[14] Rao RP (2004) Hierarchical bayesian inference in networks of spiking neurons. In: Advances in neural
information processing systems. pp. 1113 -- 1120.
[15] Wainwright MJ, Jaakkola TS, Willsky AS (2003) Tree-based reparameterization framework for analysis of
sum-product and related algorithms. Information Theory, IEEE Transactions on 49: 1120 -- 1146.
[16] Jazayeri M, Movshon JA (2006) Optimal representation of sensory information by neural populations.
Nature neuroscience 9: 690 -- 696.
[17] Beck JM, Ma WJ, Kiani R, Hanks T, Churchland AK, Roitman J, Shadlen MN, et al. (2008) Probabilistic
population codes for bayesian decision making. Neuron 60: 1142 -- 1152.
[18] Graf AB, Kohn A, Jazayeri M, Movshon JA (2011) Decoding the activity of neuronal populations in
macaque primary visual cortex. Nature neuroscience 14: 239 -- 245.
[19] Heeger DJ (1992) Normalization of cell responses in cat striate cortex. Visual neuroscience 9: 181 -- 197.
[20] Carandini M, Heeger DJ (2012) Normalization as a canonical neural computation. Nature Reviews
Neuroscience 13: 51 -- 62.
[21] Rubin DB, Van Hooser SD, Miller KD (2015) The stabilized supralinear network: A unifying circuit motif
underlying multi-input integration in sensory cortex. Neuron 85: 402 -- 417.
[22] Vasudeva Raju R, Pitkow X (2015) Marginalization in random nonlinear neural networks. In: COSYNE
Abstract.
[23] Hayden BY, Platt ML (2010) Neurons in anterior cingulate cortex multiplex information about reward and
action. The Journal of Neuroscience 30: 3339 -- 3346.
[24] Rigotti M, Barak O, Warden MR, Wang XJ, Daw ND, Miller EK, Fusi S (2013) The importance of mixed
selectivity in complex cognitive tasks. Nature 497: 585 -- 590.
[25] Raposo D, Kaufman MT, Churchland AK (2014) A category-free neural population supports evolving
demands during decision-making. Nature neuroscience 17: 1784 -- 1792.
1The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwith-
standing any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of
the authors and should not be interpreted as necessarily representing the official policies or endorsements, either
expressed or implied, of IARPA, DoI/IBC, or the U.S. Government.
9
|
1506.06717 | 2 | 1506 | 2015-11-22T11:50:07 | How we move is universal: scaling in the average shape of human activity | [
"q-bio.NC"
] | Human motor activity is constrained by the rhythmicity of the 24 hours circadian cycle, including the usual 12-15 hours sleep-wake cycle. However, activity fluctuations also appear over a wide range of temporal scales, from days to a few seconds, resulting from the concatenation of a myriad of individual smaller motor events. Furthermore, individuals present different propensity to wakefulness and thus to motor activity throughout the circadian cycle. Are activity fluctuations across temporal scales intrinsically different, or is there a universal description encompassing them? Is this description also universal across individuals, considering the aforementioned variability? Here we establish the presence of universality in motor activity fluctuations based on the empirical study of a month of continuous wristwatch accelerometer recordings. We study the scaling of average fluctuations across temporal scales and determine a universal law characterized by critical exponents $\alpha$, $\tau$ and $1/{\mu}$. Results are highly reminiscent of the universality described for the average shape of avalanches in systems exhibiting crackling noise. Beyond its theoretical relevance, the present results can be important for developing objective markers of healthy as well as pathological human motor behavior. | q-bio.NC | q-bio | Papers in Physics, vol. 7, art. 070017 (2015)
Received: 18 October 2015, Accepted: 10 November 2015
Edited by: E. Mizraji
Reviewed by: J. Lin, Department of Physics, Washington College, Maryland, USA.
Licence: Creative Commons Attribution 3.0
DOI: http://dx.doi.org/10.4279/PIP.070017
www.papersinphysics.org
ISSN 1852-4249
How we move is universal: Scaling in the average shape of human activity
Dante R. Chialvo,1 Ana Mar´ıa Gonzalez Torrado,2 Ewa Gudowska-Nowak,3
Jeremi K. Ochab,4 Pedro Montoya,2 Maciej A. Nowak,3, 4 Enzo Tagliazucchi5
Human motor activity is constrained by the rhythmicity of the 24 hours circadian cycle,
including the usual 12-15 hours sleep-wake cycle. However, activity fluctuations also ap-
pear over a wide range of temporal scales, from days to a few seconds, resulting from the
concatenation of a myriad of individual smaller motor events. Furthermore, individuals
present different propensity to wakefulness and thus to motor activity throughout the cir-
cadian cycle. Are activity fluctuations across temporal scales intrinsically different, or is
there a universal description encompassing them? Is this description also universal across
individuals, considering the aforementioned variability? Here we establish the presence of
universality in motor activity fluctuations based on the empirical study of a month of con-
tinuous wristwatch accelerometer recordings. We study the scaling of average fluctuations
across temporal scales and determine a universal law characterized by critical exponents
α, τ and 1/µ. Results are highly reminiscent of the universality described for the average
shape of avalanches in systems exhibiting crackling noise. Beyond its theoretical relevance,
the present results can be important for developing objective markers of healthy as well
as pathological human motor behavior.
1 Consejo Nacional de Investigaciones Cient´ıficas y Tec-
nol´ogicas (CONICET), Rivadavia 1917, Buenos Aires,
Argentina.
2 Institut Universitari d'Investigacions en Ci`encies de la
Salut (IUNICS) & Universitat de les Illes Balears (UIB),
Palma de Mallorca, Spain.
3 M. Kac Complex Systems Research Center and M.
Smoluchowski Institute of Physics, Jagiellonian Univer-
sity, Krak´ow, Poland.
4 Biocomplexity Department, Ma(cid:32)lopolska Center of
Biotechnology, Jagiellonian University, Krak´ow, Poland.
5 Institute for Medical Psychology, Christian Albrechts
University, Kiel, Germany.
I.
Introduction
The most obvious periodicity of human (as well as
animal) motor activity is the circadian twenty four
hours modulation. However, smaller fluctuations
are evident on a wide range of temporal scales, from
days to a few seconds. Data shows that the activity
evolves in bursts of all sizes and durations which are
known to be scale-invariant [1–8] regardless of the
origins and intended consequences of such activity.
Despite the variety of results, the mechanisms un-
derlying the scale-invariant behavior of motor ac-
tivity remain to be elucidated. Considering the in-
termittent nature of human motor activity - com-
prising brief activity excursions separated by peri-
ods of quiescence - a natural approach would be to
study the average shape of the events, following re-
cent results [9–12] which show that for a large class
of processes, the average shape is a scaling function
determined mostly by the temporal correlations of
070017-1
5
1
0
2
v
o
N
2
2
]
.
C
N
o
i
b
-
q
[
2
v
7
1
7
6
0
.
6
0
5
1
:
v
i
X
r
a
Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
the process and its nonlinearities [13].
In the present work, long time series of human
motor activity are analyzed, recorded via wrist-
watch accelerometer,
lasting approximately one
month. We establish first the presence of truncated
scale-invariance in the distribution of the durations
of the events as well as in its power spectral den-
sity, as described previously in similar type of data.
Afterwards, we uncover the average shape of the
bursts of activity and derive the scaling function
and its associated exponents. Finally, we discuss
the origins of such scaling and some possible appli-
cations.
II. Materials and methods
The recordings analyzed were part of a larger study
and included six healthy, non-smokers, drug-free
volunteers (mean age 50.1 years, S.D. = 6.8). The
study was approved by the Bioethics Commission
of the University of Isles Baleares (Spain). Par-
ticipants were informed about the procedures and
goals of the study, and provided their written con-
sent. After determining their handedness, each
subject was provided with a wristwatch-sized activ-
ity recorder (Actiwatch from Mini- Mitter Co., OR,
USA) measuring acceleration changes in the fore-
arm in any plane. Each data point of activity corre-
sponded to the number of zero crossings in acceler-
ation larger than 0.01 G (sampled at 32 Hz and in-
tegrated over a 30-second window length). Records
of several thousands of data points were kept in the
device's internal memory until being downloaded to
a personal computer every week. Subjects wore the
device in their non-dominant arm continuously for
up to several weeks (mean 28.1 days, S.D.= 4.). Af-
ter careful visual inspection of the data to exclude
sets with gaps (due to subject non-compliance), a
combined total of 280 days of data was available
for further analysis.
III. Results
Figure 1: Example data set, distribution of suc-
cessive increments and their spectral power. Panel
A: Time series of activity x(n) recorded continu-
ously from a subject during a month.
Individual
traces correspond to consecutive days. The top
subpanel depicts daily activity averaged over the
entire month. Panel B: Time series of successive
increments I(n) = x(n + 1) − x(n) (normalized by
its SD) for the same data. Panel C: Probability
density distribution of the time series of succes-
sive increments I(n) (continuous line), exhibiting
exponential tails (compare with the dotted line, a
Gaussian of the same variance). Panel D: Power
spectral density (black line) of the time series of
successive increments I(n) of panel B. This is scale
invariant S(f ) ∼ f γ with γ = 0.9 (dashed line). In
contrast, for the randomly shuffled increments, the
serial correlations vanish and a flat spectral density
is obtained (red).
For ease of presentation, we will use recordings from
a single subject to describe the main results. Nev-
ertheless, results are robust as well as similar for
the entire group of subjects in the study. A typical
recording is presented in Fig. 1. Panel A shows
a full month of continuously recorded activity from
this subject, who is particularly regular in her daily
routines. The subject wakes up with the alarm
clock at 6:45 a.m. on week days and has lunch fol-
lowed by a short nap each day (between 2:00 p.m.
and 4:00 p.m. Panel B displays the time series of
the successive increments of the signal x(n), defined
070017-2
68101214161820220246Mon12/14Tue12/15Wed12/16Thu12/17Fri12/18Sat12/19Sun12/20Mon12/21Tue12/22Wed12/23Thu12/24Fri12/25Sat12/26Sun12/27Mon12/28Tue12/29Wed12/30Thu12/31Fri01/01Sat01/02Sun01/03Mon01/04Tue01/05Wed01/06Thu01/07Fri01/08Sat01/09Sun01/10Mon01/11Tue01/12HourDate6810121416182022024600.51Activity (a.u)07142128−505Time(days)I(n)/S.D.−505−4−3−2−10I/S.D.log10P(I)/P(0)−6−5−4−3−2−6−4−20log10f(sec−1)log10S(f)BCDA1 day1 hour1 min. Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
S(f ) ∼ f β [1, 2]. Because this type of processes
are likely to be non-stationary, it is best to esti-
mate the exponents of the spectral density by doing
the calculations over the time series of successive
increments, whose density distribution is station-
ary. For instance, for Brownian motion (which is
summed white noise), the power spectrum decays
S(f ) ∼ f β with β = −2 and for white noise β = 0;
the summed time series has an exponent +2 larger
than the non-summed time series. As discussed in
[14], this can be generalized for all self-affine pro-
cesses: summing a self-affine time series shifts the
theoretical power-spectral density exponent by +2,
and the reverse process is also true: the differences
in consecutive values (the "first differences") of a
Brownian motion result in white noise, thus tak-
ing the first differences shifts the theoretical power-
spectral density exponent β by −2. In our case, the
exponent obtained for the time series of successive
increments I(n) was γ = 0.9. Thus, the exponent
of the raw data is β = γ − 2 = −1.1 [14]. For com-
parison, the spectral densities of the actual signal
and of a surrogate obtained after randomly shuf-
fling the increments are jointly displayed in Panel
D of Fig. 1.
To further study the time series from the perspec-
tive of individual bursts of activity, we introduce
the definition of an event. We consider the time
series of activity x(n) and select a threshold value
U to be vanishingly small. An event is defined by
the consecutive points starting when x(n) > U and
ending when x(n) < U . This is equivalent to the
definition of avalanches in other contexts [9,15]. In
the following part, we will be concerned primarily
with the statistics of event lifetimes T , as well as of
their average size S and shape. In all subjects, we
found that the distributions of event durations and
sizes (defined by the area, i.e., the integral of the
signal corresponding to the individual events) can
be well described, for relatively small values, by a
power-law (Fig. 2, Panel A). In contrast, the dis-
tribution of waiting times between events demon-
strated an exponential decay.
In addition to the
scale invariance, we found that the longer an event
lasted, the stronger the motor activity executed by
the subject. The plot of average event size (cid:104)S(cid:105) as
a function of duration T follows a power-law (for
small values of T ) described by (cid:104)S(cid:105)(T ) = T µ+1 with
µ + 1 = 1.59. The exponents in this power-law are
robust across subjects and to changes of threshold
Figure 2: Scaling of activity events in a single
subject (same dataset as in Fig.
1). Panel A:
The complementary cumulative distribution func-
tion (CCDF) for event durations (T) and sizes (S)
obeys power-laws with exponents α(cid:48) = 0.70 and
τ(cid:48) = 0.44, respectively (dashed lines). Note that
here the densities are cumulative, thus the expo-
nents of the respective PDFs are α = α(cid:48) + 1 and
τ = τ(cid:48) + 1. The waiting time between events
falls exponentially. Panel B: The average size of
a given duration is well described (for small T) by
(cid:104)S(cid:105)(T ) ∼ T µ+1 with µ + 1 = 1.59 (blue dashed
line) comparable with results obtained from fitting
within the scaling region (red filled symbols) giving
µ + 1 = 1.61.
as I(n) = x(n + 1) − x(n).
The large-scale statistical features of the time se-
ries presented in Fig. 1 are already well known.
The density distribution of the successive incre-
ments i(n) is non-Gaussian, as can be appreci-
ated by a joint plot with a Gaussian distribution
of the same variance (Fig. 1, Panel C). It is known
that the power spectrum of the activity decays as
070017-3
102103104T102103104105<S> 100101102103104105S,Wt,T10-310-210-1100CCDFSWtTABa't'm + 1 Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
cesses in cells [17], crackling noise [11], the statis-
tics of Barkhausen noise in permalloy thin films [10]
and plastic deformation of metals [18]. In all these
cases, the distributions obey universal functional
forms:
(1)
(2)
(3)
f (S) ∼ S−τ ,
f (T ) ∼ T −α,
(cid:104)S(cid:105)(T ) ∼ T 1/σνz,
where f denotes the probability density functions
of the size of the event S and its duration T , and
(cid:104)S(cid:105)(T ) is the expected size for a given duration.
The parameters τ , α and 1/σνz are the critical ex-
ponents of the system and are expected to be inde-
pendent of the details, being related to each other
by the scaling relation:
α − 1
τ − 1
σνz
(4)
1
.
=
We found that the empirical exponents very
closely fulfill the expression above. Using the fitting
approach introduced by Clauset [19] in the scaling
regions depicted in Panel A of Fig. 2, we found
τ = 1.44 and α = 1.70 . Thus, from Eq.
(4) a
value of 1/σνz = µ + 1 = 1.59 is expected. The
experimental data points are very close to this the-
oretical expectation (dashed line), especially for the
relatively small T values within the scaling region of
Panel A (where a linear fit estimates µ + 1 = 1.61),
while those for relatively larger T values (corre-
sponding to the cutoff of the distributions) are a
bit apart, probably due to undersampling. After
repeating this analysis for all subjects in our sam-
ple, the average exponents were all within 5% of
the reported values.
From scaling arguments, it is expected that the
average shape of an event of duration T (cid:104)x(T, t)(cid:105)
scales as :
(cid:104)x(T, t)(cid:105) = T µfshape(t/T ).
(5)
Thus, the shapes of events of different durations
T rescaled by µ should collapse on a single scaling
function given by fshape(t/T ). Note that µ corre-
sponds in this context to the wandering exponent
(i.e., the mean squared displacement) of the activ-
ity [13, 20].
Figure 3: Collapse of events of different duration
into a single functional form. Panel A: Three ex-
amples of typical events of duration T=480, 960
and 1920 sec.. Panel B: The heterogeneous events
shown in Panel A can be collapsed onto the av-
erage shape (dashed black line) by normalizing t
to t/T and (cid:104)x(t)(cid:105) to (cid:104)x(t)(cid:105)/T µ. The inset shows
the cumulative variance for a range of µ. Panel
C: The average event shape, i.e., fshape(t/T ), re-
covered from six data sets (thin lines). The best
fit using an inverted parabola is shown as a red
dashed line (µ = 0.49) as well as the one expected
from the critical exponent µ = 0.59 as a dot-dashed
blue line.
over a reasonable range of values.
This type of scaling is well known in the statisti-
cal mechanics of critical phenomena [15]. Examples
range from earthquakes [16] to active transport pro-
Examples of this collapse are presented in Pan-
els A and B of Fig. 3. Considering the number of
events here averaged (in the order of N ∼ 102), the
070017-4
00.20.40.60.81 t/T051015<x> / T m0.20.40.6m0.10.150.2e00.20.40.60.81t/T00.20.40.60.81 <x> / T m<x(t/T)> / <x(t/T)>max(4(t/T)(1-t/T))0.48(4(t/T)(1-(t/T))0.59060012001800Time(sec)0150300Activity <x(t)>TABC Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
studied using this method. The best fit disagrees
with the empirical functions near their peak, the
latter being flatter, likely an effect related to satu-
ration observed in long events.
Finally, we turn to discuss simple null models.
We consider two extreme cases, in both of them
the raw time series are randomly shuffled to remove
serial correlations. In the first case, we remove all
temporal correlations by randomly reordering x(n),
thus attaining a flat power spectral density. After
repeating the above analysis in this surrogate data
set, it becomes clear (as shown in Fig. 4) that the
scale invariance is absent in all the statistics un-
der study: size S, waiting time Wt and duration
T of events (note that the distributions are here
plotted using a logarithmic-linear scale). Results
in Panel B show that µ + 1 = 1, thus µ = 0, imply-
ing that there is not collapse, because with T µ = 1
in Eq. (5), the amplitude of the individual events
remains invariable. To consider the second case,
we need first to reorder randomly the time series of
increments I(n) and then proceed to integrate the
increments. Since each increment is now a random
variable, the power spectral density for this surro-
gate process obeys f β with β = −2 , and as shown
analytically by Baldassari et al.
[13], for this case
µ = 1/2 and the scaling function is a semicircle.
Please note that the fluctuations of human activity
described here differ from a simple auto-regressive
process: indeed successive increments I(n) are anti-
correlated and the power spectral density corre-
sponds to non-trivial power law correlations (i.e.,
β (cid:54)= −2).
Figure 4: Scaling is absent in a null model result-
ing from defining events after randomly reordering
the time series x(n). Panel A: Density distribu-
tions (CCDF) for event duration, size, and waiting
time. All the distributions are exponential (note
the logarithmic-linear scale). Panel B: The ex-
pected average size for a given duration in the null
model is a linear function of T (the dashed line rep-
resents the fit with slope 1), therefore, µ = 0 and
there is no collapse.
IV. Discussion
data collapse is quite satisfactory, while the value of
the exponent (µ = 0.48) does not exactly match the
one predicted in Eq. (4), µ = 0.59 (likely a conse-
quence of insufficient sampling). To determine the
generality of our results, we extended this analy-
sis to six other data sets. For each data set, the
value of µ was first determined. Subsequently, the
x(T, t) obtained from the events were rescaled with
T µ and their average computed. To account for
individual differences in mean activity, shape func-
tions were normalized by their mean value. The
results for the six datasets are presented in Panel
C of Fig. 3. They can be accurately described by
an inverted parabola, as in other systems previously
The present findings can be summarized by six styl-
ized facts describing bursts of human activity: I)
the spectral density of the time series of activity
x(t) obeys a power law, with exponent β ∼ 1;
II) successive increments I(n) are anti-correlated
with a spectral density obeying a power law with
exponent γ ∼ 1, which corresponds to a spec-
tral density for the raw data f β with β ∼ −1;
III) the PDF of the increments I(n) is definitely
non-gaussian; IV) the PDF of duration and sizes
of events obeys truncated power laws with expo-
nents 1 < τ < 2 and 1 < α < 2; V) the aver-
age size of the events scales with its lifetime T as
070017-5
102103T102103104<S> 020004000S,Wt,T10-410-310-210-1100CCDFSWtTAB Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
(cid:104)S(cid:105)(T ) ∼ T µ, where µ + 1 = (α − 1)/(τ − 1); VI)
the time series of individual events can be appropri-
ately rescaled via a transformation of its duration T
and amplitude x(t) onto a unique functional shape:
(cid:104)x(T, t)(cid:105) = T µfshape(t/T ).
We are aware that these observations are novel
only for human activity, because similar statistical
regularities of avalanching activity are well known
for a large variety of inanimate systems [9–12]. The
rescaling of the average shape is not surprising be-
cause, placed in the appropriate context, it can be
traced back to Mandelbrot's study of the fractal
properties of self-affine functions [21]. A curve or
a time series are said to be self-affine if a transfor-
mation can be found, such that rescaling their x, y
coordinates by k and kµ, respectively, and the vari-
ance in y is preserved (with µ = 1 corresponding
to self-similarity). In that sense, the successful col-
lapse of the events shape is a trivial consequence of
the overall self-affinity of the x(t) time series.
Thus, it is clear that the existence of the scal-
ing uncovered here is not informative per se of the
type of mechanism behind: scale-invariance can be
constructed via different processes, ranging from
critical phenomena [15] to simple stochastic auto-
regressive dynamics [13, 20]. What is then the
mechanism by which the above six facts are gen-
erated?
It seems that this question cannot be easily an-
swered by the type of experiments reported here.
Fluctuations of this type could have either an in-
trinsic (i.e., brain-born) origin but also could be
the reflection of a collective phenomena (including
humans and its environment). In either case, the
correlations observed seem to reject the case of in-
dependent random events starting and stopping hu-
man actions, because neither the distribution of the
increments I(n), nor the exponents match the case
of a random walk. In terms of brain-born process,
it is hard to accept some of the implications of the
scaling function in the activity shape. The average
parabolic shape means that the very beginning of
the motion activity contains information about how
long the activity will last, in the same sense that the
initial trajectory of a projectile predicts when and
where it will land. This proposal is hardly realistic,
because there is hardly a reasonable physiological
argument in support of any motor planning for the
length of time we are observing (∼ 103 secs). In
terms of collective processes, the results here sug-
gest that the interaction with other humans could
determine when and where, on the average, we start
and stop moving.
Despite our current relative ignorance, a possi-
bility that sounds interesting is to determine in
children, as they grow, if their behavioral product
of parental (and otherwise) education are reflected
in the shape of their individual scaling function.
This seems reasonable given the fact that "tireless
running around" is almost a definition of early age
well-being, which gives way to less hectic activity
as children mature. In the same line of thoughts,
if changes in the scaling function can be quantita-
tively traced to behavioral changes, one could also
consider to explore applications of these techniques
to monitor eventual progress in the treatment of
hyperactivity disturbances such as in the subjects
affected by the Attention Deficit Hyperactivity Dis-
order syndrome. The converse, i.e., cases in which
the average activity diminish, as in elderly subjects
shall be also explored. Further experiments and
analysis should shed light on these possibilities. In,
the meantime, the present results provide a guide
and six important constraints for the models that
should best capture the physics (and biology) of the
process.
Acknowledgements - Work supported by Na-
tional Science Center of Poland (ncn.gov.pl, grant
DEC-2011/02/A/ST1/00119); State Secretary for
Research and Development (grants PSI2010-19372
and PSI2013-48260) from Spain and by CONICET
from Argentina.)
[1] T Nakamura, K Kiyono, K Yoshiuchi, R Naka-
hara, Z R Struzik, Y Yamamoto, Universal
scaling law in human behavioral organization,
Phys. Rev. Lett. 99, 138103 (2007).
[2] T Nakamura, et al., Of mice and men - uni-
versality and breakdown of behavioral organi-
zation, PLoS ONE 3, e2050 (2008).
[3] K Hu, P C Ivanov, Z Chen, M F Hilton, H
E Stanley, S A Shea, Non-random fluctuations
and multi-scale dynamics regulation of human
activity, Physica A 337, 307 (2004).
070017-6
Papers in Physics, vol. 7, art. 070017 (2015) / D. Chialvo et al.
[4] L A N Amaral, D J B Soares, L R da Silva, L S
Lucena, M Saito, H Kumano, N Aoyagi, Y Ya-
mamoto, Power law temporal auto-correlations
in day-long records of human physical activ-
ity and their alteration with disease, Europhys.
Lett. 66, 448 (2004).
[5] C Anteneodo, D R Chialvo, Unravelling the
fluctuations of animal motor activity, Chaos
19, 033123 (2009).
[6] K Christensen, D Papavassiliou, A de
Figueiredo, N R Franks, A B Sendova-Franks,
Universality in ant behaviour, J. R. Soc. Inter-
face 12, 20140985 (2014).
[7] A Proekt, J Banavar, A Maritan, D Pfaff,
Scale invariance in the dynamics of sponta-
neous behavior, Proc Natl Acad Sci USA 109,
10564 (2012).
[8] J K Ochab, et al., Scale-free fluctuations in be-
havioral performance: Delineating changes in
spontaneous behavior of humans with induced
sleep deficiency, PLoS ONE 9, e107542 (2014).
[9] L Laurson, X Illa, S Santucci, K T Tallak-
stad, K J Maloy, M J Alava, Evolution of the
average avalanche shape with the universality
class, Nature Comm. 4, 2927 (2013).
[10] S Papanikolaou, F Bohn, R L Sommer, G
Durin, S Zapperi, J P Sethna, Universality
beyond power laws and the average avalanche
shape, Nature Phys. 7, 316 (2011).
[11] J P Sethna, K A Dahmen, C R Myers, Crack-
ling noise, Nature 410, 242 (2001).
[12] N Friedman, S Ito, B A W Brinkman, M Shi-
mono, R E L DeVille, K A Dahmen, J M
Beggs, T C Butler, Universal critical dynam-
ics in high resolution neuronal avalanche data,
Phys. Rev. Lett. 108, 208102 (2012).
[13] A Baldassarri, F Colaiori, C Castellano, Av-
erage shape of a fluctuation: Universality in
excursions of stochastic processes, Phys. Rev.
Lett. 90, 060601 (2003).
[14] B D Malamud, D L Turcotte, Self-affine time
series: I. Generation and analyses, Adv. Geo-
phys. 40, 1 (1999).
[15] P Bak. How nature works. The science of self-
organized criticality, Copernicus, New York
(1996).
[16] B Gutenberg, C F Richter, Magnitude and en-
ergy of earthquakes, Ann. Geofis. 9, 1 (1956).
[17] B Wang, J Kuo, S Granick, Burst of active
transport in living cells, Phys. Rev. Lett. 111,
208102 (2013).
[18] L Laurson, M J Alava, 1/f noise and avalanche
scaling in plastic deformation, Phys. Rev. E
74, 066106 (2006).
[19] A Clauset, C R Shalizi, M E J. Newman,
Power-law distributions in empirical data,
SIAM Rev. 51, 661 (2009).
[20] F Colaiori, A Baldassarri, C Castellano, Aver-
age trajectory of returning walks, Phys. Rev.
E 69, 041105 (2004).
[21] B B Mandelbrot, Self-affine fractals and frac-
tal dimension, Physica Scripta 32, 257 (1985).
070017-7
|
1812.00278 | 2 | 1812 | 2019-04-26T05:13:15 | Interpreting Encoding and Decoding Models | [
"q-bio.NC"
] | Encoding and decoding models are widely used in systems, cognitive, and computational neuroscience to make sense of brain-activity data. However, the interpretation of their results requires care. Decoding models can help reveal whether particular information is present in a brain region in a format the decoder can exploit. Encoding models make comprehensive predictions about representational spaces. In the context of sensory systems, encoding models enable us to test and compare brain-computational models, and thus directly constrain computational theory. Encoding and decoding models typically include fitted linear-model components. Sometimes the weights of the fitted linear combinations are interpreted as reflecting, in an encoding model, the contribution of different sensory features to the representation or, in a decoding model, the contribution of different measured brain responses to a decoded feature. Such interpretations can be problematic when the predictor variables or their noise components are correlated and when priors (or penalties) are used to regularize the fit. Encoding and decoding models are evaluated in terms of their generalization performance. The correct interpretation depends on the level of generalization a model achieves (e.g. to new response measurements for the same stimuli, to new stimuli from the same population, or to stimuli from a different population). Significant decoding or encoding performance of a single model (at whatever level of generality) does not provide strong constraints for theory. Many models must be tested and inferentially compared for analyses to drive theoretical progress. | q-bio.NC | q-bio | Interpreting encoding and decoding
models
Nikolaus Kriegeskortea & Pamela K. Douglasb
a Department of Psychology, Department of Neuroscience, Department of Electrical Engineering, Zuckerman Mind
Brain Behavior Institute, Columbia University, New York, NY
b Center for Cognitive Neuroscience, University of California, Los Angeles, CA; Modeling, Simulation, Computer
Science, UCF, USA
Abstract
Encoding and decoding models are widely used in systems, cognitive, and computational
neuroscience to make sense of brain-activity data. However, the interpretation of their results
requires care. Decoding models can help reveal whether particular information is present in a
brain region in a format the decoder can exploit. Encoding models make comprehensive
predictions about representational spaces. In the context of sensory experiments, where stimuli
are experimentally controlled, encoding models enable us to test and compare brain-
computational theories. Encoding and decoding models typically include fitted linear-model
components. Sometimes the weights of the fitted linear combinations are interpreted as
reflecting, in an encoding model, the contribution of different sensory features to the
representation or, in a decoding model, the contribution of different measured brain responses
to a decoded feature. Such interpretations can be problematic when the predictor variables or
their noise components are correlated and when priors (or penalties) are used to regularize the
fit. Encoding and decoding models are evaluated in terms of their generalization performance.
The correct interpretation depends on the level of generalization a model achieves (e.g. to new
response measurements for the same stimuli, to new stimuli from the same population, or to
stimuli from a different population). Significant decoding or encoding performance of a single
model (at whatever level of generality) does not provide strong constraints for theory. Many
models must be tested and inferentially compared for analyses to drive theoretical progress.
Highlights
1. Decoding models can reveal whether particular information is present in a brain region in
a format the decoder can exploit.
2. Encoding models make comprehensive predictions about representational spaces and
more strongly constrain computational theory.
3. The weights of the fitted linear combinations used in encoding and decoding models are
not straightforward to interpret.
4. Interpretation of encoding and decoding models critically depends on the level of
generalization achieved.
1
5. Many models must be tested and inferentially compared for analyses to drive theoretical
progress.
Encoding and decoding: concepts with caveats
The notions of encoding and decoding are rooted in the view that the brain processes
information. Information about the world enters our brains through the senses, is processed and
potentially stored, and informs our behavior over a large range of time scales. To understand
brain function, then, we must understand the information processing: what information is
processed, in what format it is coded in neural activity, and how it is re-coded across stages of
processing, so as to ultimately contribute to behavior.
The view of the brain as an information-processing device (or, equivalently, a computational
device) is closely linked to the concept of representation. A pattern of neural activity represents
information about the world if it serves the function to convey this information to downstream
regions, which use it to produce successful behavior (DeCharms & Zador 2000).
When we talk about encoding and decoding, we focus on a particular brain region X whose
function we are trying to understand. To simplify the problem, we divide the process into the
encoder (which produces the code in region X from the sensory signals) and the decoder (which
uses the code in region X to enable successful behavior). This bipartite division often provides a
useful simplification. However, we have to consider three caveats to avoid confusion.
First, encoder and decoder do not inherently differ with respect to the nature of the processing.
Where the encoder ends and the decoder begins depends on our region of interest X. When we
move our focus to the next stage of representation in region Y, the processing between X and
Y, which was part of the decoder, becomes part of the encoder. Whether a particular processing
step in the brain is to be considered encoding or decoding, thus, is in the eye of the beholder.
Second, our region of interest X may not be part of all causal paths from input to output. The
division about region X into encoder and decoder, then, misses a portion of the causal graph.
Brain regions do not in general form chains of processing stages without skipping connections
or recurrent signaling. The primate visual hierarchy is a case in point, where cortical areas
interact in a network with about a third of all possible pairwise inter-area connections (Felleman
& Van Essen 1991, Young 1992). Encoding and decoding models are nevertheless useful in
this context, providing a partial view of the encoded information and its format.
Third, the terms encoding and decoding suggest that each is the inverse of the other: The
encoder maps stimuli to brain responses; the decoder maps brain responses to stimuli.
However, if we would like to interpret our encoding and decoding models as models of brain
computation, then both encoder and decoder ought to operate in the causal direction. The
decoder should not map back to stimuli, but on to motor representations. An encoding model
might predict neural activity elicited by images (Wu et al. 2006, Kay et al. 2008, Wen et al.
2018). A decoding model might attempt to predict category labels from neural activity, thus
explicating information that might be reflected in a behavioral response (Majaj et al. 2015).
The notion that encoding is followed by its inverse begs the question why the original input is
not simply copied. A possible motivation for a sequence of an encoder and its inverse is to
temporarily compress the information, e.g., for passage through a bottleneck such as the optic
nerve. A more general notion is that information is re-coded through a sequence of stages,
possibly compressing it for more efficient representation (Simoncelli & Olshausen 2001), but
2
also discarding unneeded information and converting needed information into a format that
enables it to be exploited to control behavior (Tishby et al. 2000, Achille & Soatto 2018). When
decoding models are conceptualized as inverse encoders (mapping back to the stimuli, rather
than on to downstream representations), they cannot be interpreted as process models of brain
function. However, they can still be useful tools, revealing information present in a brain region
and giving us clues about the format of the code. For a careful discussion of the interpretation of
causal and anti-causal encoding and decoding models, see Weichwald et al. (2015).
Encoding and decoding are important concepts in both theoretical and experimental
neuroscience. This paper addresses experimentalists and focuses on the interpretation of
empirical results obtained by fitting encoding and decoding models to brain-activity data.
Decoding models: revealing information and its format
Decoding is sometimes described as "reading the brain" or "cracking the neural code" (Cox &
Savoy 2003, Haynes & Rees 2006, Tong & Pratte 2012). The underlying concept is that of
decoding as inverse encoding, where the goal is not to model brain information processing, but
to reveal the content of the code. The sense of revealing a mystery and gaining insight into
hidden information drives the imagination of scientists and lay people alike.
The intuition that something important has been learned when we can decode some perceptual
or higher cognitive content from brain activity is correct. However, "decrypting" the brain should
not be equated with understanding its computational mechanism. Decoding reveals the
products, not the process of brain computation (Box 1). However, it is a useful tool for testing
whether a brain region contains a particular kind of information in a particular format (Haxby et
al. 2001, Cox & Savoy 2003, Carlson et al. 2003, Kamitani & Tong 2005, Haynes & Rees 2006,
Kriegeskorte et al. 2006, Norman et al. 2006, Paninski et al. 2007, Mur et al. 2009, Pereira et al.
2009, Pillow et al. 2011, Kriegeskorte 2011, Tong & Pratte 2012, Hebart et al. 2015, Haynes
2015, Hebart & Baker 2017, Varoquaux et al. 2017, Paninski & Cunningham 2017).
A simple example of a decoding model is a linear classifier (Duda et al. 2012) that takes a
measured brain-activity pattern as input and outputs a class label (say "cat" and "dog"),
revealing which of two stimuli (say a particular image of a cat and a particular image of a dog)
elicited the response pattern (Figure 1). A linear classifier may achieve this by computing a
weighted sum of the measured responses and imposing a threshold to decide which label to
return (see Mur et al. 2009 for an introduction in the context of neuroimaging). The weights are
set to maximize the accuracy of the decoding on a training data set. If decoding succeeds
reliably on a test data set, then the region must contain information about the decoded variable
(the binary stimulus distinction here).
Decoding provides a statistically advantageous way of testing for stimulus
information
In the two-stimulus example, all the linear decoder shows is that the two images elicit distinct
response patterns. This means that there is mutual information between stimulus and response.
To demonstrate mutual information, we could have used an encoding model (Naselaris et al.
2011) or a multivariate test of difference between the response patterns elicited by the two
3
stimuli (e.g. multivariate analysis of variance), instead of a decoding model. However, a
univariate encoding model would in general have less sensitivity, because it does not account
for the noise correlations between different response channels (Averbeck et al. 2006). A
multivariate analysis of variance would account for the noise correlations, but might have less
specificity. In fact, it might fail to control false-positives at the nominal level if its assumptions of
multivariate normality did not hold (as is often the case), making it invalid as a statistical test
(Kriegeskorte 2011). Decoding provides a natural approach to modelling the noise correlations
(e.g. using a multivariate normal model as in the Fisher linear discriminant), without relying on
the model assumptions for the validity of the test: Violations of the decoding model assumptions
will hurt decoding performance. We, thus, err on the safe side of concluding that there is no
information about the stimuli in the responses. In sum, it is not the direction of the decoding
model ("brain reading") that makes it a compelling test for information, but the statistical nature
of the problem (noise correlations) and the fact that decoders are tested on independent data.
Figure 1 Linear encoding and decoding models. (A) Encoding and decoding model the relationship
between stimuli and responses in opposite directions. An encoding model (top) predicts brain responses
as a linear combination of stimulus properties (black circles). A decoding model (bottom) predicts stimulus
properties as a linear combination of brain responses. (B) Example of a linear decoding model using a 2-
dimensional feature space consisting of two voxels. Voxel 1 contains relevant information about which of
two classes (green, blue) the stimulus belongs to. Voxel 2 contains no information about the stimulus class.
The two dimensions jointly define the linear discriminatory boundary. Note that the weights assigned to
each voxel are defined by the vector w, which is orthogonal to the decision boundary. Because the noise is
correlated between the voxels, a linear decoder will assign significant negative weight to voxel 2, using this
voxel (which contains only noise) to cancel the noise in the voxel which contains signal. As a result,
interpreting the absolute weights of linear decoders requires care and additional analyses.
Linear decodability indicates "explicit" information
For decoding to succeed, the information must be present in the brain region in a format that the
decoder can exploit. Linear decoders, the most widely used class, require that the distributions
of patterns be linearly separable to some extent. This is a weakness in that we might fail to
detect information encoded in a more complex format. However, it is a strength in that it
provides clues to the format of the information we do detect. The simpler the decoder, the more
focused its sensitivity will be. From the perspective of understanding the brain computations, it
4
is attractive to use decoding operations that single neurons can implement. These include linear
readout, but also simple nonlinear forms of readout such as radial basis function decoding
(Poggio & Girosi 1990). Linear decodability indicates that a downstream neuron receiving input
from a sufficient portion of the pattern, could read out the information in question (DiCarlo & Cox
2007). Information amenable to direct readout by downstream neurons is sometimes referred to
as "explicit" in the code (Kriegeskorte 2011, DiCarlo et al. 2012, Hong et al. 2016).
The level of generalization beyond the training set must be considered when
interpreting a decoding result
Fitting a model always poses the risk of overfitting, i.e. of optimizing the fit to the training data at
the expense of predictive performance on independent data. Overfitting can lead to high
decoding accuracy on the training set, even if the response patterns contain no information
about the stimulus. Decoders therefore need to be tested for generalization to independent data
(Hastie et al. 2009, Kriegeskorte 2015). In our example, we might test the decoder on an
independent set of response measurements for the same two particular images of a cat and a
dog. If decoding accuracy on this independent test set is significant, we can reject the null
hypothesis that the response patterns contain no information about the stimulus (Mur et al.
2009, Pereira et al. 2009).
However, detecting information about which of two images has been presented tells us almost
nothing about the nature of the code. The two images must have distinct response patterns in
the retina and V1 (low-level representations) as well as in the visuo-semantic regions of the
ventral stream (high-level representations). We would therefore expect a linear decoder to work
on new measurements for the same images in any of these regions. This reflects the fact that all
the regions contain image information. In the retina, for example, we expect the two images to
elicit distinct response patterns, while the manifolds of response patterns corresponding to the
two categories are hopelessly entangled (DiCarlo & Cox 2007, Chung et al. 2016, Chung et al.
2018).
Given responses to just two images, we can demonstrate the presence of information, but have
no empirical basis for characterizing the nature of the code (see Kriegeskorte et al. 2007 for a
study limited by this drawback). In order to learn whether the region we are decoding from
contains a low-level image encoding or a high-level categorical encoding, we can train the
decoder on one set of cat and dog images and test it on another set of images of different cats
and dogs (Anzellotti et al. 2014, Freedman et al. 2002, Kriegeskorte 2011).
To support the interpretation that "cats" and "dogs" are linearly separable in the representation
(rather than the weaker claim that there is image information), it is not sufficient to increase the
number of particular images of cats and dogs, while training and testing on the same images.
The linear decoder has many parameters (one for each response channel) and is expected to
overfit even to a larger set of particular images. Even for the retinal representation, we therefore
expect a cat/dog decoder to generalize to new measurements performed on the same images.
We must test the linear decoder for generalization to different cats and dogs.
Note, however, that interpreting linear decodability as linear separability of the two classes in
the neuronal representational space would further require the decoding accuracy to be so high
that errors can be attributed to the measurement noise rather than the neural representation. In
practice, we typically face ambiguity. For example, decoding accuracy may be significantly
5
above chance, but far from perfect. This indicates that the code contains some linearly
decodable information, but claims of linear separability may be difficult to evaluate as it would
require attributing the substantial proportion of errors to limitations of the measurements (noise
and subsampling), rather than to a lack of linear separability of the neuronal activity patterns.
If we managed to show that cats and dogs fall in linearly separable regions in representational
space, there would still be a question about the nature of the features that support the encoding.
There may be many different visual features that can be computed from images and that lend
themselves to separating cats from dogs (e.g. the shape of the ears or the shape of the eyes).
Revealing which particular features the encoding in a given brain region employs would require
further experiments. For example, we could design artificial stimuli that contain subsets of the
distinctive features of cats and dogs. We could then test whether a linear decoder trained on
responses to natural images of cats and dogs generalizes to responses to the artificial stimuli.
Successful decoding would support the hypothesis that the brain region employs some of the
features present in the artificial images.
More generally, we can go beyond using different exemplars from the same categories in the
test set. Testing decoders for cross-generalization between different domains is useful to
address a host of questions about the invariances of the encoding, for example its consistency
between imagery and perception (Stokes et al. 2009, Cichy et al. 2012) and its stability across
latencies after stimulus onset (King & Dehaene 2014).
Stimulus reconstruction provides a richer view of the information present, but
complex priors complicate the interpretation
A decoder may discriminate more than two categories. It might decode a continuous variable
(e.g. the orientation of the stimulus; Kamitani & Tong 2005), or it might output a
multidimensional description of the stimulus. The most ambitious variant of decoding is stimulus
reconstruction, where the output is a detailed estimated rendering of the stimulus. A decoder
providing a rendering of the stimulus can potentially reveal more of the information present in
the neuronal encoding (Thirion et al. 2006, Miyawaki et al. 2008, Mesgarani et al. 2009,
Naselaris et al. 2009, Santoro et al. 2017, Parthasarathy et al. 2017).
Stimulus reconstruction is a particularly impressive feat of decoding for two related reasons.
First, the space of outputs is much richer, so the decoder reveals more of the encoded
information (discriminating among more possible stimuli). Second, a more challenging level of
generalization is typically required for the decoder to discriminate among this richer set of
possible stimuli, because there are usually severe limitations on the number of stimuli that the
decoder can be directly trained on. Successful generalization suggests that the structure of the
model captures something about the code that enables generalization far beyond the training
set.
To the extent that stimulus reconstruction, when applied to novel stimuli that the model has not
been trained on, successfully specifies close matches within a rich space of reconstruction
targets (e.g. pixel images), the analysis indicates that the code contains rich information about
the stimulus. An extreme example of this would be a decoder that can precisely reconstruct
arbitrary natural or unnatural images from their representation in V1. This would indicate that V1
encodes the image precisely and that the encoding is not restricted to natural images. In fact,
6
this has never been shown for V1 and would be puzzling, because we expect visual
representations to be specialized for natural stimuli.
Box 1: Decoding Models: Benefits and Limitations
Decoding provides an intuitive and compelling demonstration of the presence of information in a
brain region. Decoding models bring several benefits:
• They enable researchers to look into the regions and reveal whether particular information is
encoded in the responses. Encoded information might be used by downstream neurons (i.e.
the encoding might serve as a representation). Training and testing decoders has helped
move the focus from single-neuron activity and regional-mean activation to the information
being processed, making analyses more relevant to the computational function in question.
• They enable cognitive neuroscience to exploit the fine-grained multivariate information
present in modern imaging data. Before the advent of decoding, dominant brain mapping
techniques required smoothing of the data, treating fine-grained information as noise. The
multivariate modelling of local patterns also brought locally multivariate noise models to the
literature, which greatly enhance the information that can be drawn from neuroimaging data.
• Decoding analyses use independent test data to assess the presence of information.
Demonstrating significant information with independent test data ensures that violations of
model assumptions lead to conservative errors (making significant results less likely). This is
an advantage over methods (such as univariate and multivariate linear regression) that use
distributional assumptions rather than independent test sets to perform statistical inference.
Decoding models are limited in several ways:
• They cannot in general be interpreted as models of brain-information processing. In the
context of sensory systems, decoding models operate in the wrong direction, capturing the
inverse of the brain's processing of the stimuli. If the decoded variable is a behavioral
response, then direction of information flow matches between model and brain. However,
such applications are rare. Moreover, the most successful applications of decoding in the
literature rely on linear decoders, which are useful to reveal explicit information, but cannot
capture the complex nonlinear computations we wish to understand.
• Decoder weight maps are difficult to interpret (Figure 3) because voxels uninformative by
themselves can receive large weights when they help cancel noise and because the weights
are codetermined by the data and the prior, and the fitting procedure might select one but
not the other of two essentially equally informative voxels.
• Decoding results provide only weak constraints on computational mechanisms. Knowing
that a particular kind of information is present provides some evidence for and against
alternative theories about what the region does. However, it does not exhaustively
characterize the representation, let alone reveal the underlying computations.
In order to improve the appearance of reconstructions of natural stimuli, it is attractive to use a
prior over the possible outputs of the decoder. For example, we might constrain the model to
output an image whose low-level or high-level statistics match natural images. We may
constrain the decoder more strongly, by limiting it to output only actual natural images. This
constraint could be implemented using a restricted set of target images (Naselaris et al. 2009)
or a generative model of natural images (Goodfellow et al. 2014). In either case, the quality of
the reconstruction will reflect a combination of the information provided by the brain region and
the information contributed by the prior. A complicated prior therefore also complicates the
7
interpretation of the reconstruction results: good looking reconstructions no longer indicate that
all the detail they provide is encoded in the brain region. The reconstruction has to be compared
to the presented stimulus, and the complexity of the output space (which is reduced by the prior
over the outputs) needs to be considered in the interpretation.
An important question is what we can learn from stimulus reconstructions. The goal to learn
about the content and format of the code may not be ideally served by striving for the most
natural looking reconstruction.
Decoding models predicting behavioral responses from brain responses can be
interpreted as brain-computational models
Decoding is usually used as a tool of analysis that reveals aspects of the content and format of
the information encoded in a brain region. The decoding model, thus, is not interpreted as a
model of brain computation. In the context of sensory systems, a decoder maps from brain
responses to stimuli. Since stimulus processing by the brain operates in the opposite direction, it
is difficult to interpret a decoder as a model of brain information processing. However, if a
decoder is used to predict behavioral responses, e.g. judgments of categorical or continuous
stimulus variables (possibly including errors and reaction times on individual trials), then the
decoder can be interpreted as a model (at a high level of abstraction) of the brain computations
generating the behavioral responses from the encoding of the stimuli in the decoded brain
region (Shadlen et al. 1996, Williams et al. 2007, Walther et al. 2012).
Encoding models: testing comprehensive representational
predictions
Encoding models attempt to predict brain response patterns from descriptions of the
experimental conditions (Figure 1A; Paninski et al. 2007, Kay et al. 2008, Dumoulin & Wandell
2008, Mitchell et al. 2008, Naselaris et al. 2011, Khaligh-Razavi & Kriegeskorte 2014, Yamins et
al. 2014, Naselaris & Kay 2015, van Gerven 2017). Encoding models, thus, operate in the
opposite direction as decoding models.
If our goal is merely to demonstrate that a brain region contains information about the
experimental conditions, then the direction the model should operate in is a technical issue: One
direction may be more convenient for capturing the relevant statistical dependencies (e.g. noise
correlations among responses), but a model operating in either direction could support the
inference that particular information is present in the code. If our goal is to test computational
theories, however, then the direction that the model operates in matters, because it determines
whether the model can be interpreted as a brain-computational model.
Encoding models predicting brain responses from sensory stimuli can serve as
brain-computational models
Whereas a decoding model typically serves to test for the presence of particular information in a
brain region, an encoding model can provide a process model, at some level of abstraction, of
the brain computations that produce the neuronal code. An encoding model makes
8
comprehensive predictions about the representational space (Naselaris & Kay 2015,
Diedrichsen & Kriegeskorte 2017; Box 2) and, ideally, should fully explain neuronal responses in
the region in question (up to the maximum achievable given the noise in the data).
In sensory neuroscience, the experimental conditions correspond to sensory stimuli, and so an
encoding model maps sensory stimuli to their encodings in sensory regions of the brain (Kay et
al. 2008, Mitchell et al. 2008, Dumoulin & Wandell 2008, Huth et al. 2012). A sensory encoding
model, thus, operates in the direction of the information flow: from stimuli to brain responses. It
should take raw sensory stimuli as input (e.g. images or sounds; Kay et al. 2008, Naselaris et al.
2011, Khaligh-Razavi & Kriegeskorte 2014, Yamins et al. 2014, Kell et al. 2018; for recent
reviews see Kriegeskorte 2015, Yamins & DiCarlo 2016) and predict the patterns of brain
activity the stimuli elicit. We can adjudicate among computational theories by implementing
them in encoding models and testing their ability to predict brain-activity patterns (Kriegeskorte
& Diedrichsen 2016).
Brain-computational encoding models can be tested by predicting raw
measurements, representational dissimilarities, or the activity-profiles distribution
A brain-computational encoding model could consist in a set of hand-engineered features
computed from the stimuli or in a neural network model trained on some task. How can we
evaluate a high-dimensional representation in a brain-computational encoding model with brain-
activity data, when the units of the model may not correspond one-to-one to the measured
channels of brain activity?
One approach is to predict the raw measurements (Kay et al. 2008, Dumoulin & Wandell et al.
2008, Mitchell et al. 2008, Naselaris et al 2011, van Gerven 2017). To achieve this, we can fit a
linear model that maps from the outputs of the nonlinear encoder thought to capture the brain
computations (e.g. a layer of a neural network model processing the stimuli) to each measured
response channel, e.g. each neuron or fMRI voxel in the region we would like to understand.
The linear model will capture which units of the nonlinear encoding model best predict each
response channel. It can be interpreted as capturing the sampling and averaging that occurs in
the measurement of brain activity. A neuronal recording array, for example, will capture a
sample of neurons, and fMRI will give us measurements akin to local spatiotemporal averages
at regular grid locations. For each response channel, the linear model will have a parameter for
each nonlinear feature (e.g. each unit of the neural network layer). Fitting these parameters
requires substantial training data and normally also a prior on the parameters (e.g. a 0-mean
Gaussian prior as is implicit to fitting with an L2 penalty).
Instead of predicting the raw measurements, we can use a brain-computational encoding model
to predict to what extent different stimuli are dissimilar in the representation, an approach called
representational similarity analysis (RSA; Kriegeskorte et al. 2008, Nili et al. 2014). The pairwise
dissimilarities of the multivariate response patterns representing the stimuli can be summarized
in a representational dissimilarity matrix (RDM). For example, for each pair of stimuli, the
dissimilarity between the associated response patterns could be measured using the Euclidean
distance. The RDM characterizes the geometry of the set of points in the multivariate response
space that correspond to the stimuli. Noise correlations can be captured by estimating the
covariance matrix of the residuals of the response-estimation model, and replacing the
Euclidean distance by the Mahalanobis distance. To remove the positive bias associated with
9
measuring distances between noisy data points, we can use the crossnobis (crossvalidated
Mahalanobis) estimator (Nili et al. 2014, Walther et al. 2016). The resulting crossnobis RDM
provides a full characterization of the linearly decodable information in the representational
space (Diedrichsen & Kriegeskorte 2017). Comparing representations in models and brains at
the level of RDMs obviates the need for fitting a linear model to predict each measured
response (thus reducing the need for training data) and enables the analysis to naturally handle
noise correlations between responses (which are typically ignored when encoding models
separately predict each of the measured response channels).
A third approach to the evaluation of encoding models is to predict the distribution of activity
profiles. In pattern component modelling (PCM, Diedrichsen et al. 2011), this distribution is
characterized by the second moment of the activity profiles. Like the RDM, this is a stimulus-by-
stimulus summary statistic of the stimulus-response matrix. Each entry of the second-moment
matrix corresponds to the inner product between two response patterns.
All three approaches can be construed as testing hypotheses about the representational space
induced by the activity profiles (Diedrichsen & Kriegeskorte 2017). Consider a linear encoding
model using a Gaussian prior on the weights. Such a model predicts a Gaussian distribution of
activity profiles. The predicted distribution of activity profiles is captured by its second moment.
For representational similarity analysis, similarly, the RDM is a function of the second moment
of the activity profiles. This core mathematical commonality between the methods
notwithstanding, each is best suited for a particular set of questions.
Linear models predicting raw measurements lend themselves to univariate brain mapping,
revealing which voxels or neurons are accounted for by a particular nonlinear encoding model.
RSA lends itself to characterizing the geometry of the representational space, naturally handles
noise correlations among responses, and reduces the need for training data. PCM can have
greater sensitivity for adjudicating among models than the other two methods, at the expense of
relying on stronger assumptions. The three methods are best viewed as part of a single toolbox
of representational model analyses, whose elements can be combined as needed to address
particular questions.
The level of generalization beyond the training set must be considered when
interpreting an encoding result
Encoders, like decoders, are tested by evaluating how well they predict independent data,
whether the predicted quantities are the raw brain-activity measurements, the representational
dissimilarity matrix, or the second-moment matrix of the activity profiles. For encoders, as for
decoders, the interpretation depends on both the prediction accuracy and the level of
generalization beyond the training set that the model achieves.
Encoding models typically require the fitting of parameters, so overfitting needs to be accounted
for in any inferential procedure. In the simplest type of a univariate linear encoding model, we
can rely on Gaussian assumptions and perform inference without a separate test set (e.g.
Friston et al. 1994). However, more interesting models require independent test sets, for
example when parameters are fitted using priors over the weights and when the model is a
brain-computational model to be tested for generalization to new conditions.
10
A key consideration is how much flexibility to allow in fitting each model representation to a
brain representation. One extreme is to allow no flexibility and assume that the model
representation precisely predicts the geometry of the representational space (Kriegeskorte et al.
2008). This case is most naturally handled by RSA and PCM, but could also be implemented
with linear encoders by using a prior that prevents any distortion of the representational
geometry. The other extreme is to allow arbitrary linear remixing of the units of the nonlinear
encoding model. This case is most naturally handled with linear encoding models, but can also
be implemented with PCM and RSA (Diedrichsen et al. 2011, Khaligh-Razavi & Kriegeskorte
2014, Khaligh-Razavi et al. 2016). In practice, some compromise is desirable, which we can
think of as a prior on the mapping from the brain-computational model to the measured brain
responses. We might use a 0-mean Gaussian prior on the weights (e.g. Kay et al. 2008).
Alternatively, we can limit flexibility more aggressively, by allowing each unit (or each feature
map or layer) a single weight (not a separate weight for each response). Such weighted
representational models (e.g. Khaligh-Razavi et al. 2014) are naturally implemented with RSA
and PCM. Each brain-computational model in this case predicts a superset of the features
spanning the brain representational space (disallowing linear mixtures), but does not predict the
prevalence of each of the features in the neuronal population.
The lowest level of generalization beyond the training set is generalization to new
measurements for the same experimental conditions. This is sufficient, if the experimental
conditions exhaustively cover the domain we would like to draw inferences about (consider the
case of the representation of the five fingers in motor cortex: Diedrichsen et al. 2012). However,
in a domain such a sensory systems, the goal is typically to evaluate to what extent a brain-
computational model can predict brain representations of arbitrary stimuli. This requires a higher
level of generalization beyond the training set. A vision model, for example, might be trained
with responses to one sample of natural images and tested for generalization to responses to
an independent (and nonoverlapping) sample from the same distribution of natural images.
Because the set of all natural images is so rich, this is a challenging generalization task (as
illustrated by the difficulty of computer vision). An even more stringent test of the assumptions
implicit to a model is to train the model on a sample from one population of images and test it on
a sample from a different population of images (e.g. Eickenberg et al. 2017).
The prediction accuracy can be assessed in terms of whether it is significantly above chance
level, whether it significantly differs from that of competing encoding models, and how close it
comes come to the noise ceiling (the highest achievable accuracy given the noise in the data,
Nili et al. 2014).
We can generalize claims about an encoding model to the extent that its predictions generalize.
If we want to conclude that the model can predict responses for the stimuli presented, we need
not test the model with different stimuli (only with different response measurements elicited by
the same stimuli). If we want to conclude that the model can predict responses to arbitrary
natural stimuli, we need to test it with new arbitrary natural stimuli. The population of conditions
the test set is a sample of defines the scope of the claims we can make (Hastie et al. 2009,
Kriegeskorte 2015).
We focus here on encoding and decoding models that are fitted to individual subjects' brains, so
as be able to exploit fine-grained idiosyncratic patterns of activity that are unique to each
subject. Within-subject prediction accuracy may support generalization to a population of stimuli,
but it doesn't support generalization to the population of subjects. In some fields, such as low-
11
level vision, researchers draw on prior knowledge and assume that results that hold for a few
subjects will hold for the population. If we were instead to use our data to generalize our
inferences to the population of subjects, we would either need to predict results for held-out
subjects (which would make it impossible to exploit individually unique fine-grained activity
patterns) or perform inference on the within-subject prediction accuracies with subject as a
random effect.
The feature fallacy: Interpreting the success of a model as confirmation of its
basis set of features
Linear encoding models predict each measured activity profile as a linear combination of a set
of model features. When a model can explain the measured activity profiles, we might be
tempted to conclude that the model features are encoded in the measured responses. This
interpretation is problematic because the same linear space can be spanned by many
alternative basis sets of model features.
The fact that multiple sets of basis vectors can span the same space is widely appreciated.
However, it is not obvious whether the ambiguity is removed when a prior over the encoding
weights is used. An encoding model with a 0-mean isotropic Gaussian prior on the weights
(equivalent to ridge regression) predicts a Gaussian distribution of activity profiles (captured by
the second moment of the activity profiles as the sufficient statistic) and a particular
representational geometry (captured by the RDM). The use of a weight prior does reduce the
ambiguity. In the absence of a weight prior, all models spanning the same space of activity
profiles make equivalent predictions. With a weight prior, different models spanning the same
space can predict distinct distributions of activity profiles. However, substantial ambiguity
remains (Figure 2) because there are infinitely many alternative ways to express the same
distribution of activity profiles by different feature sets (each assuming a 0-mean isotropic
Gaussian prior over the weights).
The freedom to choose among feature sets generating the same distribution of activity profiles
is useful in that it enables us to implement a nonisotropic Gaussian prior on the weights of a
given model (Tikhonov regularization) by re-expressing the model such that the features induce
the same distribution of activity profiles in combination with a 0-mean isotropic Gaussian prior
(Nunez-Elizalde et al. 2018). However, this freedom to express the same model by different
feature sets needs to be kept in mind when interpreting results: Many alternative feature sets
would have given the same results.
In sum, the same distribution of activity profiles can be expressed in many ways (by different
feature sets), so the fact that a model explains brain responses does not provide evidence in
favor of the particular feature set chosen. Diedrichsen (2018) termed the interpretation in terms
of the particular features used the "feature fallacy". This fallacy is arguably somewhat persistent
in the neuroscience literature (Churchland et al. 2012).
12
Figure 2 The feature fallacy. Different linear encoding models spanning the same space of activity
profiles may not be distinguishable. There are many alternative sets of feature vectors {f1, f2, …} that span
the same space of activity profiles. In the absence of a prior on the weights of the linear model, all these
sets can equally explain a given set of brain responses. The ambiguity is reduced, but not resolved when
a prior on the weights is assumed (Diedrichsen et al. 2018). If we define a prior on the weights, then each
model predicts a probability density over the space of activity profiles. This probabilistic prediction may be
distinct for two sets of basis features, even if they span the same space. For example, if the weight prior is
a 0-mean isotropic Gaussian, then each model assigns probabilities to different activity profiles according
to a Gaussian distribution over the space of profiles. Two linear models may span the same space, but
predict distinct distributions of activity profiles. However, even with a Gaussian weight prior, there are still
(infinitely) many equivalent models that make identical probabilistic predictions. We illustrate this by
example. (a) Three models (A, B, C) each contain two feature vectors as predictors (A: {fA1, fA2}, B: {fB1, fB2},
C: {fC1, fC2}). The three models all span the same 2-dimensional space of activity profiles. For each model,
we assume a 0-mean isotropic Gaussian weight prior. (b) All three models predict the same nonisotropic
Gaussian probability density over the space of activity profiles (indicated by a single iso-probability-density
contour: the ellipse). Model A (gray) predicts the density by modeling it with two orthogonal features that
capture the principal-component axes, with features having different norms to capture the anisotropy. Model
B predicts the same density by modeling with two correlated features of similar norm. Model C falls
somewhere in between, combining feature correlation and different feature norms to capture the same
Gaussian density over the activity profiles. Note that there are many other models that span the same
space, but will not induce the same probability density over activity profiles when complemented by a 0-
mean isotropic Gaussian weight prior. A given linear encoding model's success at predicting brain
responses provides evidence for the induced distribution of activity profiles, but not for the particular
features chosen to express that distribution.
13
Box 2: Encoding Models: Benefits and Limitations
An encoding model predicts the response of each measurement channel (e.g. a neuron or fMRI
voxel) on the basis of properties of the experimental condition (e.g. sensory stimulus).
Continuous brain maps can be obtained by fitting such a model to each response in turn.
Responses are typically predicted as linear combinations of the features, rendering this
approach closely related to classical univariate brain mapping. However, encoding models have
several important benefits:
• More complex feature sets are used, often comprising thousands of descriptive features.
Fitting the models therefore requires regularization. Typically, models are fitted using a
penalty on the weights.
• Encoding models can have nonlinear components, such as the location and size of a visual
receptive field.
• Models are tested for generalization to different experimental conditions (e.g. a different
sample of visual stimuli).
• When sensory systems are studied, encoding models operate in the direction of information
flow. They can then be used as a general method for testing brain-computational models,
i.e. models that process sensory stimuli. A vision model, for example, can take a novel
image as input and predict responses to that image in cortex.
Analyses using encoding models to predict each response channel separately are limited in two
ways:
• The separate modeling of each response channel is inconsistent with the multivariate nature
of neuronal population codes and noise, and is statistically suboptimal. Separate modeling
of each response entails low power for testing and comparing models, for two reasons: (1)
Single responses (e.g. fMRI voxels) may be noisy and the evidence is not combined across
locations. (2) The analyses treat the responses as independent, thus forgoing the benefit
exploited by linear decoding approaches to model the noise multivariately. This is important
in fMRI, where nearby voxels have highly correlated noise. As spatial resolution increases,
we face the combined challenge of more and noisier voxels. This renders mapping with
proper correction for multiple testing very difficult. In addition, relating results between
individuals is not straightforward.
• When model parameters are color-coded and mapped across the cortex, the resulting
detailed maps are not straightforward to interpret. (1) Weight maps of linear models are
problematic to interpret in encoding models for the same reason they are problematic to
interpret for decoding models: because a predictor's weight depends on the other predictors
present in the model (unless each predictor is orthogonal to all others). The required
regularization further complicates weight interpretation. (2) Models are fit at many locations
and voxels highlighted on the basis of model fits. Post-selection inference on parameters is
not usually performed. Because of these complications, results of fine-grained mapping
across voxels are difficult to substantiate by formal inference.
Influential studies have met these challenges by focusing on single-subject analyses, acquiring
a large amount of data in each subject, and limiting formal inference with correction for multiple
testing to the overall explained variance of a model, with colors serving exploratory purposes.
14
Weights of linear models are not straightforward to interpret
in either encoding or decoding models
Beyond interpretation of the overall success of an encoding or decoding model, researchers
often want to dig deeper and interpret the fitted parameters of their models. In the context of a
decoding model, the weights assigned to the voxels would seem to tell us where the information
the decoder uses is located in the brain. Similarly, in the context of an encoding model, the
weights of different model features promise to reveal to what extent different model features are
encoded in a brain region.
Unfortunately, the interpretation of the weights of linear models is not as straightforward as the
simplicity of a linear relationship might suggest. A weight does not reflect the predictive power of
an individual predictor (a measured brain response in decoding, a model features in encoding).
Rather a weight reflects a predictor's contribution in the context of the rest of the model.
Uninformative predictors can receive large positive or negative weights. For example, an fMRI
voxel that does not by itself contain any information can have a large absolute weight in a linear
decoder if it improves decoding accuracy by cancelling noise that the voxel shares with other
voxels that do contain stimulus information (Figure 1B; Figure 3; Haufe et al. 2014). In an
encoding model, similarly, a model feature might contain no information about the modeled
response and still receive a large absolute weight. For example, an fMRI voxel that only
responds to faces might be explained by a set of semantic descriptors unrelated to faces if the
nonface semantic descriptors are capable, in combination, of capturing contextual variation that
is correlated with the presence of faces.
Conversely, informative predictors may receive weights that are small in absolute value or zero.
For example, in fMRI, a voxel might receive zero weight when other voxels suffice for decoding
when a weight penalty (especially a sparsity encouraging weight penalty, e.g. defined by the L1
norm of the weights) leads the fitting procedure to select among equally informative voxels
(Figure 3). Weight penalties can similarly suppress model feature weights in encoding models.
A related problem is performing statistical inference to test hypotheses about the weights
(Taylor & Tibshirani 2015). Systems and cognitive neuroscience has yet to develop a proper set
of methods for hypothesis testing in this context.
A simple remedy to the complications associated with the interpretation of fitted weights is to
interpret only the accuracy of decoding and encoding models (and its significance level) and not
the parameters of the models. In the context of decoding, this makes sense for local regions of
interest corresponding, for example, to cortical areas, which we can test one by one for the
presence of particular information. It can be generalized to brain mapping by applying the
decoder (or more generally any multivariate analysis) within a searchlight that scans the brain
for the effect of interest (Kriegeskorte et al. 2006). Like classical univariate brain mapping, this
approach derives its interpretability from the fact that each location is independently subjected
to the same analysis. However, instead of averaging local responses, the evidence is
summarized using local multivariate statistics. In the context of encoding models, similarly, we
can focus on each model's overall performance and on inferential comparisons among multiple
models.
15
Figure 3 Weights of decoding and encoding models are difficult to interpret. Three examples (rows)
illustrate the difficulty of interpreting decoder weights for a pair of voxels. In the first example (top row), only
the top voxel contains signal (stimulus information, red) and the two voxels have independent noise. This
scenario is unproblematic: both univariate mapping (second column from the left) and decoder weight maps
detect the informative voxel (red). In the second example (second row), both voxels contain the same signal.
Here, univariate mapping and weight maps often work. However, the LASSO decoder, because of its
preference for a sparse solution, may choose one of the voxels arbitrarily. In the third example (third row),
only the top voxel contains signal and both voxels contain correlated noise. Univariate mapping correctly
identifies the informative voxel. Linear decoders will give negative weight (blue) to the uninformative voxel,
so as to cancel the noise.
The single-model-significance fallacy
When our goal is merely to detect information in a brain region, we don't interpret the model as
a model of brain computation. This lowers the requirements for the model: It need not operate in
the direction of information flow and it need not be neurobiologically plausible. The model is
merely a statistical tool to sense a dependency. The choice of model in this scenario, will affect
our sensitivity, and its structure is not entirely irrelevant to the interpretation. For example,
decodability by a linear model tells us something about the format of the encoding. However, a
single model will suffice to demonstrate the presence of information in a brain region.
When our goal is to gain insight into the computations the brain performs, a model has a more
prominent role: it is meant to capture, at some level of abstraction, the computations occurring
in the brain. We then require the model to operate in the causal direction and to be
neurobiologically plausible (albeit abstracted). Examples of such models include encoding
models of sensory responses and decoding models that predict behavioral responses from
brain activity. Psychophysical models, which skip the brain entirely and predict behavioral
responses directly from stimuli, also fall into this class. In all these cases, finding that a model
explains significant variance is a very low bar and tells us little as to whether the model captures
the computational process.
16
The single-model-significance fallacy is to interpret the fact that a single model explains
significant variance as evidence in favor of the model. A simple example that is widely
understood is linear correlation. A significant linear correlation does demonstrate a dependency
between two variables, but it does not demonstrate that the dependency is linear. Similarly, the
fact that a complex encoding model explains significant variance in the responses of a brain
region to a test set of novel stimuli does demonstrate that the brain region contains information
about the stimuli, but it does not demonstrate that the encoding model captures the process that
computes the encoding.
Even a bad model can explain significant variance, especially if it has a large number of
parameters fitted to the data. In order to learn about the underlying brain computations, we need
to (a) consider multiple models, (b) assess what proportion of the explainable (i.e. nonnoise)
variance each explains at a given level of generalization, and (c) compare the models
inferentially.
Representational interpretations require additional
assumptions
Decoding and encoding models are often motivated by the goal to understand how the brain
represents the world, as well as the animal's decisions, goals, plans, actions, and motor
dynamics. Significant variance explained by encoding and decoding models demonstrates the
presence of information. Interpreting this information as a representation (Dennett 1987) implies
the additional claim that the brain activity serves the purpose to convey the information to other
parts of the brain (Kriegeskorte & Bandettini 2007, Kriegeskorte 2011, Diedrichsen &
Kriegeskorte 2017). This functional interpretation is so compelling in the context of sensory
systems that we sometimes jump too easily from findings of information to representational
interpretations (de Wit et al. 2016, Ritchie et al. 2017). In addition to the presence of the
information, its functional role as a representation implies that the information is read out by
other regions, affecting downstream processing and ultimately behavior. Combining encoding
and decoding models with stimulus- and response-based experimentation can help
disambiguate the causal implications (Weichwald et al. 2015). Ideally, experimental control of
neural activity should also be used to test whether activity has particular downstream or
behavioral consequences (Afraz et al. 2006, Raizada & Kriegeskorte 2010). To the extent that
we rely on prior assumptions to justify a representational interpretation, it is important to reflect
on these and consider if there is evidence from previous studies to support them.
17
References
1. Achille, A., & Soatto, S. (2018). Information dropout: Learning optimal representations
through noisy computation. IEEE transactions on pattern analysis and machine
intelligence, 40(12), 2897-2905.
2. Afraz, S. R., Kiani, R., & Esteky, H. (2006). Microstimulation of inferotemporal cortex
influences face categorization. Nature, 442(7103), 692.
3. Anzellotti, S., Fairhall, S. L., & Caramazza, A. (2014). Decoding Representations of Face
Identity That are Tolerant to Rotation. Cerebral Cortex, 24(8), 1988 -- 1995.
https://doi.org/10.1093/cercor/bht046
4. Averbeck, B. B., Latham, P. E., & Pouget, A. (2006). Neural correlations, population coding
and computation. Nature reviews neuroscience, 7(5), 358.
5. Carlson, T. A. et al. (2003) Patterns of activity in the categorical representation of objects. J.
Cogn. Neurosci. 15, 704 -- 717
6. Churchland, M. M., Cunningham, J. P., Kaufman, M. T., Foster, J. D., Nuyujukian, P., Ryu,
S. I., & Shenoy, K. V. (2012). Neural population dynamics during reaching. Nature.
https://doi.org/10.1038/nature11129
7. Cichy, R. M. et al. (2012) Imagery and perception share cortical representations of content
and location. Cereb. Cortex 22, 372 -- 380
8. Cox, D. D. and Savoy, R. L. (2003) Functional magnetic resonance imaging (fMRI) "brain
reading": detecting and classifying distributed patterns of fMRI activity in human visual
cortex. NeuroImage 19, 261 -- 270
9. Chung, S., Lee, D. D., & Sompolinsky, H. (2016). Linear readout of object
manifolds. Physical Review E, 93(6), 060301.
10. Chung, S., Lee, D. D., & Sompolinsky, H. (2018). Classification and geometry of general
perceptual manifolds. Physical Review X, 8(3), 031003.
11. Decharms, R. C., & Zador, A. (2000). Neural representation and the cortical code. Annual
review of neuroscience, 23(1), 613-647.
12. Dennett, D. (1987) The Intentional Stance, MIT Press
13. de-Wit, L., Alexander, D., Ekroll, V., & Wagemans, J. (2016). Is neuroimaging measuring
information in the brain? Psychonomic bulletin & review, 23(5), 1415-1428.
14. DiCarlo, J. J. and Cox, D. (2007) Untangling invariant object recognition. Trends Cogn. Sci.
11, 334-341
15. DiCarlo, J. J., Zoccolan, D., & Rust, N. C. (2012). How does the brain solve visual object
recognition?. Neuron, 73(3), 415-434.
16. Diedrichsen, J. (2018). Representational Models. In The Cognitive Neurosciences (M.S.
Gazzaniga (Ed.)).
17. Diedrichsen, J., & Kriegeskorte, N. (2017). Representational models: A common framework
for understanding encoding, pattern-component, and representational-similarity analysis.
PLOS Computational Biology, 13(4), e1005508.
https://doi.org/10.1371/journal.pcbi.1005508
• This paper explains the conceptual and mathematical relationships between linear
encoding models, pattern-component modeling, and representational similarity analysis and
18
how these methods provide complementary tools in a single toolbox for representational
analyses.
18. Diedrichsen, J., Ridgway, G. R., Friston, K. J., & Wiestler, T. (2011). Comparing the
similarity and spatial structure of neural representations: A pattern-component model.
NeuroImage, 55(4), 1665 -- 1678. https://doi.org/10.1016/j.neuroimage.2011.01.044
• This paper introduces pattern component modeling.
19. Diedrichsen, J., Wiestler, T., & Krakauer, J. W. (2012). Two distinct ipsilateral cortical
representations for individuated finger movements. Cerebral Cortex, 23(6), 1362-1377.
20. Duda, R. O., Hart, P. E., & Stork, D. G. (2012). Pattern classification. John Wiley & Sons.
21. Dumoulin, S. O., & Wandell, B. A. (2008). Population receptive field estimates in human
visual cortex. NeuroImage, 39(2), 647 -- 660.
https://doi.org/10.1016/j.neuroimage.2007.09.034
• This paper introduces nonlinear encoding models for early visual representations, which
enabled more rigorous investigations of early vision with neuroimaging data.
22. 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.
23. Felleman, D. J., & Van Essen, D. (1991). Distributed hierarchical processing in the primate
cerebral cortex. Cerebral cortex (New York, NY: 1991), 1(1), 1-47.
24. Freedman, D. J., Riesenhuber, M., Poggio, T., & Miller, E. K. (2002). Visual Categorization
and the Primate Prefrontal Cortex: Neurophysiology and Behavior. Journal of
Neurophysiology, 88(2), 929 -- 941. https://doi.org/10.1152/jn.2002.88.2.929
25. Friston, K. J., Holmes, A. P., Worsley, K. J., Poline, J. P., Frith, C. D., & Frackowiak, R. S.
(1994). Statistical parametric maps in functional imaging: a general linear approach. Human
brain mapping, 2(4), 189-210.
26. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... &
Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information
processing systems (pp. 2672-2680).
27. Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning. New
York, NY: Springer New York. https://doi.org/10.1007/978-0-387-84858-7
28. Haufe, S., Meinecke, F., Görgen, K., Dähne, S., Haynes, J.-D., Blankertz, B., & Biessmann,
F. (2014). On the interpretation of weight vectors of linear models in multivariate
neuroimaging. NeuroImage, 87, 96 -- 110. https://doi.org/10.1016/j.neuroimage.2013.10.067
• This paper provides a careful mathematical discussion of the difficulty of interpreting the
weights of linear models.
29. Haxby, J. V. et al. (2001). Distributed and Overlapping Representations of Faces and
Objects in Ventral Temporal Cortex. Science, 293(5539), 2425 -- 2430.
https://doi.org/10.1126/science.1063736
• A seminal decoding study of ventral-stream category representations that helped the field
conceptualize neuroimaging data as reflections of distributed neuronal population codes.
30. Haynes, J.-D. (2015). A Primer on Pattern-Based Approaches to fMRI: Principles, Pitfalls,
and Perspectives. Neuron, 87(2), 257 -- 270. https://doi.org/10.1016/j.neuron.2015.05.025
31. Haynes, J.-D., & Rees, G. (2006). Decoding mental states from brain activity in humans.
Nature Reviews Neuroscience, 7(7), 523 -- 534. https://doi.org/10.1038/nrn1931
19
32. Hebart, M. N., & Baker, C. I. (2017). Deconstructing multivariate decoding for the study of
brain function. Neuroimage.
33. Hebart, M. N., Görgen, K., & Haynes, J. D. (2015). The Decoding Toolbox (TDT): a versatile
software package for multivariate analyses of functional imaging data. Frontiers in
neuroinformatics, 8, 88.
34. Hong, H., Yamins, D. L. K., Majaj, N. J., & DiCarlo, J. J. (2016). Explicit information for
category-orthogonal object properties increases along the ventral stream. Nature
Neuroscience, 19(4), 613 -- 622. https://doi.org/10.1038/nn.4247
35. Huth, A. G. et al. (2012) A continuous semantic space describes the representation of
thousands of object and action categories across the human brain. Neuron 76, 1210 -- 1224
36. Kamitani, Y., & Tong, F. (2005). Decoding the visual and subjective contents of the human
brain. Nature Neuroscience, 8(5), 679 -- 685. https://doi.org/10.1038/nn1444
• A seminal decoding study of primary visual representations of oriented gratings, which
established that functional magnetic resonance imaging enables orientation decoding.
37. Kay, K. N., Naselaris, T., Prenger, R. J., & Gallant, J. L. (2008). Identifying natural images
from human brain activity. Nature, 452(7185), 352 -- 355. https://doi.org/10.1038/nature06713
• A seminal encoding and decoding study that helped the field envision how neuroimaging
data can be used to test image-computable vision models.
38. Kell, A. J. E., Yamins, D. L. K., Shook, E. N., Norman-Haignere, S. V., & McDermott, J. H.
(2018). A Task-Optimized Neural Network Replicates Human Auditory Behavior, Predicts
Brain Responses, and Reveals a Cortical Processing Hierarchy. Neuron, 98(3), 630-
644.e16. https://doi.org/10.1016/j.neuron.2018.03.044
39. Khaligh-Razavi, S.-M., Henriksson, L., Kay, K., & Kriegeskorte, N. (2016). Fixed versus
mixed RSA: Explaining visual representations by fixed and mixed feature sets from shallow
and deep computational models. Journal of Mathematical Psychology.
https://doi.org/10.1016/j.jmp.2016.10.007
40. King, J. R., & Dehaene, S. (2014). Characterizing the dynamics of mental representations:
the temporal generalization method. Trends in cognitive sciences, 18(4), 203-210.
41. Kriegeskorte, N. (2011). Pattern-information analysis: From stimulus decoding to
computational-model testing. NeuroImage, 56(2), 411 -- 421.
https://doi.org/10.1016/j.neuroimage.2011.01.061
42. Kriegeskorte, N. (2015). Crossvalidation in brain imaging analysis (No. biorxiv;017418v1).
Retrieved from http://biorxiv.org/lookup/doi/10.1101/017418
43. Kriegeskorte, N. (2015). Deep Neural Networks: A New Framework for Modeling Biological
Vision and Brain Information Processing. Annual Review of Vision Science 1, 417 -- 446.
• A review and perspective paper developing a framework for how to build neural network
models of brain information processing with constraints from brain-activity data.
44. Kriegeskorte, N., & Bandettini, P. (2007). Analyzing for information, not activation, to exploit
high-resolution fMRI. Neuroimage, 38(4), 649-662.
45. Kriegeskorte, N. et al. (2008) Representational similarity analysis -- connecting the
branches of systems neuroscience. Front. Syst. Neurosci. 2, 4.
• The paper introducing representational similarity analysis.
46. Kriegeskorte, N. et al. (2007) Individual faces elicit distinct response patterns in human
anterior temporal cortex. Proc. Natl Acad. Sci. U. S. A. 104, 20600-20605
20
47. Kriegeskorte, N., & Diedrichsen, J. (2016). Inferring brain-computational mechanisms with
models of activity measurements. Philosophical Transactions of the Royal Society B:
Biological Sciences, 371(1705), 20160278. https://doi.org/10.1098/rstb.2016.0278
48. Kriegeskorte, N., Goebel, R., & Bandettini, P. (2006). Information-based functional brain
mapping. Proceedings of the National Academy of Sciences of the United States of
America, 103(10), 3863 -- 3868. https://doi.org/10.1073/pnas.0600244103
49. Majaj, N. J., Hong, H., Solomon, E. A., & DiCarlo, J. J. (2015). Simple learned weighted
sums of inferior temporal neuronal firing rates accurately predict human core object
recognition performance. Journal of Neuroscience, 35(39), 13402-13418.
50. Mesgarani, N., David, S. V., Fritz, J. B., & Shamma, S. A. (2009). Influence of context and
behavior on stimulus reconstruction from neural activity in primary auditory cortex. Journal of
neurophysiology, 102(6), 3329-3339.
51. Mitchell, T. M. et al. (2008) Predicting human brain activity associated with the meanings of
nouns. Science 320, 1191 -- 1195
• A seminal encoding study of semantic brain representations using semantic property
descriptions to predict brain representations of novel stimuli.
52. Miyawaki, Y., Uchida, H., Yamashita, O., Sato, M., Morito, Y., Tanabe, H. C., … Kamitani, Y.
(2008). Visual Image Reconstruction from Human Brain Activity using a Combination of
Multiscale Local Image Decoders. Neuron, 60(5), 915 -- 929.
https://doi.org/10.1016/j.neuron.2008.11.004
53. Mur, M., Bandettini, P. A., & Kriegeskorte, N. (2009). Revealing representational content
with pattern-information fMRI -- an introductory guide. Social Cognitive and Affective
Neuroscience, 4(1), 101 -- 109. https://doi.org/10.1093/scan/nsn044
54. Naselaris, T., & Kay, K. N. (2015). Resolving Ambiguities of MVPA Using Explicit Models of
Representation. Trends in Cognitive Sciences, 19(10), 551 -- 554.
https://doi.org/10.1016/j.tics.2015.07.005
55. Naselaris, T., Kay, K. N., Nishimoto, S., & Gallant, J. L. (2011). Encoding and decoding in
fMRI. NeuroImage, 56(2), 400 -- 410. https://doi.org/10.1016/j.neuroimage.2010.07.073
56. Naselaris, T., Prenger, R. J., Kay, K. N., Oliver, M., & Gallant, J. L. (2009). Bayesian
Reconstruction of Natural Images from Human Brain Activity. Neuron, 63(6), 902 -- 915.
https://doi.org/10.1016/j.neuron.2009.09.006
57. Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., & Kriegeskorte, N. (2014). A
toolbox for representational similarity analysis. PLoS computational biology, 10(4),
e1003553.
58. Norman, K. A. et al. (2006) Beyond mindreading: multi-voxel pattern analysis of fMRI data.
Trends Cogn. Sci. 10, 424 -- 430;
59. Nunez-Elizalde, A. O., Huth, A. G., & Gallant, J. L. (2018). Voxelwise encoding models with
non-spherical multivariate normal priors. bioRxiv, 386318.
60. Paninski, L., & Cunningham, J. (2017). Neural data science: accelerating the experiment-
analysis-theory cycle in large-scale neuroscience. bioRxiv, 196949.
61. Paninski, L., Pillow, J., & Lewi, J. (2007). Statistical models for neural encoding, decoding,
and optimal stimulus design. Progress in brain research, 165, 493-507.
21
62. Parthasarathy, N., Batty, E., Falcon, W., Rutten, T., Rajpal, M., Chichilnisky, E. J., &
Paninski, L. (2017). Neural networks for efficient bayesian decoding of natural images from
retinal neurons. In Advances in Neural Information Processing Systems (pp. 6434-6445).
63. Pereira, F., Mitchell, T., & Botvinick, M. (2009). Machine learning classifiers and fMRI: a
tutorial overview. NeuroImage, 45(1 Suppl), S199-209.
https://doi.org/10.1016/j.neuroimage.2008.11.007
64. Pillow, J. W., Ahmadian, Y., & Paninski, L. (2011). Model-based decoding, information
estimation, and change-point detection techniques for multineuron spike trains. Neural
computation, 23(1), 1-45.
65. Poggio, T., & Girosi, F. (1990). Networks for approximation and learning. Proceedings of the
IEEE, 78(9), 1481-1497.
66. Raizada, R. D., & Kriegeskorte, N. (2010). Pattern‐information fMRI: New questions which it
opens up and challenges which face it. International Journal of Imaging Systems and
Technology, 20(1), 31-41.
67. Ritchie, J. B., Kaplan, D. M., & Klein, C. (2017). Decoding the brain: Neural representation
and the limits of multivariate pattern analysis in cognitive neuroscience. The British Journal
for the Philosophy of Science.
68. Santoro, R., Moerel, M., De Martino, F., Valente, G., Ugurbil, K., Yacoub, E., & Formisano,
E. (2017). Reconstructing the spectrotemporal modulations of real-life sounds from fMRI
response patterns. Proceedings of the National Academy of Sciences, 114(18), 4799-4804.
69. Shadlen, M. N., Britten, K. H., Newsome, W. T., & Movshon, J. A. (1996). A computational
analysis of the relationship between neuronal and behavioral responses to visual
motion. Journal of Neuroscience, 16(4), 1486-1510.
70. Simoncelli, E. P., & Olshausen, B. A. (2001). Natural image statistics and neural
representation. Annual review of neuroscience, 24(1), 1193-1216.
71. Stokes, M. et al. (2009) Top-down activation of shape-specific population codes in visual
cortex during mental imagery. J Neurosci. 29, 1565-1572
72. Taylor, J., & Tibshirani, R. J. (2015). Statistical learning and selective inference.
Proceedings of the National Academy of Sciences, 112(25), 7629-7634.
73. Thirion, B., Duchesnay, E., Hubbard, E., Dubois, J., Poline, J. B., Lebihan, D., & Dehaene,
S. (2006). Inverse retinotopy: inferring the visual content of images from brain activation
patterns. Neuroimage, 33(4), 1104-1116.
74. Tishby, N., Pereira, F. C., & Bialek, W. (2000). The information bottleneck method. arXiv
preprint physics/0004057.
75. Tong, F. and Pratte, M. S. (2012) Decoding patterns of human brain activity. Ann. Rev.
Psychol. 63, 483-509
76. van Gerven, M. A. (2017). A primer on encoding models in sensory neuroscience. Journal of
Mathematical Psychology, 76, 172-183.
77. Varoquaux, G., Raamana, P. R., Engemann, D. A., Hoyos-Idrobo, A., Schwartz, Y., &
Thirion, B. (2017). Assessing and tuning brain decoders: cross-validation, caveats, and
guidelines. NeuroImage, 145, 166-179.
78. Walther, A., Nili, H., Ejaz, N., Alink, A., Kriegeskorte, N., & Diedrichsen, J. (2016). Reliability
of dissimilarity measures for multi-voxel pattern analysis. Neuroimage, 137, 188-200.
22
79. Walther, D. B. et al. (2012) To err is human: correlating fMRI decoding and behavioral errors
to probe the neural representation of natural scene categories. In: Visual Population Codes -
toward a Common Multivariate Framework for Cell Recording and Functional Imaging,
(Kriegeskorte, N., Kreiman, G. eds.), pp. 391-416, MIT Press
80. Weichwald, S., Meyer, T., Özdenizci, O., Schölkopf, B., Ball, T., & Grosse-Wentrup, M.
(2015). Causal interpretation rules for encoding and decoding models in
neuroimaging. Neuroimage, 110, 48-59.
• This paper introduces a set of causal interpretation rules for encoding and decoding
models on the basis of the technical literature on causal inference.
81. Wen, H., Shi, J., Chen, W., & Liu, Z. (2018). Deep Residual Network Predicts Cortical
Representation and Organization of Visual Features for Rapid Categorization. Scientific
reports, 8(1), 3752.
82. Williams, M. A., Dang, S., & Kanwisher, N. G. (2007). Only some spatial patterns of fMRI
response are read out in task performance. Nature neuroscience, 10(6), 685.
83. Wu, M. C. K., David, S. V., & Gallant, J. L. (2006). Complete functional characterization of
sensory neurons by system identification. Annu. Rev. Neurosci., 29, 477-505.
84. Yamane, Y. et al. (2008) A neural code for three-dimensional object shape in macaque
inferotemporal cortex. Nat. Neurosci. 11, 1352-1360
85. Yamins, D. L. K. & DiCarlo, J. J. Using goal-driven deep learning models to understand
sensory cortex. Nature Neuroscience 19, 356 -- 365 (2016).
• A review and perspective paper describing a framework for modeling sensory processing
by training deep neural networks on tasks and testing them with brain-activity data.
86. Yamins, D. L. K. et al. Performance-optimized hierarchical models predict neural responses
in higher visual cortex. Proceedings of the National Academy of Sciences 111, 8619 -- 8624
(2014).
87. Young, M. P. (1992). Objective analysis of the topological organization of the primate
cortical visual system. Nature, 358(6382), 152.
23
|
1104.5458 | 1 | 1104 | 2011-04-28T17:56:32 | An MRI-Derived Definition of MCI-to-AD Conversion for Long-Term, Automati c Prognosis of MCI Patients | [
"q-bio.NC",
"physics.med-ph"
] | Alzheimer's disease (AD) and mild cognitive impairment (MCI), continue to be widely studied. While there is no consensus on whether MCIs actually "convert" to AD, the more important question is not whether MCIs convert, but what is the best such definition. We focus on automatic prognostication, nominally using only a baseline image brain scan, of whether an MCI individual will convert to AD within a multi-year period following the initial clinical visit. This is in fact not a traditional supervised learning problem since, in ADNI, there are no definitive labeled examples of MCI conversion. Prior works have defined MCI subclasses based on whether or not clinical/cognitive scores such as CDR significantly change from baseline. There are concerns with these definitions, however, since e.g. most MCIs (and ADs) do not change from a baseline CDR=0.5, even while physiological changes may be occurring. These works ignore rich phenotypical information in an MCI patient's brain scan and labeled AD and Control examples, in defining conversion. We propose an innovative conversion definition, wherein an MCI patient is declared to be a converter if any of the patient's brain scans (at follow-up visits) are classified "AD" by an (accurately-designed) Control-AD classifier. This novel definition bootstraps the design of a second classifier, specifically trained to predict whether or not MCIs will convert. This second classifier thus predicts whether an AD-Control classifier will predict that a patient has AD. Our results demonstrate this new definition leads not only to much higher prognostic accuracy than by-CDR conversion, but also to subpopulations much more consistent with known AD brain region biomarkers. We also identify key prognostic region biomarkers, essential for accurately discriminating the converter and nonconverter groups. | q-bio.NC | q-bio |
An MRI-Derived Definition of MCI-to-AD Conversion
for Long-Term, Automatic Prognosis of MCI Patients
for the Alzheimer's Disease Neuroimaging Initiative
$
Yaman Aksua, David J. Millerb, George Kesidisb,c, Don C. Biglerd, Qing X.
Yanga
aCenter for NMR Research, Department of Radiology, Penn State University College of
bElectrical Engineering Department, Penn State University, University Park, PA, USA
cComputer Science and Engineering Department, Penn State University, University
Medicine, Hershey, PA, USA
dCenter for Emerging Neurotechnology and Imaging, Penn State University College of
Medicine, Hershey, PA, USA
Park, PA, USA
Abstract
Alzheimer's disease (AD), and its precursor state, mild cognitive impairment
(MCI), continue to be widely studied. While there is no consensus on whether
MCIs actually "convert" to AD, this concept is widely applied, allowing sta-
tistical testing and machine learning methods to help identify early disease
biomarkers and build models for predicting disease progression. Thus, the
more important question is not whether MCIs convert, but what is the best
such definition. We focus on automatic prognostication, nominally using only
a baseline image brain scan, of whether an MCI individual will convert to AD
within a multi-year period following the initial clinical visit. This is in fact
not a traditional supervised learning problem since, in ADNI, there are no
definitive labeled examples of MCI conversion. It is not truly unsupervised,
either, since there are (labeled) AD and Control subjects, as well as clini-
cal and cognitive scores for MCIs. Prior works have defined MCI subclasses
$Data used in preparation of this article were obtained from the Alzheimer's Dis-
ease Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). As such, the in-
vestigators within the ADNI contributed to the design and implementation of ADNI
and/or provided data but did not participate in analysis or writing of this report. A
complete listing of ADNI investigators can be found at: http://adni.loni.ucla.edu/wp-
content/uploads/how to apply/ADNI Authorship List.pdf
August 7, 2018
based on whether or not clinical/cognitive scores significantly change from
baseline. There are serious concerns with these definitions, however, since
e.g. most MCIs (and ADs) do not change from a baseline CDR=0.5 at any
subsequent visit in ADNI, even while physiological changes may be occur-
ring. These works ignore rich phenotypical information in an MCI patient's
brain scan and labeled AD and Control examples, in defining conversion.
We propose an innovative conversion definition, wherein an MCI patient is
declared to be a converter if any of the patient's brain scans (at follow-up
visits) are classified "AD" by an (accurately-designed) Control-AD classifier.
This novel definition bootstraps the design of a second classifier, specifically
trained to predict whether or not MCIs will convert. This second classifier
thus predicts whether an AD-Control classifier will predict that a patient
has AD. Our results demonstrate this new definition leads not only to much
higher prognostic accuracy than by-CDR conversion, but also to subpopu-
lations much more consistent with known AD brain region biomarkers. We
also identify key prognostic region biomarkers, essential for accurately dis-
criminating the converter and nonconverter groups.
Keywords: Alzheimer's, mild cognitive impairment, AD conversion, MRI,
support vector machines, feature selection, AD biomarkers, MFE, RFE
1. Introduction
The dementing illness Alzheimer's disease (AD), and the transitional
state between normal aging and AD referred to as mild cognitive impair-
ment (MCI) continue to be widely studied. Individuals diagnosed with MCI
have memory impairment, yet without meeting dementia criteria. Annually
≈ 10-15% of people with MCI are diagnosed with AD (Petersen, 2004). More-
over, prior to symptom onset, brain abnormalities have been found in people
with MCI, as ascertained by retroactive evaluation of longitudinal MRI scans
(Davatzikos et al., 2008). There is no consensus on whether MCI patients
actually "convert" to AD. First, some MCI patients may suffer from other
neurodegenerative disorders (e.g., Lewy body dementia, vascular dementia
and/or frontotemporal dementia). Second, it is possible that all other MCI
patients already have AD, but at a preclinical stage. AD diagnosis itself may
not be considered definitive without e.g. confirming pathologies such as the
amyloid deposits detectable at autopsy. Regardless of whether MCI patients
truly "convert" to AD or not, the concept of MCI-to-AD conversion has been
2
widely applied, e.g. (Chou et al., 2009; Davatzikos et al., 2010; Misra et al.,
2009; Vemuri et al., 2009; Zhang et al., 2011) and is utilitarian -- defining
MCI (converter and nonconverter) subgroups allows use of statistical group
difference tests and machine learning methods to help identify early disease
biomarkers and to build models for predicting disease progression. For these
purposes, the more important question is not whether MCIs convert, but
rather what is the best such definition.
Accordingly, here we focus on the following Aim: automatic prognos-
tication, (nominally) using only a baseline brain scan, of whether an MCI
individual will convert to AD within a multi-year (three year) period follow-
ing an initial (baseline) clinical visit1.
While only image voxel-based features are evaluated here for use by our
classifier, our framework is extensible to incorporating other baseline clinical
information (e.g. weight, gender, education level, genetic information, and
clinical cognitive scores such as the Mini Mental State Exam (MMSE)) into
the decisionmaking. Moreover, our approach can also incorporate the recent,
promising cerebrospinal fluid (CSF) based markers (De Meyer et al., 2010).
However, as this requires an invasive spinal tap, we focus here on image scans,
which are routinely performed for subjects with MCI.
We do not hypothesize that, within ADNI, there are actually two sub-
classes of MCI subjects when evaluated over the very long term -- those that
(eventually) convert to AD, and those that do not. Even if an overwhelm-
ing majority of MCI subjects will eventually convert, identifying the sub-
group likely to convert within several years has several compelling utilities:
1) early prognosis, to assist family planning; 2) facilitating group-targeted
treatments/drug trials; 3) we identify key prognostic brain "biomarker" re-
gions, i.e. those found to be most critical for accurately discriminating our
"converter" and "nonconverter" groups. These regions may shed light on
disease etiology.
Distinguishing AD converters from nonconverters is a binary (two-class)
classification problem. Moreover, it may appear this classification problem
can be directly addressed via supervised learning methods (Duda et al., 2001).
However, it is in fact an unconventional problem, lying somewhere between su-
pervised classification and unsupervised classification (clustering), and thus
1Our system performs three-year ahead prediction because it is designed based on the
ADNI database, which followed participants for a period of up to three years.
3
requiring a unique approach. To appreciate this, consider the ADNI cohort
of MCI individuals. ADNI consists of clinical information and image scans on
hundreds of participants, taken at six-month intervals for up to three years.
A clinical label (AD, MCI, or Control) was assigned to each participant at
first visit2. Even though a probable AD definition based on CDR and MMSE
scores and NINCDS/ADRDA criteria has been used, e.g. (Leow et al., 2009;
Zhang et al., 2011), to provide follow-up assessment for MCI patients, this is
strictly a clinically driven definition, based on a clinical rating (CDR) and a
cognitive score (MMSE) whose difficulties will be pointed out shortly. This
is not a definitive (autopsy-based) determination of AD, nor is it a definition
based on physiological brain changes. Even if the probable AD definition has
very high specificity, it may not be sufficiently sensitive, i.e. there may be pa-
tients who are undergoing significant physiological brain changes consistent
with conversion, yet without clinical manifestation.
Accordingly, we will approach the conversion problem from a perspective
as agnostic and unbiased as possible, and simply state that it is not defini-
tively known which MCI participants in ADNI truly converted to AD within
three years. In conventional supervised classifier learning, one has labeled
training examples, used for designing the classifier, and labeled test exam-
ples, used to estimate the classifier's generalization accuracy. For predicting
whether MCI participants in ADNI convert to AD, we in fact have neither.
Thus, our problem is not conventional supervised learning. On the other
hand, consider unsupervised clustering (Duda et al., 2001). Here, even if one
knows the number of clusters (classes) present, there is no prior knowledge
on what is a good clustering -- one is simply looking for underlying group-
ing tendency in the data. Clearly, our problem does not fit unsupervised
clustering, either -- while we have no labeled MCI converter/nonconverter
instances per se, 1) there are two designated classes of interest (converter
and nonconverter); and 2) there are known class characteristics -- conversion
to AD should, plausibly, mean that: (i) a clinical measure such as CDR or
a cognitive measure has changed and/or (ii) there are changes in brain fea-
tures more characteristic of AD subjects than normal/healthy subjects. Note
that in ADNI we do have plentiful labeled AD and normal/healthy (Control)
2Clinicians derive the AD/MCI/Control label based on multiple criteria, which may
include Clinical Dementia Rating (CDR), whose possible values are: 0=none, 0.5=ques-
tionable, 1=mild, 2=moderate, 3=severe.
4
examples to help assess ii).
Based on the above, MCI prognosis is an interesting and novel problem,
lying somewhere between supervised and unsupervised classification. The
crux of this problem is to craft criteria through which meaningful MCI sub-
groups can be defined, well-capturing notions of "AD converter" and "non-
converter". To help guide development and evaluation of candidate defini-
tions, we state the following three desiderata: 1) The proposed definition
of AD converter should be plausible and should exploit the available, rele-
vant information in the ADNI database (e.g.
image data, labeled AD and
Control examples, and clinical information). To appreciate 1), note that
the MCI population could be dichotomized in many ways, e.g. by height,
and there might be significant clustering tendency with respect to height,
but such a grouping is likely meaningless for MCI prognosis; 2) A classifier
trained based on these class definitions should generalize well on test data
(not used for training the classifier) -- this quantifies how accurately we can
discriminate the classes that we have defined. If we create what we believe
to be good definitions, but ones that cannot be accurately discriminated,
that would not be useful clinically; 3) The class definitions should be vali-
dated using known AD conversion biomarkers (i.e., external measures) such
as measured changes from baseline both in volumes of brain regions known
to be associated with the disease (Schuff et al., 2010) and in cognitive test
scores (such as the clinical MMSE measure).
Prior Related Work
Several prior works, e.g. (Davatzikos et al., 2010; Misra et al., 2009; Wang
et al., 2010), defined converter and nonconverter classes solely according to
whether the baseline visit CDR score of 0.5 rose or stayed the same over
all visits. Change in CDR has also been used as surrogate ground-truth for
cognitive decline in a number of other papers, e.g. (Chou et al., 2010; Vemuri
et al., 2009). While CDR gives a workable conversion definition, it should
be evaluated with respect to the three desiderata above. We will evaluate 2)
and 3) in the sequel. With respect to 1), one should challenge a CDR-based
conversion definition. First, CDR is not an effective discriminator between
the AD and MCI groups, i.e.
there is very significant AD-MCI overlap,
not only with respect to CDR=1 but even 0.5 -- for ADNI, the majority of
the hundreds of AD subjects used in our experiments start (at first visit)
at CDR=0.5 and stay at 0.5 at all later visits;
likewise, nearly all MCI
subjects start at 0.5, with a large majority of these also staying at 0.5 for
all visits. This latter fact further implies difficulties in finding an adequate
5
number of conversion-by-CDR subjects in ADNI, both for accurate classifier
training and test set evaluation. For the even more stringent probable AD
definition (meeting MMSE and NINCDS/ADRDA criteria, in addition to
CDR changing from 0.5 to 1) there are necessarily even fewer MCI converters
for classifier training and testing. Second, a purely CDR-based (or "probable
AD" based) conversion definition ignores the (rich) phenotypical information
in an MCI subject's image brain scans and does not exploit the labeled
AD and normal/healthy (Control) examples in ADNI. These prior works do
treat features derived from brain scans as the covariates (the inputs) to the
classifier/predictor. However, we believe the MCI brain scans can themselves
be used, in conjunction with the labeled AD and Control examples, to help
define more accurate surrogate ground-truth.
Previous work has demonstrated that structural MRI analysis is useful
for identifying AD biomarkers in individual brain regions (Chetelat et al.,
2002; Fan et al., 2008; Fennema-Notestine et al., 2009) -- e.g., cortical thin-
ning (Lerch et al., 2005; Thompson et al., 2003), ventricle dilation and gaping
(Chou et al., 2010, 2009; Schott et al., 2005), volumetric and shape changes
in the hippocampus and entorhinal cortex (Csernansky et al., 2005; de Leon
et al., 2006; Stoub et al., 2005), and temporal lobe shrinkage (Rusinek et
al., 2004). It is important to capture interaction effects across multiple brain
regions.
(Davatzikos et al., 2008; Misra et al., 2009; Vemuri et al., 2008;
Wang et al., 2010) did jointly analyze voxels (or regions) spanning the entire
brain and did build classifiers or predictors. Moreover, as part of their work,
(Wang et al., 2010) investigated prediction of future decline in MCI subjects
working from baseline MRI scans, which is the primary subject of our cur-
rent paper. However, there are several limitations of these past works. First,
all these studies used the previously discussed CDR and cognitive measures
such as MMSE, which has been described as noisy and unreliable, as the
ground-truth prediction targets for classifier/regressor training. In (Chou et
al., 2010), the authors state: "Cognitive assessments are notoriously variable
over time, and there is increasing evidence that neuroimaging may provide
accurate, reproducible measures of brain atrophy." Even in (Wang et al.,
2010), where MMSE was treated as the measure of decline and the ground-
truth regression target, the authors acknowledged that "individual cognitive
evaluations are known to be extremely unstable and depend on a number
of factors unrelated to...brain pathology." Such factors include sleep depri-
vation, depression, other medical conditions, and medications. Even though
MMSE is widely used by clinicians, these comments (even if not universally
6
accepted), do indicate MMSE by itself may not be so reliable in quantifying
the disease state. Moreover, while (Wang et al., 2010) did build predictors of
future MMSE scores working from baseline scans, this was not a main focus
of their paper -- their paper focused on predicting the current score. Their
prognostic experiments involved a very small sample size (just 26 participants
from the ADNI database). Accordingly, it is difficult to draw definitive con-
clusions about the accuracy of their prognostic model and their associated
brain biomarkers. The main reason the authors chose such a small sample
was, as the authors state: "A large part of...ADNI...are from patients who
did not display significant cognitive decline...[these] would overwhelm the re-
gression algorithm if..used in the...experiment." While this statement (with
cognitive decline measured according to MMSE) may be true, that does not
mean many of those excluded ADNI subjects are not experiencing significant
physiological brain changes/atrophy. The novel approach we next sketch is
well-suited to identifying MCI subjects undergoing such changes.
Our Neuroimaging-Driven, Trajectory-Based Approach
Here, we propose a novel approach for prognosticating putative conver-
sion to AD driven by image-based information (and exploiting AD-Control
examples), rather than by a single, non-image-based, weakly discriminat-
ing clinical measurement such as CDR. Our solution strategy is as follows.
We first build an accurate image-based Control-AD classifier (i.e., using AD
and Control subjects, we build a Support Vector Machine (SVM) classifier)
(Vapnik, 1998). We then apply this classifier to a training population of
MCI subjects -- separately, for each subject visit, we determine whether the
subject's image is on the AD side or the Control side of the SVM's fixed
(hyperplane) decision boundary. In addition to a binary decision, the SVM
gives a "score" -- essentially the distance to the classifier's decision boundary.
Thus, for each MCI subject, as a function of visits, we get an image-based
"phenotypical" score trajectory. We fit a line to each subject's trajectory
and extend the line to the sixth visit if the sixth visit is missing. We can
then give the following trajectory-based conversion definition: if the extended
line either starts on the AD side or crosses to the AD side over the six visits,
we declare this person a "converter-by-trajectory". Otherwise, this person
is a "nonconverter-by-trajectory"3. In this fashion, we derive ground-truth
3A very small percentage of the MCI population, in our experiments often 1% and not
exceeding 5%, may unexpectedly start on the AD side and cross to the Control side. We
7
"converter" and "nonconverter" labels for an (initially unlabeled) training
MCI population. These (now) labeled training samples bootstrap the de-
sign of a second SVM classifier which uses only the first-visit training set
MCI images and is trained to predict whether or not an MCI patient is a
"converter-by-trajectory". Essentially, this second (prognostic) classifier is
predicting whether, within three years, an AD-Control classifier will predict
that a patient has AD. Via these two classifier design steps, we thus build a
classification system for our (unconventional) pattern recognition task.
SVMs are widely used classifiers whose accuracy is attributed to their
maximization of the "margin", i.e. the smallest distance from any training
point to the classification boundary. Since the SVM finds a linear discrimi-
nant function that maximizes margin, a significant change in score is generally
needed to cross from the control side to the AD side, which is thus suggestive
of conversion from MCI to AD. This is the premise underlying our approach.
The main contributions of our work are: 1) a novel image-based prognos-
ticator of MCI-to-AD conversion that we will demonstrate to achieve both
better generalization accuracy and much higher correlation with known brain
region biomarkers than the CDR-based approach; 2) The finding that MCI
subgroups that are strongly correlated with known AD brain region biomark-
ers are not so strongly correlated with "cognitive decline" as measured by
MMSE; 3) Identification of the brain regions most critical for accurately
discriminating between our "converter" and "nonconverter" groups, via ap-
plication of margin-based feature selection (MFE) (Aksu et al., 2010) to brain
image classification, and demonstration of MFE's better performance than
the well-known RFE method (Guyon et al., 2002) on this domain.
treat these individuals as outliers and omit them from our experiments.
8
2. Methods
2.1. Subjects and MRI data
We used T1-weighted ADNI images4 that have undergone image correc-
tion described at the ADNI website.5 ADNI aims to recruit and follow 800
research participants in the 55-90 age range: approximately 200 elderly Con-
trols, 400 people with MCI, and 200 people with AD. The number of Control,
MCI, and AD participants in our analysis were ≈180, 300, and 120, respec-
tively -- experiment-specific detailed descriptions will be provided in Sec. 3.
We processed the T1-weighted images as described in Appendix A, produc-
ing new images from which we then obtained the features (next discussed)
used by our statistical classifiers.
2.2. Features for classification
We chose as features the voxel intensities of a processed6 RAVENS im-
age, a type of "volumetric density" image (Davatzikos, 1998; Davatzikos et
al., 2001; Goldszal et al., 1998; Shen et al., 2003) that has been validated
for voxel-based analysis (Davatzikos et al., 2001) and applied both to AD
e.g. (Davatzikos et al., 2010; Misra et al., 2009; Wang et al., 2010) and other
studies e.g. (Fan et al., 2007)7. For each of the three processed RAVENS tis-
sue maps (gray matter (GM), white matter (WM), and ventricle), to reduce
4 Data used in the preparation of this article were obtained from the Alzheimer's Disease
Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). The ADNI was launched
in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical
Imaging and Bioengineering (NIBIB), the Food and Drug Administration (FDA), private
pharmaceutical companies and non-profit organizations, as a $60 million, 5-year public-
private partnership. The primary goal of ADNI has been to test whether serial mag-
netic resonance imaging (MRI), positron emission tomography (PET), other biological
markers, and clinical and neuropsychological assessment can be combined to measure the
progression of mild cognitive impairment (MCI) and early Alzheimer's disease (AD). De-
termination of sensitive and specific markers of very early AD progression is intended to
aid researchers and clinicians to develop new treatments and monitor their effectiveness,
as well as lessen the time and cost of clinical trials.
5ADNI image correction steps include Gradwarp, N3, and scaling for gradient drift --
see www.loni.ucla.edu/ADNI/Data/ADNI Data.shtml.
6We describe our processing of RAVENS images in Appendix A.
7Of particular interest, (Davatzikos et al., 2001) supported that voxel-based SPM statis-
tical analysis, which we perform herein for comparison with our methods, can be performed
on RAVENS images.
9
complexity for subsequent processing, we obtained a subsample by succes-
sively skipping five voxels along each of the three dimensions, and took as
feature set the union of the three subsampled maps. We will also report
results for the case of skipping only two voxels, rather than five.
Since high-dimensional nonlinear registration (warping) of all individuals
to a common atlas (via HAMMER (Shen et al., 2002)) is applied in pro-
ducing our features, they capture both volumetric and morphometric brain
characteristics, which is important since individuals with AD/MCI typically
exhibit brain atrophy (affecting both volume and shape).
2.3. Classification and feature selection for high-dimensional images
A challenge in building classifiers for medical images is the relative paucity
of available training samples, compared to the huge dimensionality of the
voxel space and, thus, to the number of parameters in the classifier model --
in general, the number of parameters may grow at least linearly with dimen-
sionality. In the case of 3D images, this could imply even millions of param-
eter values (e.g. one per voxel) need to be determined, based on a training
set of only a few hundred patient examples. In such cases, classifier over-
fitting is likely, which can degrade generalization (test set) accuracy. Here
we will apply a linear discriminant function (LDF) classifier with a built-in
mechanism to avoid overfitting and with design complexity that scales well
with increasing dimensionality - the support vector machine (SVM) (Vap-
nik, 1998). The choice of LDF achieving perfect separation (no classification
errors) for a given two-class training set is not unique. The SVM, however,
is the unique separating LDF that maximizes the margin, i.e. the minimum
distance to the classifier decision boundary, over all training samples. In this
sense, the SVM maximizes separation of the two classes. For an SVM, unlike
a standard LDF, the number of model parameters is bounded by the number
of training samples, rather than being controlled by the feature dimension-
ality. Since the number of samples is the much smaller number for medical
image domains, in this way the SVM greatly mitigates overfitting. SVMs
have achieved excellent classification accuracy for numerous scientific and
engineering domains, including medical image analysis, (Aksu et al., 2010;
Davatzikos et al., 2010; Guyon et al., 2002).
Even though SVMs are effective at mitigating overfitting, generalization
accuracy may still be improved in some cases by removing features that
contribute little discrimination power. Moreover, even if generalization ac-
curacy monotonically improves with increasing feature dimensionality, high
10
complexity (both computation and memory storage) of both classifier design
and class decisionmaking may outweigh small gains in accuracy achieved by
using a huge number of features. Most importantly here, it is often useful
to identify the critical subset of features necessary for achieving accurate
classification -- these "markers" may shed light on the underlying disease
mechanism. In our case, this will help to identify prognostic brain regions,
associated with MCI conversion.
Unfortunately, there is a huge number of possible feature subsets, with
exhaustive subset evaluation practically prohibited even for modest number
of features, M , let alone M ∼ 106. Practical feature selection techniques are
thus heuristic, with a large range of tradeoffs between accuracy and complex-
ity (Guyon et al., 2003). "Front-end" (or "filtering") methods select features
prior to classifier training, based on evaluation of discrimination power for
individual features or small feature groups. "Wrapper" methods are gen-
erally more reliable, interspersing sequential feature selection and classifier
design steps, with features sequentially selected to maximize the current sub-
set's joint discrimination power. There are also embedded feature selection
methods, e.g. for SVMs, use of (cid:96)1-regularization within the SVM design opti-
mization (Fung et al., 2004), in order to find "sparse" weight vector solutions,
which effectively eliminate many features. For wrappers, there is greedy for-
ward selection, with "informative" features added, backward elimination,
which starts from the full set and removes features, and more complex bidi-
rectional searches. In our work, due to the high feature dimension, we focus
on two backward elimination wrappers that afford practical complexity:
i)
the widely used recursive feature elimination algorithm (RFE) (Guyon et al.,
2002), where at each step one removes the feature with least weight magni-
tude in the SVM solution. RFE has been applied before to AD (Davatzikos
et al., 2010; Misra et al., 2009; Wang et al., 2010); ii) the recent margin-
based feature elimination (MFE) algorithm (Aksu et al., 2010), which uses
the same objective function (margin) for feature elimination, one consistent
with good generalization, that the SVM uses for classifier training (Vapnik,
1998). MFE was shown in (Aksu et al., 2010) to outperform RFE (Guyon et
al., 2002) and to achieve results comparable to embedded feature selection
for domains with up to 8,000 features (gene microarray classification). Here
we will also find that MFE gives better results than RFE.
11
2.4. An MRI-Derived Alternative to CDR-based MCI-to-AD Conversion
In the Introduction, we outlined our two classifier design steps for build-
ing an automatic prognosticator for an individual with MCI. In this sec-
tion, we elaborate on these two steps and give an illustrative example. Our
AD-Control classifier, used in the first step, is discussed in Sec. 2.4.1, and
our second classifier, used to discriminate converter-by-trajectory (CT) and
nonconverter-by-trajectory (NT) classes, is discussed in Sec. 2.4.2.
2.4.1. AD-Control classifier
For the AD training population, we chose individual AD visit images with
a CDR score of at least 1. For the Control training population, on the other
hand, we only chose initial visits, and only those for participants who stayed
at CDR=0 throughout all their visits. Thus, we excluded Controls with
"questionable dementia" (i.e., CDR=0.5) at any visit. By these choices, we
sought to exclude outlier examples or even possibly any mislabeled examples,
recalling that CDR for the majority of both AD and MCI participants is 0.5
throughout all visits.
2.4.2. CT-NT classifier
Fig. 1 gives an illustrative example of the phenotypical score trajecto-
ries for MCI subject described in the Introduction. A positive score is on the
Control side and a negative score is on the AD side -- the x-axis represents the
AD-Control SVM's decision boundary. Score vs. age is plotted, with each line
segment a trajectory obtained by linearly fitting an individual's phenotype
scores (and linearly extrapolating if there are missing visits). Nonconverters-
by-CDR (N-CDR) and converters-by-CDR (C-CDR) are illustrated in (a)
and (b), respectively. Green and black subjects are those whose fitted tra-
jectory stayed on the Control side and AD side, respectively, whereas gray
lines are subjects who crossed to the AD side. Thus, by our conversion-by-
trajectory definition, the green group is the nonconverters-by-trajectory, and
the black and gray groups together are the converters-by-trajectory. Subject
counts for these groups are given in the figure legends. The outlier sub-
jects are shown in orange -- there are five, making up less than 2% of the
MCI cohort. Notice, intriguingly, from the left figure that more than one
third of all (non-orange) MCI patients (106 of 298) are converters by tra-
jectory and yet nonconverters according to CDR -- i.e., there is a very large
percentage of patients for which the two converter definitions disagree, with
the neuroimage-based definition indicating disease state changes that are not
12
(a)
13
(b)
Figure 1: AD-Control SVM score trajectories for MCI subjects. (a) Nonconverters-by-
CDR. (b) Converters-by-CDR.
405060708090100(cid:239)2(cid:239)1.5(cid:239)1(cid:239)0.500.511.52agescore 144 (56.9%)42 (16.6%)64 (25.3%)3 (1.2%)405060708090100(cid:239)3(cid:239)2.5(cid:239)2(cid:239)1.5(cid:239)1(cid:239)0.500.511.5agescore 11 (22.0%)14 (28.0%)23 (46.0%)2 (4.0%) predicted using the clinical, CDR-based definition. Likewise, an additional
3% of all MCI subjects (11 of 298) "defy" their by-CDR converter label in
that they do not reach the AD side of the decision boundary.
Based on these trajectories, i.e. whether or not the AD side is visited, we
derive the ground-truth 'CT' and 'NT' labels for all MCI subjects. We then
build a CT-NT classifier using as input only the image scans at initial visit.8
3. Results and Discussion
3.1. Introductory overview
In this section we will perform 1) classification experiments to evaluate
conversion-by-trajectory and conversion-by-CDR with respect to desidera-
tum 2; 2) additional experiments to compare the two definitions with re-
spect to desideratum 3; and lastly, 3) experiments to identify prognostic
brain "biomarker" regions.
It is important to mitigate the potential confounding effect of the subject's
age. In our classification experiments, we mitigated in two ways:
1) For every classifier training, each training sample in one class was
uniquely paired via "age-matching" with a training sample in the other class
(with age separation at most one year).
2) For every linear-kernel SVM classifier, we separately adjusted each
feature for age prior to classification using linear fitting. We subtracted
the extrapolated line (computed only using "control" samples9) from the
feature's value, for all (training and test) samples. As an aside, we note
that, given the subsequent linear SVM operation, the representation power
of this linear fitting step is essentially equivalent to simply treating age as
an additional feature input to the linear SVM classifier.
Finally, prior to building classifiers, we normalized feature values to the
[0,1] range, which is suitable for the LIBSVM software (Chang et al., 2001)
we used for training SVMs.
8For a small percentage of the MCI subjects, we did not obtain the patient's first visit.
However, we did ensure that the visit we took as the "initial visit" had a CDR of 0.5.
9For the AD-Control classifier, these are the samples in the Control class. For the
CT-NT classifier we computed the line using only the NT samples.
14
(a) By-CDR.
(b) By-trajectory.
(c) Overlap.
Figure 2: A population of 298 MCI subjects in ADNI is shown here, broken up according to
the two criteria discussed in Sec. 3.2.1: (a) by-CDR criterion, (b) by-trajectory criterion;
with overlap shown in (c).
3.2. Experiments with voxel-based features
The test set (generalization) accuracy of the voxel-based AD-Control clas-
sifier, built using 70 training samples per class, was 0.89 (86 of 88 Controls,
and 16 of 27 AD subjects, were correctly classified.) This classifier, with high
specificity for Controls, was then applied to a population of MCI subjects to
determine the CT and NT subgroups.
3.2.1. Classification experiments for the MCI population
Fig. 2(a) shows the sizes of the converter-by-CDR (C-CDR) and nonconverter-
by-CDR (N-CDR) groups within the ADNI MCI cohort for a typical exper-
iment in our work. Fig. 2(b) shows the same population broken up as
converters-by-trajectory (CT) and nonconverters-by-trajectory (NT). Super-
imposing the two charts, Fig. 2(c) illustrates their overlap, where converters
by both definitions are accordingly indicated by orange. Since converters-
by-CDR are relatively scarce, we used a large majority of them (80%, i.e. 39
individuals among the 48) for the by-CDR classifier's training set, with the
rest (20%) put into the test set. We reiterate that a general disadvantage of
the by-CDR approach is its scarcity of converter examples -- by contrast, a
more balanced number of examples is available for by-trajectory training (at
least 100, rather than 39, training samples per class, as in Fig. 2(b)). Note
also that if we were to use a "probable AD", rather than a by-CDR converter
definition, there would be even fewer converter examples.
A fair performance comparison between by-trajectory and by-CDR clas-
sification requires: 1) using the same per-class training set size (i.e. 39) for
15
(a)By-CDR.(b)By-trajectory.(c)Overlap.Figure2:Apopulationof298MCIsubjectsinADNI(approximately75%oftheMCIpopulationsizeof400targetedbyADNI)isshownhere,brokenupaccordingtothetwocriteriadiscussedinSec.3.2.1:(a)by-CDRcriterion,(b)by-trajectorycriterion;withoverlapshownin(c).samesets);11ii)wecanmakethetrainingsetsofthetwoclas-sifiersidenticalratherthanmerelysame-sized,aswellasmakethetestsetsidentical.Thislatterapproach,though,willhavesomebiasbecause,inselectingsamplesfortheby-trajectoryclassifier,wewillhavetomakeuseofknowledgeofthesam-ples'conversion-by-CDRstatus(andviceversafortheby-CDRclassifier).Thefirstapproach,ontheotherhand,clearlydoesnothavethisbias.Asbothapproachesarevalidwaysofdealingwithby-CDRdatalimitations,wewillcomparegeneralizationaccuraciesofby-CDRandby-trajectoryclassifiersunderboththesedataselectionschemes,referringtotheseapproachesas"random"and"identical"inthesequel.Ourtraining/testsetselectionprocedureforthe"identicalap-proach"isasfollows.FortheC-CDR-CTgroup(Fig.2(c)),randomlyselect80%ofthegroup(theyellowstripedgroupofsize30inFig.3(a))suchthatacorrespondinggroupwithinN-CDR(whiteportioninFig.2(a))canbefoundthatisbothNTandsatisfiesage-matching.Thiscorrespondinggroupisillus-tratedinFig.3(a)asthewhitestripedgroup(ofsize30),placedoppositefromtheyellowstripedareaitispaired(matched)with.LikewisefortheC-CDR-NTgroup(Fig.2(c)),randomlyselect80%ofthegroup(theyellowstripedgroupofsize9inFig.3(a))suchthatacorrespondinggroupwithinN-CDRcanbefoundthatisbothCTandsatisfiesage-matching.Thiscor-respondinggroupisillustratedinFig.3(a)asthewhitestripedgroup(ofsize9).NoticebycomparingthisfiguretoFig.2(c)thatthetwowhitestripedareasareseparatedbytheCT-NTborder.Wetakethetrainingset -- sharedbytheby-CDRandby-trajectoryclassifiers -- tobepreciselytheunionofthesefourstripedareas.12Subsequently,wetakethetestset -- sharedbythetwoclassifiers -- tobethesubjectswhoareneitherin1)thetrainingset(stripedareas)norin2)thespecialsetofsubjectsshowninsolidgrayinFig.3(a)(alsoshownidenticallyinFig.11Notethatthismeansthatthetrainingsetsfortheconverterandnoncon-verterclassesoftheconversion-by-CDRclassifierarerandomlyselectedfromtheyellowandwhiteregionsinFig.2(a),respectively,withnoconsiderationoftrajectory-based(i.e.red/white)labelingillustratedinFig.2(b).12Fortheby-CDRclassifier,theclassmembershipofanyofthesefoursub-setsofthetrainingsetisillustratedbythecolorbeingyelloworwhiteinFig.3(a).Likewise,fortheby-trajectoryclassifier,classmembershipisillustratedbyredorwhitecolorinFig.3(b).3(b)).13Thatis,thetestsetisthetiledareasinFig.3(a)(or,identically,inFig.3(b)).Noteabovethatsomerandomselectionisbeingemployedinchoosingthetraining/testsetseveninthe"identicalapproach"(whereas,inthe"randomapproach"theselectioniscompletelyrandom).Thus,forbothapproaches,theaccuracyofperfor-mancecomparisonwillbenefitfromaveragingaccuracyresultsovermultipletraining/testsplit"trials".Resultsaveragedacross10trialsaregiveninFig.3(c)foralinear-kernelSVM14;µ±σnotationisusedtoindicatethemeanµandstandarddeviationσofquantitiesacrossthetrials.15Notethatby-trajectory'sgen-eralizationperformanceisashighas0.83,whereasby-CDR'sgeneralizationperformanceisverypoor -- aspoorasrandomguessing(see0.5and0.56tablevalues) -- duemainlytopoorperformanceonnonconverters-by-CDR.Fig.3(d)andFig.3(e)showby-trajectoryandby-CDRresults,respectively,foroneofthe10trials(forthe"randomapproach"),witheachbarindi-catingdistancetotheclassificationboundary16foranMCIsub-jectinatestpopulationofsize88andnonconverters/convertersshowninleft/rightfigures,respectively.Amongthe88subjects,by-trajectorycorrectlyclassified79whereasby-CDRcorrectlyclassifiedonly40.Inaseparateexperiment,weevaluatedusingoneof27sub-samples(ratherthanoneof216subsamples),i.e.essentiallya10-foldincreaseinthenumberof(voxel-based)features,andfoundthattheby-trajectorygeneralizationaccuracyroseto0.91inthe"random"case.Wethentriedbuilding27separateby-converterclassifiers,oneforevery1/27thsubsample(thusef-fectivelyusingthewhole3Dimage),withmajority-basedvot-ingusedtocombinethe27decisions.Thisensembleschemeagainachieved0.91accuracy,i.e.therewasnofurtheraccuracy13Weexcludethis"specialset"(ingray)fromthetestsetsothatallourexperimentsunderthe"identicalapproach"canhaveashared,fixedtestset(forfaircomparisonwitheachother),including,crucially,anexperimentthatwillincludethis"specialset"ofsamplesinthetrainingset.14Forgeneratingallclassificationresultsherein,includingthoseinFig.3(c),weusedSVMclassifiersthatwerebuiltbyemployingthecommonapproachofbootstrap-basedvalidationforselectingtheclassifier's(trial's)hyperparametervalues(Aksuetal.,2010).15µandσofintegerquantities(e.g.samplecounts)areshownroundedup.16Positive/negativedistancemeansnonconverter/convertersideofthebound-ary,respectively.7 both by-CDR and by-trajectory training, and 2) making the test set sizes
the same for both classifiers. There are several different ways in which the
data can be partitioned into training and test sets, consistent with these two
conditions: i) we can perform simple random selection on a class-by-class ba-
sis, ensuring only that the two classifiers are given the same training/test set
sizes (but not the same sets);10 ii) we can make the training sets of the two
classifiers identical rather than merely same-sized, as well as make the test
sets identical. This latter approach, though, will have some bias because, in
selecting samples for the by-trajectory classifier, we will have to make use of
knowledge of the samples' conversion-by-CDR status (and vice versa for the
by-CDR classifier). The first approach, on the other hand, clearly does not
have this bias. As both approaches are valid ways of dealing with by-CDR
data limitations, we will compare generalization accuracies of by-CDR and
by-trajectory classifiers under both these data selection schemes, referring to
these approaches as "random" and "identical" in the sequel.
Our training/test set selection procedure for the "identical approach"
is as follows. For the C-CDR-CT group (Fig. 2(c)), randomly select 80%
of the group (the yellow striped group of size 30 in Fig. 3(a)) such that
a corresponding group within N-CDR (white portion in Fig. 2(a)) can be
found that is both NT and satisfies age-matching. This corresponding group
is illustrated in Fig. 3(a) as the white striped group (of size 30), placed
opposite from the yellow striped area it is paired (matched) with. Likewise
for the C-CDR-NT group (Fig. 2(c)), randomly select 80% of the group (the
yellow striped group of size 9 in Fig. 3(a)) such that a corresponding group
within N-CDR can be found that is both CT and satisfies age-matching. This
corresponding group is illustrated in Fig. 3(a) as the white striped group (of
size 9). Notice by comparing this figure to Fig. 2(c) that the two white
striped areas are separated by the CT-NT border. We take the training set
-- shared by the by-CDR and by-trajectory classifiers -- to be precisely the
union of these four striped areas.11 Subsequently, we take the test set --
10Note that this means that the training sets for the converter and nonconverter classes
of the conversion-by-CDR classifier are randomly selected from the yellow and white re-
gions in Fig. 2(a), respectively, with no consideration of trajectory-based (i.e. red/white)
labeling illustrated in Fig. 2(b).
11For the by-CDR classifier, the class membership of any of these four subsets of the
training set is illustrated by the color being yellow or white in Fig. 3(a). Likewise, for the
by-trajectory classifier, class membership is illustrated by red or white color in Fig. 3(b).
16
shared by the two classifiers -- to be the subjects who are neither in 1) the
training set (striped areas) nor in 2) the special set of subjects shown in solid
gray in Fig. 3(a) (also shown identically in Fig. 3(b)).12 That is, the test set
is the tiled areas in Fig. 3(a) (or, identically, in Fig. 3(b)).
Note above that some random selection is being employed in choosing the
training/test sets even in the "identical approach" (whereas, in the "random
approach" the selection is completely random). Thus, for both approaches,
the accuracy of performance comparison will benefit from averaging accuracy
results over multiple training/test split "trials". Results averaged across 10
trials are given in Fig. 3(c) for a linear-kernel SVM13; µ± σ notation is used
to indicate the mean µ and standard deviation σ of quantities across the
trials.14 Note that by-trajectory's generalization performance is as high as
0.83, whereas by-CDR's generalization performance is very poor -- as poor
as random guessing (see 0.5 and 0.56 table values) -- due mainly to poor
performance on nonconverters-by-CDR. Fig. 3(d) and Fig. 3(e) show by-
trajectory and by-CDR results, respectively, for one of the 10 trials (for the
"random approach"), with each bar indicating distance to the classification
boundary15 for an MCI subject in a test population of size 88 and non-
converters/converters shown in left/right figures, respectively. Among the
88 subjects, by-trajectory correctly classified 79 whereas by-CDR correctly
classified only 40.
Recently, similarly poor by-CDR classification performance was also re-
ported in (Davatzikos et al., 2010), where it was found that the majority
of (by-CDR) nonconverters "had sharply positive SPARE-AD scores indi-
cating significant atrophy similar to AD patients". Since the SPARE-AD
score is produced by a classifier that was trained to discriminate Control and
AD patients (Fan et al., 2007, 2008), this comment and associated results
are consistent both with our conjecture in the Introduction and our above
12We exclude this "special set" (in gray) from the test set so that all our experiments
under the "identical approach" can have a shared, fixed test set (for fair comparison with
each other), including, crucially, an experiment that will include this "special set" of
samples in the training set.
13For generating all classification results herein, including those in Fig. 3(c), we used
SVM classifiers that were built by employing the common approach of bootstrap-based
validation for selecting the classifier's (trial's) hyperparameter values (Aksu et al., 2010).
14µ and σ of integer quantities (e.g. sample counts) are shown rounded up.
15Positive/negative distance means nonconverter/converter side of the boundary, re-
spectively.
17
histogram results, which suggest that there may be a significant number of
patients undergoing physiological brain changes consistent with conversion,
yet without clinical manifestation.
The results above indicate that the conversion-by-CDR definition's two
classes are not well-discriminated, and thus, clinical usefulness of this def-
inition for our prognostic Aim is expected to be poor. The much greater
generalization accuracy of the by-trajectory definition (coupled with its in-
herent plausibility as a conversion definition) indicates its greater utility.
Increasing the By-Trajectory Image-Based Feature Resolution:
In a separate experiment, we evaluated using one of 27 subsamples (rather
than one of 216 subsamples), i.e. essentially a 10-fold increase in the number
of (voxel-based) features, and found that the by-trajectory generalization ac-
curacy rose to 0.91 in the "random" case. We then tried building 27 separate
by-converter classifiers, one for every 1/27th subsample (thus effectively us-
ing the whole 3D image), with majority-based voting used to combine the
27 decisions. This ensemble scheme again achieved 0.91 accuracy, i.e. there
was no further accuracy benefit beyond that from an ≈ 10-fold increase in
the number of voxel features.
Increasing the By-Trajectory Training Set Size
Note that the converter-by-CDR sample scarcity and class-balancing (via
age-matching) in the experiments above had the effect of artificially limiting
the by-trajectory classifier training set size. Next we investigated how much
the generalization accuracy of by-trajectory classification improves when this
limitation is removed. The tiled areas in Figures 4(a) and 3(b) are identical,
illustrating that in this new experiment (Fig. 4(a)) we used the same test set
as previously, for fairness of comparison. However, as indicated by differences
in the total striped area between these two charts, we now make the training
set much larger than previously. Specifically, for the "identical" case, we used
the previous 10 trials but simply augmented a trial's training set with the
two large, previously-excluded gray sets16, as these two sets do age-match
each other. The results, averaged across the 10 trials, are given in Fig.
4(b). Notice in this figure the now larger per-class training size (on average
≈ 100 rather than 39), and that the random approach uses this size as
well. The by-trajectory results in Fig. 4(b) indicate that accuracy improved
from 0.78 ± 0.02 (Fig. 3(c)) to 0.84 ± 0.02 for the "identical approach",
16Shown in Fig. 3(b), with size 60. Note that this size can vary from trial to trial.
18
(a) Training/test set selec-
tion for by-CDR classifica-
tion.
(b) Training/test set selec-
tion for by-trajectory clas-
sification.
Sample
selection
Classifier
Converters
Test set
Count Accuracy Count Accuracy
0.84 ± 0.05
46 ± 1
0.47 ± 0.04
9 ± 1
0.78 ± 0.04
45 ± 1
9 ± 1
0.54 ± 0.04
Nonconverters
45 ± 1
82 ± 2
47 ± 2
82 ± 2
0.81 ± 0.06
0.79 ± 0.07
0.78 ± 0.04
0.74 ± 0.11
Overall
accuracy
0.83 ± 0.05
0.50 ± 0.04
0.78 ± 0.02
0.56 ± 0.04
Random By-trajectory
approach
By-CDR
By-trajectory
Identical
approach
By-CDR
(c) Average test set classification accuracy using all 11, 293 features, 39±1 per-class
training samples.
(d) By-trajectory. Left: nonconverters;
Right: converters
(e) By-CDR. Left: nonconverters; Right:
converters
Figure 3: Test set accuracy comparison of by-CDR and by-trajectory classification.
19
(a)Training/testsetselectionforby-CDRclassifi-cation.(b)Training/testsetselectionforby-trajectoryclas-sification.SampleClassifierTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyRandomBy-trajectory46±10.81±0.0645±10.84±0.050.83±0.05approachBy-CDR9±10.79±0.0782±20.47±0.040.50±0.04IdenticalBy-trajectory45±10.78±0.0447±20.78±0.040.78±0.02approachBy-CDR9±10.74±0.1182±20.54±0.040.56±0.04(c)Averagetestsetclassificationaccuracyusingall11,293features,39±1per-classtrainingsamples.02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 41, nonpositives: 3]sample indexdistance02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 6, nonpositives: 38]sample indexdistance(d)By-trajectory.Left:nonconverters;Right:con-verters050−1.5−1−0.500.51 79 test samples [positives: 33, nonpositives: 46]sample indexdistance123456789−1.5−1−0.500.51 9 test samples [positives: 2, nonpositives: 7]sample indexdistance(e)By-CDR.Left:nonconverters;Right:convertersFigure3:Testsetaccuracycomparisonofby-CDRandby-trajectoryclassification.9(a)Training/testsetselectionforby-CDRclassifi-cation.(b)Training/testsetselectionforby-trajectoryclas-sification.SampleClassifierTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyRandomBy-trajectory46±10.81±0.0645±10.84±0.050.83±0.05approachBy-CDR9±10.79±0.0782±20.47±0.040.50±0.04IdenticalBy-trajectory45±10.78±0.0447±20.78±0.040.78±0.02approachBy-CDR9±10.74±0.1182±20.54±0.040.56±0.04(c)Averagetestsetclassificationaccuracyusingall11,293features,39±1per-classtrainingsamples.02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 41, nonpositives: 3]sample indexdistance02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 6, nonpositives: 38]sample indexdistance(d)By-trajectory.Left:nonconverters;Right:con-verters050−1.5−1−0.500.51 79 test samples [positives: 33, nonpositives: 46]sample indexdistance123456789−1.5−1−0.500.51 9 test samples [positives: 2, nonpositives: 7]sample indexdistance(e)By-CDR.Left:nonconverters;Right:convertersFigure3:Testsetaccuracycomparisonofby-CDRandby-trajectoryclassification.9(a)Training/testsetselectionforby-CDRclassifi-cation.(b)Training/testsetselectionforby-trajectoryclas-sification.SampleClassifierTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyRandomBy-trajectory46±10.81±0.0645±10.84±0.050.83±0.05approachBy-CDR9±10.79±0.0782±20.47±0.040.50±0.04IdenticalBy-trajectory45±10.78±0.0447±20.78±0.040.78±0.02approachBy-CDR9±10.74±0.1182±20.54±0.040.56±0.04(c)Averagetestsetclassificationaccuracyusingall11,293features,39±1per-classtrainingsamples.02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 41, nonpositives: 3]sample indexdistance02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 6, nonpositives: 38]sample indexdistance(d)By-trajectory.Left:nonconverters;Right:con-verters050−1.5−1−0.500.51 79 test samples [positives: 33, nonpositives: 46]sample indexdistance123456789−1.5−1−0.500.51 9 test samples [positives: 2, nonpositives: 7]sample indexdistance(e)By-CDR.Left:nonconverters;Right:convertersFigure3:Testsetaccuracycomparisonofby-CDRandby-trajectoryclassification.9(a)Training/testsetselectionforby-CDRclassifi-cation.(b)Training/testsetselectionforby-trajectoryclas-sification.SampleClassifierTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyRandomBy-trajectory46±10.81±0.0645±10.84±0.050.83±0.05approachBy-CDR9±10.79±0.0782±20.47±0.040.50±0.04IdenticalBy-trajectory45±10.78±0.0447±20.78±0.040.78±0.02approachBy-CDR9±10.74±0.1182±20.54±0.040.56±0.04(c)Averagetestsetclassificationaccuracyusingall11,293features,39±1per-classtrainingsamples.02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 41, nonpositives: 3]sample indexdistance02040−2.5−2−1.5−1−0.500.511.5 44 test samples [positives: 6, nonpositives: 38]sample indexdistance(d)By-trajectory.Left:nonconverters;Right:con-verters050−1.5−1−0.500.51 79 test samples [positives: 33, nonpositives: 46]sample indexdistance123456789−1.5−1−0.500.51 9 test samples [positives: 2, nonpositives: 7]sample indexdistance(e)By-CDR.Left:nonconverters;Right:convertersFigure3:Testsetaccuracycomparisonofby-CDRandby-trajectoryclassification.9 (a) Training/test set selection for
by-trajectory classification.
Sample
selection
Converters
Count Accuracy Count Accuracy
0.86 ± 0.02
By-trajectory Random 45 ± 1
45 ± 1
0.83 ± 0.05
Identical
0.77 ± 0.01
0.85 ± 0.03
(b)
Test set
Nonconverters
47 ± 2
47 ± 2
Overall
accuracy
0.82 ± 0.01
0.84 ± 0.02
Figure 4: By-trajectory average test set classification accuracy for the larger training set
size (98 ± 1 per class) and 11, 293 features.
and modestly worsened for the "random approach" (from 0.83 ± 0.05 to
0.82 ± 0.01).
Evaluating a Third Conversion Definition
We now consider a third definition of conversion that combines the first
two definitions as follows. Let "converters" consist of individuals who con-
verted either by-trajectory or by-CDR (non-white areas in Fig. 2(c)), with
the "nonconverter" class consisting of the remaining MCI individuals (white
area). The point of view of this new definition, "conversion-by-union", is to
be more inclusive in defining an MCI subpopulation at risk, which may ben-
efit from early treatment or diagnostic testing. While from that perspective
the new definition is reasonable, the fact that grouping individuals by CDR
has a role in this definition may be its disadvantage, considering that by-CDR
classification was previously shown to perform not much better than random
guessing. Results, averaged across the same 10 trials used in Fig. 3, are
given in Table 1 and indicate that conversion-by-union generalizes somewhat
20
(a)Training/testsetselectionforby-trajectoryclas-sification.SampleTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyBy-trajectoryRandom45±10.77±0.0147±20.86±0.020.82±0.01Identical45±10.85±0.0347±20.83±0.050.84±0.02(b)Figure4:By-trajectoryaveragetestsetclassificationaccuracyforthelargertrainingsetsize(98±1perclass)and11,293features.−4−2024x 10−300.10.20.30.40.50.60.70.80.91normalized volume of hippocampus(normalized) count 250 [m=−0.015, s=1.004]48 [m=−0.654, s=0.924]155 [m=0.399, s=0.896]143 [m=−0.678, s=0.828]Figure5:Forthehippocampus,by-trajectory(red)haslargerhistogramseparationbetweenconverter(dashedline)andnonconverter(solidline)groupsthanby-CDR(blue).Toillustratethismoreclearly,alsoshownistheGaussiancurveforeachofthesefoursubjectgroups(plottedbasedongroupmean(m)andstandarddeviation(s)indicatedinthefigurelegendwiththesame0.001scalingasthex-axis).10 Sample
selection
Converters
By-union
Identical
Count Accuracy Count Accuracy
48 ± 1
0.75 ± 0.06
0.78 ± 0.05
Test set
Nonconverters
44 ± 2
Overall
accuracy
0.77 ± 0.03
Table 1: Average test set accuracy of by-union classification for 39 ± 1 per-class training
samples and 11, 293 features.
worse than conversion-by-trajectory.17
3.2.2. Validation on Known AD Conversion Biomarkers
To validate the proposed conversion definitions with respect to desider-
atum 3, we performed correlation tests on the MCI population between the
binary class variable C ∈ {0 = no conversion, 1 = conversion} and known
AD conversion biomarkers consisting of both 1) volume in reported AD-
affected regions (Table 2 in (Schuff et al., 2010)), which we measured for
each individual's final-visit MRI18 and 2) the clinical MMSE measure. The
stronger the correlation, the more accurately the biomarker is predicted from
the class variable and the greater the separation between the biomarker his-
tograms, conditioned on the two classes.
Before presenting correlation test results, we first illustrate in Fig. 5
the increased separation of the histograms of hippocampus volume for the
converter and nonconverter groups in the by-trajectory case, compared with
by-CDR. Next, we performed comprehensive statistical tests for a number of
suggested AD biomarkers. The R statistical computing package was used to
perform all tests with statistical significance set at the 0.05 level. In Table 2a
17Note that "by-union" is, by definition, an instance of the "identical approach", as
stated in Table 1. To ensure fairness of comparison with the by-trajectory definition of
conversion, our test sets, and training set sizes, in these two cases were identical. In fact,
we chose the by-union training set to be as similar to by-trajectory's, in every trial, as
possible. Referring to Fig. 3(b) (which represents a trial example), the by-union training
set was chosen to include 1) the two large striped groups (red and white); 2) the small
"special" gray group (of size seven in this trial example) and its age-matched counterpart
within small white-striped group, and; 3) a subset of the second small "special" gray group
(two of five individuals in this trial example) and its age-matched counterpart within the
small red-striped group.
18As discussed in Appendix A, we measured normalized region volume. Note also that
our regions are defined based on the atlas (Atlas2) we used. The correspondence between
the regions in (Schuff et al., 2010) and our defined regions is given in Appendix B. Finally,
note that a subject's final visit is not always the sixth visit.
21
Figure 5: For the hippocampus, by-trajectory (red) has larger histogram separation be-
tween converter (dashed line) and nonconverter (solid line) groups than by-CDR (blue).
To illustrate this more clearly, also shown is the Gaussian curve for each of these four
subject groups (plotted based on group mean (m) and standard deviation (s) indicated in
the figure legend with the same 0.001 scaling as the x-axis).
22
(a)Training/testsetselectionforby-trajectoryclas-sification.SampleTestsetselectionConvertersNonconvertersOverallCountAccuracyCountAccuracyaccuracyBy-trajectoryRandom45±10.77±0.0147±20.86±0.020.82±0.01Identical45±10.85±0.0347±20.83±0.050.84±0.02(b)Figure4:By-trajectoryaveragetestsetclassificationaccuracyforthelargertrainingsetsize(98±1perclass)and11,293features.−4−2024x 10−300.10.20.30.40.50.60.70.80.91normalized volume of hippocampus(normalized) count 250 [m=−0.015, s=1.004]48 [m=−0.654, s=0.924]155 [m=0.399, s=0.896]143 [m=−0.678, s=0.828]Figure5:Forthehippocampus,by-trajectory(red)haslargerhistogramseparationbetweenconverter(dashedline)andnonconverter(solidline)groupsthanby-CDR(blue).Toillustratethismoreclearly,alsoshownistheGaussiancurveforeachofthesefoursubjectgroups(plottedbasedongroupmean(m)andstandarddeviation(s)indicatedinthefigurelegendwiththesame0.001scalingasthex-axis).10 (and Fig. 6), the correlation coefficients for by-trajectory and by-CDR are
shown for each biomarker, along with their associated p-values (Chambers
et al., 1993). Note that for 11 out of 14 brain regions, the correlation with
by-trajectory is greater than the correlation with by-CDR (in bold), with
by-trajectory meeting the significance threshold in 10 of these 11 regions.
Further, for only two of the remaining four biomarkers - posterior cingulate
and the clinical MMSE measure - does the correlation with by-CDR meet the
significance threshold. Most notably, well-established markers for AD such
as the hippocampus, lateral ventricles, and inferior parietal exhibited strong
correlation with the by-trajectory definition.
To further assess statistical significance of the comparison between by-
trajectory and by-CDR correlations, we performed a correlated correlation
test (Meng et al., 1992), the appropriate test given that the same MCI sam-
ple population was used in measuring correlations for both by-CDR and
by-trajectory. This test (Table 2b) reveals that the larger correlation of by-
trajectory is statistically significant at the 0.05 level in nine brain regions (in
bold). By contrast, conversion-by-CDR does not achieve a statistically sig-
nificant advantage for any of the brain regions, nor with respect to MMSE.
To mitigate the confounding effect of age in the AD biomarker measurements,
we linearly adjusted them for age prior to performing the experiments de-
scribed above. We also repeated these experiments without age adjustment.
In this case, the number of regions for which by-trajectory's correlation ex-
ceeded by-CDR's rose from 11 to 12 (with the addition of the entorhinal
cortex), and the number of regions for which the larger correlation of by-
trajectory was statistically significant rose from 9 to 10 (with the addition of
the medial orbitofrontal GM).
3.2.3. Identification of prognostic brain "biomarker" regions
In the previous section, we validated conversion definitions using estab-
lished (diagnostic) AD biomarker brain regions (with volumes measured at
final visit). In this section, we will identify key prognostic biomarker brain
regions via supervised feature selection, aiming first to identify the "essen-
tial" subset of voxel features, i.e. the voxels (at initial visit) necessary for
our classifier to well-discriminate the CT and NT classes. The brain regions
(consistent with a registered brain atlas) within which these select voxels
principally reside then identify our prognostic brain biomarker regions. Sim-
ilarly, we will identify diagnostic regions, critical for discriminating between
AD and Control subjects (using our AD-Control classifier). In both cases,
23
(a)
By-trajectory
By-CDR
Correlation P-value Correlation P-value
Biomarker
Entorhinal cortex
Fusiform gyrus
Hippocampus
Inferior parietal GM
Lateral orbitofrontal GM
Lateral ventricles
Medial orbitofrontal GM
Parahippocampal gyrus
Posterior cingulate
Precentral GM
Superior frontal GM
Superior temporal GM
Total GM
Total WM
MMSE
coefficient
0.041
0.430
0.530
0.513
0.077
0.586
0.239
0.095
0.001
0.173
0.331
0.478
0.491
0.170
0.356
0.486
8.20E-15
2.00E-16
2.00E-16
0.185
2.00E-16
3.01E-05
0.102
0.981
0.003
4.50E-09
2.00E-16
2.00E-16
0.003
2.57E-10
coefficient
0.050
0.152
0.231
0.097
0.078
0.211
0.101
0.060
0.144
0.088
0.094
0.161
0.246
0.041
0.453
0.385
0.008
5.67E-05
0.094
0.178
0.000246
0.081
0.307
0.013
0.128
0.104
0.005
1.70E-05
0.477
2.00E-16
(b)
P-value
0.889
2.17E-06
2.42E-10
0.985
1.34E-09
0.048
0.616
0.045
0.230
0.001
3.10E-05
0.0001
0.069
1.33E-06
0.118
Table 2: Correlation coefficients and associated p-values: (a) Correlation test results; (b)
Correlated correlation test results for each of the regions in (a). Statistically significant
results are shown in bold.
24
Figure 6: Comparison of correlation coefficients for by-trajectory and by-CDR.
AD-Control classifier only
Amygdala left
Cingulate region right
Entorhinal cortex right
Inferior occipital gyrus right
Medial occipitotemporal gyrus left
Parahippocampal gyrus left
Temporal lobe WM right
Temporal pole right
intersection
Hippocampal formation* left
Hippocampal formation* right
Entorhinal cortex* left
Inferior temporal gyrus right
Lateral occipitotemporal gyrus* right
Parahippocampal gyrus* right
Perirhinal cortex left
Perirhinal cortex right
Middle temporal gyrus right
Uncus left
CT-NT classifier only
Superior temporal gyrus* left
Middle temporal gyrus left
Precuneus right
Lateral front-orbital gyrus* right
Insula right
Supramarginal gyrus* left
Temporal lobe WM left
Temporal pole left
Medial front-orbital gyrus* left
Table 3: Brain regions identified as biomarkers using voxel-based features and MFE.
25
00.10.20.30.40.5correlation coefficient entorhinal cortexhippocampusinferior parietal GMlateral orbitofrontal GMlateral ventriclesmedial orbitofrontal GMparahippocampal gyrusposterior cingulateprecentral GMsuperior frontal GMfusiform gyrustotal GMtotal WMtransverse temporal GMMMSEby−trajectoryby−CDRFigure6:Comparisonofcorrelationcoefficientsforby-trajectoryandby-CDR.Tofurtherassessstatisticalsignificanceofthecomparisonbetweenby-trajectoryandby-CDRcorrelations,weperformedacorrelatedcorrelationtest(Mengetal.,1992),theappropri-atetestgiventhatthesameMCIsamplepopulationwasusedinmeasuringcorrelationsforbothby-CDRandby-trajectory.Thistest(Table2b)revealsthatthelargercorrelationofby-trajectoryisstatisticallysignificantatthe0.05levelinninebrainregions(inbold).Bycontrast,conversion-by-CDRdoesnotachieveastatisticallysignificantadvantageforanyofthebrainregions,norwithrespecttoMMSE.Tomitigatethecon-foundingeffectofageintheADbiomarkermeasurements,welinearlyadjustedthemforagepriortoperformingtheexper-imentsdescribedabove.Wealsorepeatedtheseexperimentswithoutageadjustment.Inthiscase,thenumberofregionsforwhichby-trajectory'scorrelationexceededby-CDR'srosefrom11to12(withtheadditionoftheentorhinalcortex),andthenumberofregionsforwhichthelargercorrelationofby-trajectorywasstatisticallysignificantrosefrom9to10(withtheadditionofthemedialorbitofrontalGM).3.2.3.Identificationofkeyprognosticbrain"biomarker"re-gionsIntheprevioussection,wevalidatedconversiondefini-tionsusingestablished(diagnostic)ADbiomarkerbrainregions(withvolumesmeasuredatfinalvisit).Inthissection,wewillidentifykeyprognosticbiomarkerbrainregionsviasupervisedfeatureselection,aimingfirsttoidentifythe"essential"subsetofvoxelfeatures,i.e.thevoxels(atinitialvisit)necessaryforourclassifiertowell-discriminatetheCTandNTclasses.Thebrainregions(consistentwitharegisteredbrainatlas)withinwhichtheseselectvoxelsprincipallyresidethenidentifyourprognosticbrainbiomarkerregions.Similarly,wewilliden-tifydiagnosticregions,essentialfordiscriminatingbetween(a)By-trajectoryBy-CDRCorrelationP-valueCorrelationP-valueBiomarkercoefficientcoefficientEntorhinalcortex0.0410.4860.0500.385Fusiformgyrus0.4308.20E-150.1520.008Hippocampus0.5302.00E-160.2315.67E-05InferiorparietalGM0.5132.00E-160.0970.094LateralorbitofrontalGM0.0770.1850.0780.178Lateralventricles0.5862.00E-160.2110.000246MedialorbitofrontalGM0.2393.01E-050.1010.081Parahippocampalgyrus0.0950.1020.0600.307Posteriorcingulate0.0010.9810.1440.013PrecentralGM0.1730.0030.0880.128SuperiorfrontalGM0.3314.50E-090.0940.104SuperiortemporalGM0.4782.00E-160.1610.005TotalGM0.4912.00E-160.2461.70E-05TotalWM0.1700.0030.0410.477MMSE0.3562.57E-100.4532.00E-16(b)P-value0.8892.17E-062.42E-100.9851.34E-090.0480.6160.0450.2300.0013.10E-050.00010.0691.33E-060.118Table2:Correlationcoefficientsandassociatedp-values.11 the accuracy of the selected brain region biomarkers rests heavily on the ac-
curacy of the supervised feature selection algorithm we employ. In Figures
7(a) and 7(b), we compare MFE and RFE feature elimination (i.e. feature
selection via feature elimination) for both Control-AD classification and for
CT-NT classification (for one representative, example trial). The curves show
test set accuracy as a function of the number of retained features (which is
reduced going from right to left). Note that the "MFE/MFE-slack" hybrid
method (Aksu et al., 2010)) outperforms RFE for both brain classification
tasks, achieving lower test set error rates, and with much fewer retained fea-
tures. The circle, determined without use of the test set based on the rule in
(Aksu et al., 2010), marks the point at which we stopped eliminating features
by MFE, thus determining the (trial's) retained voxel set. This MFE-RFE
comparison (and the previous comparison in (Aksu et al., 2010)) supports
our use of MFE to determine brain biomarkers.
To relate the retained voxel set to anatomic regions in the brain, we
overlaid the retained voxel set onto a registered atlas space. For CT-NT
classification, to improve robustness, the final voxel set was formed from
the union of the retained voxel sets from each of ten feature elimination
trials (each using a different, randomly selected training sample subset). For
AD-Control classification, the final voxel set came from a single trial (the
only trial, from which the 10 CT-NT trials stemmed). For each of these
two cases, overlaying the final voxel set onto the co-registered atlas (Atlas2,
defined in Appendix A) yielded between 70-80 anatomic regions. For data
interpretation purposes, we then identified a subset of (biomarker) regions
using the following procedure. First, for each brain region, we measured the
percentage of the region's voxels that are retained, sorted these percentages,
and then plotted them. As shown in Fig. 8(a), the resulting curve for the
AD-Control case has a distinct knee, which we thus used as a threshold
(0.125) to select the final, retained (diagnostic) regions for AD-Control. We
used the same threshold for the CT-NT curve, shown in Fig. 8(b). This
choice of threshold yields a reasonable number of regions -- 19 for the CT-NT
(prognostic) case and 21 for the AD-Control (diagnostic) case.
The resulting sets of identified prognostic and diagnostic biomarkers are
given in Table 3, along with their intersection. The diagnostic markers in the
table include the majority of the known brain regions in the medial tempo-
ral lobe involved in AD pathology. For example, hippocampus atrophy and
lateral ventricle enlargement, particularly in its anterior aspects of the tem-
poral horn, are considered the most prominent diagnostic markers for AD.
26
Entorhinal cortical regions, including the perirhinal cortex, are presumably
the earliest sites of degeneration (Braak et al., 1997). Thus, independent
identification by our AD-control classifier of known AD diagnostic biomark-
ers establishes a reasonable basis for applying the same approach to identify
prognostic biomarkers. The brain regions listed as CT-NT prognostic mark-
ers include most known AD diagnostic markers (including 8 of the 12 regions
from (Schuff et al., 2010) (marked by *), 4 of which are also diagnostic mark-
ers), indicating that some AD-linked pathological changes in these brain re-
gions already occurred and remained active in a subset of MCI subjects who
likely progress to AD rapidly. Conversely, the brain areas appearing only on
the prognostic marker list are likely the most active areas of degeneration
during this stage of progression to dementia. These structures tend to be
the brain regions further away from the entorhinal cortex onto the parietal
(Supramarginal gyrus, Precuneus) and temporal cortex (Superior temporal
gyrus and Middle temporal gyrus) regions. All the brain structures listed
in the table are known to be involved in AD (Braak et al., 1997; Chan et
al.
, 2003; Frisoni et al., 2009). Thus, the markers in Table 3 suggest an
interesting anatomic pattern of trajectory for MCI conversion to AD which
conforms with the Brak and Brak hypothesis and previous imaging findings
(Chan et al.
, 2003; Frisoni et al., 2009). Moreover, the CT-NT regions
uniquely found by our MFE-based procedure in Table 3 may be viewed as
"putative" prognostic markers, and may warrant further investigation.
Finally, we note that we have used a particular criterion (percentage of a
region's voxels that are retained) to identify biomarker regions, starting from
MFE-retained voxels. While our identified regions are plausible, it is possible
that other (equally reasonable) criteria may produce different biomarker re-
gion results. Thus, the biomarkers we identify should be viewed as anecdotal,
identifying regions that figure prominently in our classifier's decisionmaking
and also potentially assisting researchers in forming hypotheses about MCI-
to-AD disease progression. However, we do not view the identified regions
as definitive.
3.2.4. Comparison with an SPM-based biomarker identification approach
In the previous section, we used MFE to identify voxels as biomarkers
for the CT and NT classes. Here, using the same CT-NT training and
test populations, we will alternatively identify voxel-based biomarkers us-
ing statistical testing with SPM5 (see: (SPM website)). Subsequently we
will present a classifier generalization accuracy comparison (where accuracy
27
(a)
(b)
Figure 7: Test set misclassification rate during the course of feature elimination, for (a)
the AD-Control classifier and (b) the CT-NT classifier.
(a)
(b)
Figure 8: Sorted retained voxel percentages for initial regions (a) AD-Control; b) CT-NT,
used to select final regions (Sec. 3.2.3).
28
02000400060008000100001200000.10.20.30.40.50.60.70.8number of features retained (starting at 11293)test set classification error rate RFEMFE/MFE(cid:239)slack02000400060008000100001200000.10.20.30.40.5number of features retained (starting at 11293)test set classification error rate RFEMFE/MFE(cid:239)slack(a)(b)Figure7:[ABOVEARENOTTHEACTUALIMAGES]Imagedisplayofbrainregionsidentifiedasbiomarkersusingvoxel-basedfeaturesandMFE.02000400060008000100001200000.10.20.30.40.50.60.70.8number of features (starting at 11293)test set classification error rate RFEMFE/MFE−slack(a)02000400060008000100001200000.10.20.30.40.5number of features (starting at 11293)test set classification error rate RFEMFE/MFE−slack(b)Figure8:Testsetmisclassificationrateduringthecourseoffeatureelimination,for(a)theAD-Controlclassifierand(b)theCT-NTclassifier.02040608010000.10.20.30.40.50.60.70.8ratioindex(a)0102030405060708000.050.10.150.20.250.30.350.40.450.5ratioindex(b)Figure9:Sortedretainedvoxelpercentagesforinitialregions(a)AD-Control;b)CT-NT,usedtoselectfinalregions(Sec.3.2.3).13(a)(b)Figure7:[ABOVEARENOTTHEACTUALIMAGES]Imagedisplayofbrainregionsidentifiedasbiomarkersusingvoxel-basedfeaturesandMFE.02000400060008000100001200000.10.20.30.40.50.60.70.8number of features (starting at 11293)test set classification error rate RFEMFE/MFE−slack(a)02000400060008000100001200000.10.20.30.40.5number of features (starting at 11293)test set classification error rate RFEMFE/MFE−slack(b)Figure8:Testsetmisclassificationrateduringthecourseoffeatureelimination,for(a)theAD-Controlclassifierand(b)theCT-NTclassifier.02040608010000.10.20.30.40.50.60.70.8ratioindex(a)0102030405060708000.050.10.150.20.250.30.350.40.450.5ratioindex(b)Figure9:Sortedretainedvoxelpercentagesforinitialregions(a)AD-Control;b)CT-NT,usedtoselectfinalregions(Sec.3.2.3).13 is again measured on the previous section's CT-NT (test) population) for
these two biomarker detection methods. We determined SPM biomarkers
as follows. The CT-NT training set population, being age-matched, is read-
ily suitable for a paired t-test, an appropriate statistical test for determining
SPM-identified biomarkers, i.e. voxels that discriminate between the CT and
NT groups. In contrast with MFE's use of only one (out of 216) RAVENS
subsamples (taken jointly from the GM, WM, ventricle maps), we performed
t-tests on whole RAVENS maps (without subsampling), which makes the
SPM-MFE comparison favorably biased towards SPM. More specifically our
steps were as follows. First, for the GM and WM maps separately, we found
using SPM that a large portion of each of these two tissues was statistically
significant at the 0.05 level when correction for multiple comparisons was
not applied. Next, we used SPM's FDR-based correction for multiple com-
parisons -- based on an SPM FDR cluster size of 5 voxels we found that the
spatial extent of the statistically significant regions, at each of the levels 0.05,
0.01, and 0.005, was approximately a subset of the above-mentioned spatial
support found in the uncorrected case. Given that the number of significant
voxels in any of these SPM experiments is, again due to no subsampling,
much larger than the 11,293 voxels started from in the MFE case, we sim-
ply 1) chose as our SPM result the result for 0.01 (FDR-corrected), 2) took
from among those significant voxels the most significant 11,434 voxels in or-
der to be able to compare MFE and SPM for ≈ the same number of voxels
(biomarkers). To obtain the generalization accuracy for this SPM-identified
biomarker voxel set, using the same training/test set as in the MFE experi-
ment, we trained an SVM classifier and measured its generalization accuracy,
which was found to be 0.76. This accuracy is somewhat lower than the 0.8
accuracy of the previous section's CT-NT SVM classifier. Recalling that this
comparison is actually favorably biased towards SPM, and further noticing
the fact that MFE was able to maintain the 0.8 accuracy all the way down to
2000 features (cf. Fig. 7(b)), this experimental comparison provides another
validation (beyond the comparison with RFE given earlier) for MFE-based
feature/biomarker selection, applied to brain images.
4. Conclusions
We have presented an automated prognosticator of MCI-to-AD conver-
sion based on brain morphometry derived from high resolution ADNI MR
images. The primary novel contributions of our work are:
i) casting MCI
29
prognostication as a novel machine learning problem lying somewhere be-
tween supervised and unsupervised learning; ii) our proposal of a conversion
definition which, unlike previous methods, exploits both rich phenotypical
information in neuroimages and AD and control examples; iii) correlation
testing and classifier accuracy evaluations to validate candidate conversion
definitions; iv) prognostic biomarker discovery based on our conversion defi-
nition. We demonstrated that our method achieved both better generaliza-
tion accuracy and stronger, statistically significant, correlations with known
brain region biomarkers than a predictor based on the clinical CDR score,
the approach used in several past works. The brain structures identified
as AD-control diagnostic markers and MCI conversion prognostic markers
well conform with known brain atrophic patterns and progression trajecto-
ries occurring in AD-afflicted brains. While the noisy nature of cognitive
assessments, including MMSE, has been acknowledged in past works, in fu-
ture we may extend our methodology to consider cognitive assessment data,
both potentially as additional (baseline) input features and as additional or
alternative prediction targets to our "conversion-by-trajectory" labels. We
may also consider alternative ways to adjust for confounding effects of age,
noting that (Schuff et al., 2010) has characterized the nonlinear dependence
of age on brain region volumes. Finally, while we have focused on the MCI
subpopulation here, our system could also potentially be used to detect, as
possible misdiagnoses, subjects diagnosed as "Control" who are classified as
MCI converters by our system.
5. Acknowledgement
Funding supporting this work has been provided in part through NIH
R01 AG02771 and the Pennsylvania Department of Health. We thank Dr.
Michelle Shaffer for assistance with statistical testing and Jianli Wang and
Zachary Herse for assistance with segmentation.
Data collection and sharing for this project was funded by the Alzheimer's
Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant
U01 AG024904). ADNI is funded by the National Institute on Aging, the
National Institute of Biomedical Imaging and Bioengineering, and through
generous contributions from the following: Abbott, AstraZeneca AB, Bayer
Schering Pharma AG, Bristol-Myers Squibb, Eisai Global Clinical Devel-
opment, Elan Corporation, Genentech, GE Healthcare, GlaxoSmithKline,
Innogenetics, Johnson and Johnson, Eli Lilly and Co., Medpace, Inc., Merck
30
and Co., Inc., Novartis AG, Pfizer Inc, F. Hoffman-La Roche, Schering-
Plough, Synarc, Inc., as well as non-profit partners the Alzheimer's Asso-
ciation and Alzheimer's Drug Discovery Foundation, with participation from
the U.S. Food and Drug Administration. Private sector contributions to
ADNI are facilitated by the Foundation for the National Institutes of Health
(www.fnih.org). The grantee organization is the Northern California In-
stitute for Research and Education, and the study is coordinated by the
Alzheimer's Disease Cooperative Study at the University of California, San
Diego. ADNI data are disseminated by the Laboratory for Neuro Imaging at
the University of California, Los Angeles. This research was also supported
by NIH grants P30 AG010129, K01 AG030514, and the Dana Foundation.
Appendix A. Image processing
The input to our processing is a T1-weighted MR image of the head.
First, using rigid-body registration implemented in FSL's linear image reg-
istration tool FLIRT (Jenkinson et al., 2001), we coarsely aligned this (3d)
image with the MNI/ICBM atlas resampled to 1mm isotropic voxel dimen-
sions (aka Atlas1). We used FSL version 3 (see: (FSL website)). Next,
we removed non-brain anatomy from the aligned head image, using FSL's
brain extraction tool BET (Smith et al., 2002). We then segmented the
resulting brain-only 3d image into the following five segments required by
HAMMER(Shen et al., 2002)19 : WM, GM, ventricles, (non-ventricle) CSF,
(non-anatomy) background. Next, using as inputs an MNI atlas distributed
with HAMMER (aka Atlas220) and the ADNI participant's five-segment im-
age, we performed two different HAMMER operations, generating: 1) the
three "volumetric density" Ravens images: GM, WM, ventricles. The union
of these three images (following some preprocessing) forms the set of features
used by our classifier; 2) the 3d region-segmented image, whose region vol-
umes are used for by-trajectory statistical validation in section 3.2.2 and for
region biomarker identification in section 3.2.3.21
19We used a 2006 version of HAMMER, downloaded on November 8, 2006.
20Since the region-segmented version of the atlas distributed with HAMMER had much
better segmentation quality than the five-segment version, as a pre-processing step we
used the former to create a replacement for the latter (by simply taking the union of all
the regions). The resulting five-segment atlas is dubbed 'Atlas2'.
21We measured normalized region volumes. We normalized by dividing by the sum of all
(98) region volumes. This sum is essentially intracranial volume minus cerebellum volume,
31
Each of these two HAMMER operations performs a different type of atlas-
based nonlinear registration -- the one that generates the RAVENS tissue
maps performs less aggressive warping between the five-segment image and
the atlas in order both to combat noise inherent in the registration process
and to preserve volume on a tissue-by-tissue basis (so as to properly detect
and encode brain atrophy). Consequently, there is considerable individual
variability in RAVENS images, which we mitigate in a standard way by
smoothing with a Gaussian filter with an FWHM of 5mm.
The sum of the voxel intensities over the three RAVENS maps is on the
order of 106 and varies across individuals. Thus, prior to smoothing, we
normalized each individual's Ravens maps, such that each individual has the
same total volume. However, after the normalization and smoothing (aka
"nsRAVENS" images), some areas of poor registration, manifesting as areas
with very low voxel values (including zero), will remain.22 For a population
of nsRAVENS images, these areas are considered to be outlier areas and are
thus removed from each image in the population by thresholding 23. The
resulting images are the features used by our classification system.
Appendix B. Correspondence Between Atlas-defined Regions and
Those Defined in (Schuff et al., 2010)
as our "intracranial region" list includes CSF in addition to brain regions and excludes
the cerebellum.
22A substantial portion of the very low, nonzero voxel intensities are in fact introduced
by the smoothing itself as it calculates each new voxel intensity as a weighted average of
many neighboring voxels (thereby switching some voxels from zero (non-anatomy) to very
low intensity values (anatomy), which essentially slightly grows the anatomy boundaries
outward).
23We calculated the threshold solely using the training population of our AD-Control
classifier, and then applied the thresholding operation on the entire population of AD,
MCI, and Control individuals. In this way, we were careful to exclude test examples from
all phases of classifier training.
32
Entorhinal cortex left/right
Lateral occipitotemporal gyrus right/left
Hippocampal formation right/left
Supramarginal gyrus left/right, Angular gyrus right/left
Lateral ventricle left/right
Entorhinal cortex
Fusiform gyrus
Hippocampus
Inferior parietal GM
Lateral orbitofrontal GM Lateral front-orbital gyrus right/left
Lateral ventricles
Medial orbitofrontal GM Medial front-orbital gyrus right/left
Parahippocampal gyrus
Posterior cingulate
Precentral GM
Superior frontal GM
Superior temporal GM
Parahippocampal gyrus right/left
Cingulate region left/right
Precentral gyrus right/left
Superior frontal gyrus left/right
Superior temporal gyrus right/left
Table B.4: Correspondence between the regions in (Schuff et al., 2010) (left) (except "Total
GM" and "Total WM") and our defined regions (right).
References
Y. Aksu, D. J. Miller, G. Kesidis, and Q. X .Yang, "Margin-Maximizing
Feature Elimination Methods for Linear and Nonlinear Kernel-Based Dis-
criminant Functions", IEEE Transactions on Neural Networks, vol. 25,
no.10, pp.701-717, 2010.
H. Braak and E. Braak, "Frequency of Stages of Alzheimer-Related Lesions
in Different Age Categories", Neurobiology of Aging 18(4):351-357, July 8
1997.
C. Chang and C. Lin, "LIBSVM : a library for support vector machines,"
software available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm, 2001.
J. M. Chambers, "Linear models", in: Statistical Models in S, J. M. Chambers
and T. J. Hastie (Eds.), Chapman & Hall, New York, 1993.
D. Chan, J.C. Janssen, J.L. Whitwell, H.C. Watt, R. Jenkins, C. Frost, M.N.
Rossor and N.C. Fox, "Change in rates of cerebral atrophy over time in
early-onset Alzheimer's disease: longitudinal MRI study", Lancet 362, pp.
1121-1122, 2003.
G. Chetelat, B. Desgranges, V. de la Sayette, F. Viader, F. Eustache, J-C.
Baron, "Mapping gray matter loss with voxel-based morphometry in mild
cognitive impairment", NeuroReport vol.13. no.15, 28 October 2002.
Y.-Y. Chou, N. Lepor´e, C. Avedissian, S. K. Madsen, X. Hua, C. R. Jack
Jr., M. W. Weiner, A. W. Toga, P. M. Thompson, and the Alzheimer's
33
Disease Neuroimaging Initiative, "Mapping Ventricular Expansion and its
Clinical Correlates in Alzheimer's Disease and Mild Cognitive Impairment
using Multi-Atlas Fluid Image Alignment", Proc. SPIE, vol.7259, 725930,
2009.
Y.-Y. Chou, N. Lepor´e, P. Saharan, S. K. Madsen, X. Hua, C. R. Jack, L. M.
Shaw, J. Q. Trojanowski, M. W. Weiner, A. W. Toga, P. M. Thompson,
and the Alzheimer's Disease Neuroimaging Initiative, "Ventricular maps
in 804 ADNI subjects: correlations with CSF biomarkers and clinical de-
cline", Neurobiology of Aging 31, pp.1386-1400, 2010.
J. G. Csernansky, L. Wang, J. Swank, J. P. Miller, M. Gado, D. McKeel,
M. I. Miller, J. C. Morris, "Preclinical detection of Alzheimer's disease:
hippocampal shape and volume predict dementia onset in the elderly",
NeuroImage 25, pp.783-792, 2005.
C. Davatzikos, "Mapping image data to stereotaxic spaces: Applications to
brain mapping," Hum. Brain Mapp., 6:334-338, 1998.
C. Davatzikos, A. Genc, D. Xu, and S. M. Resnick, "Voxel-Based Morphom-
etry Using the RAVENS Maps: Methods and Validation Using Simulated
Longitudinal Atrophy," NeuroImage 14, pp.1361-1369, 2001.
C. Davatzikos, Y. Fan, X. Wu, D. Shen, S. M. Resnick, "Detection of pro-
dromal Alzheimer's disease via pattern classification of magnetic resonance
imaging", Neurobiology of Aging 29, pp.514-523, 2008.
C. Davatzikos, P. Bhatt, L. M. Shaw, K. N. Batmanghelich, J. Q.
Trojanowski, "Prediction of MCI to AD conversion, via MRI, CSF
biomarkers, and pattern classification", Neurobiology of Aging, 2010,
doi:10.1016/j.neurobiolaging.2010.05.023.
M. J. de Leon, S. DeSanti, R. Zinkowski, P. D. Mehta, D. Pratico, S. Segal,
H. Rusinek, J. Li, W. Tsui, L. A. Saint Louis, C. M. Clark, C. Tarshish,
Y. Li, L. Lair, E. Javier, K. Rich, P. Lesbre, L. Mosconi, B. Reisberg, M.
Sadowski, J. F. DeBernadis, D. J. Kerkman, H. Hampel, L. -O. Wahlund,
P. Davies, "Longitudinal CSF and MRI biomarkers improve the diagnosis
of mild cognitive impairment," Neurobiology of Aging 27, pp.394-401, 2006.
34
G. De Meyer, F. Shapiro, H. Vanderstichele, E. Vanmechelen, S. Engel-
borghs, P. P. De Deyn, E. Coart, O. Hansson, L. Minthon, H. Zetter-
berg, K. Blennow, L. Shaw, J. Q. Trojanowski, for the Alzheimers Dis-
ease Neuroimaging Initiative, "Diagnosis-Independent Alzheimer Disease
Biomarker Signature in Cognitively Normal Elderly People", Arch. Neu-
rol., vol. 67, no. 8, pp.949-956, Aug 2010.
R. Duda, P. Hart, and G. Stork, Pattern Classification. Second Edition, John
Wiley and Sons, New York, 2001.
Y. Fan, D. Shen, R. C. Gur, R. E. Gur, C. Davatzikos, "COMPARE: Clas-
sification of Morphological Patterns Using Adaptive Regional Elements",
IEEE Transactions on Medical Imaging, vol. 26, no. 1, pp.93-105, 2007.
Y. Fan, N. Batmanghelich, C. M. Clark, C. Davatzikos, "Spatial patterns
of brain atrophy in MCI patients, identified via high-dimensional pat-
tern classification, predict subsequent cognitive decline", NeuroImage 39,
pp.1731-1743, 2008.
C. Fennema-Notestine, D. J. Hagler Jr., L. K. McEvoy, A. S. Fleischer, E.
H. Wu, D. S. Karow, A. M. Dale, the Alzheimer's Disease Neuroimaging
Initiative, "Structural MRI biomarkers for preclinical and mild Alzheimer's
disease", Human Brain Mapping, vol.30, issue 10, pp.3238-3253, 2009.
G.B. Frisoni, A. Prestia, P.E. Rasser, M. Bonetti, P.M. Thompson, "In
vivo mapping of incremental cortical atrophy from incipient to overt
Alzheimer's disease", Neurol. 256(6):916-24, Feb 28, 2009.
FSL
website:
FSL
http://www.fmrib.ox.ac.uk/fsl
(FMRIB
Software
Library),
G. Fung, O. L. Mangasarian, "A feature selection Newton method for support
vector machine classification," Computational Optimization and Applica-
tions, vol. 28, no.2:185-202, July 2004.
A. F. Goldszal, C. Davatzikos, D. L. Pham, M. X. H. Yan, R. N. Bryan, S. M.
Resnick, "An Image-Processing System for Qualitative and Quantitative
Volumetric Analysis of Brain Images," J. Comput. Assist. Tomogr., 22:827-
837, 1998.
35
I. Guyon, J. Weston, S. Barnhill, and V. Vapnik, "Gene selection for
cancer classification using support vector machines," Machine Learning,
46(1):389-422, 2002.
I. Guyon and A. Elisseeff, "An introduction to variable and feature selection,"
J. Mach. Learn. Res., vol. 3, pp.1157-1182, 2003.
M. Jenkinson and S.M. Smith, "A global optimisation method for robust
affine registration of brain images," Med. Image Anal., 5(2):143-156, 2001.
J. P. Lerch, A. C. Evans, "Cortical thickness analysis examined through
power analysis and a population simulation", NeuroImage 24, pp.163-173,
2005.
A. D. Leow, I. Yanovsky, N. Parikshak, X. Hua, S. Lee, A. W. Toga, C. R.
Jack Jr., M. A. Bernstein, P. J. Britson, J. L. Gunter, C. P. Ward, B.
Borowski, L. M. Shaw, J. Q. Trojanowski, A. S. Fleisher, D. Harvey, J.
Kornak, N. Schuff, G. E. Alexander, M. W. Weiner, P. M. Thompson,
Alzheimers Disease Neuroimaging Initiative. "Alzheimers Disease Neu-
roimaging Initiative: A One-Year Follow-up Study Using Tensor-Based
Morphometry Correlating Degenerative Rates, Biomarkers and Cogni-
tion," Neuroimage 45, pp.645-55, 2009.
X. -L. Meng, R. Rosenthal, D. B. Rubin, "Comparing Correlated Correlation
Coefficients", Psychological Bulletin 111, pp.172-175, 1992.
C. Misra, Y. Fan, C. Davatzikos, "Baseline and longitudinal patterns of brain
atrophy in MCI patients, and their use in prediction of short-term conver-
sion to AD: Results from ADNI", NeuroImage 44, pp.1415-1422, 2009.
R. C. Petersen, Mild cognitive impairment: aging to Alzheimer's Disease,
Oxford University Press, 2004.
H. Rusinek, Y. Endo, S. De Santi, D. Frid, W.-H. Tsui, S. Segal, A. Convit,
"Atrophy rate in medial temporal lobe during progression of Alzheimer's
disease", Neurology, vol.63, issue 12, 2354-2359, 2004.
J. M. Schott, S. L. Price, C. Frost, J. L. Whitwell, et al, "Measuring atro-
phy on Alzheimer disease - A serial MRI study over 6 and 12 months,"
Neurology, 65(1), pp.119-124, 2005.
36
N. Schuff, D. Tosun, P. S. Insel, G. C. Chiang, D. Truran, P. S. Aisen, C.
R. Jack, Jr., M. W. Weiner, the Alzheimer's Disease Initiative, "Nonlinear
time course of brain volume loss in cognitively normal and impaired elders",
Neurobiology of Aging, 2010, doi:10.1016/j.neurobiolaging.2010.07.012.
S. M. Smith, "Fast robust automated brain extraction," Human Brain Map-
ping, 17(3):143-155, 2002.
D. Shen and C. Davatzikos, "HAMMER: hierarchical attribute match-
ing mechanism for elastic registration," IEEE Trans. Medical Imaging,
21(11):1421-1439, 2002.
D. Shen, C. Davatzikos, "Very High-resolution Morphometry Using Mass-
preserving Deformations and HAMMER Elastic Registration," NeuroIm-
age 18, pp.28-41, 2003.
SPM
website:
Statistical
http://www.fil.ion.ucl.ac.uk/spm
Parametric
Mapping
(SPM),
T. R. Stoub, M. Bulgakova, S. Leurgans, D. A. Bennett, D. Fleischman, D. A.
Turner, L. deToledo-Morrell, "MRI predictors of risk of incident Alzheimer
disease," Neurology 64, pp. 1520-1524, 2005.
P. M. Thompson, K. M. Hayashi, G. de Zubicaray, A. L. Janke, S. E. Rose,
et al, "Dynamics of gray matter loss in Alzheimer's disease," J. Neurosci.
23, pp.994-1005, 2003.
V. Vapnik, Statistical Learning Theory. John Wiley & Sons, 1998.
P. Vemuri, J. L. Gunter, M. L. Senjem, J. L. Whitwell, K. Kantarci, D. S.
Knopman, B. F. Boeve, R. C. Petersen, C. R. Jack Jr., "Alzheimer's disease
diagnosis in individual subjects using structural MR images: Validation
studies," NeuroImage 39, pp.1186-1197, 2008.
P. Vemuri, H. J. Wiste, S. D. Weigand, L. M. Shaw, J. Q. Trojanowski,
M. W. Weiner, D. S. Knopman, R. C. Petersen, C. R. Jack Jr., and On
behalf of the Alzheimer's Disease Neuroimaging Initiative, "MRI and CSF
biomarkers in normal, MCI, and AD subjects: Predicting future clinical
change," Neurology 73:294-301, 2009.
37
Y. Wang, Y. Fan, P. Bhatt, C. Davatzikos, "High-dimensional pattern regres-
sion using machine learning: From medical images to continuous clinical
variables," NeuroImage 50, pp.1519-1535, 2010.
Y. Zhang, M. Brady, and S. Smith, "Segmentation of brain MR images
through a hidden Markov random field model and the expectation maxi-
mization algorithm," IEEE Trans. Medical Imaging, 20(1):45-57, 2001.
D. Zhang, Y. Wang, L. Zhou, H. Yuan, D. Shen,
and the
Alzheimer's Disease Neuroimaging Initiative, "Multimodal classification
of Alzheimer's disease and mild cognitive impairment," NeuroImage, 2011,
doi:10.1016/j.neuroimage.2011.01.008.
38
|
1208.5350 | 4 | 1208 | 2013-02-19T17:08:00 | Impact of intrinsic biophysical diversity on the activity of spiking neurons | [
"q-bio.NC"
] | We study the effect of intrinsic heterogeneity on the activity of a population of leaky integrate-and-fire neurons. By rescaling the dynamical equation, we derive mathematical relations between multiple neuronal parameters and a fluctuating input noise. To this end, common input to heterogeneous neurons is conceived as an identical noise with neuron-specific mean and variance. As a consequence, the neuronal output rates can differ considerably, and their relative spike timing becomes desynchronized. This theory can quantitatively explain some recent experimental findings. | q-bio.NC | q-bio |
Impact of intrinsic biophysical diversity on the activity of spiking neurons
Bernstein Center Freiburg & Faculty of Biology, University of Freiburg, Freiburg, Germany
Man Yi Yim, Ad Aertsen, Stefan Rotter∗
(Dated: May 7, 2018)
We study the effect of intrinsic heterogeneity on the activity of a population of leaky integrate-
and-fire neurons. By rescaling the dynamical equation, we derive mathematical relations between
multiple neuronal parameters and a fluctuating input noise. To this end, common input to het-
erogeneous neurons is conceived as an identical noise with neuron-specific mean and variance. As
a consequence, the neuronal output rates can differ considerably, and their relative spike timing
becomes desynchronized. This theory can quantitatively explain some recent experimental findings.
I.
INTRODUCTION
In statistical physics, it is often assumed that indi-
viduals are intrinsically identical. In neuroscience also,
identical parameters are typically assumed for all neu-
rons in the study of neuronal population activity and
correlation transmission. Real neurons, though, even if
they are of the same type and located in the same brain
area, exhibit intrinsic differences. Their morphologies
and the intracellular concentrations of ions, to name just
two examples, can differ widely, although in principle
they have been generated by the same mechanisms [1].
As a consequence, neuronal spike patterns can differ al-
though neurons receive identical inputs [2, 3]. Recently,
in vitro intracellular recordings of isolated mitral cells in
the mouse olfactory bulb were conducted while they re-
sponded to identical input [3] (Fig. 1(a)). The neurons
displayed diverse output firing rates and pairwise corre-
lations. Specifically, the spike correlation coefficient ob-
tained with a 1 ms observation window co-varied with the
rate difference of the neuron pairs: small differences re-
sulted in a wide range of different spike correlations, but
large differences led always to small spike correlation.
In homogeneous network models, additional indepen-
dent Gaussian white noises or independent Poisson spikes
are very often added to every constituent identical neu-
ron to account for their diverse spike timing.
In real
brain networks, not only the spike timing but also the
spiking rate of neurons differ due to their intrinsic bio-
physical diversity. Therefore, it is of great interest to
understand how the biophysical heterogeneity of a neu-
ronal population contributes to neural coding and infor-
mation processing in neuronal networks. Research work
has been conducted on the coding properties [4, 5] and
synchronous responses [6 -- 8] in a network of heteroge-
neous neurons.
In many cases, neuronal heterogeneity
was implemented simply by replacing one or more fixed
neuronal parameters, such as the offset current [6, 7], the
spiking threshold [5], or the synaptic conductance [4], by
a Gaussian- or uniformly-distributed random variable.
Here we investigated more fundamental questions, us-
ing both theoretical analysis and simulations: how neu-
∗ [email protected]
ronal heterogeneity can be represented appropriately in
theory and how it can affect the neuronal dynamics and
the spiking statistics in a population of simple leaky
integrate-and-fire (LIF) neurons. The limitations of the
existing approaches are addressed first. Then we suggest
a more general scheme to implement biophysical diver-
sity when either rate or correlation is of interest. By
rescaling the dynamical equation, we derive mathemati-
cal relations between multiple neuronal parameters and
the input noise. The main impact of common input to
heterogeneous neurons on rate and correlation can be
realized by an identical (frozen) noise current injection
with different values of mean and variance, whereas the
complete effect is captured by additionally drawing dis-
tributed values of the membrane time constant and the
refractory period. In this scheme, the rate difference of
heterogeneous LIF neurons can be treated analytically.
As for correlation, we utilize alternative correlation mea-
sures to illustrate that spikes from heterogeneous neurons
may be desynchronized by several milliseconds, thus es-
caping detection by a 1 ms observation window.
II. MODEL
We consider a population of isolated leaky integrate-
and-fire (LIF) neurons, each of which has its membrane
potential V (t) governed by
τm V (t) = −V (t) + RI(t),
(1)
where the input synaptic current
RI(t) = τmJE Xj
δ(t − tj) − τmJI Xk
δ(t − tk). (2)
τm = RC is the membrane time constant. R and C are
the membrane resistance and capacitance, respectively.
JE (JI ) is the amplitude of an excitatory (inhibitory)
post-synaptic potential, whereas tj (tk) represents the
time of the jth (kth) excitatory (inhibitory) input spike.
When V (t) = θ, V (t) is reset to Vr and a pause for synap-
tic integration τr is imposed to mimic the refractory pe-
riod. In the high-input regime, the sum of synaptic in-
puts to a neuron can be approximated by a fluctuating
input noise [9, 10]
I(t) ≡ τm[µ + ση(t)],
(3)
a
b
1.0
0.5
r
r
o
C
e
k
p
S
i
0.0
0
40%
20%
10%
20
60
Rate difference (Hz)
40
80
FIG. 1. (Color online). (a) Correlation coefficient (1 ms win-
dow) of 589 spike trains from mitral cells in vitro receiving an
identical input as a function of rate difference, adapted with
permission from [3]. (b) The same measure for 100 simulated
heterogeneous LIF neurons using uniformly distributed τm,
τr, C and θ with different distribution widths.
where
µ = JEνE − JI νI ,
σ = qJ 2
EνE + J 2
I νI .
(4)
(5)
is a white noise random process
η(t)
such that
hη(t)η(t′)i = δ(t − t′). νE (νI ) is the firing rate of the
excitatory (inhibitory) input.
integration of Eq. 1 in our simula-
tions was performed using the fourth-order Runge-Kutta
method with a time step of 0.01 ms.
The numerical
III. HETEROGENEITY
Intrinsic diversity of a population of neurons can be
directly imposed by drawing neuronal parameters from a
distribution. Here, we tested the response of 100 isolated
heterogeneous LIF neurons to an identical fluctuating in-
put current in the form of Eq. (3). Neuronal heterogene-
ity is implemented by drawing four uniformly distributed
parameters: τm, τr, C (or R) and θ, which, together with
Vr, represent all the independent parameters of a LIF
neuron in response to a current input. The mean values
of the uniform distribution are 20 ms, 2 ms, 1 and 1 (in
arbitrary units) respectively, and Vr is fixed to 0. These
values are used throughout this work. We maintain the
temporal scale of the dynamics of a typical neuron and
rescale the potential by setting the mean reset to zero and
the mean threshold to 1. The correlation coefficient as a
function of the output firing rate difference of all possible
pairs with different distribution widths (percentage with
respect to the mean) from 100 s of simulations (an exam-
ple with µ = 0.03 and σ = 0.3 shown in Fig. 1(b)) highly
resemble the experimental findings in [3] (Fig. 1(a)).
Diverse neuronal spike timing in a network has very of-
ten been achieved by adding independent random inputs
to individual neurons. We provide every identical neuron
with a common input as the input signal plus an indepen-
dent input with the same statistics among neurons , in
a
100
D
I
n
o
r
u
e
N
50
0
30
t
n
u
o
C
0
100
120
Time (ms)
140
1.0
0.5
r
r
o
C
e
k
p
S
i
0.0
0
2
b
100
50
0
30
0
1.0
0.5
100
120
Time (ms)
140
c
100
50
0
30
0
1.0
0.5
100
120
Time (ms)
140
60
20
Rate difference (Hz)
40
0.0
0
80
20
60
Rate difference (Hz)
40
0.0
0
80
20
60
Rate difference (Hz)
40
80
FIG. 2. Rasterplot, peri-stimulus time histogram (PSTH)
and correlation coefficient as a function of rate difference for
the following input into 100 LIF neurons: (a) µ + σ(cid:2)√cη(t) +
√1 − cξi(t)(cid:3) (c = 0.9); (b) identical input µ + ση(t) but dis-
tributed θi; (c) our scheme using µi + σiη(t) together with
distributed τm and τr, consistent with both the experimental
findings [3] and the mathematical analysis. The subscript i
denotes "independent".
the form of µ + σ(cid:2)√cη(t) + √1 − cξi(t)(cid:3) where η(t) and
ξi(t) are independent Gaussian white noises. Fig. 2(a)
displays the raster and the correlation coefficient as func-
tions of the rate difference when µ = 0.06, σ = 0.2 and
c = 0.9. The rate difference is close to zero and the cor-
relation coefficient between any pair is nearly the same
[11]. Decreasing c leads to a drop in spike correlation but
has no effect on the rate difference. These observations
are very distinct from both the experimental (Fig. 1(a))
and simulation (Fig. 1(b)) results. In view of some previ-
ous work on the reliability of single neurons in response
to a repeated input [2, 3, 12], this implementation may
be adopted to account for trial-to-trial variability of the
same neuron.
It is common practice to implement heterogeneity of
neurons by drawing random variables for a single neu-
ronal parameter. To test its validity, we provide every
identical neuron with a common input plus a random
value of the spiking threshold drawn from a uniform dis-
tribution θ ∈ [0.5, 1.5]. Fig. 2(b) displays the raster and
the correlation coefficient as functions of the rate differ-
ence. Another example with distributed values of the in-
put offset current µ instead of θ is shown in Fig. 3(a). In
either case, neurons exhibit dispersive firing rates. How-
ever, the spike correlation distribution at different values
of rate difference is too narrow compared with Fig. 1(a)
and (b), and the region of small rate difference and small
spike correlation cannot be reached. Thus, the impact
of neuronal heterogeneity is only partially accounted for.
We describe our alternative approach in the next section.
IV. MATHEMATICAL RELATIONS
In addition to the five parameters mentioned above,
two additional parameters correspond to synaptic inputs:
JE and JI . We analyze the contribution of the hetero-
geneity in the seven independent neuronal parameters in
a population of LIF neurons in the high-input regime.
Regarding synaptic inputs, heterogeneity in the pa-
rameters JE and JI can be captured by heterogeneity
in µ and σ, according to Eqs. (4) and (5). For exam-
ple, if JE is uniformly distributed, µ is also uniformly
distributed, whereas σ2 is distributed as the square of a
uniformly distributed variable.
The other five parameters are present in the neuronal
dynamics irrespective of the type of inputs (current or
spikes). First, R (or C, depending on the form of writing)
can be absorbed into I(t) as shown in Eq. (1) so any
distribution of R can be accounted for by a corresponding
distribution of µ and σ.
The difference between θ and Vr, which is the potential
difference a neuron has to traverse, is a quantity relative
to the synaptic strengths JE and JI . For instance, lifting
θ, or lowering Vr, is equivalent to reducing JE and JI
together by the same ratio. Thus, heterogeneity in θ and
Vr can be included in µ and σ by means of rescaling.
Unlike the above five parameters related to the poten-
tial, the remaining two shaping the neuronal response in
the temporal scale, τm and τr, cannot be rescaled or cap-
tured by µ and σ. Their distributions among neurons
have to be accounted for separately. Therefore, in the
high input regime when the approximation of a fluctuat-
ing input noise is valid, the seven independent neuronal
parameters (and their distributions) can be reduced to
four: µ, σ, τm and τr. Based on this analysis, we suggest
using distributed values of these four parameters together
with an identical noise η(t) to account for the effects of
all the parameters in a population of LIF neurons re-
ceiving identical inputs. This is in contrast to the com-
mon practice of using independent noises as shown in
Fig. 2(a). We draw the parameters from uniform distri-
butions µ ∈ [0.015, 0.105], σ ∈ [0.1, 0.3], τm ∈ [16, 24] ms
In Fig. 2(c), both the rate dif-
and τr ∈ [1.5, 2.5] ms.
ference and the correlation coefficient, as well as their
relation, are consistent with the experimental and our
simulation results.
The respective contributions of the four parameters are
investigated. Fig. 3 shows the correlation as functions of
the rate difference for µ, σ, τm and τr separately. Each
realization is drawn from a uniform distribution of 10 %,
20 % and 50 % around their mean values, which are 0.06,
0.2, 20 ms and 2 ms, respectively. The firing rate of a
neuron is largely shaped by µ, whereas the distributed
values of the variance give rise to different degrees of
imprecise spiking. The wider the distribution of τm and
τr, the larger is the rate difference and the lower the
correlation. When τr ≪ 1/ν, the effect of τr is small.
1.0
0.5
a
r
r
o
C
e
k
p
S
i
0.0
0
c
1.0
0.5
r
r
o
C
e
k
p
S
i
0.0
0
20
m
20
Rate difference (Hz)
3
20
r
50%
20%
10%
40
50%
20%
10%
20
Rate difference (Hz)
40
b
1.0
0.5
50%
20%
10%
0.0
0
40
50%
20%
10%
d
1.0
0.5
0.0
0
40
FIG. 3. (Color online). Correlation coefficient as a function of
rate difference when neurons receive identical noise with uni-
formly distributed (a) µ, (b) σ, (c) τm and (d) τr separately.
a
80
60
40
)
z
H
(
e
t
a
R
20
=0
=0.1
=0.3
b
80
60
40
20
=0
=0.05
=0.1
c
80
60
40
20
0
0
0.05
0.1
0
0
0.1
0.2
0.3
0
10
20
m
30
FIG. 4. (Color online). Output rate as a function of (a) µ, (b)
σ and (c) τm from theory (black) and simulation (colored).
V. RATE DIFFERENCE
The firing rate of a LIF receiving a Gaussian dis-
tributed noise is known analytically [13, 14]:
= τr + τm√πZ
θ−µτm
σ√τm
Vr −µτm
σ√τm
θ
1
ν
1
ν
dueu2
(cid:0)1 + erf(u)(cid:1), when σ > 0;
= τr − τmln(1 −
µτm
), when σ = 0.
(6)
Only six parameters µ, σ, τm, τr, θ and Vr influence the
firing rates, of which only the first four are independent.
Changing any of them can result in a rate difference as
shown in Fig. 4, and this explains the wide distribution
of firing rates in a heterogeneous neuronal population.
VI.
IMPRECISE SPIKING
The raster plot in Fig. 2(c) shows population synchrony
with spike-time jitter. Low average correlation coeffi-
cients in a population of neurons do not necessarily im-
ply an asynchronous state of the population. Removing
spikes from one of the two identical spike trains can also
reduce the correlation coefficient significantly. When we
compute the correlation coefficient in our data with a
a
1.0
Spike corr
b
0.10
rxy(
)
c
1.0
VII. DISCUSSION
4
0.5
0.0
0
0.05
0.00
20
40
60
Bin (ms)
80
100
10
5
0
5
10
0.5
0.0
0
20
60
Rate difference (Hz)
40
80
FIG. 5. (Color online). (a) Correlation coefficient as a func-
tion of window size (colored lines indicate five examples; the
black thick line indicates the average over all pairs).
(b)
Cross-correlation. Positive correlation at negative τ indicates
that the higher-rate neuron is leading.
(c) κ, which takes
jittered spikes into account, as a function of rate difference.
larger bin size, larger values for the the correlation co-
efficient are obtained, as shown in Fig. 5(a). This is be-
cause some spikes become "coincident" only for larger bin
sizes. The output spike trains behave like jittered spikes
as discussed in earlier theoretical studies [15]. Whether
the decorrelation due to neuronal heterogeneity is signifi-
cant depends critically on the bin size, or the integration
window of the neurons receiving such inputs.
In view of a significant number of spikes jittering out-
side a 1 ms bin, we look into the cross-correlation function
rxy(τ ) =
cxy(τ )
σxσy
= hx(t)y(t + τ )i − hx(t)ihy(t)i
σxσy
, (7)
where x(t) and y(t) denote two output spike trains in dis-
crete time, consisting of 0 and 1 with bin size of 0.1 ms.
x(t) is assigned to be the spike train with lower spike
count. cxy(τ ) is the covariance function and σx and σy
denote the standard deviation of the two spike trains,
considered as discrete signals. Fig. 5(b) shows that the
mean rxy(τ ) over all pairs is positive in a small neighbor-
hood of τ , indicating a higher than chance level to ob-
serve spikes. Spikes are jittered, instead of asynchronous.
In addition, rxy(τ ) is asymmetric, and its area is skewed
towards negative τ , indicating that the higher-firing-rate
neuron is more likely to lead in terms of spiking [16, 17].
the normalized total cross-
We further
look at
covariance κ [15, 18]
We remark that the plots of the correlation coefficient
as a function of rate difference from our simulations (such
as Figs. 1(b) and 2(c) can fit the experimental results in
[3] more satisfactorily by introducing a random delay of
up to 1 ms to every incoming input spike. The spike
correlation distribution then becomes broader, and the
regime of low correlated output at small rate differences
can also be reached (data not shown). This effect could
be due to some (unknown) biological mechanism not cap-
tured in a simple LIF neuron model.
We emphasize that common input into heterogeneous
neurons is better realized by a shared noise with dis-
tributed mean and variance and, more completely, with
additionally distributed values of the membrane time
constant and the refractory period. This insight is based
on both the mathematical analysis presented here and
the in vitro experimental findings [3]. As far as firing
rates and spike correlations are concerned, the distribu-
tions of mean and variance of the input account for most
of the experimental observations. Diversity of τm and τr
has a smaller effect on the quantities in question, never-
theless, including them can account for the full degree of
heterogeneity.
In the raster plot of a neuronal population with sim-
ilar rate differences and spike correlations, synchrony is
obvious, although spike times are not precise. Spikes, if
present, are jittered in the millisecond range, which can-
not be captured by the 1 ms temporal window used for
analysis. This is why using larger bins leads to larger
values for spike correlation. We emphasize that neu-
ronal heterogeneity alone does impose an appreciable
decorrelation effect on the population activity. However,
whether decorrelation is functionally significant depends
on the readout of the downstream neurons. On top of
that, a network of heterogeneous neurons may give rise
to richer network dynamics. It remains to be explored
whether such intrinsic heterogeneity can facilitate other
decorrelation mechanisms to increase the amount of in-
formation flow [3, 19, 20]. Given the significant reduc-
tion in spike correlation among heterogeneous neurons,
research concerning correlation transmission must take
neuronal heterogeneities into consideration.
κ =
−∞ dτ cxy(τ )
R ∞
qR ∞
−∞ dτ cxx(τ )R ∞
−∞ dτ ′cyy(τ ′)
,
(8)
ACKNOWLEDGMENTS
which is an overall measure for the fraction of the spikes
that are correlated above chance level. Fig. 5(c) shows
that κ is quite close to unity, and has a weak depen-
dence on the rate difference. This indicates that spikes
would not be decorrelated when a larger time window is
considered.
We thank Volker Pernice, Moritz Deger, and Tom Tet-
zlaff for discussions. The present work was supported by
the German Federal Ministry of Education and Research
(BMBF Grant No. 01GQ0420 to "BCCN Freiburg" and
BMBF Grant No. 01GW0730 "Impulse Control") and
the EU (INTERREG-V Grant to Neurex: TriNeuron).
[1] E. Marder and J.-M. Goaillard, Nat. Rev. Neurosci. 7,
[2] Z. F. Mainen and T. J. Sejnowski, Science 268, 1503
563 (2006).
(1995).
(2008).
[3] K. Padmanabhan and N. N. Urban, Nat. Neurosci. 13,
[13] N. Brunel and V. Hakim, Neural Comput. 11, 1621
1276 (2010).
(1999).
[4] M. I. Chelaru and V. Dragoi, PNAS 105, 16344 (2008).
[5] J. Mejias and A. Longtin, Phys. Rev. Lett. 108 (2012).
[6] M. Tsodyks,
and H. Sompolinsky,
I. Mitkov,
[14] A. Siegert, Phys. Rev. 81, 617 (1951).
[15] T. Tetzlaff, S. Rotter, E. Stark, M. Abeles, A. Aertsen,
and M. Diesmann, Neural Comput. 20, 2133 (2008).
5
[7] X.
Phys. Rev. Lett. 71, 1280 (1993).
and
J. Neurosci. 16(20), 6402 (1996).
Wang
J.
[8] R. Brette, PLoS Comput. Biol. 8, e1002561 (2012).
[9] L.
Ricciardi
and
M.
L.
Sacerdote,
G.
Buzs´aki,
[16] S. Ostojic,
N. Brunel,
J. Neurosci. 29, 10234 (2009).
and V. Hakim,
[17] T. Tchumatchenko, A. Malyshev, T. Geisel, M. Volgu-
shev, and F. Wolf, Phys. Rev. Lett. 104, 058102 (2010).
and W. T. Newsome,
[18] W. Bair, E. Zohary,
Biol. Cybern. 35, 1 (1979).
J. Neurosci. 21(5), 1676 (2001).
[10] A. Kuhn, A. Aertsen, and S. Rotter, J. Neurosci. 24,
2345 (2004).
[11] J. de la Rocha, B. Doiron, E. Shea-Brown, K. Josi´c, and
A. Reyes, Nature 448, 802 (2007).
[19] A. Renart, J. de la Rocha, P. Bartho, L. Hollender,
N. Parga, A. Reyes, and K. D. Harris, Science 327, 587
(2010).
[20] M. T. Wiechert, B. Judkewitz, H. Riecke, and R. W.
[12] J.-n. Teramae and T. Fukai, Phys. Rev. Lett. 101, 248105
Friedrich, Nat. Neurosci. 13, 1003 (2010).
|
1509.01913 | 3 | 1509 | 2016-02-18T08:20:46 | Fractal Fluctuations in Human Walking: Comparison between Auditory and Visually Guided Stepping | [
"q-bio.NC"
] | In human locomotion, sensorimotor synchronization of gait consists of the coordination of stepping with rhythmic auditory cues (auditory cueing, AC). AC changes the long-range correlations among consecutive strides (fractal dynamics) into anti-correlations. Visual cueing (VC) is the alignment of step lengths with marks on the floor. The effects of VC on the fluctuation structure of walking have not been investigated. Therefore, the objective was to compare the effects of AC and VC on the fluctuation pattern of basic spatiotemporal gait parameters. Thirty-six healthy individuals walked 3 x 500 strides on an instrumented treadmill with augmented reality capabilities. The conditions were no cueing (NC), AC, and VC. AC included an isochronous metronome. In VC, projected stepping stones were synchronized with the treadmill speed. Detrended fluctuation analysis assessed the correlation structure. The coefficient of variation (CV) was also assessed. The results showed that AC and VC similarly induced a strong anti-correlated pattern in the gait parameters. The CVs were similar between the NC and AC conditions but substantially higher in the VC condition. AC and VC probably mobilize similar motor control pathways and can be used alternatively in gait rehabilitation. However, the increased gait variability induced by VC should be considered. | q-bio.NC | q-bio | Fractal Fluctuations in Human Walking: Comparison between Auditory and
Visually Guided Stepping
PHILIPPE TERRIER
1,2
1IRR, Institute for Research in Rehabilitation, Sion, Switzerland; 2Clinique romande de
réadaptation SUVACare, Sion, Switzerland
Address correspondence to Philippe Terrier, Clinique romande de réadaptation SUVACare,
Av. Gd-Champsec 90, 1951 Sion Switzerland. Tel.: +41-27-603-20-77. Electronic mail:
Philippe.Terrier[at]crr-suva.ch.
Notice: this is the author's version of a work that was accepted for publication in the Annals of
Biomedical Engineering. Changes resulting from the publishing process, such as final peer review,
editing, corrections, structural formatting, and other quality control mechanisms are not reflected in
this document. The final publication is available at Springer via http://dx.doi.org/10.1007/s10439-
016-1573-y .
Fractal fluctuations of guided stepping
2
Abstract-In human locomotion, sensorimotor synchronization of gait consists of the coordination of
stepping with rhythmic auditory cues (auditory cueing, AC). AC changes the long-range correlations
among consecutive strides (fractal dynamics) into anti-correlations. Visual cueing (VC) is the
alignment of step lengths with marks on the floor. The effects of VC on the fluctuation structure of
walking have not been investigated. Therefore, the objective was to compare the effects of AC and
VC on the fluctuation pattern of basic spatiotemporal gait parameters. Thirty-six healthy individuals
walked 3 × 500 strides on an instrumented treadmill with augmented reality capabilities. The
conditions were no cueing (NC), AC, and VC. AC included an isochronous metronome. In VC,
projected stepping stones were synchronized with the treadmill speed. Detrended fluctuation
analysis assessed the correlation structure. The coefficient of variation (CV) was also assessed. The
results showed that AC and VC similarly induced a strong anti-correlated pattern in the gait
parameters. The CVs were similar between the NC and AC conditions but substantially higher in the
VC condition. AC and VC probably mobilize similar motor control pathways and can be used
alternatively in gait rehabilitation. However, the increased gait variability induced by VC should be
considered.
Key terms-Human locomotion, Motor control, Sensorimotor synchronization, Gait variability,
Auditory cueing, Visual cueing, Long-range correlations.
Abbreviations:
AC:
Auditory cueing
NC:
No cueing
VC:
Visual cueing
DFA: Detrended fluctuation analysis
CV:
Coefficient of variation
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
3
INTRODUCTION
In the 1970s, Benoit Mandelbrot laid the foundations of a new method for understanding the
geometry of nature. He coined the term "fractal" to describe geometric objects that look identical,
regardless of the scale at which they are observed (self-similarity).20 He also developed an analogous
concept about particular time series that present self-similarities.21 In this case, a constant statistical
distribution exists across time scales. In other words, the statistical features of the parts of the series
are comparable, even if the time interval during which the observations are made changes. The
corollary is that these fractal time series exhibit serial correlations (or autocorrelations) between
consecutive samples that slowly decrease under a power law (long-range correlations).
Biological systems are inherently complex. They are constituted of multiple sub-parts-ranging
from the molecular to the population level-that interact nonlinearly on large spatial and temporal
scales. Consequently, signals measured from living organisms are most often fractal. Although the
classical approach considers the fluctuations in a physiological signal as random noise around the
true average, the fractal analysis postulates that the structure of fluctuations enlightens us about the
underlying processes that produced the signal.10 Fractal analysis has been used on a wide variety of
signals, such as heartbeat time series29 or electroencephalograms.1
In human walking, the muscles of the lower limbs cyclically propel the body forward over a certain
distance (step length) during a certain time (step time), thus maintaining ambulatory speed
(length/time). Gait control expends low energy by delivering an optimal combination of step length
and step time.48 Furthermore, an active control of lateral foot placement is needed to provide a base
of support that minimizes fall risks.3 In short, the control of human walking is a highly complex
process implying intricate interactions of feed-forward and feed-back mechanisms,11, 35 which is a
condition that is conducive for the emergence of fractal patterns in gait fluctuations. Actually, in
1995, by analyzing the walk of healthy individuals, Hausdorff and colleagues observed that the time
series of stride time (i.e., the time between two consecutive heel strikes of the same limb) exhibited
fractal fluctuations16: Deviations above and below the average tend to persist over several decades of
later strides (long memory process). In order to highlight the fractal patterns, Hausdorff et al. used
detrended fluctuation analysis (DFA), which was designed to assess the scaling properties of time
series with nonstationarities.28 In subsequent studies using DFA, they observed that, in patients with
neurological gait disorders, the fractal pattern tends to be replaced by a random pattern (i.e., no
correlations among successive strides).12, 14 Thus, DFA has often been adopted to study locomotor
function in patients with neurological disorders.13
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
4
During the decade after these seminal works, further studies using DFA extended the knowledge
about fractal fluctuations in human locomotion. In 2003, West and Scaffetta46 proposed a nonlinear
stochastic dynamical model of walking that accounted for the presence of long-range correlations in
stride time series. In 2005, Terrier et al.42 observed that the time series of stride time, as well as the
time series of stride length and stride speed were fractal. The presence of fractal patterns in several
spatiotemporal gait parameters was later confirmed by Jordan et al. in a treadmill experiment.18 In
the meantime, theoretical considerations highlighted some
issues with the DFA method,22
questioning the presence of true long-range correlations in gait time series. However, in 2009,
Delignières and Torre5 tested the effective presence of long-range correlations in human walking
with an alternative methodology43: by comparing different synthetic signals with actual stride time
series, they concluded that artificial time series that incorporated long-range correlation by design
(ARFIMA models) best fit human time series.
An interesting discovery has been that sensorimotor synchronization substantially alters the
fractal structure of human walking. Sensorimotor synchronization is the coordination of movements
with external rhythms or cues. This ability is responsible for unique human behavior, such as dancing
or performing music.32 In locomotion, sensorimotor synchronization implies guided stepping with
external cues. The simplest expression of this synchronization behavior consists of the voluntary
adjustment of heel strikes to the beat of an isochronous metronome, hereafter simply referred to as
auditory cueing (AC). AC is important in gait rehabilitation. For example, in stroke patients, a recent
meta-analysis of seven randomized controlled trials has concluded that the gait training with AC
improves walking speed, stride length and cadence.25 AC is also a key tool for improving gait among
patients with Parkinson's disease.45 AC has a strong effect on the fractal fluctuations of stride time
but no effect on stride length and stride speed.36, 42 The long-range correlated pattern is then
replaced by anti-correlations. A value above the mean is more likely to be followed by a value below
the mean (anti-persistence): the voluntary control of step duration induces a continual shift around
the target (over-correction).8, 41
Another type of externally-driven synchronization of human movement is treadmill walking.
Indeed, a motorized treadmill imposes a constant speed upon the walker. Whereas treadmill walking
has only a marginal effect on the fractal fluctuation of stride time compared to overground walking,4,
39 treadmill walking changes the serial correlations of stride speed into anti-correlations, as AC does
for stride time.8, 34, 41 Moreover, when AC is combined with treadmill walking, all the gait parameters
(stride time, stride length, and stride speed) exhibit anti-correlated patterns.34, 41 The constant speed
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
5
of the treadmill (speed cueing) requires coordinated adjustments between the stride time and the
stride length to maintain an appropriate speed and thus avoid falling off the treadmill. As a result,
when cadence and speed are voluntarily adjusted to external cues (dual cueing), stride length must
be coordinated accordingly (the "loss of redundancy" paradigm).8, 34, 41
In addition to the temporal adjustments of steps to auditory cues, another possible type of
sensorimotor synchronization consists of adjusting step length to visual cues (hereafter referred to as
visual cueing, VC). In this case, a walker anticipates the position of his or her next step to coincide
with a mark on the floor. Like AC, VC has applications in gait rehabilitation. In Parkinson's disease,
gait training with VC might have a long-term beneficial effect on walking capabilities.37 Recent
technical advances have led to the development of treadmills equipped with projection devices that
draw visual targets on the treadmill belt (augmented reality). The use of projected visual targets on a
treadmill resembles the more conventional, real, marks on the ground, but greatly facilitates the
application of VC for research and rehabilitation. Promising results have been obtained with such
treadmills, in particular for stroke rehabilitation.17 In patients in the chronic stage of stroke,
adaptability training using visually guided stepping improved obstacle-avoidance performance.44
Although the field is still in its infancy, by offering complementary solutions to AC, the VC method has
substantial potential for growth.
In summary, numerous studies have analyzed the effects of AC and treadmills on the variability8,
15, 38, 39 and the fractal pattern5, 34, 42 in human locomotion. This has led to interesting hypotheses
about the neurological basis of gait control, such as the existence of a specific neural structure that
generates fractal noise (the super central pattern generator hypothesis),5, 46 or the implication of the
minimum intervention principle.8, 41 However, whether or not VC walking supports these hypotheses
has yet to be investigated.
The first objective of the present study was to determine, in healthy individuals walking on a
treadmill, the effects of VC on stride-to-stride fluctuations and to compare the results with the
effects of AC. The hypothesis was that VC and AC similarly alter the fractal fluctuations normally
present in the time series of stride time and stride length (loss of redundancy) and that the long-
range correlations are assumed to be replaced by anti-correlations. The second objective was to
measure the fluctuation magnitude (the coefficient of variation, CV) of the spatiotemporal gait
parameters under different cueing conditions.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
6
MATERIALS AND METHODS
Subjects
Thirty-six healthy volunteers (14 men, 22 women) with no orthopedic or neurological problems
participated in the study. The mean and standard deviation (SD) of their characteristics were as
follows: age 33 years (10), body height 1.72 m (0.08), and body mass 66 kg (13). The subjects had had
no previous experience with walking following visual cues. All subjects signed an informed consent
form according to the guidelines of the local ethic committee (Commission Cantonale Valaisanne
d'Ethique Médicale, [CCVEM]), that had approved the protocol.
Material and Experimental Procedure
The instrumented treadmill was a C-mill model (ForceLink BV, Culemborg, The Netherlands),17
which is equipped with an embedded vertical force platform of 70 × 300 cm. The platform recorded
the vertical force and the position of the center of pressure at a sampling rate of 500 Hz. A projection
system displayed visual objects on the walking area from the right side of the treadmill. Ad hoc
control software (CueFors®) was used to compute the preliminary values of basic gait parameters
(stride length and duration) and to control the projection of the visual cue drawings, the stepping
stones."17
During all procedures, an elastic band was placed in front of the participant (1.40 m behind the
beginning of the belt), hanging perpendicular to the handrails, at hip level. The participant was
instructed to stay 10 cm behind the band, ensuring consistent placement on the walking area
(approximately in the middle). The reasons were the increased safety and the standardization of the
number of incoming stepping stones seen by the participant. Firstly, the preferred walking speed
(PWS) of each participant was assessed using a standardized procedure,9 which consisted of (1) a
progressive increase in the treadmill speed from a low speed (2 km h–1) until the participant reported
a comfortable pace and then (2), similarly, a progressive decrease of the treadmill speed from a high
speed (6 km h–1) to a comfortable pace. The PWS was defined as the average of four tests: two with
increasing speeds and two with decreasing speeds.
Then, the participant performed a 2 min walking test at PWS. The last 30 steps were analyzed to
measure the average preferred stride length and stride time (i.e., the duration of one gait cycle).
Then, the participant walked continuously for about 30 min at his or her PWS. Three conditions were
presented in a random order. Before each condition, the experimenter described the task to the
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
7
subject while he or she continued to walk. The conditions were as follows: (1) no cueing, i.e., normal
walking (NC); (2) AC, i.e., walking while synchronizing the heel strike to the beep of an electronic
metronome set to the preferred previously measured cadence; (3) VC, walking while adjusting steps
to stepping stones, which were 20 × 30 cm moving rectangles projected on the walking area that
went back at the same speed as the treadmill belt. The longitudinal distance along the belt between
the successive stepping stones was set to correspond to half of the preferred stride length. The
instruction was to aim accurately for each stepping stone with the foot. Given his or her position in
the middle of the treadmill, the participant could see in advance two stepping stones. For a better
understanding of the VC method, see the short movie provided in the supplemental material (S1).
Thirty seconds of familiarization with the cueing task was given to the volunteer before the recording
began. One thousand steps (500 gait cycles) were then recorded, which correspond to 10 min of
walking at a step rate of 100 steps min–1.
Data Analysis
One hundred and eight files (36 subjects × 3 conditions), each containing the position of the
center of pressure recorded at 500 Hz, were exported for the subsequent analyses that were
performed using MATLAB (MathWorks, Natick, MA). The stride time and length of each gait cycle
(500 per file) were computed using a custom algorithm that implemented a validated method.33 The
reasoning behind the method is to detect heel strikes in the longitudinal signal and then to compute
the distance and time between subsequent heel strikes of the same foot taking into account the
treadmill speed. The average speed of each gait cycle was defined as speed = stride length/stride
time. As a result, three time series (NC, AC, VC) of 500 gait cycles were obtained for stride time,
stride length, and stride speed (total 3 × 3 × 36 = 324). A typical result for one participant is shown in
Fig. 1.
To characterize the dispersion of the values around the mean in the time series (fluctuation
magnitude), the CVs were computed, which were defined as the SD normalized by mean, and were
expressed in percentages.
A fractal time series with long-range correlations exhibits an autocorrelation function C(s) that
declines following a power law (cid:2)(cid:3)(cid:4)(cid:5)∝ (cid:4)(cid:7)(cid:8),0<(cid:12)<1. The DFA is a method designed to assess the
scaling exponent α (α = 1 – γ / 2) in a time series with nonstationarities.29 Although DFA may not
always be appropriate to evidence the existence of long-range correlations in time series5, 22, 43, here,
I took advantage of its proved capacity to efficiently distinguish between statistical persistence
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
8
(short- or long-range correlations) and anti-persistence (anti-correlations).7, 8, 41 First, the time series
of length N was integrated. Then, it was divided into non-overlapping boxes of equal length n. In each
box, a linear fit, using the least squares method, was performed. The average fluctuation F(n) for that
box
length
(cid:14)(cid:3)(cid:15)(cid:5)=(cid:17)1(cid:18)(cid:19)(cid:20)(cid:21)(cid:3)(cid:22)(cid:5)−(cid:21)(cid:24)(cid:3)(cid:22)(cid:5)(cid:25)(cid:26)
(cid:27)
(cid:28)(cid:29)(cid:30)
was:
where yn(k) was the y-coordinate of the kth point of the straight line resulting from the linear fit, and
y(k) was the corresponding point in the original time series. The procedure was repeated for
increasing box sizes. Box sizes (n) ranging from 12 to 125 (i.e., N / 4) were used, with exponential
spacing to avoid a bias toward larger box sizes.23 A statistical self-affinity at different scales implies
that F(n) » nα . Therefore, because logF(n) » α·log(n), a linear fit is realized in a log-log plot between n
and F(n) to compute the scaling exponent α. If α lies between 0.5 and 1, a long-range correlation is
likely. A random process (white noise) induces α value of 0.5. In the case of anti-correlation, a small α
is expected (α < 0.5).40
Statistics
Six dependent variables were included in the statistical analysis, namely the CV and the scaling
exponent α for each gait parameter: stride time, stride length and stride speed. The independent
variable was the cueing condition (NC, AC, and VC). Boxplots were used to describe the distribution
of the individual results (Figs. 2 and 3). Means and SDs are shown in Tables 1 and 2. As inferential
statistics, one-way, repeated-measures analyses of variance (ANOVA) were used. The explained
variance was assessed with partial η2. Post hoc analyses with Tukey's honestly significant difference
(HSD) tests were used to highlight specific differences between the cueing conditions; only the p-
values of the significant differences have been included in Tables 1 and 2, with associated relative
differences.
RESULTS
Fluctuation Magnitude / Gait Variability
The distribution of the variability results (CV) is presented in Fig. 2. The presence of some outliers can
be observed. A substantially higher CV of stride time, length, and speed is observed for the VC
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
9
condition. Inferential statistics (Table 1) confirm that cueing had a significant effect on the CV of the
gait parameters. The ANOVA results show that cueing conditions explain a substantial part of the
variance (partial η
2: 0.56–0.72). In other words, VC induced a substantial increase in the gait
parameters' CV compared to NC, whereas for AC such an increase in fluctuation magnitude was not
observed.
Fractal Analysis
The boxplots of the fractal analysis results (DFA) show that stride time and stride length exhibited
long-range correlations under NC conditions (α > 0.5), while stride speed was constantly anti-
correlated (α < 0.5). An obvious anti-correlated pattern was observed for all gait parameters in the
AC and VC conditions. The ANOVA results showed significant differences among the cueing
conditions for all gait parameters, with the exception of stride speed. Concerning stride length and
stride time, a very large part of the variance is explained by the model (partial η
2 > 0.79), due to the
change in the scaling exponent from correlation to anti-correlation. Post hoc analyses confirmed
large changes in the correlation structure for stride length and stride time when AC and VC are
compared to NC. For these parameters, no difference existed between the AC and VC conditions
(similar anti-correlations).
DISCUSSION
Based on the analysis of a large number of gait cycles (54,000), the objective of the present study
was to characterize the effects of external cues on the stride-to-stride fluctuations in gait
parameters, in structural (fractal pattern) and magnitude (CV) terms. As hypothesized, concerning
fractal fluctuations, spatiotemporal parameters responded similarly to AC and VC, namely, the
emergence of strong anti-correlations was observed. However, AC and VC did not affect the
fluctuation magnitude in the same way: VC induced a substantial increase in the gait parameters' CVs
compared to the control condition, whereas no differences were observed for AC.
When individuals walk in a constant environment, basic spatiotemporal parameters fluctuate
from stride-to-stride in a narrow window, very likely framed by energetic, biomechanical,26, 48 and
stability27 constraints. In overground walking, typical CV values between 2.5% and 3% have been
reported in healthy young subjects.39, 42 Similar results have been observed in treadmill walking,39, 41
which were confirmed in the present study (CV 1.8% to 2.8%, Table 1). The effect of AC on the
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
10
fluctuation magnitude has been assessed in experimental and clinical studies. In elderly subjects, in
two studies, which analyzed overground walking with or without AC given at a preferred cadence, a
small increase in gait variability15 or no effect47 was found. In a previous study that combined
treadmill and AC (N = 20 young adults, 5 min walking), when the AC and NC conditions were
compared, a small decrease in the CVs for the stride time (–19%), stride length (–17%), and stride
speed (–7%) has been observed.41 Compared to that study, the present study included more
participants (N = 36) and longer trials (10 min). The (not significant) changes induced by AC were
stride time, –13%; stride length, –7%; and stride speed, 0% (Table 1). In summary, AC has little
impact on stride-to-stride fluctuation magnitude. As illustrated in Fig.1, the erratic wandering of the
parameter around the mean, which is typical of fractal fluctuation (first column), is replaced by high
frequency noise around the imposed frequency, which is typical of anti-correlated patterns (second
column). Thus, these concomitant changes in the fluctuation pattern triggered by AC, which
decreased low frequency wanderings and increased high frequency oscillations, nearly compensate
for each other to give a comparable fluctuation magnitude. On the other hand, VC had a profound
effect on the fluctuation magnitude. The stride time CV increased by 51%, stride length CV by 73%,
and stride speed CV by 63% (Table 1). Although the participants were able to target the stepping
stones, their stride length oscillated in an extended range (CV: 5%, 7 cm for a 1.3 m stride length).
Due to the interdependence among gait parameters, the stride time and the stride speed were also
affected. This suggests that VC was a challenging task. A cause of the difficulty was perhaps that only
two incoming steps were seen in advance by the subject, which could be too short a warning to
anticipate precise foot placement. Peper et al.30, using the same treadmill as in the present study,
showed that individuals spontaneously chose a lower walking speed in VC condition: this
corroborates the hypothesis that a sufficient reaction time to the incoming targets should be allowed
to help the subject to be comfortable with VC. However, further studies are needed to analyze gait
variability at different speeds under VC condition. As far as I know, this strong effect on fluctuation
magnitude when VC is combined with treadmill walking has not been described in the literature and
has to be confirmed independently. Interestingly, high variability has been observed in the
pathological gait of patients with Parkinson's disease (CV of stride length: 5.32%).19 Patients with
Parkinson's disease rely more on vision while walking.2 The fact that healthy individuals, who guide
their steps based on vision, also exhibited high variability may have important implications in
fundamental research and clinical application, which deserve further investigations.
Walking on a treadmill requires coordinated control of stride time and stride length to match the
treadmill speed. As evidence of this coordinated regulation, several studies have shown that stride
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
11
time and stride length are actually cross-correlated:7, 34, 41 In other words, stride time and stride
length vary over time in a similar way (positive correlations).34 A model that explains how motor
control manages speed regulation has been proposed by Dingwell and collaborators.6, 8 A key feature
of this model is that an infinite combination of stride time and stride length is possible to meet the
goal of maintaining a constant speed. In other words, redundancy exists between spatial and
temporal control of walking speed: if a deviation occurs in the stride length, the deviation can be
compensated for by a correction in the stride time, and vice versa. As a result, deviations can persist
over consecutive strides, which may explain long-range correlations and fractal fluctuations (the
minimum intervention principle).6 If two goals are simultaneously imposed on a walking individual,
for instance, treadmill walking (speed goal) and AC (stride time goal), the redundancy disappears,
and tight regulation of stride length is also performed, which leads to the loss of the fractal
fluctuations, as demonstrated in previous studies8, 34, 41 and in the present study (Fig. 3). The main
goal of the present study was to provide further evidence to support this model. As expected,
imposing a spatial goal (stepping to visual cues) modified the fluctuation pattern of stride length as
well as stride time (Fig. 3 and Table 2). Actually, AC and VC altered fractal fluctuations in a similar way
(Table 2). As observed previously, strong anti-correlations appeared.41, 42
As illustrated in Table 3, the results of the present study provide new information to complement
what is already known about voluntary synchronization of gait to external cues. The main gap that
remains to be filled is the fluctuation pattern of overground walking under the VC condition. It can be
assumed that stride length would then be anti-correlated, while stride time and stride speed would
remain correlated. However, conducting such an experiment is technically challenging.
Taken together, these findings highlight a stereotyped response of gait control to external cueing,
the neurobiological basis of which remains to be elucidated. What is already known is that
sensorimotor synchronization requires anticipation to align motor response with external cues. The
delivery of sensory signals and central processing delay the motor response. Thus, to get in time,
motor command must anticipate future steps based on internal models and past sensory inputs and
not react to current stimuli.24 This anticipation leads to a well-known phenomenon in rhythmic
tapping experiments: the taps precede the metronome (negative mean asynchrony).31 Likewise, in
gait experiments with AC, the heel-strike slightly precedes (50 ms on average) the next occurrence of
the metronome sound.23 The anticipation of movement based on visual cues, for example, reaching a
moving object, is also a basic task of sensorimotor coordination that implies feedforward
mechanisms.24 In the present experiment, the subject saw two stepping stones in advance to
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
12
anticipate the correct step lengths. In short, guided stepping requires voluntary adaptation to
external cues and anticipation to produce a timed motor response. It can be assumed that visual,
auditory, and somatosensory afferents converge to modulate a central pacemaker that triggers the
appropriate rhythmic gait behavior under voluntary control. How exactly this sensory feedback feeds
this hypothetical pacemaker, and how an anti-correlated pattern is produced remain to be further
investigated.
In conclusion, because it can be assumed that AC and VC mobilize the same motor pathways, they
can probably be used alternatively in gait rehabilitation. The efficiency of VC to enhance walking
abilities in patients with neurological gait disorders needs further studies. However, the high gait
variability induced by VC might have detrimental effects, for instance, a lower dynamic balance. This
should be taken into account in the development of VC rehabilitation methods.
ACKNOWLEDGMENTS
The author warmly thanks Emilie Du Fay de Lavallaz for his valuable support in
bibliographical research and Vincent Bonvin for his help in data collection. The study was
funded by the Swiss accident insurance company SUVA, an independent, non-profit
company under public law, and by the clinique romande de réadaptation. The IRR (Institute
for Research in Rehabilitation) is funded by the State of Valais and the City of Sion. Study
funders had no role in the collection, analysis, and interpretation of data; in the writing of
the manuscript; and in the decision to submit the manuscript for publication.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
13
REFERENCES
Azulay J.-P., S. Mesure, B. Amblard, O. Blin, I. Sangla and J. Pouget. Visual control of
Dingwell J. B. and J. P. Cusumano. Identifying Stride-To-Stride Control Strategies in Human
Dingwell J. B. and J. P. Cusumano. Re-interpreting detrended fluctuation analyses of stride-
Dingwell J. B., J. John and J. P. Cusumano. Do humans optimally exploit redundancy to control
Dingwell J. B. and L. C. Marin. Kinematic variability and local dynamic stability of upper body
Eke A., P. Herman, L. Kocsis and L. Kozak. Fractal characterization of complexity in temporal
Chang M. D., S. Shaikh and T. Chau. Effect of treadmill walking on the stride interval dynamics
Accardo A., M. Affinito, M. Carrozzi and F. Bouquet. Use of the fractal dimension for the
Bauby C. E. and A. D. Kuo. Active control of lateral balance in human walking. J. Biomech.
Delignieres D. and K. Torre. Fractal dynamics of human gait: a reassessment of the 1996 data
1.
analysis of electroencephalographic time series. Biol. Cybern. 77:339-350, 1997.
2.
locomotion in Parkinson's disease. Brain 122:111-120, 1999.
3.
33:1433-1440, 2000.
4.
of human gait. Gait Posture 30:431-435, 2009.
5.
of Hausdorff et al. J. Appl. Physiol. 106:1272-1279, 2009.
6.
Treadmill Walking. PloS ONE 10:e0124879, 2015.
7.
to-stride variability in human walking. Gait Posture 32:348-353, 2010.
8.
step variability in walking? PLoS Comput. Biol. 6:e1000856, 2010.
9.
motions when walking at different speeds. J. Biomech. 39:444-452, 2006.
10.
physiological signals. Physiol. Meas. 23:R1, 2002.
11.
a systematic review. Neurosci. Biobehav. R. 57:310-327, 2015.
12.
stride length, gait variability, and fractal-like scaling. Chaos 19:026113, 2009.
13.
fluctuations of human walking. Hum. Mov. Sci. 26:555-589, 2007.
14.
Hausdorff J. M., A. Lertratanakul, M. E. Cudkowicz, A. L. Peterson, D. Kaliton and A. L.
Goldberger. Dynamic markers of altered gait rhythm in amyotrophic lateral sclerosis. J. Appl. Physiol.
88:2045-2053, 2000.
15.
Hausdorff J. M., J. Lowenthal, T. Herman, L. Gruendlinger, C. Peretz and N. Giladi. Rhythmic
auditory stimulation modulates gait variability in Parkinson's disease. Eur. J. Neurosci. 26:2369-2375,
2007.
16.
Hausdorff J. M., C. K. Peng, Z. Ladin, J. Y. Wei and A. L. Goldberger. Is walking a random walk?
Evidence for long-range correlations in stride interval of human gait. J. Appl. Physiol. 78:349-358,
1995.
Heeren A., M. W. van Ooijen, A. C. Geurts, B. L. Day, T. W. Janssen, P. J. Beek, M. Roerdink
17.
and V. Weerdesteyn. Step by step: a proof of concept study of C-Mill gait adaptability training in the
chronic phase after stroke. J. Rehabil. Med. 45:616-622, 2013.
18.
Gait Posture 26:128-134, 2007.
Hamacher D., F. Herold, P. Wiegel, D. Hamacher and L. Schega. Brain activity during walking:
Hausdorff J. M. Gait dynamics in Parkinson's disease: common and distinct behavior among
Jordan K., J. H. Challis and K. M. Newell. Walking speed influences on gait cycle variability.
Hausdorff J. M. Gait dynamics, fractals and falls: finding meaning in the stride-to-stride
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
14
Lewis G. N., W. D. Byblow and S. E. Walt. Stride length regulation in Parkinson's disease: the
O'Connor S. M., H. Z. Xu and A. D. Kuo. Energetic cost of walking with increased step
Peng C. K., S. Havlin, H. E. Stanley and A. L. Goldberger. Quantification of scaling exponents
Peper C. L. E., M. J. de Dreu and M. Roerdink. Attuning one's steps to visual targets reduces
Repp B. H. Sensorimotor synchronization: a review of the tapping literature. Psychon. Bull.
Maraun D., H. W. Rust and J. Timmer. Tempting long-memory - on the interpretation of DFA
Marmelat V., K. Torre, P. J. Beek and A. Daffertshofer. Persistent fluctuations in stride
Owings T. M. and M. D. Grabiner. Variability of step kinematics in young and older adults.
Peng C.-K., S. V. Buldyrev, S. Havlin, M. Simons, H. E. Stanley and A. L. Goldberger. Mosaic
Miall R. C. and D. M. Wolpert. Forward models for physiological motor control. Neural
Mandelbrot B. B. Les objets fractals: forme, hasard et dimension. Flammarion, 1975.
Mandelbrot B. B. and J. W. Van Ness. Fractional Brownian motions, fractional noises and
19.
use of extrinsic, visual cues. Brain 123 ( Pt 10):2077-2090, 2000.
20.
21.
applications. SIAM review 10:422-437, 1968.
22.
results. Nonlinear Proc. Geoph. 11:495-503, 2004.
23.
intervals under fractal auditory stimulation. PLoS ONE 9:e91949, 2014.
24.
networks 9:1265-1279, 1996.
25.
Nascimento L. R., C. Q. de Oliveira, L. Ada, S. M. Michaelsen and L. F. Teixeira-Salmela.
Walking training with cueing of cadence improves walking speed and stride length after stroke more
than walking training alone: a systematic review. J. Physiother. 61:10-15, 2015.
26.
variability. Gait Posture 36:102-107, 2012.
27.
Gait Posture 20:26-29, 2004.
28.
organization of DNA nucleotides. Phys. Rev. E 49:1685, 1994.
29.
and crossover phenomena in nonstationary heartbeat time series. Chaos 5:82-87, 1995.
30.
comfortable walking speed in both young and older adults. Gait Posture 41:830-834, 2015.
31.
Rev. 12:969-992, 2005.
32.
2012). Psychon. Bull. Rev. 20:403-452, 2013.
33.
detection using a large force platform embedded in a treadmill. J. Biomech. 41:2628-2632, 2008.
34.
Persistent Fractal Metronome without Falling Off the Treadmill? PLoS ONE 10:e0134148, 2015.
35.
Physiol. Rev. 86:89-154, 2006.
36.
Temporal Dynamics of Human Gait. PLoS ONE 7:e43104, 2012.
37.
training using visual cues in an individual with Parkinson disease. Phys. Ther. 86:186-194, 2006.
38.
Cueing. PLoS ONE 7:e47171, 2012.
39.
treadmill walking. J. Neuroeng. Rehabil. 8:12, 2011.
40.
auditory cueing on local dynamic stability. Front. Physiol. 4:230, 2013.
41.
treadmill walking: influence of rythmic auditory cueing. Hum.Mov. Sci. 31:1585-1597, 2012.
42.
Terrier P., V. Turner and Y. Schutz. GPS analysis of human locomotion: further evidence for
long-range correlations in stride-to-stride fluctuations of gait parameters. Hum. Mov. Sci. 24:97-115,
2005.
Sidaway B., J. Anderson, G. Danielson, L. Martin and G. Smith. Effects of long-term gait
Terrier P. Step-to-Step Variability in Treadmill Walking: Influence of Rhythmic Auditory
Repp B. H. and Y.-H. Su. Sensorimotor synchronization: A review of recent research (2006-
Roerdink M., B. H. Coolen, B. H. Clairbois, C. J. Lamoth and P. J. Beek. Online gait event
Rossignol S., R. Dubuc and J. P. Gossard. Dynamic sensorimotor interactions in locomotion.
Sejdic E., Y. Fu, A. Pak, J. A. Fairley and T. Chau. The Effects of Rhythmic Sensory Cues on the
Terrier P. and O. Dériaz. Kinematic variability, fractal dynamics and local dynamic stability of
Terrier P. and O. Dériaz. Nonlinear dynamics of human locomotion: effects of rhythmic
Terrier P. and O. Dériaz. Persistent and anti-persistent pattern in stride-to-stride variability of
Roerdink M., A. Daffertshofer, V. Marmelat and P. J. Beek. How to Sync to the Beat of a
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
15
Torre K., D. Delignieres and L. Lemoine. Detection of long-range dependence and estimation
van Wegen E. E., M. A. Hirsch, M. Huiskamp and G. Kwakkel. Harnessing Cueing Training for
43.
of fractal exponents through ARFIMA modelling. Brit. J. Math. Stat. Psy. 60:85-106, 2007.
44.
van Ooijen M. W., A. Heeren, K. Smulders, A. C. Geurts, T. W. Janssen, P. J. Beek, V.
Weerdesteyn and M. Roerdink. Improved gait adjustments after gait adaptability training are
associated with reduced attentional demands in persons with stroke. Exp. Brain Res. 233:1007-1018,
2015.
45.
Neuroplasticity in Parkinson Disease. Top. Geriatr. Rehabil. 30:46-57, 2014.
46. West B. J. and N. Scafetta. Nonlinear dynamical model of human gait. Phys. Rev. E 67:051917,
2003.
47. Wittwer J. E., K. E. Webster and K. Hill. Music and metronome cues produce different effects
on gait spatiotemporal measures but not gait variability in healthy older adults. Gait Posture 37:219-
222, 2013.
48.
walking. Eur. J. Appl. Physiol. Occup. Physiol. 33:293-306, 1974.
Zarrugh M. Y., F. N. Todd and H. J. Ralston. Optimization of energy expenditure during level
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
16
Figures
FIGURE 1. Typical results. One participant walked approximately 3X10min on an instrumented
treadmill under 3 conditions: no cueing (NC), auditory cueing (AC) and visual cueing (VC). In each
condition, 500 strides were recorded (x-axes, #stride). Basic spatiotemporal gait parameters (y-axes),
i.e., stride time (ST), stride length (SL), and stride speed (SS), were assessed.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
17
FIGURE 2. Boxplots of the fluctuation magnitude results. Thirty-six subjects walked on a treadmill
under 3 conditions: NC = no cueing (normal walking); AC = auditory cueing (metronome walking); VC
= visual cueing (visually guided stepping). In each condition for each subject, 500 gait cycles were
recorded, from which the relative fluctuation magnitude (variability) of the spatiotemporal gait
parameters was assessed (i.e. CV = SD / mean * 100). Boxplots show the quartiles, the medians and
the ranges of the individual results. Outliers are indicated with + signs.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
18
FIGURE 3. Boxplots of the fractal fluctuation results. Thirty-six subjects walked on a treadmill under
3 conditions: NC = no cueing (normal walking); AC = auditory cueing (metronome walking); VC =
visual cueing (visually guided stepping). In each condition for each subject, 500 gait cycles were
recorded, from which the serial correlations of the spatiotemporal gait parameters were assessed
(scaling exponent α, DFA). Boxplots show the quartiles, the medians and the ranges of the individual
results. Outliers are indicated with + signs.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
19
Tables
TABLE 1. Fluctuation magnitude of the gait parameters
Descriptive statistics: Mean
(SD)
ANOVA
Post hoc comparisons
N=36
NC
AC
VC
F
VC vs NC
VC vs AC
p
η2
CI
p
D %
p
D %
Stride time CV (%) 1.6 (0.6) 1.4
(0.5) 2.3
(0.9) 45.1 <0.001 0.56
0.45
0.70 <0.001
51
<0.001
66
Stride length CV (%) 2.8 (0.8) 2.6
(0.7) 4.8
(1.7) 88.5 <0.001 0.72
0.62
0.83 <0.001
73
<0.001
82
Stride speed CV (%) 1.8 (0.5) 1.8
(0.5) 2.8
(0.9) 83.6 <0.001 0.70
0.60
0.83 <0.001
63
<0.001
62
Thirty-six subjects walked on a treadmill under 3 conditions: NC = no cueing (normal walking); AC =
auditory cueing (metronome walking); and VC = visual cueing (visually guided stepping). In each
condition for each subject, 500 gait cycles were recorded, from which the relative fluctuation
magnitude (variability) of the spatiotemporal gait parameters was assessed (i.e. CV = SD/mean *
100). Descriptive and inferential statistics and significant post hoc comparisons are shown. % is the
relative difference: (VC-NC)/NC *100. Significant results are shown in bold.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
20
TABLE 2. Long-range correlations (fractal fluctuations) of the gait parameters
Descriptive statistics: Mean (SD)
ANOVA
Post hoc comparisons
AC vs NC
VC vs NC
N=36
NC
AC
VC
F
p
η2
η2 CI
p
D %
p
D %
Stride time α 0.90
(0.13) 0.32
(0.17) 0.34 (0.14) 209 <0.001 0.86 0.80 0.91 <0.001
-64
<0.001
-62
Stride length α 0.67
(0.12) 0.31
(0.12) 0.33 (0.15) 135 <0.001 0.79 0.73 0.86 <0.001
-54
<0.001
-51
Stride speed α 0.32
(0.11) 0.32
(0.11) 0.29 (0.15) 1.43
0.25 0.04 0.00 0.21
-
-
-
-
Thirty-six subjects walked on a treadmill under 3 conditions: NC = no cueing (normal walking); AC =
auditory cueing (metronome walking); VC = visual cueing (visually guided stepping). In each condition
for each subject, 500 gait cycles were recorded, from which the serial correlations of the
spatiotemporal gait parameters were assessed (scaling exponent α, detrended fluctuation analysis).
Descriptive and inferential statistics and significant post hoc comparisons are shown. % is the
relative difference: (VC-NC)/NC *100. Significant results are shown in bold.
Author's accepted version
Rev2, Feb. 2016
Fractal fluctuations of guided stepping
21
TABLE 3. Overview of the cueing effects on fractal dynamics and fluctuations magnitude in human
walking.
Stride time
Stride length
Stride speed
Overground walking16, 42
Overground walking + AC42
α >0.5 / fl
α <0.5 / fl
Overground walking + VC
?
Treadmill walking8, 34, 41
Treadmill walking + AC34, 41
Treadmill walking + VC
α >0.5 / fl
α <0.5 / fl
α <0.5 / ›
α >0.5 / fl
α >0.5 / fl
?
α >0.5 / fl
α <0.5 / fl
α <0.5 / ›
α >0.5 / fl
α >0.5 / fl
?
α <0.5 / fl
α <0.5 / fl
α <0.5 / ›
Each row shows the results for a particular cueing condition on the different gait parameters
(columns). AC: auditory cueing. VC: visual cueing. α >0.5 (green): DFA indicates the presence of long-
range correlations. α <0.5 (red): DFA indicates the presence of anti-correlations. fl
fluctuation magnitude. ›
(red): high fluctuation magnitude. The new findings brought by the present
(green): low
study are shown in bold. The numbers refers to the bibliography at the end of the article.
Author's accepted version
Rev2, Feb. 2016
›
›
›
›
›
›
›
›
›
|
1810.05077 | 1 | 1810 | 2018-10-10T09:22:21 | On the Brain Networks of Complex Problem Solving | [
"q-bio.NC",
"cs.CV"
] | Complex problem solving is a high level cognitive process which has been thoroughly studied over the last decade. The Tower of London (TOL) is a task that has been widely used to study problem-solving. In this study, we aim to explore the underlying cognitive network dynamics among anatomical regions of complex problem solving and its sub-phases, namely planning and execution. A new brain network construction model establishing dynamic functional brain networks using fMRI is proposed. The first step of the model is a preprocessing pipeline that manages to decrease the spatial redundancy while increasing the temporal resolution of the fMRI recordings. Then, dynamic brain networks are estimated using artificial neural networks. The network properties of the estimated brain networks are studied in order to identify regions of interest, such as hubs and subgroups of densely connected brain regions. The major similarities and dissimilarities of the network structure of planning and execution phases are highlighted. Our findings show the hubs and clusters of densely interconnected regions during both subtasks. It is observed that there are more hubs during the planning phase compared to the execution phase, and the clusters are more strongly connected during planning compared to execution. | q-bio.NC | q-bio |
ON THE BRAIN NETWORKS OF COMPLEX PROBLEM SOLVING
Abdullah Alchihabi † Omer Ekmekci † Baran B. Kivilcim † Sharlene D. Newman(cid:63) Fatos T. Yarman Vural †
Middle East Technical University, {abdullah.alchihabi, oekmekci, baran.kivilcim, vural}@ceng.metu.edu.tr
† Department of Computer Engineering
(cid:63) Department of Psychological and Brain Sciences
Indiana University , [email protected]
ABSTRACT
Complex problem solving is a high level cognitive process
which has been thoroughly studied over the last decade. The
Tower of London (TOL) is a task that has been widely used
to study problem-solving. In this study, we aim to explore
the underlying cognitive network dynamics among anatomi-
cal regions of complex problem solving and its sub-phases,
namely planning and execution. A new brain network con-
struction model establishing dynamic functional brain net-
works using fMRI is proposed. The first step of the model
is a preprocessing pipeline that manages to decrease the spa-
tial redundancy while increasing the temporal resolution of
the fMRI recordings. Then, dynamic brain networks are es-
timated using artificial neural networks. The network prop-
erties of the estimated brain networks are studied in order to
identify regions of interest, such as hubs and subgroups of
densely connected brain regions. The major similarities and
dissimilarities of the network structure of planning and execu-
tion phases are highlighted. Our findings show the hubs and
clusters of densely interconnected regions during both sub-
tasks. It is observed that there are more hubs during the plan-
ning phase compared to the execution phase, and the clusters
are more strongly connected during planning compared to ex-
ecution.
Index Terms -- fMRI, Machine Learning, Brain Net-
works, Tower of London, Complex Problem Solving.
1. INTRODUCTION
Complex problem solving has been the focus of numerous
studies in the field of neuroscience for over 30 years given the
large number of high-level cognitive tasks that fall under its
umbrella. To name a few, complex problem solving includes,
strategy formation, coordination, sequencing of mental func-
tions, and holding information on-line. These complex high-
level cognitive sub-processes make revealing the inner work-
ings of problem solving difficult.
The Tower of London (TOL) task, designed by Tim Shal-
lice in (1982) [1], has been one of the standard tools in the lit-
erature to study complex problem solving. It consists of three
bins having different capacities with colored balls placed in
the bins; the aim is to rearrange the balls from their initial
state to a predetermined goal state while moving one ball at
a time and taking into consideration the limited capacity of
each bin.
TOL also has been used to investigate the effect of various
clinical disorders on functions associated with the prefrontal
cortex such as planning. It has been utilized to identify exec-
utive dysfunction in children and adolescents suffering from
epilepsy and seizures [2]. It also has been used to analyze the
cognitive activation patterns of individuals suffering from de-
pression [3]. Additionally, it has been employed to examine
cognitive impairment in patient's diagnosed with Parkinson
disease [4]. In another study, TOL along with fMRI have been
employed to study the differences in the neural basis of plan-
ning and executive function between first-episode schizophre-
nia patients and healthy subjects [5].
Besides clinical disorders, TOL has also been employed
to study the effect of various parameters on complex prob-
lem solving performance in healthy subjects. The predictive
power of working memory, inhibition, and fluid intelligence
on TOL performance has been explored [6, 7]. Also, the ef-
fect of physical activity and exercise, age, gender, and im-
pairment in the executive function on planning and problem-
solving ability and its underlying neural basis have all been
studied [8, 9, 10, 11, 12, 13, 14].
TOL itself has many variations due to its large number of
parameters such as goal hierarchy, demand for subgoal gen-
eration, number of solution paths, and the existence of sub-
optimal alternatives. Several studies have examined the ef-
fect of the aforementioned structural parameters along with
numerous other non-structural parameters including instruc-
tions, experience, environment and problem-solving strategy
on the performance of subjects when solving TOL puzzles,
where performance is measured by pre-planning time and ac-
curacy [15, 16, 17, 18, 19, 20, 21].
Despite the popularity of the TOL problem in the litera-
ture and the wide range of problems covered in these stud-
ies, relatively few researchers have explored the underlying
network structures. In [22], the involvement of the parietal
cortex, prefrontal cortex, basal ganglia and anterior cingu-
late in complex problem solving was reported. The activa-
tion patterns of the dorsolateral and rostrolateral subregions
of the prefrontal cortex during planning has been examined
[23]. The focus of another study has been the hemispheric
differences in the pre-frontal cortex during planning and exe-
cution as well as the contribution of the superior parietal re-
gion to spatial working memory [24]. In addition, some work
has been done to investigate the variance in the neural basis
of planning between standard and expert subjects [25]. Other
works have investigated the cognitive activation patterns dur-
ing planning and execution subtasks [26, 27, 28, 29]. Given
this brief literature review, a holistic understanding of the ac-
tive anatomical brain regions, their respective roles and their
interactions during complex problem-solving is an important,
yet lacking research study.
Numerous studies have proposed various computational
models in order to build brain networks from fMRI measure-
ments, both during cognitive tasks or during resting state.
These studies represent a shift in the literature towards brain
decoding algorithms that are based on the connectivity pat-
terns in the brain motivated by the findings that these patterns
provide more information about cognitive tasks than the iso-
lated behavior of individual voxel groups or anatomical re-
gions [30, 31, 32, 33, 34].
Some of these studies focused on the pairwise relation-
ships between voxels or brain regions. For example, Pear-
son correlation has been used in order to construct undirected
functional connectivity graphs at different frequency resolu-
tions in [35]. Also, pairwise correlations and mutual informa-
tion have also been used in order build functional brain net-
works in various studies aiming to investigate the network dif-
ferences between patients with Schizophrenia or Alzheimers
disease and healthy subjects [36, 37, 38]. Partial correlation
along with constrained linear regression was also used to gen-
erate brain networks in [39].
Other studies take advantage of the locality property of the
brain by constructing local mesh networks around each brain
region then representing the entire brain network as an ensem-
ble of local meshes. In such studies, the Blood-Oxygenation
Level Dependent (BOLD) response of each brain region is es-
timated as a linear combination of the responses of its closest
neighboring regions. Levinson-Durbin recursion has been ap-
plied in several studies in order to estimate the edge weights
of each local star mesh, where the nodes are the neighboring
regions of the seed brain region [40, 41]. Ridge regression
has also been used to estimate edge weights while construct-
ing local mesh networks across windows of time [42, 32].
Other works suggested various methods to prune the
constructed brain networks. In [43], researchers established
resting-state brain networks as sparse constrained networks
using both L1 and L2 regularization to introduce sparsity
and control for across-subject variability. Another study
used a two-step model to build functional brain networks,
where at first a sparse multivariate autoregressive model was
employed with penalized regression to estimate the brain
networks. Then, false discovery rate (FDR) is used to prune
low probability connections to present sparsity in the brain
network [44].
After constructing brain networks, the statistical proper-
ties of the established networks are studied in order to ob-
tain neuroscientific insights related to the experiment at hand.
Several measures of centrality have been proposed that aim
to identify potential hubs that are central to the flow of infor-
mation in the network such as: node degree, node strength
and node betweenness centrality.
In addition, measures of
functional segregation have been proposed that aim to detect
subgroups or modules of densely interconnected anatomic re-
gions such as: local efficiency, clustering coefficient and tran-
sitivity [45, 46, 47]. Furthermore, the node degree distribution
of constructed brain networks has been measured and com-
pared with power law and truncated power law distributions
[48]. Also, the small-world property of brain networks has
been studied extensively in numerous studies [48, 46]. Some
studies, such as [49], have further extended the literature by
defining the null model for weighted undirected functional
brain networks. Further work have focused on controlling
for family-wise error (FWE) that complements false discov-
ery rate (FDR) [50].
Several studies have compared the network properties of
functional brain networks across different age groups [51, 47]
and under different developmental factors [52]. Other stud-
ies have performed similar analysis to compare the network
properties of healthy individuals against those suffering from
several diseases related to cognitive impairment (Alzheimer,
epilepsy, Schizophrenia) [53].
In this study, we propose a dynamic functional brain net-
work model, extracted from fMRI measurements using arti-
ficial neural networks. The decoding power of the suggested
brain network model is examined by distinguishing the two
phases of complex problem solving, namely, planning and
execution. Then, the network properties of the dynamic brain
networks are studied in order to identify the active anatomical
regions during both planning and execution phases of com-
plex problem solving. Potential hubs and clusters of densely
connected brain regions are identified for both subtasks. Fur-
thermore, the distinctions and similarities between planning
and execution networks are highlighted. The results identify
both active, inactive, hub regions as well as clusters of densely
connected anatomical regions during complex problem task.
In addition, results show that there are potential hubs during
planning phase compared to execution phase, also the clus-
ters of densely interconnected regions are significantly more
strongly connected during planning compared to execution.
2. TOL EXPERIMENT PROCEDURE
In this section, we introduce the details of the experiment as
well as data collection and preprocessing methods.
2.1. Participants & Stimuli
18 college students aged between 19 and 38 participated in
the experiment after signing informed, written consent doc-
uments approved by the Indiana University Institutional Re-
view Board. The subjects solved a computerized version of
TOL problem, two configurations were presented at the be-
ginning of each puzzle:
the initial state and the goal state.
The subjects were asked to transform the initial state into the
goal state using the minimum number of moves. However,
the subjects were not informed of the minimum number of
moves needed to solve a given puzzle nor of the existence of
multiple solution paths.
2.2. Procedure
Each subject underwent a practice session before entering the
scanning session to acquaint subjects with the TOL problem.
The subjects were given the following instructions: You will
be asked to solve a series of puzzles. The goal of the puz-
zle is to make the start or current state match the goal state
(They were shown an example). Try to solve the problems
in the minimum number of moves by planning ahead. Work
as quickly and accurately as possible, but accuracy is more
important than speed.
The scanning session consisted of 4 runs, each run in-
cluded 18 timed puzzles, with a 5 second planning only time
slot during which subjects were not allowed to move the balls.
However, they were allowed to continue planning after the 5
seconds planning only time slot if they chose to do so. Fol-
lowing every puzzle, there was a 12-second rest period where
subjects focused on a plus sign in the center of the screen.
Each run was also followed by a 28-second fixation period.
2.3. fMRI Data Acquisition & Preliminary Analysis
The fMRI images were collected using a 3 T Siemens TRIO
scanner with an 8-channel radio frequency coil located in the
Imaging Research Facility at Indiana University. The images
were acquired in 18 5 mm thick oblique axial slices using the
following set of parameters: TR=1000 ms, TE=25 ms, flip
angle=60◦, voxel size=3.125 mm×3.125 mm×5 mm with a 1
mm gap.
The statistical parametric mapping toolbox was used to
perform the preliminary data analysis that included:
image
correction for slice acquisition timing, resampling, spatial
smoothing, motion correction and normalization to the Mon-
treal Neurological Institute (MNI) EPI template. Further
details concerning the procedure and data acquisition can be
found in [22].
3. MESH BRAIN NETWORKS FOR COMPLEX
PROBLEM SOLVING
3.1. Data Processing
Given the small number of subjects in the TOL dataset, and
the large number of voxels in each brain volume (185,405
voxels per time instant), voxel selection is used to reduce the
spatial resolution of the collected brain images and dampen
the noise that is inherent in the data. Furthermore, due to the
short duration of each puzzle (max 15 seconds) and the rela-
tively low sampling rate (TR = 1 sec), temporal interpolation
is needed in order to increase the number of brain volumes
for each puzzle. Finally, Gaussian noise is used in order to
regularize the data and improve generalization.
3.1.1. Voxel Selection
First, an ANOVA feature selection method is used to choose
the most discriminative subset of voxels and discard the re-
maining ones [54, 55, 56]. For this purpose, we calculate the
f-value score of each voxel vi as shown in equation 1:
f scorei =
M SB(vi, ylabel)
M SW (vi, ylabel)
(1)
where ylabel is the label indicating the subtask (Planning
or Execution). M SB(vi, ylabel) is the mean square value be-
tween voxel i and the label vector ylabel which is calculated
by equation 2
M SB(vi, ylabel) =
SSB(vi, ylabel)
dfbetween
,
(2)
SSB(vi, ylabel) is the sum of squares between ylabel
and vi, dfbetween is the number of groups minus one.
M SW (vi, ylabel) is the mean square value within voxel i
and the label vector ylabel and it is calculated by 3
M SW (vi, ylabel) =
SSW (vi, ylabel)
dfwithin
(3)
where, SSW (vi, ylabel) is the sum of squares within
group and dfwithin is the degree of freedom within (total
number of elements in vi and ylabel minus the number of
groups).
We order the voxels according to their f-value scores.
Then, the distribution of f-value scores of all voxels is plot-
ted in order to determine the appropriate number of voxels to
retain. Voxel selection is applied to the voxels of all brain
regions except the ones located in the cerebellum, which we
exclude during network extraction.
Voxel selection successfully manages to significantly re-
duce the number of voxels in each brain volume, thus, making
the space and time complexity of the analysis on the dataset
feasible given the large total number of voxels in each brain
volume, 185,405 voxels per time instant. It is also a neces-
sary step for decreasing the curse of dimensionality problem
for decoding the planning and execution phases.
The BOLD response of the selected voxels is then aver-
aged into their corresponding brain regions defined by the
automated anatomical labeling (AAL) atlas [57] as shown in
equation 4:
(cid:80)
rj =
i∈ζ[j] vi
ζ[j]
(4)
where rj is the BOLD response of region j, vi is the
BOLD response of voxel i and ζ[j] is the set of selected vox-
els located in region j. Averaging the selected voxels in their
corresponding anatomical regions smooths the noise embed-
ded in fMRI signal to a certain degree, and further reduces
the dimensionality of each brain volume. As a result, each
region is represented by a BOLD response, thus, enabling us
to investigate the role and contribution of each region to the
planning and execution phases of the problem solving task.
3.1.2. Interpolation
It is well-known that in spite of its high spatial resolution,
fMRI signal has very low temporal resolution compared to
EEG signal. In this study, we interpolate the fMRI signal in
order to compensate for this drawback and study the effect of
interpolation on decoding planning and execution phases of
TOL.
In the TOL study subjects solved a puzzle in at most 15
seconds and the sampling rate, TR, is 1000 ms. Interpolation
is used to increase the temporal resolution by estimating z
extra brain volumes between each two consecutive measured
brain volumes. As a result, the total number of available brain
volumes for each puzzle becomes n+z∗(n−1), where n is the
number of measured brain volumes of a given puzzle. We use
the cubic spline interpolation function rather than linear inter-
polation methods in order to prevent edge effects and smooth-
ing out the spikes between the measured brain volumes [58].
In order to analyze the effect of time interpolation and to
estimate an acceptable number of inserted brain volumes z,
we compare the Fourier Transform of the fMRI signal com-
puted before and after interpolation so that the frequency con-
tent of the signal is not distorted by interpolation. The origi-
nal single-sided amplitude of the signal and the one obtained
after interpolation are compared in order to ensure that in-
terpolation is preserving the smooth peaks of the data in the
frequency domain [59, 60].
3.1.3. Injecting Gaussian Noise
When modeling a deterministic signal by a probabilistic
method, adding noise to the signal decreases the estimation
error in most of the practical applications. The final phase of
preprocessing is adding a Gaussian noise to the interpolated
time series of the BOLD response in each anatomical region.
For this purpose, instead of just injecting white noise, a rather
informed noise, colorful Gaussian noise, is added. In order to
reflect the corresponding brain region's properties, for each
sample the additive noise sample is generated from a Gaus-
sian distribution having mean and variance of that anatomical
region. This newly generated samples not only act like a
natural regularizer to improve generalization performance of
brain decoding, but, also help making local mesh estima-
tion algorithms more stable when generating brain networks
[61, 62].
Given a representative time series from a particular
brain region, i represents the index of an anatomical re-
gion. The new samples are generated with vector addition
of noise while preserving the signal-to-noise ratio (SNR) as
in rj = rj + τj, where τj is a noise vector sampled from
N (αnoise µ(rj), βnoise σ2(rj)), αnoise and βnoise are the
scaling factors which are set empirically, to optimize the
decoding performance.
3.2. Building Dynamic Brain Network with Neural Net-
works
After applying the preprocessing pipeline, we construct dy-
namic functional brain networks. In order to do that, we par-
tition each time series, which represents an anatomical region
into fixed-size windows, where each window, win(t), is cen-
tered at the measured brain volume at time instance, t. The
size of each window is W in Size = z + 1 brain volumes,
where z is the number of interpolated brain volumes in each
window. Equation 5 shows the time instances included in
each window.
(cid:34)
t −(cid:106) z
(cid:107)
2
(cid:109)(cid:35)
(cid:108) z
2
win(t) =
, .., t, .., t +
(5)
We construct a dynamic brain network, N (t) = (V, W (t)),
for each time window win(t), where V is the set of nodes
of the graph corresponding to the brain anatomical regions,
while W (t) = {wt,j,i∀i, j ∈ V } is the directed weighted
edges between the nodes of the graph within time window
win(t). The nodes of the graph are the AAL defined brain
regions[57], except for the regions located in the cerebellum.
The nodes are then pruned using voxel selection, as some
anatomical regions contribute no voxels at all and get deleted
from the set of nodes of the graph V .
Note that our aim is to label the BOLD responses mea-
sured at each brain volume as it belongs to one of the two
phases of complex problem solving, namely, planning and
execution. For this purpose, we represent each brain volume
measured at a time instant t by a network, which shows the
relationship among the anatomical regions. This dynamic net-
work representation will allow us to investigate the network
properties of planning and execution subtasks. In this section,
we describe how we estimate the weights of the edges, W (t),
of the brain network, N (t), for each time instance, t, where
we employ the method proposed in [63].
For each window win(t), we define the functional neigh-
borhood matrix, Ωt. The entries of Ωt are binary, either 1 or
0, indicating if there is a connection between two regions or
not. The size of the matrix is M xM, where M is the num-
ber of brain anatomical regions. The functional neighborhood
matrix contains no self-connections, thus, Ωt(i, i) = 0∀i ∈
[1, M ]. Also, the brain regions pruned by voxel selection
contributing no voxels have no in/out connections, thus, the
corresponding entries in Ωt are all zeros. The connectivity
of each region to the rest of the regions is determined using
Pearson correlation, as follows: first, for every region i, we
measure the Pearson correlation between its BOLD response
ri,t and the BOLD responses of all the other remaining re-
gions as shown below:
cor(ri,t, rj,t) =
cov(ri,t, rj,t)
σ(ri,t)σ(rj,t)
,
(6)
where ri,t is the BOLD response of region i across time
window win(t), cov(ri,t, rj,t) is the covariance between the
corresponding BOLD responses of regions i and j. σ is the
standard deviation of the BOLD response of a given region.
Thus, the higher the Pearson correlation between two regions
the closer they are to each other in the functional neighbor-
hood system.
Then, we select p of the regions with the highest correla-
tion scores with region i. Thus, obtaining the neighborhood
set ηp[i], which contains the p closest brain regions to region
i. Finally, we define the Ωt(i, j) as the connectivity between
the regions i and j, using the constructed neighborhood sets
as follows:
(cid:40)
Ωt(i, j) =
if j ∈ ηp[i]
1,
0, otherwise.
(7)
Note that each anatomical region is connected to its p
closest functional neighbors. This approach forms a star mesh
around each anatomical region. The ensemble of all of the
local meshes creates a brain network at each time instance.
Note, also, that Pearson correlation values are not used as the
weights between two regions. They are just used to identify
the nodes of each local mesh formed around an anatomical re-
gion. The estimated brain network becomes sparser as p gets
smaller. When p is set to the number of anatomic regions, M,
the network becomes fully connected. The selection method
for the degree of neighborhood, p, is explained in the next
section. This approach of defining the connectivity matrix
not only makes the network representation sparse for small p
values, but, it also constructs a network which is connected in
functionally closest regions, satisfying the locality property
of the human brain.
After having determined the edges of the brain graph us-
ing the functional neighborhood matrix Ωt, all that is left is to
estimate the weights of these edges at each local mesh. In or-
der to do that, we represent the response of each region i (ri,t)
as a linear combination of its closest p-functional neighbors
as shown in equation 8,
(cid:88)
j∈ηp[i]
ri,t =
wt,j,irj,t + i,t.
(8)
In equation 8, ri,t represents the representative time se-
ries of the BOLD response of region i within the time window
win(t), wt,j,i is the estimated edge weight between node (re-
gion) i and j at time instance t. ηp[i] is the p closest functional
neighbors of region i.
Onal et.al. [64] estimated the arc-weights for each mesh
formed around region i for each time window win(t) by min-
imizing the mean-squared error loss function using Ridge re-
gression. In this approach, the mean-squared error loss func-
tion is minimized with respect to wt,j,i, for each region, inde-
pendent of the other regions, where the expectation is taken
over the time-instances, in window win(t) as shown in equa-
tion 9 .
E[(i,t)2] = E[(ri,t − (cid:88)
j∈ηp[i]
wt,j,irj,t)2] + λwt,j,i2, (9)
where λ is the L2 regularization parameter whose value
is optimized using cross-validation. L2 regularization is
used in order to improve the generalization of the con-
structed mesh networks. Note that the estimated arc-weights,
wt,j,i (cid:54)= wt,i,j. Therefore, the ensemble of meshes yields a
directed brain network.
In this work, we define an artificial neural network to es-
timate the values of mesh arc-weights for all anatomical re-
gions jointly in each time window, as proposed in [63]. In this
method, we estimate the mesh arc-weights matrix W (t) =
{wt,j,ij, i ∈ V } using a feedforward neural network. The
architecture of this network consists of an input layer and an
output layer, both containing M nodes corresponding to each
brain region. The edges of this network are constructed us-
ing the neighborhood matrix Ωt. There is an edge between
node i of the output layer and node j from the input layer, if
Ωt(i, j) = 1.
The loss function of the suggested artificial neural net-
work is given in equation 10, where W is the weight matrix
of the entire neural network that corresponds to directed edge
weights of the brain graph and Wi is the row of matrix W
corresponding to region i:
Loss(Outputi) = E[(i,t)2] + λW T
i Wi
wt,j,irj,t)2] + λW T
i Wi.
= E[(ri,t − (cid:88)
j∈ηp[i]
We train the aforementioned artificial neural network in
order to obtain the weights of the brain network at each time
instance t that minimize the loss function by applying a gra-
dient descent optimization method as shown in equation 11,
(10)
t,j,i = w(κ−1)
w(κ)
t,j,i − αlearning
∂E[(i,t)2]
∂wt,j,i
,
(11)
where w(κ)
t,j,i is the weight of the edge from node j to node
i at epoch (iteration) κ, αlearning is the learning rate. The
number of epochs and learning rate used to train the network
are optimized empirically using cross-validation.
Finally, the weights of the above artificial neural network,
computed for each win(t), correspond to the edge weights
of the dynamic brain network, N (t) = (V, W (t)), at each
time instant t. Thus, we refer to the brain networks using
their window indices in order to obtain a set of dynamic brain
networks T = {N (1), N (2), ...N (tot win)}, where N (t) is
the brain network for time window win(t) and tot win is the
total number of time windows.
3.3. Network Metrics for Analyzing Brain Networks
In this section, we introduce some measures which we will
use to investigate the network properties of each phase of the
complex problem solving task, namely, planning and execu-
tion, using the estimated dynamic brain functional networks.
The connectivity patterns of anatomical regions are analyzed
by the set of network measures, given below. Two separate
sets of measures are used, namely, measures of centrality and
segregation. Since our estimated brain networks are directed,
we distinguish the incoming and outgoing edges in the net-
work, while defining the measures.
Recall that the suggested brain network N (t) = (V, W (t))
consist of a set of nodes, V , each of which corresponding to
one of the M anatomical regions. W (t) is the dynamic edge
weight matrix with the entries, wi,j, representing the weight
of the edge from node i to node j. For the sake of simplicity,
we omit the time dependency parameter t, since we compute
the network properties at each time instant. Matrix A is the
binarized version of W (t) matrix, where ai,j takes value 0 if
(wi,j == 0) and takes value 1 otherwise.
3.3.1. Measures of Centrality
Measures of centrality aim to identify brain regions that play
a central role in the flow of information in the brain network,
or nodes that can be identified as hubs. It is commonly mea-
sured using node degree, node strength and node betweenness
centrality, which are defined below.
3.3.2. Node Degree
The degree of a node is the total number of its edges as shown
in equation 12, where degreei is the degree of node i, V is
the set of all nodes in the graph and ai,j is the edge between
node i and node j.
(cid:88)
j∈V
ai,j
degreei =
(12)
In the case of a directed graph, we distinguish two dif-
i and node out-degree
ferent metrics: node in-degree degreein
i metrics which are shown in equations 13 and 14
degreeout
respectively where aj,i = 1, if there is a directed edge from
node j to node i.
(cid:88)
(cid:88)
j∈V
j∈V
aj,i
ai,j
degreein
i =
degreeout
i =
(13)
(14)
Node degree is a measure of centrality of the given nodes,
where it aims to quantify the hub brain regions interacting
with a large number of brain regions. Thus, a node with high
degree indicates its central role in the network.
3.3.3. Node Strength
Node strength is the sum of the weights of edges connected to
a given node 15, where wi,j is the weight of the edge between
node i and node j.
(cid:88)
j∈V
wi,j
strengthi =
(15)
Similar to node degree, node strength, also, distinguishes
two metrics in the case of directed graphs, namely, node in-
strength strengthin
shown in
equations 16 and 17 respectively, where wj,i is the weight of
the edge from node j to node i.
i and out-strength strengthout
i
(cid:88)
j∈V
wj,i
strengthin
i =
(16)
(cid:88)
j∈V
strengthout
i =
wi,j.
(17)
Node strength is a node centrality measure that is sim-
ilar to node degree, which is used in the case of weighted
graphs. Nodes with large strength values are tightly con-
nected to other nodes in the network forming hub nodes.
3.3.4. Node Betweenness Centrality
Betweenness centrality of node i is the fraction of the shortest
paths in the network that pass through node i as shown in
equation 18
betweennessi =
1
(M − 1)(M − 2)
(cid:88)
j,k∈V
ρi
j,k
ρj,k
,
(18)
where ρj,k is the number of shortest paths betweens nodes
j and k, ρi
j,k is the number of shortest paths between nodes j
and k that pass through node i, nodes i, j and k are distinct
nodes.
Before measuring the betweenness centrality of a node,
we need to change our perspective from connection weight
matrix to connection length matrix since betweenness central-
ity is a distance-based metric. In connection weights matrix,
larger weights imply higher correlation and shorter distance
while it is the opposite in the case of length matrix. Connec-
tion length matrix is obtained by inverting the weights of the
connection weight matrix. Then, the algorithm suggested in
[65] is employed in order to calculate the node betweenness
centrality for each anatomical region.
Nodes with high betweenness centrality are expected to
participate in many of the shortest paths of the networks.
Thus, taking a crucial role in the information flow of the
network.
proposed in [66]. It is defined as the fraction of the neighbors
of node i that are also neighbors of each other.
Ci =
[(dout
i + din
i )(dout
i − 1) − 2(cid:80)
χi
i + din
j∈V ai,jaj,i]
.
(19)
where din
i
is the in-degree of node i and dout
is the out-
degree of node i. χi is the weighted geometric mean of trian-
gles around node i that is calculated by equation 20. Recall
that aj,i = 1 , if there is a directed edge from node j to node
i and aj,i = 0, otherwise.
i
(cid:88)
j,h∈V
1
2
χi =
(wi,jwi,hwj,h)1/3.
(20)
The clustering coefficient of a node is the fraction of trian-
gles around the node. It is defined as the fraction of the neigh-
bors of the node that are also the neighbors of each other.
3.3.7. Transitivity
Transitivity of a node is similar to its clustering coefficient.
However, transitivity is normalized over all nodes while clus-
ter coefficient for each node is normalized independently
which makes clustering coefficient biased towards nodes with
low degree. Transitivity can be expressed as the ratio of tri-
angles to triplets in the network. It is calculated by equation
21 , as suggested in [66]:
(cid:80)
Ti =
j − 1) − 2(cid:80)
χi
j + din
j∈V [(dout
j + din
j )(dout
,
h∈V aj,hah,j]
(21)
3.3.5. Measures of Segregation
Measures of segregation aim to quantify the existence of sub-
groups within brain networks, where the nodes are densely
interconnected. These subgroups are commonly referred to as
clusters or modules. The existence of such clusters in func-
tional brain networks is a sign of interdependence among the
nodes forming the cluster. Measures of segregation include
clustering coefficient, transitivity and local efficiency. While
global efficiency is a measure of functional integration repre-
senting how easy it is for information to flow in the network.
3.3.6. Clustering Coefficient
The clustering coefficient of a node i is the fraction of tri-
angles around node i which is calculated by equation 19 as
is the in-degree of node j and dout
where din
is the out-degree
j
of node j. χi is the weighted geometric mean of triangles
around node i that is calculated by equation 20. Note that
ah,jaj,h = 1 , if there exits an edge in both directions.
j
3.3.8. Global & Local Efficiency
The global efficiency of a brain network is a measure of its
functional integration. It measures the degree of communica-
tion among the anatomical regions. Thus, it is closely related
to the small-world property of a network. Formally speak-
ing, global efficiency is defined as the average of the inverse
shortest path lengths between all pairs of nodes in the brain
network. Equation 22 shows how to calculate the global effi-
ciency of a brain network, where w
i,j is the weighted shortest
path length between two distinct nodes i and j [45].
(cid:88)
i∈V
1
M
Eglobal =
(cid:80)
i,j)−1
j∈V (w
M − 1
(22)
On the other hand, the local efficiency of a network is
defined as the global efficiency calculated over the neighbor-
hood of a single node. The local efficiency is, thus, a measure
of segregation rather than functional integration as it is closely
related to clustering coefficient. While global efficiency is
calculated for the entire network, local efficiency is calculated
for each node in the network [45].
4. EXPERIMENTS & RESULTS
In this section, we explore the validity of the suggested net-
work model by applying it to the TOL dataset. First, we an-
alyze the effect of the preprocessing step on the brain decod-
ing performance of planning and execution phases of com-
plex problem solving. Then, we investigate the validity of
the dynamic functional brain network model proposed in this
study. Finally, we analyze the network properties of the con-
structed functional brain networks for planning and execution
subtasks.
4.1. Voxel Selection
At the first step of the proposed computational model, we dis-
carded all of the voxels located in the cerebellum anatomical
regions. Then, we calculated the f-score for each one of the
remaining voxel and order the obtained f-scores of the vox-
els. Following that, we plotted the ordered f-scores of the
voxels in order to determine the appropriate number of voxels
to retain. Figure 1 shows the ordered f-scores of the voxels
averaged across all subjects. It can be observed from this fig-
ure that a relatively small number of voxels is crucial for dis-
criminating the subtasks of problem solving while the remain-
ing voxels do not have significant information concerning the
subtasks of problem solving. Based on the f-score distribu-
tion shown in Figure 1, we kept the 10,000 voxels with the
highest f-scores given the clear the elbow point whereas we
discarded the remaining ones.
After selecting the 10,000 voxels with the highest f-
scores of each session, we computed the number of selected
voxels contained in each one of the 90 anatomical regions.
We also calculated the percentage of selected voxels to the
total number of voxels located in each anatomical region.
Figure 2a shows the average number of voxels contributed by
each region across all subjects with its corresponding stan-
dard deviation, Figure 2b shows the average percentage of
voxels contributed by each region across all subjects with its
corresponding standard deviation.
It is clear from these figures that a large number of re-
gions contribute little to no voxels, such as the amygdala,
Fig. 1: Ordered f-scores of voxels.
caudate, heschl gyrus, hippocampus, pallidum, putamen, tem-
poral pole, superior temporal cortex, thalamus and parahip-
pocampus. A small number of regions contribute a signifi-
cantly large number of voxels (over 300 voxels each) during
complex problem solving, such as occipital, precentral, pre-
cuneus and parietal regions.
Furthermore, Figure 2b ensures that there is no bias
against tiny anatomical regions with small number of voxels
by normalizing the number of voxels selected from each re-
gion by its total number of voxels. Figure 2b clearly shows
that in the left prefrontal and inferior occipital regions a
significant percentage of voxels are active during complex
problem solving. Both figures also show high standard de-
viations across subjects, which indicates high inter-subject
variability.
4.2. Interpolation
After selecting the most discriminative voxels and averag-
ing their BOLD responses with respect to their corresponding
brain anatomical regions, we employed temporal interpola-
tion to increase the temporal resolution of the TOL dataset.
As a result, the total number of obtained brain volumes is
equal to n + z ∗ (n − 1) where n is the number of mea-
sured brain volumes of a given puzzle and z is the number of
estimated brain volumes plugged between each pair of mea-
sured brain volumes. The optimal value of z is equal to 8
which is determined empirically using cross-validation. Fig-
ure 3 shows the interpolated BOLD response of a randomly
selected anatomical region from the given subjects, where the
blue dots represent the measured BOLD response of the re-
gion and the orange dashes are the interpolated values.
It
is clear from Figure 3 that the interpolated points using cu-
bic spline function do not introduce sharp edges nor do they
smooth out the spikes between measured brain volumes.
Furthermore, Figure 4 shows the single-sided amplitude
spectrum of a randomly selected anatomical region from a
given subject before interpolation, after interpolation and fi-
nally after adding Gaussian noise. The figure clearly demon-
(a) Average number of voxels selected from each anatomical region across all subjects.
(b) Average percentage of voxels selected from each anatomical region across all subjects.
Fig. 2: Distribution of selected voxels across anatomical regions, measured by number of selected voxels (top) and percentage
of selected voxels (bottom) from each anatomical region.
[0.025, 0.05, 0.075, 0.1].
4.4. Brain Decoding
We use brain decoding in order to quantify the effect of our
proposed preprocessing steps on the TOL dataset. We aim
to distinguish the two phases of complex problem solving
namely: planning and execution. At first, we used ANOVA
to select the 10,000 voxels with the highest f-scores then we
averaged the selected voxels into their corresponding anatom-
ical regions defined by AAL [57]. Following that, we em-
ployed temporal interpolation to increase the temporal res-
olution of each puzzle by estimating z = 8 brain volumes
between each pair of measured brain volumes. Finally, we
added Gaussian noise in order to regularize the BOLD re-
sponses of each region to improve the generalization perfor-
mance of the classifiers. We used k-fold Cross validation for
each subject in all of the experiments introduced in this sec-
tion, with k = 8. After we obtained the results, we averaged
them across the different fold, then we calculated the average
and standard deviation across all subjects. We used both su-
pervised and unsupervised brain decoding methods, a linear
support-vector machine (SVM) [67] was used for supervised
brain decoding while k-means clustering was used for unsu-
pervised brain decoding.
Table 1 shows the effect of our preprocessing pipeline on
the brain decoding of complex problem solving subtasks. The
first row shows the performances of brain decoding on the raw
dataset without any preprocessing, simply averaging all of the
Fig. 3: Interpolated BOLD response.
strates that both interpolation and injecting Gaussian noise
preserve the smooth peaks of the signal in the frequency do-
main.
4.3. Gaussian Noise
In order to control the signal-to-noise ration (SNR), we used
cross-validation to choose the optimal pair of values for
αnoise and βnoise, the ratios of mean and standard devia-
tion of the added noise respectively. As a result, the op-
timal values obtained are αnoise = 0.025 and βnoise =
0.075 from the following set of values αnoise, βnoise ∈
Fig. 4: Single-Sided amplitude spectrum.
voxels into their corresponding anatomical regions. While the
second row shows the results of applying voxel selection then
averaging the selected voxels into their anatomical regions.
The third row shows the results of brain decoding after ap-
plying temporal interpolation, while the forth row shows the
results after injecting the data with Gaussian noise.
Preprocessing
Raw data
Voxel Selection
Interpolation
Noise addition
SVM
0.60 ± 0.11
0.74 ± 0.12
0.81 ± 0.08
0.82 ± 0.08
k-Means
0.63 ± 0.09
0.85 ± 0.06
0.84 ± 0.06
0.85 ± 0.06
Table 1: Decoding performances of preprocessing pipeline
after each step.
From the results of the preprocessing experiments, it is
observed that voxel selection improves the brain decoding
performance for both supervised and unsupervised methods
from %60 to %74 and from %63 to %85 respectively. This
can be attributed to voxel selection retaining only the most
discriminative voxels and trashing the remaining less infor-
mative ones. In addition, voxel selection manages to sparsify
the representation of the data since some brain regions con-
tribute no voxels at all thus have a flat BOLD response.
The table also shows that temporal interpolation further
improves the supervised brain decoding performance from
%74 to %81, this significant increase is due to increasing the
number of brain volumes thus increasing the number of train-
ing samples for the classifier. However, temporal interpola-
tion slightly reduces the performance of unsupervised meth-
ods from %85 to %84 which can be partially attributed to
the estimated brain volumes during the transitions between
the two phases of problem solving, planning and execution
which reduces the separation between the two natural sub-
groups. This is due to the method used to label the estimated
brain volumes, where each estimated brain volume is given
the labels of its closest neighboring measured brain volume.
Finally, the addition of Gaussian noise slightly boosts the
performance of both supervised and unsupervised methods
from %81 to %82 and from %84 to %85 respectively. The ta-
ble also shows high standard deviation across subjects, which
is consistent with voxel selection plots, revealing high inter-
subject variability.
4.5. Building Brain Networks
In this section, we compare our model for building dynamic
functional brain networks with some of the popular methods
proposed in the literature in terms of their brain decoding
power. Brain decoding can verify whether the constructed
brain networks are good representatives of the underlying
cognitive subtasks or not.
For this purpose, we built brain networks as explained
in the previous sections after having successfully applied the
preprocessing pipeline. The optimal values for learning rate
αlearning and number of epochs were chosen empirically us-
ing cross-validation obtaining the following values respec-
tively 1 ∗ 10−8 and 10. As for p, the number of neighbors
used to represent each anatomical region, we chose p equal to
the total number of regions which is 90, in this way, a fully-
connected brain network is obtained at each time window.
However, the total number of nodes is less than 90 given that
some regions have flat BOLD responses therefore they were
pruned along with all their edges from the brain network.
We also constructed brain networks using Pearson corre-
lation and ridge regression as proposed in [35, 64] and [32,
42] respectively in order to compare the performance of our
methods with other works in the literature.
In the case of
Pearson correlation, the functional brain networks were con-
structed using Pearson correlation scores between each pair
of brain regions [35, 64]. As for the case of ridge regression,
the mesh arc-weight descriptors were estimated using ridge
regression in order to represent each region as a linear combi-
nation of its neighbors [32, 42].
Table 2 shows the brain decoding results of the aforemen-
tioned brain network construction methods compared against
the results of multi-voxel pattern analysis (MVPA). The first
row shows the brain decoding results of MVPA, while the sec-
ond and third rows show the results of Pearson correlation and
ridge regression methods respectively. The last row shows
the brain decoding results of our proposed neural network
model. The table clearly shows that both Pearson correlation
and ridge regression fail to construct valid brain networks that
are good representatives of the underlying cognitive tasks.
However, our model managed to get brain decoding results
similar or slightly better than those obtained from MVPA both
in the cases of supervised and unsupervised methods. This
can be attributed to the challenging nature of the TOL dataset,
Pearson correlation does not manage to capture the interde-
pendencies between the anatomical regions over short time
windows. While ridge regression fails to correctly estimate
the mesh arc-weights as it estimates the arc-weights for each
region independently of the other ones. Our proposed model,
with a relatively small number of epochs manages to obtain
mesh arc-weight values that capture the activation patterns of
anatomical regions and their relationships.
Algorithm
MVPA
Pearson
Ridge Regression
Neural Networks
SVM
0.82 ± 0.08
0.58 ± 0.05
0.56 ± 0.05
0.82 ± 0.10
k-Means
0.85 ± 0.06
0.57 ± 0.04
0.55 ± 0.02
0.87 ± 0.06
Table 2: Results of proposed model.
5. BRAIN NETWORK PROPERTIES
In this section, we aim to analyze the network properties of
the constructed functional brain networks. We investigate the
network properties for each anatomical brain region during
both planning and execution subtasks in order to understand
which regions are most active and which regions work to-
gether during each one of the two subtasks of complex prob-
lem solving.
Given that the constructed brain functional networks are
both weighted, directed, fully-connected and contain both
negative and positive weights, we preprocessed the networks
before measuring their network properties. Firstly, we got rid
of all the negative weights by shifting all the mesh arc-weights
values by a positive quantity equal to the absolute value of
the largest negative arc-weight. Then, we normalized the
mesh arc-weights to ensure that all of weights are within the
range of [0, 1]. Finally, we measured the network properties
on the pruned brain graph, where the brain regions (nodes)
contributing no voxels (have a flat BOLD response) and all
of their corresponding arc-weights (edges) were deleted from
the brain graph. Thus, the networks contained less than
90 regions with their corresponding edges. We used brain
connectivity toolbox to calculate the investigated network
properties [45].
In order to measure for centrality, the number of neigh-
bors for each anatomical region (P) was chosen to be equal to
89, which is equal to the total number of neighbors for any
given node as the total number of brain anatomical regions
defined by the AAL atlas [57] after deleting the regions resid-
ing in the cerebellum equals 90. In addition, since we pruned
the nodes that correspond to regions from which no voxels
were selected, our constructed brain networks were weighted
directed fully-connected networks. Therefore, the in-degree,
out-degree and total degree of all nodes in the graph were
equal to the total number of anatomical regions retained after
voxel selection.
Therefore, we used node strength and node betweenness
centrality to identify nodes with high centrality which are po-
tential hubs in the brain networks controlling the flow of in-
formation in the network. In our proposed model, the node
in-strength of node i is the sum of the mesh arc-weight values
which is estimated using our proposed neural network method
in order to minimize the reconstruction error of the BOLD re-
sponse of anatomical region i using its neighbors. Thus, node
in-strength is not used as part of our network properties analy-
ses, we rather used node out-strength to measure the centrality
of all anatomical regions.
As for measures of segregation, quantifying the existence
of subgroups within brain networks is based on densely in-
terconnected nodes. These subgroups are commonly referred
to as clusters or modules. The existence of such clusters in
functional brain networks is a sign of interdependence among
the nodes forming the cluster. Therefore, clustering coeffi-
cient, transitivity and local efficiency were measured in order
to identify potential clusters with dense interconnections in
the brain networks.
5.1. Planning & Execution Brain Networks
In this section, we discuss the network properties of the plan-
ning and execution networks. For each aforementioned net-
work metric, we ranked the brain regions in descending order
according to their score on that network measure for all sub-
jects across all sessions. Then, we retained the 10 anatomi-
cal regions with the highest scores. Following that, we mea-
sured the frequency of occurrence of each brain region among
the top 10 regions across all sessions in order to identify the
shared regions and patterns across all subjects for both plan-
ning and execution subtasks. The results of the analysis are
shown in tables:
table 3 shows the brain regions that have
high scores for the reported network properties during plan-
ning subtask, and table 4 shows the brain regions that have
high scores during execution subtask.
There are a number of processes taking place during plan-
ning and execution. Plan generation involves a series of re-
cursive events including: 1) problem encoding; 2) decision-
making in order to decide which ball to move and where to
move it; 3) mental imagery to imagine the ball moving; and 4)
working memory to maintain the intermediate steps as well as
the move number. During plan execution there is 1) retrieval
of the steps from memory; 2) confirming the correct steps are
being performed; and 3) the motor execution of those steps.
As the results demonstrate the networks for planning and ex-
ecution are overlapping. These results are similar to the ac-
tivation results reported in [22] in that the regions that were
found to be active during the task are also regions that are
most prominently found with the highest network measures.
These regions include the right and left middle frontal gyrus,
anterior cingulate cortex, precentral cortex, and superior pari-
etal cortex.
Previous work has suggested that the regions found in the
current study to show high network measures are directly re-
lated to the sub-tasks associated with TOL performance. For
example, both the left and right prefrontal cortex have been
found to be involved in the TOL task with the two regions per-
forming distinguishable functions. The right prefrontal cortex
is involved in constructing the plan for solving the TOL prob-
lem while the left prefrontal cortex is involved in supervising
the execution of that plan [22, 24]. The anterior cingulate has
been linked to error detection and is particularly involved in
the TOL when the number of moves is higher or the problem
difficulty is manipulated. The right superior parietal cortex
and precentral cortex have been linked to visuo-spatial atten-
tion necessary for planning [24] and the left parietal cortex
has been linked to visuo-spatial working memory processing
[24]. The overlap between the regions with the highest net-
work measures and those that have been linked to the task is
an important feature and is not due to the voxel selection pro-
cess. Many regions that passed threshold were not in the top
ranked list of network measures. For example, the basal gan-
glia including the caudate has been found in previous studies
to be involved in TOL performance [22, 29, 68, 69, 26]; how-
ever, the region appears to not be an important network hub.
Figures 5a and 5b visualize the reported brain regions in
tables 3 and 4 respectively using Brain Net Viewer [70]. In
figures 5a and 5b, the color of the node (brain region) implies
the following: red indicates that the region has high transi-
tivity, clustering coefficient or local efficiency. Green indi-
cates that the node has high node centrality measured by node
out-strength and node betweenness. As for blue, it shows the
nodes that have high node centrality and is part of subgroup
of densely interconnected regions.
5.2. Differences between Planning and Execution Net-
works
In this section, we explore the network differences between
planning and execution by calculating the difference between
the network property scores for planning and execution for
each session. To achieve that, we took the difference be-
tween the network property scores for brain anatomical re-
gions during planning and the network property scores for
brain anatomical regions during execution for each session.
Then, we counted the frequency of times a given anatomi-
cal region is more active during planning than execution and
vice-versa in order to identify consistent patterns of the dis-
agreements between planning brain networks and execution
brain networks across all subjects. Results showed, generally,
that the network measures were higher for planning than ex-
ecution. This, too, mirrors the findings from [22] in which
planning resulted in greater activation than execution.
Node out-strength is a measure of how connected the node
is to other nodes in the network. Planning showed greater out-
strength than execution in the following regions: occipital
regions (calcarine, cuneus), parietal regions (bilateral supe-
rior parietal cortex and precunues), the right superior frontal
cortex, and inferior occipito-temporal regions (fusiform and
lingual gyri). The left angular gyrus and bilateral medial su-
perior frontal cortex showed greater out-strength for execu-
tion. As for node betweenness, the following brain regions
had higher node betweenness during planning than execution:
occipital regions (calcarine, cuneus, right middle, right supe-
rior); inferior occipito-temporal (fusiform, lingual); parietal
(bilateral superior parietal, left postcentral, precuneus). Bi-
lateral medial superior frontal had higher node betweenness
during execution than planning.
These results suggest that there is greater information flow
during planning than execution. This matches our expecta-
tions. Planning is more computationally demanding than ex-
ecution. Again, during planning participants must explore
the problem space which requires generating and manipu-
transitivity
Angular
Calcarine
Cingulum Ant
Cingulum Mid
Cuneus
Frontal Inf Oper L
local efficiency
clustering coefficient
Calcarine
Cuneus
Frontal Mid R
Frontal Sup
Fusiform
Occipital Inf R
Precentral
Supp Motor Area R
Temporal Inf R
Calcarine
Cuneus
Frontal Mid R
Frontal Sup
Fusiform
Occipital Inf R
Parietal Sup R
Precentral
Supp Motor Area R
Temporal Inf R
betweenness
Cueneus R
Frontal Sup L
Fusiform R
Paracentral Lobule L
Parietal Sup R
Precuneus L
Supp Motor Area R
Temporal Inf R
Temporal Mid R
out-strength
Cueneus R
Frontal Sup L
Fusiform R
Paracentral Lobule L
Supp Motor Area R
Temporal Inf R
Temporal Mid R
Table 3: Planning: Anatomical regions with the highest network measures across subjects, regions are painted if they overlap
with execution.
transitivity
Angular
Calcarine
Cingulum Ant
Cingulum Mid
Cuneus
local efficiency
Calcarine L
Cuneus
Frontal Sup L
Fusiform
Occipital Inf R
Frontal Inf Oper
Supp Motor Area R
Temporal Inf R
clustering coefficient
Calcarine L
Cuneus
Frontal Mid R
Frontal Sup
Fusiform
Occipital Inf R
Parietal Sup R
Precentral R
Supp Motor Area R
Temporal Inf R
betweenness
Cueneus R
Frontal Sup L
Fusiform R
Paracentral Lobule L
Precuneus L
Supp Motor Area R
Temporal Inf R
Temporal Mid R
out-strength
Frontal Sup
Fusiform
Paracentral Lobule L
Supp Motor Area R
Temporal Inf R
Temporal Mid R
Table 4: Execution: Anatomical regions with the highest network measures across subjects, regions are painted if they overlap
with planning.
lating a mental representation of the problem. The regions
that show greater information flow during planning are all
regions involved in that generation and manipulation partic-
ularly parietal, occipital and inferior occipito-temporal. On
the other hand, execution requires recall of the plan gener-
ated and stored and therefore, greater information flow from
frontal regions related to memory retrieval is observed.
Clustering coefficient, local efficency and transitvity are
measures of segregation which aim to identify sub-networks.
Each of these measures were larger for planning than exe-
cution with no regions showing larger measures for execu-
tion. The regions that showed higher clustering coefficient
in planning included: the cuneus, left middle occipital cor-
tex, and right precuneus. Local efficiency was higher in a
similar set of regions (the cuneus, left middle occipital cor-
tex, and right precuneus). The clustering coeffiencent and lo-
cal efficiency identified a visual-spatial sub-network that is
more strongly connected during planning. Transitivity iden-
tified an overlapping but more extensive set of regions that
included: bilateral angular gyrus, calcarine sulcus, cuneus,
bilateral middle frontal cortex, bilaterial superior frontal cor-
tex, bilateral fusiform and lingual gyri, bilateral occipital cor-
tex, bilatral superior parietal cortex, postcentral and precen-
tral cortex, precuneus, supplementary motor area, right supra-
marginal gyrus, and right inferior and middle temporal cortex.
Figures 6a , 6b visualize the brain regions with higher be-
tweenness during planning and during execution respectively.
Figures 7a , 7b visualize the brain regions with higher
node out-strength during planning and during execution, re-
spectively.
Figure 8a visualizes the brain regions with higher local
efficiency and higher clustering coefficient during planning
phase compared to execution phase. While Figure 8b visual-
izes the brain regions with higher transitivity during planning
than during execution phase.
5.3. Global Efficiency
Since global efficiency is measured over the entire brain net-
work, not for a given node in the network, we measured the
global efficiency for all planning and execution networks
within all sessions across subjects. Then, global efficiency of
planning is compared against that of execution. Results show
that the majority of sessions had higher global efficiency
scores during planning than execution, 43 out of 72 sessions
had higher global efficiency during planning than execution.
Furthermore, table 5 shows the number of sessions where
global efficiency was higher during planning and during ex-
ecution across all subjects for all 4 sessions of each subject.
The first column shows the number of subjects that had a
higher global efficiency score during planning than during
execution. The second column shows the number of subjects
that had a higher global efficiency score during execution
than during planning.
Although there was no significant difference in global
efficiency between planning and execution, from the table,
it is clear that the majority of subjects had a higher global
efficiency for planning for the first session. Some subjects
switched from having higher global efficiency during plan-
ning to having higher global efficiency during execution.
A potential explanation for this change across sessions is a
switch from pre-planning to on-line planning, or planning
intermixed with execution. Although there is a dedicated
planning phase in the current study, that does not mean that
planning is not taking place during execution. In fact, it has
been debated as to whether efficient pre-planning is possible
in the TOL or whether TOL performance is controlled by
on-line planning [71, 72, 73, 6]. According to Phillips et.al.
[71, 72] pre-planning the entire sequence is not natural, but
that people instead plan the beginning sequence of moves and
then intersperse planning and execution.
If this is the case
then it may be expected that some participants will switch to
on-line planning. This intermixing of planning and execu-
tion is also likely to impact the performance of the machine
learning algorithms to detect planning and execution phases.
The relationship between global efficiency and behavioral
performance was examined. Global efficiency was found to
be positively correlated with the mean number of extra moves
(a measure of error) during problem-solving (for execution
r=0.73, p=0.0006). Previous studies have shown a relation-
ship between global efficiency and task performance [74].
This suggests that the variance in global efficiency is in-
dicative of individual differences in neural processing and fur-
ther suggests that the changes in global efficiency across ses-
sions are also likely indicative of changes in neural processing
related to changing strategy. Further research using a larger
sample is necessary to explore this hypothesis.
Session Number
Planning
Execution
1
2
3
4
15
9
10
9
3
9
8
9
Table 5: Global Efficiency.
6. CONCLUSION
In this paper, we proposed a model to construct brain func-
tional networks during a complex problem solving task. Our
model successfully identified the two phases of complex
problem solving. In addition, the network properties of the
constructed brain networks during planning and execution
phases were studied in order to identify essential nodes within
the brain networks related to problem solving, potential hubs,
and densely connected clusters. Furthermore, the differences
between planning networks and execution networks were
highlighted and discussed.
There are some limitations to the study. Although the
primary aim of this study was to demonstrate the feasibility
of the methods, the sample size is somewhat small, making
the interpretation of the results difficult. Second, a goal of
this method is to identify brain states that are interspersed
with each other. In the current study planning was expected
to occur both prior to execution as well as during execution
therefore planning states are interspersed within the execution
phase. The temporal sampling rate of the fMRI data may be
a limiting factor. Alternatively, the sluggish and blurred un-
derlying hemodynamic response may be the factor preventing
the ability to detect brain states. We plan to explore this factor
in future work.
7. ACKNOWLEDGMENT
The work is supported by TUBITAK (Scientific and Tech-
nological Research Council of Turkey) under the grant No:
116E091 as well as the Indiana METACyt Initiative of Indi-
ana University, funded in part through a major grant from the
Lilly Endowment, Inc.
8. REFERENCES
[1] Tim Shallice, "Specific impairments of planning," Phil.
Trans. R. Soc. Lond. B, vol. 298, no. 1089, pp. 199 -- 209,
1982.
[2] William S MacAllister, H Allison Bender, Lind-
say Whitman, Antoinette Welsh, Shari Keller, Yael
Granader, and Elisabeth MS Sherman, "Assessment of
executive functioning in childhood epilepsy: the tower
of london and brief," Child Neuropsychology, vol. 18,
no. 4, pp. 404 -- 415, 2012.
[3] Ingeborg Goethals, Kurt Audenaert, Filip Jacobs,
Christophe Van de Wiele, Hamphrey Ham, Hanneke
Pyck, Andr´e Vandierendonck, Cees Van Heeringen, and
Rudi Dierckx,
"Blunted prefrontal perfusion in de-
pressed patients performing the tower of london task,"
Psychiatry Research: Neuroimaging, vol. 139, no. 1, pp.
31 -- 40, 2005.
[4] Irena Rektorova, Hana Srovnalova, Radka Kubikova,
and Jiri Prasek, "Striatal dopamine transporter imag-
ing correlates with depressive symptoms and tower of
london task performance in parkinson's disease," Move-
ment disorders, vol. 23, no. 11, pp. 1580 -- 1587, 2008.
[5] Paul E Rasser, Patrick Johnston, Jim Lagopoulos,
Philip B Ward, Ulrich Schall, Renate Thienel, Stefan
Bender, Arthur W Toga, and Paul M Thompson, "Func-
tional mri bold response to tower of london performance
of first-episode schizophrenia patients using cortical pat-
tern matching," Neuroimage, vol. 26, no. 3, pp. 941 --
951, 2005.
[6] Josef M Unterrainer, Benjamin Rahm, Christoph P
Kaller, Rainer Leonhart, K Quiske, K Hoppe-Seyler,
C Meier, C Muller, and U Halsband, "Planning abili-
ties and the tower of london: is this task measuring a
discrete cognitive function?," Journal of clinical and
experimental neuropsychology, vol. 26, no. 6, pp. 846 --
856, 2004.
[7] Nancy A Zook, Deana B Davalos, Edward L DeLosh,
and Hasker P Davis, "Working memory, inhibition, and
fluid intelligence as predictors of performance on tower
of hanoi and london tasks," Brain and cognition, vol.
56, no. 3, pp. 286 -- 292, 2004.
[8] Yu-Kai Chang, Chia-Liang Tsai, Tsung-Min Hung, Ed-
mund Cheung So, Feng-Tzu Chen, and Jennifer L Et-
nier, "Effects of acute exercise on executive function: a
study with a tower of london task," Journal of Sport and
Exercise Psychology, vol. 33, no. 6, pp. 847 -- 865, 2011.
[9] Dustin Albert and Laurence Steinberg, "Age differences
in strategic planning as indexed by the tower of london,"
Child development, vol. 82, no. 5, pp. 1501 -- 1517, 2011.
[10] Nancy Zook, Marilyn C Welsh, and Vanessa Ewing,
"Performance of healthy, older adults on the tower of
london revised: Associations with verbal and nonverbal
abilities," Aging, Neuropsychology, and Cognition, vol.
13, no. 1, pp. 1 -- 19, 2006.
[11] A Boghi, R Rasetti, F Avidano, C Manzone, L Orsi,
F D'agata, P Caroppo, M Bergui, P Rocca, L Pulvirenti,
et al., "The effect of gender on planning: An fmri study
using the tower of london task," Neuroimage, vol. 33,
no. 3, pp. 999 -- 1010, 2006.
[12] Josef M Unterrainer, Christian C Ruff, Benjamin Rahm,
Christoph P Kaller, J Spreer, Ralf Schwarzwald, and Ul-
rike Halsband, "The influence of sex differences and
individual task performance on brain activation during
planning," Neuroimage, vol. 24, no. 2, pp. 586 -- 590,
2005.
[13] Christoph P Kaller, Katharina Heinze, Irina Mader,
Josef M Unterrainer, Benjamin Rahm, Cornelius
Weiller, and Lena Kostering, "Linking planning per-
formance and gray matter density in mid-dorsolateral
prefrontal cortex: moderating effects of age and sex,"
Neuroimage, vol. 63, no. 3, pp. 1454 -- 1463, 2012.
[14] Manuel Desco, Francisco J Navas-Sanchez, Javier
Sanchez-Gonz´alez, Santiago Reig, Olalla Robles, Car-
Juan A Guzm´an-De-Villoria, Pedro
olina Franco,
Garc´ıa-Barreno, and Celso Arango,
"Mathematically
gifted adolescents use more extensive and more bilat-
eral areas of the fronto-parietal network than controls
during executive functioning and fluid reasoning tasks,"
Neuroimage, vol. 57, no. 1, pp. 281 -- 292, 2011.
[15] Christoph P Kaller, Josef M Unterrainer, Benjamin
Rahm, and Ulrike Halsband, "The impact of problem
structure on planning: Insights from the tower of lon-
don task," Cognitive Brain Research, vol. 20, no. 3, pp.
462 -- 472, 2004.
[16] W Keith Berg, Dana L Byrd, Joseph PH McNamara,
and Kimberly Case, "Deconstructing the tower: Param-
eters and predictors of problem difficulty on the tower
of london task," Brain and Cognition, vol. 72, no. 3, pp.
472 -- 482, 2010.
[17] JM Unterrainer, B Rahm, U Halsband, and CP Kaller,
"What is in a name: Comparing the tower of london
with one of its variants," Cognitive Brain Research, vol.
23, no. 2, pp. 418 -- 428, 2005.
[18] Sharlene D Newman and Gregory Pittman, "The tower
of london: A study of the effect of problem structure on
planning," Journal of Clinical and Experimental Neu-
ropsychology, vol. 29, no. 3, pp. 333 -- 342, 2007.
[19] JM Unterrainer, B Rahm, R Leonhart, CC Ruff, and
U Halsband, "The tower of london: The impact of in-
structions, cueing, and learning on planning abilities,"
Cognitive Brain Research, vol. 17, no. 3, pp. 675 -- 683,
2003.
[20] Zachariah Campbell, Konstantine K Zakzanis, Diana Jo-
vanovski, Steve Joordens, Richard Mraz, and Simon J
Graham, "Utilizing virtual reality to improve the eco-
logical validity of clinical neuropsychology: an fmri
case study elucidating the neural basis of planning by
comparing the tower of london with a three-dimensional
navigation task," Applied Neuropsychology, vol. 16, no.
4, pp. 295 -- 306, 2009.
[21] Sharlene D Newman, Benjamin Pruce, Akash Rusia,
and Thomas Burns Jr, "The effect of strategy on prob-
lem solving: an fmri study," The Journal of Problem
Solving, vol. 3, no. 1, pp. 2, 2010.
[22] Sharlene D Newman, John A Greco, and Donghoon
Lee, "An fmri study of the tower of london: a look
at problem structure differences," Brain research, vol.
1286, pp. 123 -- 132, 2009.
[23] Gerd Wagner, Kathrin Koch, Jurgen R Reichenbach,
Heinrich Sauer, and Ralf GM Schlosser, "The special in-
volvement of the rostrolateral prefrontal cortex in plan-
ning abilities: an event-related fmri study with the tower
of london paradigm," Neuropsychologia, vol. 44, no. 12,
pp. 2337 -- 2347, 2006.
[24] Sharlene D Newman, Patricia A Carpenter, Sashank
Varma, and Marcel Adam Just, "Frontal and parietal
participation in problem solving in the tower of london:
fmri and computational modeling of planning and high-
level perception," Neuropsychologia, vol. 41, no. 12, pp.
1668 -- 1682, 2003.
[25] F Cazalis, R Valabregue, M P´el´egrini-Issac, S Asloun,
TW Robbins, and S Granon,
"Individual differences
in prefrontal cortical activation on the tower of london
planning task: implication for effortful processing," Eu-
ropean Journal of Neuroscience, vol. 17, no. 10, pp.
2219 -- 2225, 2003.
[26] Odile A van den Heuvel, Henk J Groenewegen, Fred-
erik Barkhof, Richard HC Lazeron, Richard van Dyck,
and Dick J Veltman, "Frontostriatal system in planning
complexity: a parametric functional magnetic resonance
version of tower of london task," Neuroimage, vol. 18,
no. 2, pp. 367 -- 374, 2003.
[27] R Nathan Spreng, W Dale Stevens, Jon P Chamberlain,
Adrian W Gilmore, and Daniel L Schacter, "Default
network activity, coupled with the frontoparietal control
network, supports goal-directed cognition," Neuroim-
age, vol. 53, no. 1, pp. 303 -- 317, 2010.
[28] Judy A Kipping, Daniel S Margulies, Simon B Eick-
hoff, Annie Lee, and Anqi Qiu, "Trade-off of cerebello-
cortical and cortico-cortical functional networks for
planning in 6-year-old children," NeuroImage, vol. 176,
pp. 510 -- 517, 2018.
[29] MH Beauchamp, A Dagher, JAD Aston, and J Doyon,
"Dynamic functional changes associated with cognitive
skill learning of an adapted version of the tower of lon-
don task," Neuroimage, vol. 20, no. 3, pp. 1649 -- 1660,
2003.
[30] WR Shirer, S Ryali, E Rykhlevskaia, V Menon, and
MD Greicius,
"Decoding Subject-Driven Cognitive
States with Whole-Brain Connectivity Patterns," Cere-
bral Cortex, vol. 22, no. 1, pp. 158 -- 165, 2012.
[31] Matthias Ekman, Jan Derrfuss, Marc Tittgemeyer, and
Christian J Fiebach, "Predicting Errors from Reconfigu-
ration Patterns in Human Brain Networks," Proceedings
of the National Academy of Sciences, vol. 109, no. 41,
pp. 16714 -- 16719, 2012.
[32] Itir Onal, Mete Ozay, Eda Mizrak, Ilke Oztekin, and
Fatos T Yarman Vural, "A new representation of fmri
signal by a set of local meshes for brain decoding," IEEE
Transactions on Signal and Information Processing over
Networks, vol. 3, no. 4, pp. 683 -- 694, 2017.
[33] Martin A Lindquist, "The Statistical Analysis of fMRI
Data," Statistical Science, pp. 439 -- 464, 2008.
[34] Jonas Richiardi, Sophie Achard, Horst Bunke, and Dim-
itri Van De Ville,
"Machine Learning with Brain
Graphs: Predictive Modeling Approaches for Func-
tional Imaging in Systems Neuroscience," IEEE Signal
Processing Magazine, vol. 30, no. 3, pp. 58 -- 70, 2013.
[35] Jonas Richiardi, Hamdi Eryilmaz, Sophie Schwartz, Pa-
trik Vuilleumier, and Dimitri Van De Ville, "Decoding
brain states from fmri connectivity graphs," Neuroim-
age, vol. 56, no. 2, pp. 616 -- 626, 2011.
[36] Mary-Ellen Lynall, Danielle S Bassett, Robert Kerwin,
Peter J McKenna, Manfred Kitzbichler, Ulrich Muller,
and Ed Bullmore, "Functional connectivity and brain
networks in schizophrenia," Journal of Neuroscience,
vol. 30, no. 28, pp. 9477 -- 9487, 2010.
[37] Vinod Menon, "Large-scale brain networks and psy-
chopathology: a unifying triple network model," Trends
in cognitive sciences, vol. 15, no. 10, pp. 483 -- 506, 2011.
[38] Anvar Kurmukov, Marina Ananyeva, Yulia Dodonova,
Boris Gutman, Joshua Faskowitz, Neda Jahanshad, Paul
Thompson, and Leonid Zhukov, "Classifying pheno-
types based on the community structure of human brain
networks," in Graphs in Biomedical Image Analysis,
Computational Anatomy and Imaging Genetics, pp. 3 --
11. Springer, 2017.
[39] Hyekyoung Lee, Dong Soo Lee, Hyejin Kang, Boong-
Nyun Kim, and Moo K Chung, "Sparse brain network
IEEE Transac-
recovery under compressed sensing,"
tions on Medical Imaging, vol. 30, no. 5, pp. 1154 -- 1165,
2011.
[40] Orhan Fırat, Mete Ozay, Itır Onal, Ilke Oztekiny, and
Fatos¸ T Yarman Vural, "Functional mesh learning for
pattern analysis of cognitive processes," in Cognitive
Informatics & Cognitive Computing (ICCI* CC), 2013
12th IEEE International Conference on. IEEE, 2013, pp.
161 -- 167.
[41] Abdullah Alchihabi, Baran B Kivilicim, Sharlene D
Newman, and Fatos T Yarman Vural, "A dynamic net-
work representation of fmri for modeling and analyz-
ing the problem solving task," in Biomedical Imaging
(ISBI 2018), 2018 IEEE 15th International Symposium
on. IEEE, 2018, pp. 114 -- 117.
[42] Itir Onal, Mete Ozay, and Fatos T Yarman Vural, "Mod-
eling voxel connectivity for brain decoding," in Pat-
tern Recognition in NeuroImaging (PRNI), 2015 Inter-
national Workshop on. IEEE, 2015, pp. 5 -- 8.
[43] Chong-Yaw Wee, Pew-Thian Yap, Daoqiang Zhang,
Lihong Wang, and Dinggang Shen,
"Constrained
sparse functional connectivity networks for mci classi-
fication," Medical Image Computing and Computer-
Assisted Intervention -- MICCAI 2012, pp. 212 -- 219,
2012.
[44] Pedro A Vald´es-Sosa, Jose M S´anchez-Bornot, Agust´ın
Lage-Castellanos, Mayrim Vega-Hern´andez,
Jorge
Bosch-Bayard, Lester Melie-Garc´ıa, and Erick Canales-
Rodr´ıguez, "Estimating brain functional connectivity
with sparse multivariate autoregression," Philosophical
Transactions of the Royal Society of London B: Biolog-
ical Sciences, vol. 360, no. 1457, pp. 969 -- 981, 2005.
[45] Mikail Rubinov and Olaf Sporns, "Complex network
measures of brain connectivity: uses and interpreta-
tions," Neuroimage, vol. 52, no. 3, pp. 1059 -- 1069,
2010.
[46] Hae-Jeong Park and Karl Friston, "Structural and func-
tional brain networks: from connections to cognition,"
Science, vol. 342, no. 6158, pp. 1238411, 2013.
[47] Jonathan D Power, Damien A Fair, Bradley L Schlaggar,
and Steven E Petersen, "The development of human
functional brain networks," Neuron, vol. 67, no. 5, pp.
735 -- 748, 2010.
[48] Danielle Smith Bassett and ED Bullmore, "Small-world
brain networks," The neuroscientist, vol. 12, no. 6, pp.
512 -- 523, 2006.
[49] Mikail Rubinov and Olaf Sporns, "Weight-conserving
characterization of complex functional brain networks,"
Neuroimage, vol. 56, no. 4, pp. 2068 -- 2079, 2011.
[50] Andrew Zalesky, Alex Fornito, and Edward T Bull-
more, "Network-based statistic: identifying differences
in brain networks," Neuroimage, vol. 53, no. 4, pp.
1197 -- 1207, 2010.
[51] Sophie Achard and Ed Bullmore, "Efficiency and cost
of economical brain functional networks," PLoS com-
putational biology, vol. 3, no. 2, pp. e17, 2007.
[52] Manuel Blesa, Gemma Sullivan, Devasuda Anblagan,
Emma J Telford, Alan J Quigley, Sarah A Sparrow,
Ahmed Serag, Scott I Semple, Mark E Bastin, and
James P Boardman, "Early breast milk exposure modi-
fies brain connectivity in preterm infants," NeuroImage,
2018.
[53] Urs Braun, Sarah F Muldoon, and Danielle S Bassett,
"On human brain networks in health and disease," eLS,
2015.
[54] Arman Afrasiyabi, Itir Onal, and Fatos T Yarman Vu-
ral, "A sparse temporal mesh model for brain decod-
ing," in Cognitive Informatics & Cognitive Computing
(ICCI* CC), 2016 IEEE 15th International Conference
on. IEEE, 2016, pp. 198 -- 206.
[55] Francisco Pereira, Tom Mitchell,
and Matthew
Botvinick, "Machine learning classifiers and fmri: a
tutorial overview," Neuroimage, vol. 45, no. 1, pp.
S199 -- S209, 2009.
[56] David D Cox and Robert L Savoy, "Functional magnetic
resonance imaging (fmri)brain reading: detecting and
classifying distributed patterns of fmri activity in human
visual cortex," Neuroimage, vol. 19, no. 2, pp. 261 -- 270,
2003.
[57] Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri
Papathanassiou, Fabrice Crivello, Olivier Etard, Nico-
las Delcroix, Bernard Mazoyer, and Marc Joliot, "Au-
tomated anatomical labeling of activations in spm using
a macroscopic anatomical parcellation of the mni mri
single-subject brain," Neuroimage, vol. 15, no. 1, pp.
273 -- 289, 2002.
[58] Sky McKinley and Megan Levine, "Cubic spline inter-
polation," College of the Redwoods, vol. 45, no. 1, pp.
1049 -- 1060, 1998.
[59] Matteo Frigo and Steven G Johnson, "Fftw: An adaptive
software architecture for the fft," in Acoustics, Speech
and Signal Processing, 1998. Proceedings of the 1998
IEEE International Conference on. IEEE, 1998, vol. 3,
pp. 1381 -- 1384.
[60] William T Cochran, James W Cooley, David L Favin,
Howard D Helms, Reginald A Kaenel, William W Lang,
George C Maling, David E Nelson, Charles M Rader,
and Peter D Welch,
"What is the fast fourier trans-
form?," Proceedings of the IEEE, vol. 55, no. 10, pp.
1664 -- 1674, 1967.
[61] Kiyotoshi Matsuoka,
"Noise injection into inputs in
back-propagation learning," IEEE Transactions on Sys-
tems, Man, and Cybernetics, vol. 22, no. 3, pp. 436 -- 440,
1992.
[62] Russell Reed, Seho Oh, and RJ Marks, "Regularization
using jittered training data," in Neural Networks, 1992.
IJCNN., International Joint Conference on. IEEE, 1992,
vol. 3, pp. 147 -- 152.
[63] Baran Baris Kivilcim, Itir Onal Ertugrul, and Fatos
T Yarman Vural, "Modeling brain networks with ar-
tificial neural networks," in Graphs in Biomedical Im-
age Analysis and Integrating Medical Imaging and Non-
Imaging Modalities, pp. 43 -- 53. Springer, 2018.
[64] Itir Onal, Mete Ozay, and Fatos Tunay Yarman Vural,
"A hierarchical multi-resolution mesh network for brain
decoding," arXiv preprint arXiv:1607.07695, 2016.
[65] Ulrik Brandes, "A faster algorithm for betweenness cen-
trality," Journal of mathematical sociology, vol. 25, no.
2, pp. 163 -- 177, 2001.
[66] Giorgio Fagiolo, "Clustering in complex directed net-
works," Physical Review E, vol. 76, no. 2, pp. 026107,
2007.
[67] Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-
Rui Wang, and Chih-Jen Lin, "Liblinear: A library for
large linear classification," Journal of machine learning
research, vol. 9, no. Aug, pp. 1871 -- 1874, 2008.
[68] Alain Dagher, Adrian M Owen, Henning Boecker, and
David J Brooks, "Mapping the network for planning: a
correlational pet activation study with the tower of lon-
don task," Brain, vol. 122, no. 10, pp. 1973 -- 1987, 1999.
[69] JB Rowe, AM Owen, IS Johnsrude, and RE Passing-
ham, "Imaging the mental components of a planning
task," Neuropsychologia, vol. 39, no. 3, pp. 315 -- 327,
2001.
[70] Mingrui Xia, Jinhui Wang, and Yong He,
"Brainnet
viewer: a network visualization tool for human brain
connectomics," PloS one, vol. 8, no. 7, pp. e68910,
2013.
[71] Louise H Phillips, "The role of memory in the tower of
london task," Memory, vol. 7, no. 2, pp. 209 -- 231, 1999.
[72] Fred Phillips, "A research note on accounting students'
epistemological beliefs, study strategies, and unstruc-
tured problem-solving performance," Issues in Account-
ing Education, vol. 16, no. 1, pp. 21 -- 39, 2001.
[73] KL Kafer and Michael Hunter,
"On testing the face
validity of planning/problem-solving tasks in a normal
population," Journal of the International Neuropsycho-
logical Society, vol. 3, no. 2, pp. 108 -- 119, 1997.
[74] Matthew L Stanley, Sean L Simpson, Dale Dagenbach,
Robert G Lyday, Jonathan H Burdette, and Paul J Lauri-
enti, "Changes in brain network efficiency and working
memory performance in aging," PLoS One, vol. 10, no.
4, pp. e0123950, 2015.
(a) Planning Brain Network.
Fig. 5: Regions with the highest network measures for Planning (Top) and Execution (Bottom).
(b) Execution Brain Network.
(a) Anatomical regions with higher node betweenness during planning.
(b) Anatomical regions with higher node betweenness during execution.
Fig. 6: Anatomical regions with higher node betweenness during planning (Top) and during execution (Bottom).
(a) Anatomical regions with higher node out-strength during planning.
(b) Anatomical regions with higher node out-strength during execution.
Fig. 7: Anatomical regions with higher node out-strength during planning (Top) and during Execution (Bottom).
(a) Anatomical regions with higher local efficiency and clustering coefficient during planning.
(b) Anatomical regions with higher transitivity during planning.
Fig. 8: Anatomical regions with higher local efficiency and clustering coefficient (Top) and higher transitivity (Bottom) during
planning.
|
1509.01677 | 1 | 1509 | 2015-09-05T08:31:44 | Combinatorics of Place Cell Coactivity and Hippocampal Maps | [
"q-bio.NC"
] | It is widely accepted that the hippocampal place cells' spiking activity produces a cognitive map of space. However, many details of this representation's physiological mechanism remain unknown. For example, it is believed that the place cells exhibiting frequent coactivity form functionally interconnected groups---place cell assemblies---that drive readout neurons in the downstream networks. However, the sheer number of coactive combinations is extremely large, which implies that only a small fraction of them actually gives rise to cell assemblies. The physiological processes responsible for selecting the winning combinations are highly complex and are usually modeled via detailed synaptic and structural plasticity mechanisms. Here we propose an alternative approach that allows modeling the cell assembly network directly, based on a small number of phenomenological selection rules. We then demonstrate that the selected population of place cell assemblies correctly encodes the topology of the environment in biologically plausible time, and may serve as a schematic model of the hippocampal network. | q-bio.NC | q-bio | Combinatorics of Place Cell Coactivity and Hippocampal Maps
Andrey Babichev1,2, Daoyun Ji3, Facundo M´emoli4 and Yuri Dabaghian1,2∗
1Jan and Dan Duncan Neurological Research Institute,
Baylor College of Medicine, Houston, TX 77030,
2Department of Computational and Applied Mathematics,
Rice University, Houston, TX 77005
3Department of Neuroscience,
Baylor College of Medicine, Houston, TX 77030,
4Department of Mathematics,
Ohio State University, Columbus, OH 43210
E-mail: [email protected]
(Dated: July 16, 2021)
It is widely accepted that the hippocampal place cells' spiking activity produces a cognitive map
of space. However, many details of this representation's physiological mechanism remain unknown.
For example, it is believed that the place cells exhibiting frequent coactivity form functionally inter-
connected groups -- place cell assemblies -- that drive readout neurons in the downstream networks.
However, the sheer number of coactive combinations is extremely large, which implies that only a
small fraction of them actually gives rise to cell assemblies. The physiological processes responsi-
ble for selecting the winning combinations are highly complex and are usually modeled via detailed
synaptic and structural plasticity mechanisms. Here we propose an alternative approach that allows
modeling the cell assembly network directly, based on a small number of phenomenological selection
rules. We then demonstrate that the selected population of place cell assemblies correctly encodes
the topology of the environment in biologically plausible time, and may serve as a schematic model
of the hippocampal network.
5
1
0
2
p
e
S
5
]
.
C
N
o
i
b
-
q
[
1
v
7
7
6
1
0
.
9
0
5
1
:
v
i
X
r
a
2
I.
INTRODUCTION
The mammalian hippocampus plays a major role in spatial learning by encoding a cognitive map of
space -- a key component of animals' spatial memory and spatial awareness [1, 2]. A remarkable property
of the hippocampal neurons -- the place cells -- is that they become active only in discrete spatial regions --
their respective place fields [3] (Fig. 1A). A number of studies have demonstrated that place cell activity
can represent the animal's current location [4, 5], its past navigational experience [6, 7], and even its future
planned routes [8, 9]. Numerical simulations suggest that a population of place cells can also encode a
global spatial connectivity map of the entire environment [10 -- 12]. Hence, it is believed that the large-scale
hippocampal representation of space emerges from integrating the information provided by the individual
place cells, although the details of this process remain poorly understood.
Experimental studies point out that the hippocampal map is topological in nature, i.e., it is more similar
to a subway map than to a topographical city map [13 -- 16]. In [10] we proposed a computational approach
for modeling its structure based on several remarkable parallels between the notions of hippocampal physi-
ology and algebraic topology. For example, Cech's theorem asserts that the topological structure of a space
X can be deduced from the pattern of overlaps between regions that cover it (for details see [17] and Methods
in [10]). The argument is based on building a special simplicial complex N, each n-dimensional simplex of
which corresponds to a nonempty overlap of n + 1 covering regions, and demonstrating that the topological
signatures of N and X are same [17]. Since the place cells' spiking activity induces a covering of the envi-
ronment by the place fields, called a place field map (Fig. 1B), Cech's theorem suggests that the place cells'
coactivity (Fig. 1C), which marks the overlaps of the place fields, may be used by the brain to represent the
topology of the environment. In [10 -- 12] it was demonstrated that place cell coactivity can in fact be used to
construct a "temporal" analogue of the nerve complex, T , the simplexes of which, σ = [c1, c2, ..., ck], cor-
respond to the combinations of coactive place cells, c1, c2, ..., ck (Fig. 1D). Using the methods of persistent
homology [18, 19] it was shown that the topological structure of T captures the topological properties of
the environment, if the range of place cell spiking rates and place field sizes happen to parallel biological
values derived from animal experiments [10, 11].
However, it remained unclear whether it is possible to physically implement this information in the
(para)hippocampal network. On the one hand, electrophysiological studies suggest that place cells showing
repetitive coactivity tend to form so-called cell assemblies -- functionally interconnected neuronal groups
that synaptically drive a readout neuron in the downstream networks [20 -- 23] -- which may be viewed as
"physiological simplexes" implementing T . On the other hand, the place cell combinations of T are much
too numerous to be implemented physiologically. In a small environment, c.a. 1 × 1 m, thousands of place
cells are active and the activity of 50 -- 300 of them is near maximal level at every given location [21]. The
number of combinations of hundreds of coactive cells in an ensemble of thousands is unrealistically large,
3000 ∼ 10200. The number of cells in most parahippocampal regions, which may potentially
comparable to C100
serve as readout neurons, is similar to the number of place cells [24]. This implies that only a small fraction
of coactive place cell groups may be equipped with readout neurons, i.e., that the cell assemblies may
encode only a small part of the place cell coactivities -- those which represent a certain "critical mass" of
spatial connections.
Physiologically, the place cell assemblies emerge from dynamically changing constellations of synaptic
connections and are commonly studied in terms of the synaptic and structural plasticity mechanisms [25 --
29]. For a better qualitative understanding of the qualitative properties of the cell assembly network and
practical modeling of the hippocampus' functions, we propose a biologically plausible empirical approach
that allows selecting the most prominent combinations of coactive place cells directly and demonstrate that
the resulting population of cell assemblies is sufficient for representing the topology of the environment.
3
FIG. 1: Place fields and place cells. (A) The blue, green and brown dots, corresponding to the spikes produced by
three different place cells, form well-defined spatial clusters, which represent their respective place fields. Spikes are
positioned in space according to the animal's coordinates at the time of spiking. (B) A place field map produced by
an ensemble of 300 place cells with mean peak firing rate f = 20 Hz and mean place field size s = 14 cm located in
a 1 × 1 m environment. (C) A short time segment of the spike trains produced by three place cells. The periods of
the cells' coactivity, marked by dashed lines, indicate overlap of their respective place fields (panel A): cells c1 and
c2 are coactive in the region 12, cells c1, c2 and c3 are co-active in the region 123. (D) A simplex σ123 represents
schematically the spatial connectivity encoded by the coactivity of cells c1, c2 and c3. Its 1D edges correspond to
pairwise coactivity, e.g. σ12 represents the coactivity of cells c1 and c2.
II. THE METHODS
Mathematically, the task of identifying a subpopulation of coactive place cell combinations corresponds
to selecting according to biologically motivated criteria a subcomplex T0 of the full coactivity complex T .
The cell assemblies correspond to the maximal simplexes of T0, (i.e., the ones that are not subsimplexes
of any other simplex), in contrast with the maximal simplexes of the coactivity complex, T , which can
represent any largest combinations of coactive cells. The "cell assembly complex," T0, should satisfy
several general requirements:
I. Effectiveness. In the readercentric approach [21], each cell assembly drives a coincidence detector
readout neuron in the downstream brain regions. Since the number of the readout neurons is comparable to
the number of place cells, the total number of the maximal simplexes in T0, Nmax(T0), should be comparable
to the number of its vertexes, Nc(T0),
Nmax(T0) ≈ Nc(T0).
However, the algorithm for selecting T0 should reduce only the number of coactive place cell combinations
and not the place cells themselves, meaning that the number of vertexes in T and in T0 should not differ
significantly. In mathematical literature, the number of k-dimensional simplexes of a simplicial complex is
usually denoted as fk, and the list f = ( f0, f1, , fd) is referred to as the complex's f -vector [30]. However,
since in neuroscience literature the letter f is often used to denote firing rates, we denote the number of
k-dimensional simplexes by Nk. As a shorthand notation, we use Nmax to denote the number of the maximal
simplexes and Nc the number of 0-dimensional simplexes in a given complex.
II. Parsimony. To avoid redundancy, only a few cell assemblies should be active at a given location.
Conversely, the rat's movements should not go unnoticed by the hippocampal network, i.e., the periods
during which all place cell assemblies are inactive should be short.
III. Contiguity. A transition of the spiking activity from one cell assembly σi to another σi+1 occurs
when some cells in σi shut off and a new group of cells activates in σi+1 (see Suppl. Movies). The larger is
the subassembly σi,i+1 = σi ∩ σi+1 that remains active during this transition (i.e., the more cells are shared
by σi and σi+1) the more contiguous is the representation of the rat's moves and hence of the space in which
it moves. The overlap between a pair of consecutively active simplexes can be characterized by a contiguity
4
FIG. 2: Topological loops: each horizontal bar represents the timeline
of a topological cycle in T (T): 0D loops (connectivity components) and
the 1D loops. Most cycles last over a short time before disappearing. A
few remaining, persistent loops express stable topological information
that may correspond to physical obstacles in the rat's environment. The
time required for the correct number of cycles to appear is interpreted
as the minimal time Tmin required for the rat to learn the environment.
The environment used in these simulations (Figure 1B) is topologically
connected (b0 = 1), and has one central hole (b1 = 1). Thus, the topo-
logical barcode of this environment -- the list of Betti numbers (b0, b1,
b2,...) -- is (1,1, 0, 0, ). The last spurious loop (blue 1D loop) disappears
at about Tmin = 4.6 minutes, which is the learning time in this case.
index
ξ =
dim(σi ∩ σi+1)
√
dim(σi) dim(σi+1)
,
which assumes the maximal value ξ = 1 for coinciding cell assemblies and ξ = 0 for disjoint ones. In
constructing a cell assembly complex, we expect that the mean contiguity over the simplexes in T0 should
not be lower than in T .
IV. Completeness. The cell assembly complex T0 should capture the correct topological signatures of
the environment, such as obstacles, holes, and boundaries. For example, the lowest dimensional 0D and
1D loops in T0 represent, respectively, the piecewise and the path connectivity of the environment, as they
are captured by the place cell coactivity. This information should emerge from the "topological noise" in
a biologically plausible time period, comparable to the time required to obtain this information via the full
complex, T (see [10, 11] and Fig. 2).
Place cell spiking is modeled as a time-dependent Poisson process with spatially localized rate
λc(r) = fce
− (r−rc)2
s2
c
,
where r is a point in the environment, fc is the maximal firing rate of a place cell c, and sc defines the
size of the corresponding place field centered at rc [31]. In a familiar environment, the place fields are
stable, that is, the parameters fc, sc and rc remain constant [32, 33]. In our simulations, all computations
were performed for ten place cell ensembles, each containing 300 neurons with an ensemble mean maximal
firing rate of 20 Hz and a mean place field size of 30 cm. The place field centers in each ensemble were
randomly scattered across the environment and most quantities reported in the Results were averaged over
ten place field configurations.
Spatial map. We simulated the rat's movements through a small (1× 1 m) planar environment (Fig. 11),
similar to the arenas used in typical electrophysiological experiments (see Methods in [10]) over T = 25
minutes -- the duration of a typical "running session." The spatial occupancy rate of the rat's trajectory (i.e.,
the histogram of times spent at a particular location) and the frequency of the place cells' activity are shown
on Fig. 3A,B. The mean speed of the rat is 20 cm/sec, so that turning around the central obstacle takes about
7 seconds.
By analogy with the place fields, we designate the spatial domain where a combination of place cells
comprising a simplex σ is active as its simplex field, sσ (Fig. 3C). If the simplex corresponds to a cell
assembly, then sσ may also be referred to as the cell assembly field. Similarly to the place fields and
the place field map (Fig.1B), the collection of all simplex fields forms a simplex field map and the cell
assembly fields form a cell assembly map ((Fig. 3C). These maps provide a better "geometric proxy" for the
rat's cognitive map because they illustrate both the activity and the coactivity of the individual place cells
5
FIG. 3: Spatial maps. (A) Occupancy of spatial locations in a 1 × 1 m environment -- a 2D histogram of the time
spent by the animal in different locations. (B) Frequency of place cells' spiking: each dot marks the location of a
place cell's center rc and indicates the corresponding appearance rate according to the colorbar. Higher appearance
rates appear in the domain where the spatial occupancy is higher. (C) Simplex field map. The place field map for the
same place cell ensemble is shown in Fig. 1B. (D) Spatial distribution of the frequency of the maximal simplexes'
appearances. Notice that, since place cells with higher appearance rates tend to produce higher order cell assemblies,
which, in turn, have lower appearance rates, the spatial distribution of rates on B. and D. are complementary.
((Fig. 3C-D). In the following, the structure of these maps will be used to discuss our selection algorithms.
If the distinction between a cell assembly map and a simplex map is not essential, it will be referred to as a
space map.
Population activity. To define the population code [34] of place cell combinations, we construct place
cell activity vectors by binning spike trains into 1/4 sec long time bins (for a physiological justification of
this value see [11, 35]). If the time interval T splits into n such bins, then the activity vector of a cell c is
mc(T) = [mc,1, , mc,n],
where mc,k specifies how many spikes were fired by c in the kth time bin. The components of mc, normalized
by the total number of spikes, Mc, define spiking probabilities, pc,k = mk/Mc [36]. A stack of activity vectors
forms an activity raster illustrated on Fig. 4.
Two cells, c1 and c2, are coactive over a certain time period T, if the dot product of their activity vectors
does not vanish,
mc1(T) · mc2(T) (cid:44) 0.
The component-wise or Hadamard product of two activity vectors
mc1,c2
= mc1 (cid:12) mc2
= [mc1,1mc2,1, mc1,2mc2,2, , mc1,nmc2,n]
defines the coactivity vector of cells c1 and c2, which can also be viewed as the activity vector of the
corresponding 1D simplex σ12 = [c1, c2], mσ12 ≡ mc1,c2. Similarly, the Hadamard product of k vectors,
mσ12...k
= mc1,c2,...,ck
= mc1 (cid:12) mc2 (cid:12) (cid:12) mck ,
defines the activity vector of the simplex σ12...k = [c1, c2, , ck].
For each activity vector, mσ, we also define its bit array mapping into a binary appearance vector, aσ,
which indicates during which time-bins the corresponding simplex σ has made its appearance, i.e., aσ,i = 1
iff mσ,i > 0. The appearance rate, fσ(T), of a simplex σ over a time interval T, is defined as the L1 norm
of its appearance vector, averaged over that time interval,
fσ(T) = (1/T)Σiaσ,i.
These appearance vectors and appearance rates allow distinguishing the intrinsic physiological characteris-
tics of place cells' spiking, e.g., their maximal firing rate, from the frequency with which these cells activate
due to the rat's movements through their respective place fields. While the maximal firing rate of a typical
place cell is about 15 Hz [1], the frequency of their activation is much lower.
6
FIG. 4: An activity raster of a population of 20
place cells over 250 time bins. Each row defines the
activity vector of the corresponding place cell. The
color of the ticks indicates the number of spikes con-
tained in the corresponding bin, according to the col-
orbar on the right. At every time step, the nonempty
bins in the vertical column define the list of cur-
rently active cells, i.e., the active simplex σt. During
the time interval T, marked by the two vertical blue
lines, cell c16 is coactive with c18 but not coactive
with c14.
III. RESULTS
The simulated ensembles of 300 place cells in the environment shown on Fig. 1B produced a coactivity
complex T with about Nmax = 1000 maximal simplexes. Despite the high dimensionality of these simplexes
(up to D = 35, mean ¯D = 17), the characteristic dimensionality of a facet shared by two consecutively active
simplexes, σi and σi+1, is relatively low, so that the mean contiguity of T is ξ = 0.6. This implies that,
geometrically, if the simplexes of T are viewed as multidimensional tetrahedrons, the selected complex,
T0(θ), assumes a highly irregular shape (Fig. S1A).
More importantly, nearly 100% of the maximal simplexes appeared only once during the entire 25
minute period of navigation, i.e., a typical maximal simplex's appearance rate is low, fσ ∼ 10−3 Hz. How-
ever, a typical vertex activated about 200 times or every seven seconds, suggesting that some of the lower
dimensional subsimplexes may be better candidates for forming cell assemblies. Is it then possible to build
a cell assembly complex T0 by discarding the high-dimensional maximal simplexes with low appearance
rates and retaining their subsimplexes that appear more frequently? We tested this hypothesis by identifying
the combinations σ whose coactivity exceeds a certain threshold fσ > θ, and studied the properties of the
resulting simplicial complex as a function of θ (Fig. 5A).
First we observed that, as soon as the appearance threshold is introduced (θ (cid:38) 10−3 Hz), the high dimen-
sional simplexes start braking up, releasing large numbers of lower dimensional subsimplexes: the number
of k-dimensional subsimplexes in a n-dimensional simplex grows as combinatorial coefficient Ck+1
n+1, e.g.,
18 ≈ 44, 000. As a result, the complex T0(θ) rapidly inflates. As θ increases further
for n = 17 and k = 7, C8
(θ > 0.04), the number of "passing" simplexes decreases, and T0(θ) begins to shrink in all dimensions (i.e.,
ND(θ1) > ND(θ2) for θ1 < θ2, for all D, (Fig. 5B). Despite this, the number of maximal simplexes remains
high: Nmax = 30× Nc at θ = 0.04 Hz, Nmax = 7× Nc at θ = 0.07 Hz and Nmax = 3× Nc for the highest tested
threshold, θ = 0.1 Hz (Fig. 5B), while their characteristic dimensionality drops from D = 17 to D = 7 at
θ = 0.04 Hz and to D = 3 at θ = 0.1 Hz. The mean contiguity index for this range of thresholds remains
close to ξ = 0.7, indicating that the degree of overlap between the selected combinations of place cells is
higher than in the original coactivity complex.
However, raising the passing threshold θ quickly destroys the geometric integrity of the resulting com-
plex's spatial map. As shown on Fig. S2, for θ = 0.05 Hz, only ∼ 50% of the environment is covered by the
remaining simplex fields, and for θ = 0.07 Hz the simplex map barely retains its one-piece connectedness:
in some cases the complex T0 splits in two (the corresponding Betti numbers, b0, are listed in Table S1,
for an illustration see Fig. S3). For θ = 0.1 Hz, the complex fragments into multiple components (mean
b0 ∼ 7) that are riddled with holes: the Betti numbers bn>0 indicate the presence of hundreds of stable loops
in higher dimensions. Thus, even if the coactive place cell combinations selected at θ ≥ 0.05 Hz could be
supplied with readout neurons and would form cell assemblies, the resulting cell assembly network would
not encode the correct spatial connectivity.
An additional problem is that reducing the order of the assemblies violates the "assembly code" for
spatial locations: every time several subsimplexes σi are selected from a high-order maximal simplex σ,
7
FIG. 5: A direct selection of the simplexes by appearance rates. (A) In the original coactivity complex T (θ =
0), the maximal simplexes σmax appear on average but once during the entire observation period, resulting in low
appearance rates ( fσ < 10−3 Hz, blue line). Imposing four different thresholds θ (color coded) raises the appearance
rates of the selected maximal simplexes almost uniformly in all dimensionalities. (B) Cumulative distribution of the
number of maximal simplexes Nmax over the selected simplexes' dimension. In the T (θ = 0) case Nmax exceeds the
number of vertexes (Nc = 300, black horizontal line) by almost an order of magnitude. Small threshold values result
in an explosive increase of Nmax which then begins to decrease for θ > 0.04 Hz, remaining significantly higher than
Nc for all four tested values of θ. (C) The histograms of the maximal simplexes' dimensionalities fit with normal
distribution. The high mean dimensionality ( ¯D = 17) observed in the T (θ = 0) case reduces to ¯D = 2.2 for θ = 0.1
Hz. The width of the distributions is about 50% of ¯D. (D) The histograms of the number of the coactive maximal
simplexes, fit to an exponential distribution demonstrate that the typical number of coactive simplexes is large, β > 10.
All values are averaged over ten place field maps generated by ten place cell ensembles with the same mean peak firing
rate and mean place field size.
several overlapping simplex fields sσi are produced in place of a single sσ. As a result, the parsimony of
the representation is compromised: a location that was previously represented by a single simplex becomes
represented by a few of its subsimplexes (Fig. S4A-B). Fig. 5D shows a histogram of the numbers of
simultaneously active maximal simplexes in T0: although most of the time only a few maximal simplexes
are active, a coactivity of many of them (n > 25) is not uncommon. Conversely, while most of the time -- on
average 84% for the selected place cell ensembles -- at least one simplex is active, longer inactivity periods
are observed as described by double exponential distributed with the rate β ≈ 3.5 sec (Fig. S2).
Overall, since most of the T0-requirements listed in the Methods fail, we are led to conclude that the
most straightforward selection rule, based on selecting high appearance rates, does not produce the de-
sired tradeoff between the order of the assemblies, the frequency of their appearances, and the quality of
topological representation of the environment. This failure motivates the search for alternative methods.
Method I. To produce a more detailed approach to selecting coactive cell combinations, we observe that
place fields are typically convex planar regions, and hence the existence of higher order overlaps between
them actually follows from the lower order overlaps. According to Helly's theorem, a collection of n >
D + 1 convex D-dimensional regions in Euclidean space RD will necessarily have a nonempty common
intersection, if the intersection of every set of D + 1 regions is nonempty (see [37, 38] and Fig. S4D). From
the perspective of Cech theory, this implies that if n convex regions which cover a D-dimensional space
contribute all the combinatorially possible D-dimensional simplexes to the nerve complex, then they also
provide all the higher (up to n−1) dimensional simplexes to it. In a planar (D = 2) environment, this implies
that a set of four or more place fields has a common intersection, if any three of them overlap. Moreover,
although mathematically it is possible that three place fields exhibit pairwise, but not triple overlap, the
probability of such an occurrence is low (Fig. S4C). A direct computational verification shows that if a
triple of place cells demonstrates pairwise coactivity, then, in over 90% of cases, it also correctly encodes a
triple spatial overlap. In other words, a "clique" of pairwise coactivities indicates the overlaps of all higher
orders, which implies that the spatial connectivity graph GN whose vertexes correspond to the place fields
8
FIG. 6: Selecting maximal simplexes via the pairwise coactivity threshold (Method I). (A) The appearance rates
of the maximal simplexes computed for four different pairwise appearance rate thresholds θ decrease as a function of
their dimensionality. The values at D = 1 correspond to the value of the threshold imposed on the links' appearance
rate. (B) Cumulative distribution of the numbers of maximal simplexes, Nmax, over the selected simplexes' dimension.
The numbers of cells Nc for each threshold value are shown by horizontal lines. The tendency of the maximal
simplexes to outnumber the vertexes Nmax > Nc, characteristic for small values of θ, is reversed around θ = 0.07
Hz, where Nmax and Nc level out. (C) The histograms of the maximal simplexes' dimensionalities fit with normal
distribution. The mean dimensionalities are similar to the ones produced by the previous selection method. The width
of the distributions is about 50% of ¯D. (D) The histogram of the number of coactive maximal simplexes, fit to an
exponential distribution, shows that the expected number of coactive simplexes (β ∼ 4) is significantly lower than in
the previous selection method. The procedure of averaging over the place field maps is the same.
and links represent pairwise overlaps, encodes most simplexes in the nerve complex N.
As a reminder, a clique of an undirected graph is a set of pairwise connected vertices. From the combi-
natorial perspective, a clique and a simplex have the same defining property: any subset of a simplex is its
subsimplex and any subset of a clique is its subclique; a maximal clique is the one that is contained in no
other clique. Hence, each graph defines its own "clique complex," the k-dimensional simplexes of which
corresponds to the graph's cliques with k + 1 vertices [39].
The observation that the nerve complex induced from the place field map can be approximated by the
clique complex of the place field pairwise connectivity graph, suggests that the corresponding coactivity
complex T can also be built based only on pairwise, rather than higher-order, coactivities. This approach is
well justified physiologically, since pairwise coactivity detector pairs of synapses are commonly observed
[40, 41]. The rule for defining the temporal analogue of GN -- the relational graph GT -- is straightforward:
a pair of vertexes is connected in GT if the corresponding cells ci and c j are coactive. Thresholding pairwise
coactivity rates according to the rule
if fci,c j ≥ θ
if fci,c j < θ.
(1)
1
0
Ci j =
allows constructing a family of relational graphs GT (θ) over the pairs of place cells with high coactivity.
The higher the threshold is, the sparser its connectivity matrix Ci j and the smaller the number of maximal
cliques and hence of maximal simplexes in the corresponding clique complex. Since in the following the
graph GN will not be used we will suppress the subscript "T " in the notation for GT .
We studied the relational graphs G(θ) and their respective clique complexes T0(G(θ)) ≡ T0(θ) as a func-
tion of θ. First, we observed that the appearance rates of the maximal simplexes in T0(θ) become sensitive
to the simplexes' dimensionality (Fig. 6A), implying that this selection procedure in effect attributes differ-
ent thresholds to simplexes of different dimensions by using only one free parameter θ. Second, the size
of T0(θ) is not as large as before. As shown on Fig. 6B, even for a relatively low threshold θ = 0.05 Hz,
the number of maximal simplexes exceeds the number of cells only marginally. For higher thresholds, this
number steadily decreases: ND(θ1) < ND(θ2) for θ1 > θ2 > 0.04 and D > 3, though in lower dimensions
9
FIG. 7: Figure 7. Correction algorithms. (A) A spatial projection of the 2D skeleton of T (θ) shows gaps and holes
that compromise, respectively, the piecewise and path connectivity of T (θ). If the links across the gaps and holes of
T (θ) are restored, then its correct connectivity structure may be regained. (B) A 'hole' produced by five connected
vertexes is closed by restoring some of the previously discarded crosslinks. (C) A projection of the relational graph G
into the environment, shown in grey. The edges added across the gaps are shown in red and the edges added to fill the
holes are shown in green. The vertexes that are left disconnected due to low appearance rates of the edges connecting
them to other vertexes are shown by blue dots. (D) A spatial map of the resulting "patched" 2D skeleton of T (θ). The
parameter values are ng = 15, mh = 10, and the lowered threshold for reintroducing the missing links is θ = 50.
(1 ≤ D ≤ 3) this number may increase. The characteristic contiguity ranges between ξ = 0.65 at θ = 0.05
Hz to ξ = 0.72 at θ = 0.14 Hz, which is higher than the value produced by the direct simplex selection
method. Geometrically, this implies that the collection of maximal simplexes selected by pairwise thresh-
old selection is more aggregated than the collection produced via direct simplex selection, i.e., the resulting
complex T0(θ) is geometrically more similar to a "simplicial quasimanifold" (see Fig. S1B). However, the
number of place cells Nc drops as a result of discarding too many links with low appearance rate: Nc = 290
at θ = 0.05 Hz and Nc = 100 at θ = 0.14 Hz. At θ = 0.1 Hz number of cells levels out with the number of
maximal simplexes, Nmax ∼ Nc = 260.
As before, raising the coactivity threshold degrades the spatial map. At θ > 0.07 Hz the simplex fields
no longer cover the environment and at θ > 0.1 Hz the map fragments into pieces (Fig. S5). However the
resulting complex exhibits a much more regular topological behavior: the correct signature (b0 = 1, b1 = 1,
b2 = 0, b3 = 0, ...) in T0(θ) appears at θ = 0.05 Hz. The higher order Betti numbers (bn≥2) remain trivial
at still higher θs (Table S2A), even though the connectedness and path connectivity of the environment (b1
and b0) become misrepresented.
This improvement of the behavior of T0(θ) suggests that, despite all the shortcomings, the link-selection
strategy may lead to a successful model of the place cell assembly network. After all, it is not surprising
that a single selection rule does not resolve all the aspects of the cell assembly formation. Yet if it captures
the essence of the process, it should be possible to correct or to adjust its outcome. For example, one of
the difficulties faced by the coactivity selection algorithm is that, for high θ, T0(θ) may brake into several
pieces. However, the gaps between them are small. Thus, if a few discarded edges of the relational graph
that originally bridged these gaps are retained, then the connectedness of T0(θ) may be spared (Fig. 7A).
Similarly, a "hole" in the relational graph is a linear chain of edges, connected tail to tail, with no shortcuts.
However, if the links with the lower appearance rate ( f ≥ θh, θh < θ) that span across the hole exist at θ = 0,
then they also can be restored (Fig. 7B). This may remove the non-contractible chains of 1D simplexes
in T0(θ) that compromised its path connectivity (Fig. 7C,D). Thus, we implemented the following two
rectification algorithms:
1. Filling gaps: find pairs of vertexes va and vb separated in G(θ) by more than ng edges and then
test whether these vertexes are connected directly by links (from G(θ = 0)) whose appearance rate
exceeds a lower threshold θg < θ. If such links exist, add them to G(θ) (red lines on Fig. 7C).
2. Closing holes: A closed chain containing mh ≥ 4 edges in G(θ), with no shortcuts, is likely to
10
FIG. 8: Statistics of the vertex degrees in relaitonal graphs. (A) The histogram of the vertex degrees k in the
neighbor-controlled relational graph G(n0), computed for four different n0 (Method II) and fitted to a power law
distribution P(k) ∼ k−γ. The graph demonstrates that G(n0) is a scale-free network. (B) The same distribution on the
log-log scale and an independent linear fit of the powers γ. The confidence intervals of the two fits, ranging between
±0.15 and ±0.3, overlap for each case. (C) In the pairwise coactivity threshold (Method I), the histogram of the
relational graph's vertex degrees is fit by negative binomial distribution, suggesting that G(θ) is similar to a random
network.
produce a hole in T0(θ). We identified such chains and restored the discarded cross-links whose
appearance rate exceeds a lower threshold θh < θ.
Thus, both rectification algorithms depend on two parameters: the length of the involved chains (ng for
gaps and mh for holes) and the value of the reduced threshold θg and θh. In our numerical experiments,
we found that the optimal value for the thresholds is θh = θg = 50, and the parameters range between 5
and 10 (mh) and 10 and 15 (ng). Typically, each rectification procedure is applied once or twice before
the right signature of T0(θ) is achieved, and this without producing significant changes of the complex's
structure, such as altering the appearance of its simplexes or increasing its size Nmax. As illustrated in the
Table S2B, the correct signature in the "repaired" complex is achieved for all cases at θ = 0.07 Hz. In
particular, at θ = 0.07 Hz we obtain a simplicial complex T0 with the correct signature, having Nc = 260
vertexes and about the same number of maximal simplexes, Nmax ≈ Nc. These maximal simplexes appear
on average at a rate of fσ ≥ 0.07 Hz, at least during every other run of the rat around the environment, and
have dimensionality D = 6. As a result, the requirements to T0 are met and the maximal simplexes of T0
may represent hippocampal place cell assemblies that together encode a map of the environment, and hence
T0 itself can be viewed as the "cell assembly complex."
Method II. A common feature of the appearance-rate-based selection rules is that the resulting simplicial
complex reflects biases of its spatial occupancy: higher dimensional maximal simplexes concentrate over
the parts of the environment where the rat appears more frequently. In particular, the relational graph shows
a higher concentration of edges over the eastern segment of the environment (Fig. 7C) where the occupancy
rate is highest (Fig. 3A). On the one hand, this is natural since the frequency of the place cells' spiking
activity certainly does depend on the frequency of the rat's visits to their respective place fields, which
therefore affects the hippocampal network's architecture [28, 29]. In fact, this argument is at the core of the
classical "hippocampus as a cognitive graph" model [42, 43], which proposes that the architecture of the
hippocampal network is an epiphenomenon of the place cell coactivity. On the other hand, the physiological
processes that produce synaptic connections may be more autonomous. For example, the CA3 region of
the hippocampus is anatomically a recurrent network of place cells whose spiking activity and synaptic
architecture are dominated by the network's attractor dynamics [44 -- 46].
These considerations lead us to test an alternative method of constructing the relational graph based on
selecting, for every cell, its n0 closest neighbors as defined by the pairwise coactivity rate fci,c j. Note that
the resulting number of connections may be different for different cells: a cell c1 may be among the n0
closest neighbors of a cell c2, and hence c1 and c2 become connected, but the set of n0 closest neighbors of
11
FIG. 9: Figure 9. Selecting maximal simplexes via best neighbor selection (Method II). (A) The appearance rates
of the maximal simplexes in the simplicial complex T (n0), computed for four different values of n0 (color coded),
decrease as a function of their dimensionality. (B). Cumulative distribution of the number of maximal simplexes Nmax
over the selected simplexes' dimension. For the tested values of n0, the fixed number of vertexes Nc = 300, indicated
by the horizontal black line, is close to the number of maximal simplexes. For n0 = 7, the values Nmax and Nc come
closest. (C). The histograms of the maximal simplexes' dimensionalities, fit to the normal distribution, indicate that
for the relational graph with a similar number of links, the mean dimensionalities of the maximal simplexes are smaller
than in in the complex built via the threshold-selection method. The width of the distributions is about 40% of ¯D. (D)
The histogram of the number of coactive maximal simplexes, fit to an exponential distribution. An expected number
of coactive simplexes ranges between β = 2 and β = 6. The procedure of averaging over the place field maps is the
same.
a cell c1 may not include c2, which bears a certain resemblance to the preferential attachment models [47].
As a result, the vertex degrees k of the (undirected) relational graph may differ from one another and from
n0. A direct computational verification shows that k is distributed according to a power law, P(k) ∼ k−γ,
where γ ranges, for different n0, between γ ∼ 2 and γ ∼ 4 (Fig. 8), which implies that G(n0) demonstrates
scale-free properties [47, 48] characteristic of the hippocampal network [49, 50]. In contrast, the histogram
of the vertex degrees in the threshold-controlled relational graph G(θ) may be fit with the negative binomial
distribution (Fig. 8B), which indicates that G(θ) is similar to a random graph.
This neighbor-selection method for building the relational graph G(n0) has a number of other immediate
advantages over the threshold-controlled construction of G(θ). For example, no cells are excluded from T0
due to the low appearance of the edges connecting to them. As a result, the simplex fields are distributed
more uniformly (Fig. S6), which helps capture the correct piecewise connectedness of the environment.
By studying the properties of the clique complexes produced by the relational graphs G(n0) for n0 = 2,
4, 7 and 12 -- parameters chosen to produce similar numbers of edges as in the previous method -- we found
that the number of maximal cliques in T0(G(n0)) is typically lower than in T0(G(θ)). The appearance rates
of maximal cliques in G(n0) are more scattered and less sensitive to dimensionality than in G(θ) (Fig. 8A
and Fig. S6). The number of maximal simplexes in T0(n0) remains close to the number of cells (Fig. 8B)
and their dimensionality is lower than in the threshold-based selection approach (Fig. 8C). The contiguity
index in all complexes ranges between to ξ = 0.67 and ξ = 0.71. The coverage of the space with the
simplex fields improves with growing n0 (see Fig. S6) -- for n0 > 2 the complex T0(n0) is connected, while
the behavior of b0 is more regular (see Table S3). However, the path connectivity of the complex T0(n0)
remains deficient for all n0 because the number of stable spurious 1D loops remains high (Table S3). After
filling the gaps and closing the holes, most complexes constructed for n0 ≥ 7 acquire correct topological
signatures (Table S3), and the requirements to T0 are satisfied. Thus, the simplicial complex obtained by
the neighbor selection method for n0 ≥ 7 can also be viewed as a "cell assembly complex," meaning it can
serve as a formal model of the place cell assembly network with mean contiguity ξ = 0.7.
IV. DISCUSSION
12
The proposed approach allows creating schematic models of the place cell assembly network -- the cell
assembly complex T0 -- by controlling the basic phenomenological parameters of place cell coactivity and
then relating the network's architecture to the net topological information it encodes. The coactivity com-
plex T was previously used for representing the pool of place cell coactivities [10 -- 12]. Specifically, the
low order (pair and triple) coactivity events were used to construct the 2D skeleton of T , and then its 0D
and 1D topological loops were matched with the corresponding topological signatures of the environment.
In the current study, the entire pool of place cell coactivities is used to model the full place cell assembly
network, including the higher order assemblies representing both the low-dimensional spatial environment
as well as high-dimensional memory space [21, 51]. The learning times Tmin estimated from the dynamics
of the 0D and 1D loops in T0 remain close to the learning times computed for the full coactivity complex T
(see Table S4). This implies that the selected, "core" pool of coactive place cell combinations captures the
topological structure of the environment as fast and as reliably as the entire set of the place cell coactivities.
We view the proposed algorithms as basic models of a more general "phenomenological" approach,
one which can be further developed along several broad lines. First, the structure of the relational graph is
currently deduced from the activity vectors defined over the entire navigation period T = 25 minutes. A
biologically more plausible selection algorithm should be adaptive: the structure of the relational graph at
a given moment of time t < T should be based only on the spiking information produced before t. Hence,
in a more advanced model, the structure of the relational graph should develop in time, and in general
the cell assemblies comprising T0 should be derived using synaptic and structural plasticity mechanisms.
Second, the selection criteria in Methods I and II above may be individualized: the appearance threshold
used to construct the relational graph can be assembly-specific, i.e. θ = θ(σ), so that the properties of
the resulting network would be described in terms of the probability distribution of the threshold values
across the cell assembly population. Similarly, the number of closest neighbors can be made cell-specific,
n0 = n0(ci), which should permit better control over the topological properties both of the network and of
the cell assembly complex. Third, threshold control can be implemented using different coactivity metrics,
for instance via the pairwise correlation coefficient
ρ(c1, c2) = (m1 · m2)/m1m2,
(2)
which would connect cells with correlated spiking (irrespective of their firing rates), in contrast with the
metric (2), which does the opposite. In general, two metrics ρ and ρ(cid:48), produce relational graphs with differ-
ent topologies. Nevertheless, they may produce similar or identical large-scale effects, such as generating
topologically identical cell assembly complexes T0, or exhibit similar learning times, Tmin. Identifying
classes of metrics that produce topologically similar results will be examined in future research.
V. ACKNOWLEDGMENTS
The work was supported in part by Houston Bioinformatics Endowment Fund, the W. M. Keck Founda-
tion grant for pioneering research and by the NSF 1422438 grant.
VI. REFERENCES
13
Ann. Rev Neurosci 24: 459-486.
Press. xiv, 570 pp. p.
95: 2717-2719.
[1] Best PJ, White AM, Minai A (2001) Spatial processing in the brain: the activity of hippocampal place cells.
[2] O'Keefe J, Nadel L (1978) The hippocampus as a cognitive map. New York: Clarendon Press; Oxford University
[3] Best PJ, White AM (1998) Hippocampal cellular activity: a brief history of space. Proc. Natl Acad. Sci. USA
[4] Zhang K, Ginzburg I, McNaughton BL, Sejnowski TJ (1998) Interpreting neuronal population activity by re-
construction: unified framework with application to hippocampal place cells. J Neurophysiol. 79: 1017-1044.
[5] Brown EN, Frank LM, Tang D, Quirk MC, Wilson MA (1998) A statistical paradigm for neural spike train
decoding applied to position prediction from ensemble firing patterns of rat hippocampal place cells. J Neurosci
18: 7411-7425.
[6] Carr MF, Jadhav SP, Frank LM (2011) Hippocampal replay in the awake state: a potential substrate for memory
consolidation and retrieval. Nat Neurosci 14: 147-153.
[7] Derdikman D, Moser MB (2010) A dual role for hippocampal replay. Neuron 65: 582-584.
[8] Pfeiffer BE, Foster DJ (2013) Hippocampal place-cell sequences depict future paths to remembered goals. Na-
[9] Dragoi G, Tonegawa S (2011) Preplay of future place cell sequences by hippocampal cellular assemblies. Nature
ture advance online publication.
469: 397-401.
[10] Dabaghian Y, Mmoli F, Frank L, Carlsson G (2012) A Topological Paradigm for Hippocampal Spatial Map
Formation Using Persistent Homology. PLoS Comput. Biol. 8: e1002581.
[11] Arai M, Brandt V, Dabaghian Y (2014) The Effects of Theta Precession on Spatial Learning and Simplicial Com-
plex Dynamics in a Topological Model of the Hippocampal Spatial Map. PLoS Comput. Biol. 10: e1003651.
[12] Curto C, Itskov V (2008) Cell groups reveal structure of stimulus space. PLoS Comput. Biol. 4: e1000205.
[13] Gothard KM, Skaggs WE, McNaughton BL (1996) Dynamics of mismatch correction in the hippocampal en-
semble code for space: interaction between path integration and environmental cues. J Neurosci 16: 8027-8040.
[14] Leutgeb JK, Leutgeb S, Treves A, Meyer R, Barnes CA, et al. (2005) Progressive transformation of hippocampal
neuronal representations in "morphed" environments. Neuron 48: 345-358.
[15] Alvernhe A, Sargolini F, Poucet B (2012) Rats build and update topological representations through exploration.
Anim Cogn 15: 359-368.
10.7554/eLife.03476.
[16] Dabaghian Y, Brandt VL, Frank LM (2014) Reconceiving the hippocampal map as a topological template; eLife
[17] Hatcher A (2002) Algebraic topology. Cambridge; New York: Cambridge University Press.
[18] Ghrist R (2008) Barcodes: The persistent topology of data. Bulletin of the American Mathematical Society 45:
61-75.
243 p. p.
719 p. p.
429-441.
[19] Zomorodian AJ (2005) Topology for computing. Cambridge, UK ; New York: Cambridge University Press. xiii,
[20] Harris KD, Csicsvari J, Hirase H, Dragoi G, Buzsaki G (2003) Organization of cell assemblies in the hippocam-
pus. Nature 424: 552-556.
[21] Buzsaki G (2010) Neural syntax: cell assemblies, synapsembles, and readers. Neuron 68: 362-385.
[22] Huyck C, Passmore P (2013) A review of cell assemblies. Biol. Cybern. 107: 263-288.
[23] Harris KD (2005) Neural signatures of cell assembly organization. Nat Rev Neurosci 6: 399-407.
[24] Shepherd GM (2004) The synaptic organization of the brain. Oxford ; New York: Oxford University Press. xiv,
[25] Wennekers T, Palm G (2009) Syntactic sequencing in Hebbian cell assemblies. Cognitive Neurodynamics 3:
[26] Ghalib H, Huyck C. A Cell Assembly Model of Sequential Memory; 2007 2007. pp. 625-630.
[27] Itskov V, Curto C, Pastalkova E, Buzsaki G (2011) Cell assembly sequences arising from spike threshold adap-
tation keep track of time in the hippocampus. J Neurosci 31: 2828-2834.
[28] Caroni P, Donato F, Muller D (2012) Structural plasticity upon learning: regulation and functions. Nat Rev
14
Neurosci 13: 478-490.
[29] Chklovskii DB, Mel BW, Svoboda K (2004) Cortical rewiring and information storage. Nature 431: 782-788.
[30] Gromov ML (1968) On the number of simplexes of subdivisions of finite complexes. Mathematical notes of the
[31] Barbieri R, Frank LM, Nguyen DP, Quirk MC, Solo V, et al. (2004) Dynamic analyses of information encoding
Academy of Sciences of the USSR 3: 326-332.
in neural ensembles. Neural Comput. 16: 277-307.
[32] Wilson MA, McNaughton BL (1993) Dynamics of the hippocampal ensemble code for space. Science 261:
[33] Brown EN, Nguyen DP, Frank LM, Wilson MA, Solo V (2001) An analysis of neural receptive field plasticity
by point process adaptive filtering. Proc. Natl Acad. Sci. USA 98: 12261-12266.
[34] Pouget A, Dayan P, Zemel R (2000) Information processing with population codes. Nat Rev Neurosci 1: 125-
1055-1058.
132.
49-86.
[35] Mizuseki K, Sirota A, Pastalkova E, Buzsaki G (2009) Theta oscillations provide temporal windows for local
circuit computation in the entorhinal-hippocampal loop. Neuron 64: 267-280.
[36] Perkel DH, Gerstein GL, Moore GP (1967) Neuronal Spike Trains and Stochastic Point Processes: I. The Single
Spike Train. Biophys. J 7: 391-418.
[37] Avis D, Houle ME (1995) Computational aspects of Helly's theorem and its relatives. International Journal of
Computational Geometry and Applications 05: 357-367.
[38] Eckhoff J (1993) CHAPTER 2.1, Helly, Radon, and Carathodory Type Theorems. In: Wills PMGM, editor.
Handbook of Convex Geometry. Amsterdam: North-Holland. pp. 389-448.
[39] Bandelt HJ, Chepoi V (2008) Metric graph theory and geometry: a survey. Contemporary Mathematics 453:
[40] Katz Y, Kath WL, Spruston N, Hasselmo ME (2007) Coincidence detection of place and temporal context in a
network model of spiking hippocampal neurons. PLoS Comput. Biol. 3: e234.
[41] Brette R (2012) Computing with Neural Synchrony. PLoS Comput. Biol. 8: e1002561.
[42] Muller RU, Stead M, Pach J (1996) The hippocampus as a cognitive graph. J Gen Physiol. 107: 663-694.
[43] Burgess N, O'Keefe J (1996) Cognitive graphs, resistive grids, and the hippocampal representation of space. J
Gen Physiol. 107: 659-662.
[44] Colgin LL, Leutgeb S, Jezek K, Leutgeb JK, Moser EI, et al. (2010) Attractor-map versus autoassociation based
attractor dynamics in the hippocampal network. J Neurophys. 104: 35-50.
[45] Wills TJ, Lever C, Cacucci F, Burgess N, O'Keefe J (2005) Attractor dynamics in the hippocampal representation
of the local environment. Science 308: 873-876.
[46] Tsodyks M (2005) Attractor neural networks and spatial maps in hippocampus. Neuron 48: 168-169.
[47] Barabsi A-L, Albert R (1999) Emergence of Scaling in Random Networks. Science 286: 509-512.
[48] Albert R, Barabsi A-L (2002) Statistical mechanics of complex networks. Reviews of Modern Physics 74: 47-97.
[49] Li X, Ouyang G, Usami A, Ikegaya Y, Sik A (2010) Scale-Free Topology of the CA3 Hippocampal Network: A
Novel Method to Analyze Functional Neuronal Assemblies. Biophys. J 98: 1733-1741.
[50] Bonifazi P, Goldin M, Picardo MA, Jorquera I, Cattani A, et al. (2009) GABAergic Hub Neurons Orchestrate
Synchrony in Developing Hippocampal Networks. Science 326: 1419-1424.
[51] Eichenbaum H, Dudchenko P, Wood E, Shapiro M, Tanila H (1999) The hippocampus, memory, and place cells:
is it spatial memory or a memory space? Neuron 23: 209-226.
[52] Singh G, Memoli F, Ishkhanov T, Sapiro G, Carlsson G, et al. (2008) Topological analysis of population activity
in visual cortex. J Vis 8: 11 11-18.
VII. SUPPLEMENTARY FIGURES
15
FIG. S1: Simplicial complexes. (A) A schematic representation of an irregular simplicial complex K, in which
the number of maximal simplexes is larger than the number of vertexes (black dots). The maximal 1D simplexes
are shown as blue segments, the 2D simplexes as gray triangles and the 3D simplexes as pink tetrahedrons. (B) A
simplicial "quasi-manifold," Q, which has a similar number of vertexes and maximal simplexes of different dimen-
sionalities. If each maximal simplex, e.g. (1, 2, 3) or (10, 7, 8, 9), corresponds to an assembly of place cells driving a
readout neuron (red dots), then Q is a cell assembly complex. Dotted lines represent synaptic connections from the
place cells to the readout neuron.
16
FIG. S2: The simplicial complexes T0(θ) constructed by direct selection of coactive combinations, for four
different values of θ. (A) The appearance rates of simplexes, arranged from left to right according to their dimension.
Each dot corresponds to a maximal simplex whose dimension is color-coded according to the colorbar on the right.
(B) Spatial distribution of the dimensionalities of the selected simplexes. (C) Spatial distribution of the appearance
rates of the selected simplexes. (D) The histograms of the lapse times, fit to double exponential distribution (blue
line), and the value of the fitted distribution's rate β. (E) Spatial projections of the 2D skeletons of the T0(θ). Data for
all panels is computed for a specific place field map for illustrative purposes.
17
FIG. S3: Low-dimensional topological features of the cell assembly complexes captured by place field map.
(A) A place filed map corresponding to a singly-connected (b0 = 1, marked by the dot) complex with no non-
contractible loops (b1 = 0). (B) A place field map corresponding to a complex with correct list of Betti numbers
(correct topological barcode [18]): the physical hole (see Figures 1 and 3) produces one non-contractible loop (red
circle, b1 = 1). (C) A map containing a spurious hole in the upper-right corner produces an extra persistent loop (the
additional small circle, net b1 = 2). (D) A map containing two disconnected pieces marked by the black dots (net
b0 = 2) and having no non-contractible loops (b1 = 0). (E) A map containing two pieces and one persistent spurious
1D loop. (F) The green and the brown place fields at the top connect, yielding another persistent spurious 1D loop.
Compare these illustrations and the topological barcodes to the illustrations and Suppl. Movies provided in [52]
18
FIG. S4: Schematic illustrations of spatial maps. (A) Schematic representations of 3d order simplex fields, a, b,
c and d, encoded by third order maximal simplexes, σa = (v1, v2, v6), σc = (v2, v3, v5), etc. (B) If the 2D simplexes
are discarded and their 1D faces are retained, then the second-order simplex fields are produced, shown here as
overlapping shaded regions. The original simplex fields a, b, c and d are now represented by the coactivity of three
pairs, e.g., a is represented by σa,1 = (v1, v2), σa,2 = (v2, v6) and σa,3 = (v1, v6). (C). The three place fields on the left
exhibit pairwise, but not triple overlap. In the generic spatial configuration shown on the right, pairwise overlapping
place fields also produce a triple overlap. (D) Four pairwise overlapping convex regions in 2D produce all the higher
order (triple and quadruple) overlaps.
19
FIG. S5: Simplicial complexes T0(θ) constructed via pairwise coactivity selection (Method I) for four different
threshold values. (A) The appearance rates of the maximal simplexes, arranged according to their dimension, demon-
strate remarkably tight, graph-like distribution. The color of the dots corresponds to the dimension of the simplexes,
as indicated by the colorbar on the right. (B) Spatial distribution of the dimensionalities of the selected simplexes.
(C) Spatial distribution of the appearance rates of the selected simplexes. (D) The histograms of the lapse times, fit
to double exponential distribution, and the value of the fitted distribution's parameter β. (E) Spatial projections of the
2D skeletons of T0(θ). Data for all panels is computed for a specific place field map for illustrative purposes.
20
FIG. S6: The selected simplicial complexes T0(n0) constructed via the closed-neighbor selection algorithm
(Method II) for four different values of n0. (A) The appearance rates of simplexes, arranged according to their
dimensions, color-coded as indicated by the colorbar on the right. (B) Spatial distribution of the dimensionalities of
the selected simplexes. (C) Spatial distribution of the appearance rates of the selected simplexes. (D) The histograms
of the lapse times, fit to double exponential distribution, and the value of the fitted distribution's parameter β. (E)
Spatial projections of the 2D skeletons of T0(n0).
VIII. SUPPLEMENTARY TABLES
21
θ = 0.07 Hz
θ = 0.05 Hz
θ = 0.04 Hz
θ = 0.1 Hz
1 139 509 569 1 171 393 410 2 198 324 214 5 131 108 15
1 126 409 480 1 175 353 286 1 203 247 144 7 147 71 17
1 156 580 853 2 211 518 571 2 233 330 419 6 139 152 41
1 135 407 551 1 176 369 333 1 201 296 152
7 139 50 7
1 149 422 585 2 172 342 453 2 185 312 320 13 122 107 73
1 130 485 775 1 186 447 392 1 190 335 250 4 164 84 27
1 106 459 655 1 202 352 603 1 177 395 353 8 179 117 25
1 120 519 787 1 194 397 449 1 215 333 237 8 160 109 63
1 131 441 757 1 165 459 499 1 203 339 197 8 149 65 15
1 160 364 575 1 226 313 368 1 175 270 222 8 117 92 14
TABLE S1: Topological signature of the complex selected by simplex rate thresholding. Each cell contains a list
of four Betti numbers, (b0, b1, b2, b3). For the low rate, fσ = 0.04 Hz, T0(θ) encodes the correct spatial connectedness
of the environment (b0 = 1). For intermediate rates, 0.05 ≤ fσ ≤ 0.07 Hz, T0(θ) may occasionally break into two
pieces (b0 = 2) and for fσ ≥ 0.10 Hz and higher, T0(θ) fragments into multiple components. In higher dimensions
D ≥ 1, T0(θ) contains over a hundred noncontractible topological loops.
A. Original
B. Corrected
θ = 0.05 θ = 0.07 θ = 0.1 θ = 0.14 θ = 0.05 θ = 0.07 θ = 0.1 θ = 0.14
1 1 0 0 1 0 0 0 3 0 0 0
1 1 0 0
1 1 0 0 1 0 0 0 2 0 0 0
1 2 0 0
1 1 0 0 1 0 0 0 1 0 0 0
1 2 0 0
1 1 0 0 1 0 0 0 1 0 0 0
1 1 0 0
1 1 0 0 1 0 0 0 3 0 0 0
1 1 0 0
1 1 0 0 1 0 0 0 1 0 0 0
1 3 0 0
1 1 0 0 2 0 0 0 3 0 0 0
2 1 0 0
1 1 0 0 1 0 0 1 2 0 0 0
1 2 0 0
1 1 1 0 1 1 0 0 4 0 0 0
1 3 0 0
1 1 0 0 2 0 0 0 3 0 0 0
1 3 0 0
2 1 0 0 5 1 0 0 3 0 0 0
1 5 0 0 7 2 0 0 7 1 0 0
2 1 0 0 6 2 0 0 7 2 0 0
1 2 0 0 7 7 0 0 7 1 0 0
2 3 0 0 13 1 0 0 8 1 0 0
1 0 0 0 4 6 0 0 11 0 0 0
1 5 0 0 8 1 0 0 7 5 0 0
1 2 0 0 8 2 0 0 9 1 0 0
1 4 0 0 8 2 0 0 9 2 0 0
1 5 0 0 8 2 0 0 7 6 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
1 1 0 0
TABLE S2: Topological signature of the complex selected by link-rate thresholding. (A) For the low rate fσ =
0.05 Hz, T0(θ) occasionally produces the correct topological signature (b0 = b1 = 1, bn>1 = 0, shown in bold).
For intermediate rates fσ ∼ 0.07 Hz, T0(θ) may occasionally break into two pieces (b0 = 2) and produce extra
noncontractible loops in 1D. For high thresholds, fσ ≥ 0.10 Hz, T0(θ) fragments into multiple components. However,
the connectivity of T0(θ) in higher dimensions, D ≥ 2, is correct for all cases, which implies that T0(θ) contracts into
2D. (B) After applying the correction algorithms, the selected complexes acquire correct topological signature for
fσ ≤ 0.07 for all maps. The corresponding learning times Tmin are listed in Suppl. Table 4.
A. Original
B. Corrected
22
n0 = 2 n0 = 5
29 0 1 31 0 0 1 11 0 0 1 4 1 0
29 0 1 30 0 0 1 16 0 0 1 3 0 0
32 0 1 28 0 0 1 19 0 0 1 3 0 0
31 0 1 31 0 0 1 14 0 0 1 2 0 0
31 0 2 34 0 0 1 20 0 0 1 4 0 0
34 0 1 26 0 0 1 12 0 0 1 2 0 0
28 0 1 41 0 0 1 11 0 0 1 2 0 0
28 0 1 40 0 0 1 15 0 0 1 1 1 0
31 0 1 32 0 0 1 17 0 0 1 4 0 0
43 0 1 33 0 0 1 16 4 0 1 4 0 0
n0 = 7 n0 = 12 n0 = 2 n0 = 5 n0 = 7 n0 = 12
1 1 3 0 1 1 1 0 1 1 1 1
1 2 0 0 1 1 1 0 1 1 1 0
1 2 0 0 1 1 0 1 1 1 1 0
1 4 0 0 1 1 1 0 1 1 1 0
1 2 0 0 1 1 1 0 1 2 2 1
1 5 0 0 1 1 1 0 1 2 0 1
1 2 0 0 1 2 1 0 1 2 0 0
1 1 0 0 1 1 2 0 1 1 0 2
1 2 0 0 1 3 1 0 1 1 0 0
1 2 0 0 1 1 5 0 1 2 0 0
1 2
1 1
1 4
1 3
1 3
1 2
1 4
1 3
1 3
1 3
TABLE S3: Topological signatures of the complexes selected by the neighbor-selection algorithm. (A) If but one
pair of closest vertexes is selected n0 = 2, T0(n0) breaks into multiple components. For n0 ≥ 5, T0(n0) has only one
component, but path connectivity is compromised (b0 = 1, b1 (cid:29) 1). In higher dimensions, T0(n0) is contractible,
bn>1 = 0. (B) After applying the correction algorithms, the selected complexes for almost all maps acquire correct
topological signature in 1D and 2D (shown in boldface) for n0 ≥ 7. The corresponding learning times, Tmin, are listed
in Suppl. Table 4.
7
8
9
5
6
2
3
4
T0
1
10 mean/std
4.4 2.7 2.3 2.7 2.8 2.7 2.1 3.8 10.7 2.7 3.7/2.5
T0(θ = .05) 3.8 2.7 1.9 2.7 3.6 3.8 3.8 3.8 2.5 3.7 3.2/0.7
T0(θ = .07) 2.1 3.8 3.8 2.8 2.8 2.3 2.4 3.7 2.5 3.8 3.0/0.7
T0(n0 = 7) 1.9 4.6 2.4 2.7 2.8 3.8 ∞ 3.8 ∞ 3.8 3.2/0.9
T0(n0 = 12) 3.8 2.7 1.9 2.7 ∞ ∞ ∞ 3.8 3.8 ∞ 3.1/0.8
TABLE S4: The learning times, Tmin (in minutes) computed for the selected simplicial complexes T0(θ) and T0(n0)
are similar to the learning times computed via the full temporal nerve complex T . Thus, the information about
the topological structure of the environment emerges from the cell assembly activity as fast as from the entire pool
of place cell coactivities. However, note that the complex T0(n0) sometimes fails to produce a finite leaning time;
non-convergent cases are marked by ∞.
IX. SUPPLEMENTARY MOVIES ILLUSTRATING THE FIRST OF THE TEN TESTED MAPS
Suppl. Movie 1. The grey dots represent centers of the place fields, viewed from above, similar to
Figure 3B. The centers of the place fields that correspond to the coactive place cells are shown in red. The
resulting activity packet moves in the environment following the simulated rat's trajectory.
Suppl. Movie 2. Selection of the cell assemblies by Method I (θ = 100) and assigning readout neurons
to the cell assemblies.
Suppl. Movie 3. A side projection view of the activity packet propagating in the cell assembly network,
selected via Method I. To emphasize that the coactive place cell combinations comprise a cell assembly
complex T0(θ), the corresponding place field centers are schematically connected to the readout neurons.
Suppl. Movie 4. The same system shown in the same projection as the Figure 3B and Suppl. Movie 1.
|
1812.11937 | 1 | 1812 | 2018-12-31T18:18:46 | Two "correlation games" for a nonlinear network with Hebbian excitatory neurons and anti-Hebbian inhibitory neurons | [
"q-bio.NC",
"cs.NE"
] | A companion paper introduces a nonlinear network with Hebbian excitatory (E) neurons that are reciprocally coupled with anti-Hebbian inhibitory (I) neurons and also receive Hebbian feedforward excitation from sensory (S) afferents. The present paper derives the network from two normative principles that are mathematically equivalent but conceptually different. The first principle formulates unsupervised learning as a constrained optimization problem: maximization of S-E correlations subject to a copositivity constraint on E-E correlations. A combination of Legendre and Lagrangian duality yields a zero-sum continuous game between excitatory and inhibitory connections that is solved by the neural network. The second principle defines a zero-sum game between E and I cells. E cells want to maximize S-E correlations and minimize E-I correlations, while I cells want to maximize I-E correlations and minimize power. The conflict between I and E objectives effectively forces the E cells to decorrelate from each other, although only incompletely. Legendre duality yields the neural network. | q-bio.NC | q-bio |
Two "correlation games" for a nonlinear network
with Hebbian excitatory neurons and
anti-Hebbian inhibitory neurons
Neuroscience Institute and Computer Science Dept.
H. Sebastian Seung
Princeton University
Princeton, NJ 08544
June 13, 2021
Abstract
A companion paper introduces a nonlinear network with Hebbian excitatory
(E) neurons that are reciprocally coupled with anti-Hebbian inhibitory (I) neurons
and also receive Hebbian feedforward excitation from sensory (S) afferents. The
present paper derives the network from two normative principles that are mathe-
matically equivalent but conceptually different. The first principle formulates un-
supervised learning as a constrained optimization problem: maximization of S−E
correlations subject to a copositivity constraint on E − E correlations. A com-
bination of Legendre and Lagrangian duality yields a zero-sum continuous game
between excitatory and inhibitory connections that is solved by the neural network.
The second principle defines a zero-sum game between E and I cells. E cells want
to maximize S−E correlations and minimize E −I correlations, while I cells want
to maximize I − E correlations and minimize power. The conflict between I and
E objectives effectively forces the E cells to decorrelate from each other, although
only incompletely. Legendre duality yields the neural network.
A companion paper [Seung, 2018] introduces a nonlinear neural network for unsuper-
vised learning in which a population of excitatory (E) neurons is reciprocally connected
with a population of inhibitory (I) neurons (Fig. 1, left). The E neurons also receive
feedforward excitation from a population of sensory (S) afferents. The reciprocal E − I
connections allow the E neurons to inhibit each other through disynaptic E → I → E
pathways mediated by I neurons. This network motif will be called "disynaptic re-
current inhibition," or just "disynaptic inhibition." Excitatory connections (S → E and
E → I) are modified by Hebbian plasticity, and inhibitory connections (I → E) by anti-
Hebbian plasticity.
The companion paper investigates the computational properties of the neural net-
work through mathematical analysis and numerical simulations. It would be helpful to
attain a deeper understanding of the network as arising from some normative principle
of unsupervised learning. As emphasized by Pehlevan and Chklovskii [2015], the long
1
Figure 1: Disynaptic inhibition versus direct all-to-all inhibition. Green arrows indi-
cate excitatory connections, which change via Hebbian plasticity. Red arrows indicate
inhibitory connections (note minus sign), which change via anti-Hebbian plasticity.
(Left) E neurons are reciprocally coupled with I neurons, and receive input from S af-
ferents. E → I connections (A) and I → E connections (−A(cid:62)) have equal but opposite
strengths. (Right) The E and I populations collapse to a single population of neurons
directly connected by all-to-all inhibition (−L).
history of neural network models based on Hebbian and anti-Hebbian plasticity has
primarily relied on numerical simulations to demonstrate interesting self-organizing
behaviors, and lacked normative principles that can be mathematically formalized as
optimizations of some objective function. Normative principles are not only helpful
for understanding, but also have the practical consequence of suggesting optimization
algorithms other than neural networks, which could be useful in certain settings.
The companion paper takes one step towards a normative principle by "deriving"
the network with disynaptic inhibition (Fig. 1, left) as an approximation to a network
with all-to-all inhibition (Fig. 1, right). The latter network contains a single population
of neurons that directly inhibit each other and also receive feedforward excitation from
sensory afferents, as in the Seung and Zung [2017] variant of the Földiák [1990] model.
The derivation involves a factorized approximation of a Lagrange multiplier matrix,
which is not fully justified in the companion paper. The present paper is intended to
provide the missing theoretical foundation by deriving the network from two normative
principles that are mathematically equivalent but suggest different interpretations.
The first normative principle is a modification of a principle previously used by
Seung and Zung [2017] to derive the network with all-to-all inhibition. They defined
unsupervised learning as the maximization of input-output correlations subject to a
bound constraint on output-output correlations. The bound constraint can be regarded
as enforcing a decorrelated output representation. The principle led to a zero-sum
"correlation game" between excitation and all-to-all inhibition in a nonlinear network
that was a variant of the Földiák [1990] model.1 The inhibitory connections of the
1Zylberberg et al. [2011] claimed that the Földiák [1990] model learns by fitting a linear generative model
to sensory stimuli. The claim is only an approximation at best, and invalid at worst. It is conceptually very
different from Seung and Zung [2017] and Pehlevan et al. [2018], who view excitation and inhibition as
players in a zero-sum game or minimax problem. According to the linear generative model idea, excitation
2
u1u2u3x1x2x3y1y2WA -- ATS cellsE cellsI cellsu1u2u3x1x2x3W -- LS cells network were conjugate to output-output correlations via Lagrangian duality. The ex-
citatory connections were conjugate to input-output correlations via Legendre duality.
A similar approach was taken by Pehlevan et al. [2018], who derived linear networks
performing variants of principal component analysis from the similarity matching prin-
ciple.
For the disynaptic inhibition network, the first normative principle is a constrained
optimization problem: maximization of input-output correlations subject to a coposi-
tivity constraint on output-output correlations. The constraint can be regarded as forc-
ing an imperfect form of decorrelation. The principle leads to a zero-sum game solved
by the disynaptic inhibition network. The S → E connections are conjugate to the
input-output correlations via Legendre duality. The I − E connections are Lagrange
multipliers enforcing the copositivity constraint.
The first normative principle is based on the activities of the E cells only; the activi-
ties of the I cells do not appear until after the duality transforms. The second normative
principle is a zero-sum game between E and I cells. E cells want to maximize S− E
correlations and minimize E − I correlations. I cells want to maximize E − I correla-
tions and minimize power. There is some conflict inherent in the fact that I and E cells
want to drive E − I correlations in opposite directions. The compromise is that E cells
tend to decorrelate from each other. Both excitatory and inhibitory connections arise
from Legendre duality.
A theatrical metaphor describes the difference between the normative principles.
The first normative principle places the E cells center stage while the I cells lurk back-
stage. In the second normative principle, E and I cells are both on stage; E cells are
leading actors while I cells are supporting actors. No connections appear in either
normative principle; the network only emerges after duality transforms.
Both normative principles are mathematically equivalent to each other, but they
invite different viewpoints. The first normative principle is simple because it contains
E cell activity only, but complex because copositivity is not a terribly intuitive idea. The
second normative principle is complex because it contains both E and I cell activity,
but simple because there is no need for copositivity.
The second normative principle is similar to two algorithms for dimensionality
reduction (hard-thresholding of covariance eigenvalues and equalizing thresholded co-
variance eigenvalues) defined by Pehlevan and Chklovskii [2015]. The neural network
implementation of the algorithms involved separate populations of principal neurons
and interneurons, but the neurons did not obey Dale's Law and so could not be inter-
preted as excitatory and inhibitory neurons.
and inhibition minimize reconstruction error.
3
1 Network model with disynaptic inhibition
(cid:33)(cid:35)+
The disynaptic inhibition network (Fig. 1, left) has the activity dynamics,
(cid:34)
(cid:32) m
xi :=
(1− dt)xi + dt λ−1
i
Wiaua − r
∑
α=1
∑
a=1
yα =
n
∑
i=1
Aαixi
yαAαi
(1)
(2)
Here dt is a step size parameter, which can be set at a small constant value or adjusted
adaptively. The activation function [z]+ = max{z,0} is half-wave rectification. After
the activities converge to a steady state, update the connection matrices via
∆Wia ∝ xiua − γWia − κ ∑
∆Aα j ∝ yαx j −(cid:0)q2 − p2(cid:1)Aα j − p2∑
Wib
b
i
Aαi
(3)
(4)
where γ > 0, κ > 0, and q2 > p2 . After the updates (3) and (4), any negative elements
of W and A are zeroed to maintain nonnegativity. The divisive factor λi > 0 in Eq. (1)
is updated via
(5)
Intuitions behind the model definitions are explained in the companion paper [Seung,
2018]. The goal of the present paper is show how the network can be interpreted as a
method of solving a zero-sum game.
∆λi ∝ x2
i − q2
2 Correlation game between connections
2.1 Formulation as constrained optimization
The first normative principle concerns transformation of a sequence of input vectors
u(1), . . . ,u(T ) into a sequence of output vectors x(1), . . . ,x(T ). Both input and output
are assumed nonnegative. Define the input matrix U = [u(1), . . . ,u (T )] as the matrix
containing input vectors u(t) as its columns. The element Uat is the ath component
of u(t). Similarly, define the output matrix X = [x (1) , . . . ,x (T )] as containing output
vectors x(t) as its columns. Define the output-input correlation matrix is
XU(cid:62)
T =
1
T
T
∑
t=1
(cid:62)
x (t)u (t)
Its ia element is the time average of xiua, or (cid:104)xiua(cid:105). Similarly, define the output-output
correlation matrix
XX(cid:62)
T =
1
T
T
∑
t=1
(cid:62)
x (t)x (t)
4
Its i j element is the time average of xix j, or (cid:104)xix j(cid:105). Note that "correlation matrix" is
used to mean second moment matrix rather than covariance matrix. In other words,
the correlation matrix does not involve subtraction of mean values. This is natural
for sparse nonnegative variables, but covariance matrices may be substituted in other
settings.
Problem 1 (Constrained optimization). Define the goal of unsupervised learning as
the constrained optimization
subject to copositivity of D− XX(cid:62)
max
X≥0
(6)
where D is a fixed matrix and Φ∗ is a scalar-valued function that is assumed monotone
nondecreasing as a function of every element of its matrix-valued argument.
T
T
Φ∗(cid:18) XU(cid:62)
(cid:19)
Monotonicity is an important assumption because it allows us to interpret the ob-
jective of Eq. (6) as maximization of input-output correlations.
2.2 Copositivity vs. nonnegativity
Seung and Zung [2017] introduced the principle
Φ∗(cid:18) XU(cid:62)
(cid:19)
T
max
X≥0
subject to nonnegativity of D− XX(cid:62)
T
(7)
which differs from Eq.
(6) only by the substitution of "nonnegativity" for "copos-
itivity." (Here nonnegativity of a matrix is defined to mean nonnegativity of all its
elements.) While the formalisms here are valid for arbitrary D, a convenient choice is
to set diagonal elements of D to be q2 and off-diagonal elements of D to be p2,
(cid:40)
Di j =
q2,
p2,
i = j,
i (cid:54)= j
(8)
If p is much smaller than q, the nonnegativity constraint (cid:104)xix j(cid:105) ≤ Di j in Eq. (7) amounts
to decorrelation.
A symmetric matrix S is said to be copositive when v(cid:62)Sv ≥ 0 for every nonnegative
vector v ≥ 0. This constraint is analogous to positive semidefiniteness but is more
complex because it cannot be reduced to a single eigenvalue constraint. Hahnloser et al.
[2003] give sufficient and necessary conditions for copositivity involving eigenvalues
of submatrices.
Nonnegativity of S is a sufficient condition for copositivity of S, but it is not a neces-
sary condition. In particular, copositivity of S−XX(cid:62)/T does not require nonnegativity,
so a solution of Problem 1 may have (cid:104)xix j(cid:105) > Di j for some i and j.
A necessary condition for copositivity of S is nonnegativity of its diagonal ele-
ments, since e(cid:62)
i Sei < 0 if Sii < 0 where e1, . . . ,en denotes the standard basis for Rn.
In particular copositivity of S = D− XX(cid:62)/T requires that (cid:104)x2
i (cid:105) ≤ Dii for all i. These
inequalities will be called "power constraints," because they limit the power in the
outputs.
5
If either of the diagonal elements Sii and S j j vanish, then a necessary condition
for copositivity is nonnegativity of the off-diagonal element Si j. Therefore (cid:104)xix j(cid:105) may
exceed Di j in a solution of Problem 1 only if the power constraints for i and j are not
saturated.
2.3 Correlation game from Legendre-Lagrangian duality
The copositivity constraint in Eq. (6) can be enforced by introducing Lagrange multi-
pliers A and Λ,
(cid:26)
Φ∗(cid:18) XU(cid:62)
(cid:19)
T
(cid:19)
(cid:18)
D− XX(cid:62)
T
+
1
2
TrA
AT +
1
2
TrΛ
(cid:19)(cid:27)
(cid:18)
D− XX(cid:62)
T
(9)
max
X≥0
min
A,Λ≥0
The Lagrange multiplier A is a nonnegative r × n matrix. The outer maximum must
choose X so that D− XX(cid:62)/T is copositive because otherwise the minimum with re-
spect to A is −∞. The Lagrange multiplier Λ = diag{λ1, . . . ,λn} is a nonnegative
diagonal matrix. The outer maximum must choose X so that the diagonal elements of
D− XX(cid:62)/T are nonnegative because otherwise the minimum with respect to Λ is −∞.
As mentioned above, copositivity of D − XX(cid:62)/T by itself already implies that
the diagonal elements are nonnegative. It follows that the Lagrange multiplier Λ is
redundant for the primal problem, though it does affect the dual problem. Similarly,
adding extra rows to the Lagrange multiplier A does not change the primal problem. For
enforcing the copositivity constraint, it would be sufficient for A to be 1× n. However,
making r > 1 does affect the dual problem.
Problem 2 (Game between cells and connections). Switching the order of min and
max in Eq. (1) yields the dual problem,
(cid:26)
Φ∗(cid:18) XU(cid:62)
(cid:19)
T
(cid:18)
D− XX(cid:62)
T
(cid:19)(cid:16)
+
1
2
Tr
min
A,Λ≥0
max
X≥0
(cid:17)(cid:27)
A(cid:62)A + Λ
(10)
This is an upper bound for Eq. (9) by the minimax inequality.
At this point, it is convenient to define the objective function Φ∗ as the convex
conjugate (Legendre-Fenchel transform) of a function Φ,
(cid:40)
∑
ia
(cid:41)
Φ∗(C) = max
W≥0
WiaCia − Φ (W )
.
(11)
(11) guarantees that Φ∗(C) is monotone
The nonnegativity constraint on W in Eq.
nondecreasing as a function of every element of C. The function Φ can be interpreted
as a regularizer or prior for the weight matrix W .
With Legendre duality, a maximization with respect to W is implicit in Eq. (11).
Switching the order of W and X maximizations yields the following equivalent prob-
lem.
6
Problem 3 (Game between connections). The Lagrangian dual of the constrained op-
timization in Problem 1 is
min
A,Λ≥0
max
W≥0
R(W,A(cid:62)A + Λ)
with payoff function defined by
(cid:26) 1
T
R(W,L) = max
X≥0
TrW(cid:62)XU(cid:62) − Φ(W ) +
(cid:19)
(cid:18)
D− XX(cid:62)
T
(cid:27)
L
1
2
Tr
(12)
(13)
The min-max problem can be interpreted as a zero-sum game between W on the one
hand and A and Λ on the other.
Problem 3 is closely related to the correlation game previously introduced by Seung
and Zung [2017],
(14)
Problem 3 constrains L = A(cid:62)A + Λ for some nonnegative A and Λ, so it is an upper
bound for Eq. (14). This is the mathematical interpretation of choosing a parametrized
form A(cid:62)A + Λ for the Lagrange multiplier L, as was done by Seung [2018].
R(W,L)
max
W≥0
min
L≥0
The network model of Section 1 follows by setting
Φ(W ) =
γ
2 ∑
ia
W 2
ia +
κ
2 ∑
i
(cid:19)2
(cid:18)
∑
a
Wia
in Eq. (13) and applying online projected gradient ascent to perform the maximizations
in Eq. (12) and online projected gradient descent to perform the minimizations. For a
more general choice of Φ(W ), Eq. (3) should be replaced by
∆Wia ∝ xiua − ∂Φ
∂Wia
3 Correlation game between cells
The second normative principle concerns transformation of a sequence of nonnega-
tive input vectors u(1), . . . ,u(T ) into two sequences of nonnegative output vectors
x(1), . . . ,x(T ) and y(1), . . . ,y(T ). Define the input matrix U = [u(1), . . . ,u (T )] and
the two output matrices X = [x (1) , . . . ,x (T )] and Y = [y (1) , . . . ,y (T )].
Problem 4 (Game between cells). Define the goal of unsupervised learning as the
zero-sum game between X and Y
(cid:26)
Φ∗(cid:18) XU(cid:62)
(cid:19)
− Ψ∗(cid:18)Y X(cid:62)
(cid:19)
(cid:27)
1
2
YY(cid:62)
T
min
Y≥0
max
X≥0
(15)
where Φ∗ and Ψ∗ are scalar-valued functions assumed monotone nondecreasing as a
function of every element of their matrix-valued arguments.
Tr
+
T
T
7
Note that only nonnegativity constraints remain in Problem 4; the copositivity con-
straint of Problem 1 is completely hidden. This correlation game can be interpreted as
follows. The E cells would like to maximize S− E correlations (make Φ∗ large) and
minimize E −I correlations (make Ψ∗ small). The I cells would like to maximize I−E
correlations (make Ψ∗ large) and minimize power (make TrYY(cid:62)/T small). There is
conflict between the E and I cells because E cells would like to minimize E − I cor-
relations while I cells would like to maximize them. The compromise is that E cells
incompletely decorrelate from each other.
Problem 4 is equivalent to Problem 1 given the definition of
Ψ∗(C) = max
A≥0
as the Legendre transform of
(cid:41)
AαiCαi − Ψ(A)
(cid:40)
∑
iα
Ψ(A) =
TrADA(cid:62)
1
2
(16)
(17)
(cid:27)
(cid:26)
Φ∗(cid:18) XU(cid:62)
(cid:19)
(cid:26)
Φ∗(cid:18) XU(cid:62)
T
T
(cid:19)
Proof. Substituting the definitions of Eqs. (16) and (17) into Eq. (15) yields
max
X≥0
min
A,Y≥0
− 1
T
TrA(cid:62)Y X(cid:62) +
TrADA(cid:62) +
1
2
1
2
Tr
YY(cid:62)
T
This is minimized when Y = AX, attaining the value
max
X≥0
min
A≥0
− 1
2T
TrA(cid:62)AXX(cid:62) +
1
2
TrADA(cid:62)(cid:27)
This is identical to Eq. (9), except for the omission of the Lagrange multiplier Λ which,
as mentioned previously, is redundant in the primal problem.
4 Discussion
The second normative principle is also interesting because it can be generalized to
include E − E and I − I connections. This will be the subject of future work.
Acknowledgments
The author is grateful for helpful discussions with J. Zung, C. Pehlevan and D. Chklovskii.
The research was supported in part by the Intelligence Advanced Research Projects
Activity (IARPA) via DoI/IBC contract number D16PC0005, and by the National In-
stitutes of Health via U19 NS104648 and U01 NS090562.
8
References
Peter Földiák. Forming sparse representations by local anti-hebbian learning. Biologi-
cal cybernetics, 64(2):165 -- 170, 1990.
Richard HR Hahnloser, H Sebastian Seung, and Jean-Jacques Slotine. Permitted and
forbidden sets in symmetric threshold-linear networks. Neural computation, 15(3):
621 -- 638, 2003.
Cengiz Pehlevan and Dmitri Chklovskii. A normative theory of adaptive dimension-
ality reduction in neural networks. In Advances in neural information processing
systems, pages 2269 -- 2277, 2015.
Cengiz Pehlevan, Anirvan M Sengupta, and Dmitri B Chklovskii. Why do similarity
matching objectives lead to hebbian/anti-hebbian networks? Neural computation,
30(1):84 -- 124, 2018.
H Sebastian Seung. Unsupervised learning by a nonlinear network with hebbian exci-
tatory and anti-hebbian inhibitory neurons. arXiv, 2018.
H Sebastian Seung and Jonathan Zung. A correlation game for unsupervised learning
yields computational interpretations of hebbian excitation, anti-hebbian inhibition,
and synapse elimination. arXiv preprint arXiv:1704.00646, 2017.
Joel Zylberberg, Jason Timothy Murphy, and Michael Robert DeWeese. A sparse
coding model with synaptically local plasticity and spiking neurons can account for
the diverse shapes of v1 simple cell receptive fields. PLoS Comput Biol, 7(10):
e1002250, 2011.
9
|
1102.0566 | 2 | 1102 | 2011-07-20T15:57:22 | Model Cortical Association Fields Account for the Time Course and Dependence on Target Complexity of Human Contour Perception | [
"q-bio.NC"
] | Can lateral connectivity in the primary visual cortex account for the time dependence and intrinsic task difficulty of human contour detection? To answer this question, we created a synthetic image set that prevents sole reliance on either low-level visual features or high-level context for the detection of target objects. Rendered images consist of smoothly varying, globally aligned contour fragments (amoebas) distributed among groups of randomly rotated fragments (clutter). The time course and accuracy of amoeba detection by humans was measured using a two-alternative forced choice protocol with self-reported confidence and variable image presentation time (20-200 ms), followed by an image mask optimized so as to interrupt visual processing. Measured psychometric functions were well fit by sigmoidal functions with exponential time constants of 30-91 ms, depending on amoeba complexity. Key aspects of the psychophysical experiments were accounted for by a computational network model, in which simulated responses across retinotopic arrays of orientation-selective elements were modulated by cortical association fields, represented as multiplicative kernels computed from the differences in pairwise edge statistics between target and distractor images. Comparing the experimental and the computational results suggests that each iteration of the lateral interactions takes at least 37.5 ms of cortical processing time. Our results provide evidence that cortical association fields between orientation selective elements in early visual areas can account for important temporal and task-dependent aspects of the psychometric curves characterizing human contour perception, with the remaining discrepancies postulated to arise from the influence of higher cortical areas. | q-bio.NC | q-bio |
Model Cortical Association Fields Account for the Time Course and Dependence on
Target Complexity of Human Contour Perception
Vadas Gintautas,1, 2, ∗ Michael I. Ham,3 Benjamin Kunsberg,4 Shawn Barr,4 Steven P. Brumby,5 Craig
Rasmussen,4 John S. George,6 Ilya Nemenman,7 Lu´ıs M. A. Bettencourt,1 and Garrett T. Kenyon4, †
1Center for Nonlinear Studies and T-5, Theoretical Division,
Los Alamos National Laboratory, Los Alamos, New Mexico, USA
2Physics Department, Chatham University, Pittsburgh, Pennsylvania, USA
3P-21, Physics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA
4New Mexico Consortium, Los Alamos, New Mexico, USA
5Space and Remote Sensing Sciences, Los Alamos National Laboratory, Los Alamos, New Mexico, USA
6P-21 Applied Modern Physics (Biological and Quantum Physics),
Los Alamos National Laboratory, Los Alamos, New Mexico, USA
7Departments of Physics and Biology and Computational and Life Sciences Initiative, Emory University, Atlanta, Georgia, USA
(Dated: November 5, 2018)
Abstract
Author Summary
Can lateral connectivity in the primary visual cortex
account for the time dependence and intrinsic task diffi-
culty of human contour detection? To answer this ques-
tion, we created a synthetic image set that prevents sole
reliance on either low-level visual features or high-level
context for the detection of target objects. Rendered
images consist of smoothly varying, globally aligned con-
tour fragments (amoebas) distributed among groups of
randomly rotated fragments (clutter). The time course
and accuracy of amoeba detection by humans was mea-
sured using a two-alternative forced choice protocol with
self-reported confidence and variable image presentation
time (20-200 ms), followed by an image mask optimized
so as to interrupt visual processing. Measured psychome-
tric functions were well fit by sigmoidal functions with
exponential time constants of 30-91 ms, depending on
amoeba complexity. Key aspects of the psychophysical
experiments were accounted for by a computational net-
work model, in which simulated responses across retino-
topic arrays of orientation-selective elements were mod-
ulated by cortical association fields, represented as mul-
tiplicative kernels computed from the differences in pair-
wise edge statistics between target and distractor im-
ages. Comparing the experimental and the computa-
tional results suggests that each iteration of the lateral
interactions takes at least 37.5 ms of cortical processing
time. Our results provide evidence that cortical asso-
ciation fields between orientation selective elements in
early visual areas can account for important temporal
and task-dependent aspects of the psychometric curves
characterizing human contour perception, with the re-
maining discrepancies postulated to arise from the influ-
ence of higher cortical areas.
∗Electronic address: [email protected]
†Electronic address: [email protected]
Current computer vision algorithms reproducing the
feed-forward features of the primate visual pathway still
fall far behind the capabilities of human subjects in de-
tecting objects in cluttered backgrounds. Here we inves-
tigate the possibility that recurrent lateral interactions,
long hypothesized to form cortical association fields, can
account for the dependence of object detection accuracy
on shape complexity and image exposure time. Corti-
cal association fields are thought to aid object detection
by reinforcing global image features that cannot easily
be detected by single neurons in feed-forward models.
Our implementation uses the spatial arrangement, rela-
tive orientation, and continuity of putative contour el-
ements to compute the lateral contextual support. We
designed synthetic images that allowed us to control ob-
ject shape and background clutter while eliminating un-
intentional cues to the presence of an otherwise hidden
target. In contrast, real objects can vary uncontrollably
in shape, are camouflaged to different degrees by back-
ground clutter, and are often associated with non-shape
cues, making results using natural image sets difficult to
interpret. Our computational model of cortical associa-
tion fields matches many aspects of the time course and
object detection accuracy of human subjects on statis-
tically identical synthetic image sets. This implies that
lateral interactions may selectively reinforce smooth ob-
ject global boundaries.
Introduction
The perception of closed contours is fundamental to
object recognition, as revealed by the fact that com-
mon object categories can be rapidly detected in black
and white line drawings in which all shading and lumi-
nance cues have been removed [1]. Cortical association
fields, hypothesized to capture spatial correlations be-
tween local image features via long-range lateral synap-
tic interactions, provide a natural substrate for rapid
contour perception [2]. The link between cortical as-
sociation fields and contour perception has been inves-
tigated through a variety of behavioral, experimental,
and theoretical techniques [3 -- 6]. Psychophysical mea-
surements reveal that the detection of implicit contours,
defined by sequences of Gabor-like elements presented
against randomly oriented backgrounds, becomes more
difficult as the local curvature increases and as the indi-
vidual Gabor elements are spaced further apart or their
alignment is randomly perturbed. This dependence on
proximity and relative orientation implies that, in early
visual areas, cortical association fields are primarily lo-
cal and aligned along smooth trajectories [2, 7, 8]. In re-
lated studies, collinear Gabor patches have been shown to
both increase and decrease the contrast detection thresh-
old of a central Gabor patch in a manner that depends
on the relative timing, orientation and spatial separation
of the flanking elements [9 -- 11], providing further psy-
chophysical evidence that lateral influences act at early
cortical processing stages, although the contribution of
collinear facilitation to contour integration remains con-
troversial [12].
In primary visual cortex (V1), electro-
physiological recordings indicate that the responses to
optimally oriented and positioned stimuli can be facili-
tated by flanking stimuli placed outside the classical re-
ceptive field center [5, 6, 10, 13], although these effects
have also been ascribed to elongated central receptive
fields [14, 15] and facilitation has been attributed to in-
creases in baseline activity [16]. Nonetheless, collinear
facilitation is consistent with anatomical studies indi-
cating that orientation columns are laterally connected
to surrounding columns with similar orientation prefer-
ence [17 -- 19].
Because extensive association fields are present in the
primary visual cortex [17 -- 19], lateral interactions may
be key to discriminating smooth object boundaries at
very fast time scales (of the order of tens of ms), as ob-
served in numerous speed of sight psychophysical exper-
iments [1, 20 -- 23]. Correspondingly, theoretical models
have proposed that V1 cortical association fields can be
described mathematically on the basis of cocircularity,
and that relaxation dynamics based on cocircular asso-
ciation fields can extract global contours by suppress-
ing local variation [24]. Such models are qualitatively
consistent with human judgments as to whether pairs of
short line segments belong to the same or separate con-
tours, with human judgments closely following the pair-
wise statistics of edge segments extracted from natural
scenes [25]. Further, model cortical association fields,
when used to detect implicit contours, can predict key
aspects of human psychophysics, particularly the mea-
sured dependence on the density of foreground elements
relative to background elements [8, 26].
In this paper, we extend the above studies by inves-
tigating whether model cortical association fields can
account not only for dependence of contour perception
on intrinsic task difficulty, a relationship that has been
previously explored [8, 26], but also for the detailed
2
time course of human contour detection, an aspect that
has heretofore not been modeled explicitly, although the
time-dependent influence of lateral interactions has been
determined for several theoretical models [27, 28]. In this
work, we employ multiplicative relaxational dynamics to
estimate the time course of contour detection from a com-
putational model employing optimized kernels. Model re-
sults are then compared to speed-of-sight measurements
from human subjects performing the same contour de-
tection task. To obtain optimized cortical association
fields, we design lateral connectivity patterns using a
novel method that exploits the global statistical prop-
erties of salient contours relative to background clut-
ter. Our procedure, which can be generalized beyond
the present application, can be summarized as follows.
We begin by generating a large training corpus, di-
vided into target and distractor images, from which we
obtain estimates of the pairwise co-occurence probabil-
ity of oriented edges conditioned on the presence or ab-
sence of globally salient contours. From the difference in
these two probability distributions, we construct Object-
Distractor Difference (ODD) kernels, which are then con-
volved with every edge feature to obtain the lateral con-
textual support at each location and orientation across
the entire image. Edge features that receive substantial
contextual support from the surrounding edges are pre-
served, indicating they are likely to belong to a globally
salient contour, whereas edge features receiving minimal
contextual support are suppressed, indicating they are
more likely to be part of the background clutter. The
lateral contextual support is applied in a multiplicative
fashion, so as to prevent the appearance of illusory edges,
and the process is iterated several times, mimicking the
exchange of information along horizontal connections in
the primary visual cortex. Our method is thus intended
to capture the essential computational elements of cor-
tical association fields that are hypothesized to mediate
the pop-out of salient contours against cluttered back-
grounds.
To obtain a large number of training images and to
better isolate the role of cortical association fields linking
low-level visual features, we employ abstract computer-
generated shapes consisting of short, smooth contour seg-
ments that could either be globally aligned to form wig-
gly, nearly closed objects (amoebas), or else randomly
rotated to provide a background of locally indistinguish-
able contour fragments (clutter). Amoeba targets lack
specific semantic content, presumably reducing the in-
fluence of high level cortical areas, such as IT. However,
our computer-generated images would not be expected to
eliminate the contribution to contour perception from ex-
trastriate areas [29 -- 32]. Thus, our model of lateral inter-
actions between orientation-selective neurons is designed
to account for just one of several cortical mechanisms
that likely contribute to contour perception.
Our amoeba/no-amoeba image set differs from stim-
uli used in previous psychophysical experiments that
employed sequences of Gabor-like elements to repre-
sent salient contours against randomly oriented back-
grounds [2, 7, 8]. An advantage of contours represented
by random Gabor fields is that the target and distrac-
tor Gabor elements can be distributed at approximately
equal densities, thereby precluding the use of local den-
sity operators as surrogates for global contour percep-
tion [2]. However, our amoeba/no-amoeba image set
is more akin to the natural image sets used in previ-
ous speed-of-sight object detection tasks [33], particu-
larly with respect to studies employing line drawings de-
rived from natural scenes [1]. Humans can detect closed
contours, whether defined by aligned Gabor elements or
by continuous line fragments, in less than 200 ms [1, 20],
which is shorter than the mean interval between sac-
cadic eye movements [34], thus mitigating the contribu-
tion from visual search. Like Gabor defined contours,
our amoeba/no-amoeba image set implements a pop-out
detection task involving readily perceived target shapes
whose complexity can be controlled parametrically.
To benchmark the accuracy and the time course of the
ODD kernel-based procedure applied to the amoeba/no-
amoeba task, we compare our model results to the per-
formance of human subjects on a 2AFC speed-of-sight
task in which amoeba/no-amoeba images are presented
very briefly side by side, followed by a mask designed
to limit the time the visual system is able to process
the sensory input [1, 20 -- 23]. Since it takes an estimated
100− 300 ms for activation to spread through the ventral
stream of the visual cortex [21], an effective mask pre-
sented within this time frame can potentially degrade ob-
ject detection performance by interfering with the neural
processing mechanisms underlying recognition [22, 35].
By plotting task performance as a function of the stimu-
lus onset asynchrony (SOA) -- the interval between image
and mask presentation onsets -- the resulting psychometric
curves are hypothesized to estimate the neural processing
time required to reach a given level of classification ac-
curacy. Amoeba targets of low to moderate complexity
were found to reliably pop-out against the background
clutter, allowing subjects to achieve near perfect perfor-
mance at SOAs less than 250 ms, even when followed by
an optimized mask consisting of rotated versions of the
target and distractor images [20]. Our model cortical as-
sociation fields were able to account for the dependence
of human performance on amoeba complexity as well
as for aspects of the time course of contour perception
as measured by the improvement in human performance
with increasing SOA. Thus, we present the first network-
level computational model to simultaneously account for
spatial and temporal aspects of contour perception, as
measured in human subjects performing the same con-
tour detection task. Aspects of the experimental data
for which our model fails to account, particularly data
showing that human subjects require longer processing
times to detect more complex targets, may indicate the
possible involvement of extrastriate areas, which may be
essential for the perception of more complex shapes.
3
FIG. 1: Examples of targets and distractors from
the amoeba/no-amoeba image set for different K.
From top to bottom: K = 2, 4, 6, 8. Left column: Targets;
amoeba complexity increases with increasing numbers of ra-
dial frequencies. Clutter was constructed by randomly ro-
tating groups of amoeba contour fragments. Right column:
Distractors; only clutter fragments are present.
Results
To investigate low-level cortical mechanisms for detect-
ing smooth, closed contours presented against cluttered
backgrounds with statistically similar low-level features,
we designed an amoeba/no-amoeba detection task using
a novel set of synthetic images (Figure 1). Amoebas are
radial frequency patterns [36] constructed via superposi-
tion of periodic functions described by a discrete set of
radial frequencies around a circle. In addition, we added
clutter objects, or distractors, that were locally indis-
tinguishable from targets. Both targets and distractors
were composed of short contour fragments, thus elimi-
nating unambiguous indicators of target presence or ab-
sence, such as total line length, the presence of line end-
points, and the existence of short gaps between opposed
line segments. To keep the bounding contours smooth,
only the lowest K radial frequencies were included in the
linear superposition used to construct amoeba targets.
To span the maximum range of contour shapes and sizes,
the amplitude and phase of each radial frequency com-
ponent was chosen randomly, under the restriction that
the minimum and maximum diameters could not exceed
lower and upper limits. When only 2 radial frequencies
were included in the superposition, the resulting amoe-
bas were very smooth. As more radial frequencies were
included, the contours became more complex. Thus, K,
the number of radial frequencies included in the superpo-
sition, provided a control parameter for adjusting target
complexity. Figure 1 shows target and distractor images
generated using different values of K.
lateral
Human subjects are able to infer whether a two iso-
lated line segments extracted from a natural scene are
from the same or from separate contours using only
distance, direction and relative orientation of the two
segments as cues [25, 37]. The performance of human
subjects is well predicted by differences in the empiri-
cally calculated co-occurrence statistics of short line seg-
ments drawn from either the same or from different con-
tours. To explore the ability of cortical association fields
to account for the perception of smooth contours, we
developed a network-level computational model of lat-
eral interactions between orientation-selective elements
governed by sigmoidal (piecewise linear) input/output
synaptic transfer functions. To model
inter-
actions, we constructed "Object-Distractor Difference
(ODD) kernels" for the amoeba/no-amoeba task by com-
puting coactivation statistics for the responses of pairs of
orientation-selective filter elements, compiled separately
for target and distractor images (Figure 2). Because the
amoeba/no-amoeba image set was translationally invari-
ant and isotropic, the central filter element may without
loss of generality be shifted and rotated to a canonical po-
sition and orientation. Thus the canonical ODD kernel
was defined relative to filter elements at the origin with
orientation π/16 (to mitigate aliasing effects). Filter el-
ements located away from the origin can be accounted
for by a trivial translation. To account for filter elements
with different orientations, separate ODD kernels were
computed for 8 orientations then rotated to a common
orientation and averaged to produce a canonical ODD
kernel. The canonical kernel was then rotated in steps
between 0 and π (offset by π/16) and then interpolated
to Cartesian x−y axes by rounding to the nearest integer
coordinates.
The resulting ODD kernels were generally consistent
with the predictions of cocircular constructions [24], ex-
cept that support was mostly limited to line elements
lying along low curvature contours, which follows natu-
rally from the prevalence of low curvatures in our amoeba
training set.
Curiously, the largest differences in the coactivation
statistics occur close to the center of the kernel, where
4
targets and distractors are presumably most similar.
However, even at short distances, amoeba segments are
still more likely to be aligned than clutter elements.
Moreover, nearby pairs occur much more frequently than
more distant pairs, amplifying their contribution to the
difference map. Since, by design, the individual clutter
fragments were locally indistinguishable from the tar-
get fragments, co-occurrence statistics of oriented frag-
ments were necessary to solve the amoeba/no-amoeba
task. The simplest solution, adopted here, was to fo-
cus on pairwise co-occurrences. Notably, in some neural
preparations, pairwise interactions have been shown to
be sufficient to account for a large fraction of all higher-
order correlations [38, 39].
At the retinal stage, target and distractor images were
represented as 256 × 256 pixel monochromatic, binary
line drawings. At the next stage, corresponding to an
early cortical processing area such as V1, a set of filters
was used to represent 8 orientations, uniformly-spaced
and centered at each pixel, with the axes rotated slightly
(by π/16) to mitigate aliasing artifacts. The bottom-
up responses of each orientation-selective element were
computed via linear convolution using filters composed
of a central excitatory subunit flanked by two inhibitory
subunits. Each subunit was an elliptical Gaussian with
an aspect ratio of 7 : 1, consistent with the aspect ra-
tios of V1 simple cell receptive fields measured experi-
mentally [40] and similar to values employed in previ-
ously published models of V1 responses [41]. Likewise,
we estimate that each image pixel subtended a visual an-
gle of approximately 0.025◦ (see Methods), so that each
orientation-selective element in the model subtended a vi-
sual angle of approximately 0.2◦, consistent with physio-
logical estimates of V1 receptive field sizes at small eccen-
tricities [42]. All subunits had the same total integrated
strength (to within a sign), whose magnitude was ad-
justed to yield relatively clean representations of the orig-
inal image in terms of oriented edges. The synaptic trans-
fer function was piecewise-linear with a minimum value
of 0.0 and a maximum value of 1.0 and a fixed threshold
of 0.5. A finite threshold and saturation level were essen-
tial in order to allow non-supported contour fragments
to be suppressed while preventing well-supported frag-
ments from growing without bound. The precise values
used for threshold and saturation were not critical, as
responsiveness was controlled independently by adjust-
ing the overall integrated strength of the bottom-up and
lateral interaction kernels (see Methods).
Orientation-selective responses were modulated by 4
successive applications of the multiplicative ODD kernel.
Lateral support was first computed via linear convolu-
tion of the ODD kernel with the surrounding orientation-
selective elements, out to a radius of 32 pixels. Given
that images were approximately 7◦ × 7◦ in extent (see
Methods), ODD kernels spanned a total visual angle of
approximately 1.75 degrees, roughly in correspondence
with the estimated visuotopic extent of horizontal pro-
jections in V1 [42]. The previous activity of each cell
5
FIG. 2: ODD kernels. Top Row: For a single short line segment oriented approximately horizontally at the center (not
drawn), the co-occurrence-based support of other edges at different relative orientations and spatial locations is depicted. Axes
were rotated by (180◦/16) from vertical to mitigate aliasing effects. The color of each edge was set proportional to its co-
occurrence-based support. The color scale ranges from blue (negative values) to white (zero) to red (positive values). Left
panel: Co-occurrence statistics compiled from 40, 000 target images. Center panel: Co-occurrence statistics compiled from
40, 000 distractor images. Right panel: ODD kernel, given by the difference in co-occurrence statistics between target and
distractor kernels. Bottom Row: Subfields extracted from the middle of the upper left quadrant (as indicated by black boxes in
the top row figures), shown on an expanded scale to better visualize the difference in co-occurrence statistics between target and
distractor images. Alignment of edges in target images is mostly cocircular whereas alignment is mostly random in distractor
images, accounting for the fine structure in the corresponding section of the ODD kernel.
was multiplied by the current lateral support, passed
through the piecewise-linear synaptic transfer function,
and the process repeated for up to 4 iterations. Contour
segments that received insufficient lateral support were
thereby suppressed, whereas strongly supported elements
were either enhanced or remained maximally activated.
When applied to the amoeba/no-amoeba image set, the
ODD kernels typically suppressed clutter relative to tar-
get segments (Figure 3, left column).
When applied in a similar manner to a natural gray-
scale image to which a hard Difference-of-Gaussians
(DoG) filter has been applied to maximally enhance lo-
cal contrast (see Figure 3, right column), ODD-kernels
tended to preserve long, smooth lines while suppressing
local spatial detail. Although ODD kernels were trained
on a narrow set of synthetic images, the results exhibit
some generalization to natural images due to the overlap
between the cocircularity statistics (see Figure 2) of the
synthetic image set and those of natural images.
To quantify the ability of the model to discriminate be-
tween amoeba/no-amoeba target and distractor images,
we used the total activation summed over all orientation-
selective elements after k iterations of the ODD kernel.
A set of 2, 000 target and distractor images was used
for testing; test images were generated independently
from the training images. Histograms of the total ac-
tivation show increasing separability between target and
distractor images as a function of the number of itera-
tions (Figure 4). To maximize the range of shapes and
sizes spanned by our synthetic targets and distractors,
we did not require that the number of ON retinal pixels
be constant across images. Rather, the retinal repre-
sentations of both target and distractor images encom-
passed a broad range of total activity levels, although
the two distributions strongly overlapped and there was
no evident bias favoring one or the other. At the next
processing stage, prior to any lateral interactions, there
was likewise little or no bias evident in the bottom-up
responses of the orientation-selective elements. Each it-
eration of the multiplicative ODD kernel then caused the
distributions of total activity for target and distractor
images to become more separable, implying correspond-
ing improvements in discrimination performance on the
amoeba/no-amoeba task.
The general principles governing the operation of
our model cortical association fields are conceptually
straightforward. ODD kernels, which capture differences
in the coactivation statistics of edge segments belong-
6
ing to amoebas relative to edge segments belonging to
the background clutter, are used to determine the lateral
contextual support for individual edge segments in an
image. Edge segments receiving sufficiently strong sup-
port are preserved, indicating they are likely to be part
of an amoeba, whereas edge segments receiving insuffi-
cient support are suppressed, indicating they are likely
to belong to the background clutter.
To assess the ability of the model cortical association
fields to account for the time course of human contour
perception, we measured the stimulus presentation time
required for human subjects to reach a given level of accu-
racy on an amoeba/no-amoeba task. The psychophysical
experiment was implemented using a speed-of-sight pro-
tocol employing a two-alternative forced choice (2AFC)
design, with subjects using a slider bar to indicate
which of two images, presented side-by-side, contained
an amoeba (Figure 5). The distance the bar was dis-
placed to the left or to the right was used to indicate
confidence, see Methods. To effectively interrupt visual
processing at a given SOA, both target and distractor
images were replaced by an optimized mask, constructed
by combining randomly rotated amoeba and clutter seg-
ments [20]. Our optimized masks were designed to render
the amoeba targets virtually invisible in the fused target-
mask composite.
As a measure of human performance on the
amoeba/no-amoeba task, we constructed receiver oper-
ating characteristic (ROC) curves [43] (Figure 6), using
each subject's reported confidence (slider bar location rel-
ative to the center position) as a noisy signal for estimat-
ing which side, either left or right, contained the target
on a given trial. True positives corresponded to trials
on which the subject reported the target was on the left
(relative to threshold) and the target was actually on the
left (relative to threshold). False positives corresponded
to trials on which the subject reported the target was
on the left whereas the target was actually on the right
(relative to threshold). To construct each ROC curve,
the confidence scale along the slider bar was divided into
6 discrete threshold values. For each threshold value, a
cumulative proportional true positive rate was calculated
by considering only those trials as true positives in which
the confidence value was above threshold. The cumu-
lative proportional false positive rate for each threshold
value was calculated similarly. Each threshold value thus
contributed one point on the ROC curve, with true pos-
itive rate plotted as the ordinate and the false positive
rate as the abscissa. The complete set of points were con-
nected by straight lines to guide the eye (Figure 6), with
a separate ROC curve computed for each combination of
SOA and target complexity.
ROC curves for quantifying the performance of the
model on the amoeba/no-amoeba task were computed
similarly, using the difference in total luminance between
the left and right images as the raw signal for estimat-
ing which side contained the target on a given trial. If
the total luminance of the left image was higher than
FIG. 3: The effect of lateral interactions on exam-
ple images. Left column: black and white amoeba-target
image (K = 4). Right column: Gray-scale natural image
(the standard computer vision test image "Lena") after ap-
plying a hard Difference of Gaussians (DoG) filter to enhance
edges. Top row: Raw retinal input. Second row: Responses of
orientation-selective elements before any lateral interactions
(k = 0). To aid visualization, the activity of the maximally
responding orientation-selective element at each pixel loca-
tion is depicted as a gray-scale intensity. Rows 3-6: Activity
after k = 1, 2, 3, 4 iterations of the multiplicative ODD kernel,
as labeled. For each iteration, activity was multiplied by the
local support, computed via linear convolution of the previ-
ous output activity with the ODD kernel. Lateral interac-
tions tended to support smooth contours, particularly those
arising from amoeba segments, while suppressing clutter or
background detail.
7
that of the right (relative to threshold), the response of
the model would be reported as target on the left. Ide-
ally, after several iterations of the ODD kernel, no seg-
ments would remain in the distractor image and only
amoeba segments would remain in the target image; in
practice, the total luminance served as a measure of con-
fidence. Given the much larger number of trials (1000)
available for assessing model performance, 100 equally
spaced threshold values were used to calculate the cor-
responding ROC curves. As with the ROC curves con-
structed from the confidence values reported by the hu-
man subjects, the ROC curves computed from the confi-
dence values reported by the model give the cumulative
proportional true positive rate as a function of cumula-
tive proportional false positive rate, with the confidence
threshold varied from zero to maximum. Graphically,
the area under the ROC curves is given by the amount
of overlap between the total luminance histograms (see
figure 4) for the target and distractor images [44].
ROC curves for human subjects show performance in-
creasingly above chance, indicated by a diagonal line of
slope 1, as a function of both increasing SOA and de-
creasing target complexity. For amoeba targets of low
to moderate complexity, ROC curves obtained from hu-
man subjects were well matched to those generated by
the model cortical association fields, consistent with the
hypothesis that lateral interactions between orientation-
selective neurons contribute to human contour percep-
tion, at least for simple targets.
The area under the ROC curve (AUC) gives the prob-
ability that a randomly chosen target image will be cor-
rectly classified relative to a randomly chosen distractor
image, and thus provides a threshold-independent assess-
ment of performance on the 2AFC task. Both the aver-
age over human subjects and the model cortical associa-
tion fields exhibited qualitatively similar performance on
the 2AFC amoeba/no-amoeba task (Figure 7). Perfor-
mance declined as a function of increasing target com-
plexity, both for human subjects, measured at a fixed
SOA, and for the model, measured at a fixed number
of iterations, implying that K was an effective control
parameter for adjusting task difficulty. At 20 ms SOA,
the performance of human subjects was indistinguishable
from chance, suggesting that our optimized masks effec-
tively prevented the development of bottom-up cortical
responses, even for the simplest targets (K = 2). Al-
though some studies report that line drawings are pro-
cessed more rapidly than natural images, with above
chance performance being observed at short SOA val-
ues [1, 26], the fact that performance on the amoeba/no-
amoeba task was no better than chance at a 20 ms SOA
implies that our optimized masks effectively interrupted
visual processing of the amoeba targets. Since the model
used here did not include any account for the time course
of bottom-up retinocortical dynamics, we assumed that
the performance of human subjects at 20 ms SOA should
be equated to model performance at 0 iterations (prior to
any lateral interactions), a time frame consistent with the
FIG. 4: Histograms of total luminance in target and
distractor images as a function of the number of it-
erations. Red bins: Total activity histograms for all 1, 000
test target images. Blue bins: Total activity histograms for
all 1, 000 test distractor images. The degree that the two
distributions overlap is shown as the gray shaded area, which
provides a measure of whether total luminance can be used to
distinguish targets from distractors. The percentage in each
shaded area shows the approximate lower bound amount of
overlap of the two histograms, for comparison. Top row: To-
tal summed activity over all retinal pixels. Little, if any bias
between target and distractor images was evident in the in-
put black and white images as there is nearly complete over-
lap between the distributions. Subsequent rows: Total activ-
ity histograms summed over all orientation-selective elements.
Second row: Bottom-up responses prior to any lateral inter-
actions. Third - sixth rows: Total activity histograms after 1
- 4 iterations of the multiplicative ODD kernel, respectively.
Total summed activity became progressively more separable
with additional iterations, as evinced by a decrease in the
overlapping areas.
8
dial frequencies are processed by different cortical chan-
nels [46]. Model performance might have been improved
by training a new set of ODD kernels specifically for tar-
gets containing K ≥ 6 radial frequencies, thereby utiliz-
ing a hypothetical sub-population of orientation-selective
neurons optimized for detecting high-curvature contours.
Here, our model was limited to a single multiplicative
kernel for detecting all predominately smooth contours.
To quantify how average human performance on the
2AFC amoeba/no-amoeba task varied with SOA, and to
compare with the dependence of model performance on
the number of iterations of the ODD kernel, areas un-
der both sets of ROC curves were fit to a monotonically
increasing function of the following sigmoidal form:
f (t) =
F∞
1 − (1 − 2F∞)e−λ(t−t0)
.
(1)
For human experiments, the parameter t corresponds to
the SOA in ms. Since we expect humans to perform close
to 100% accuracy for very long SOA, we set F∞ = 1.
Since humans perform essentially at chance (50%) for 20
ms SOA, we set t0 = 20 ms. Thus λ was the only free
parameter; fits to the average human data were denoted
by λH ; λH has units of 1/ms. Likewise, model perfor-
mance was fit to a curve with the same functional form,
with t = k measuring the number of iterations; λM was
used to denote curve fits to the model data. However,
visual inspection of the model data suggests that its per-
formance saturates at less than 100% accuracy even after
an infinite number of iterations, thus we forced the sig-
moidal curve fit to the model results to asymptote at the
final measured value of AUC: F∞ = AU Ck=4(K). Since
the model performs better than chance after only 1 iter-
ation, we set t0 = 0. For both the human experiments
and the model performance, the functional form of f (t)
ensures that f (t0) = 1
2 , corresponding to a minimal per-
formance equal to chance.
We find that 1/λH and 1/λM behave quite differently
as a function of K, the number of radial frequencies used
in amoeba generation (Figure 7). As anticipated for a re-
laxational process governed by a single kernel, the model
data was well described by a single value of λM (in units
of 1/iterations), equal to 1.26. For the human subjects
data, values of λH increased from 0.034 to 0.011 as a
function of amoeba complexity, corresponding to lateral
processing times of 29.8 to 90.6 ms, respectively. If hu-
man performance depended on only a single set of lateral
connections, then, at least in the linear approximation
case, we might expect human performance to be well de-
scribed by a single dominant time constant, represent-
ing the dominant eigenmode of the horizontal interac-
tions [47, 48]. Multiple time scales in the human perfor-
mance case may emerge from any number of physiological
mechanisms not included in the present model, including
additional non-linearities in the action of the horizontal
connections and/or contributions to contour perception
from extrastriate areas. Our data do not allow us to
make a firm distinction between these possibilities.
FIG. 5: Psychophysical experiment schematic. The
stimulus consisted of one target image and one distractor im-
age (randomly positioned with equal probability on the left
or right), presented simultaneously for an SOA between 20
ms and 200 ms, followed by an optimized 100 ms mask gen-
erated from randomly rotated groups of target and distractor
segments. Subjects indicated which side contained the target
object (amoeba) using a computer mouse to click along a hor-
izontal slider bar. Clicking far to the left or right indicated
strong confidence that the corresponding side contained the
target; clicking close to the center indicated weak confidence.
A narrow gap in the center forced subjects to choose between
left and right.
distribution of the shortest measured response latencies
recorded in primary visual cortex [45].
Overall, average human performance improved as a
function of increasing SOA in a manner analogous to
the improvement in model performance as a function of
the number of iterations of the ODD kernel. This cor-
respondence was especially evident for amoebas of low
to moderate complexity (K ≤ 4 ). For more complex
targets, model performance lagged well behind that of
human subjects. Studies suggest that low and high ra-
9
FIG. 6: ROC curves comparing human and model performance on the amoeba/no amoeba task. Top two rows:
ROC curves averaged over four different human test subjects using reported confidence (points). The dashed diagonal line
in each plot indicates the curve corresponding to chance. Red, blue, green, black correspond to K = 2, 4, 6, 8, respectively.
Bottom two rows: ROC curves for model cortical association fields computed from total activity histograms.
However, one possible interpretation of the present re-
sults is that the perception of simple contours is domi-
nated by relatively fast lateral interactions placed early in
the visual processing pathway, thereby accounting for the
good fit between the model and experimental results for
targets of low to moderate complexity. Building on this
interpretation, we postulate that the perception of more
complex contours requires more extensive, and therefore
slower, processing mechanisms involving higher cortical
areas, thus explaining the discrepancy between model
and experimental performance as target complexity in-
creases. Under the assumption that human perception of
simple amoeba targets (K ≤ 4) depends primarily on re-
current lateral interactions between orientation-selective
neurons, we can estimate the time required for each it-
eration of the multiplicative ODD kernel. This rate is
estimated using the K = 2 time constants from the fits:
λM,K=2/λH = 37.5 ms per iteration, a value consistent
with estimates of lateral conduction delays within the
same cortical area [13].
Having shown that the lateral interactions based on
multiplicative ODD kernels can account for both spa-
tial and temporal aspects human contour perception, we
seek to identify model details that are essential to the
performance reported here. First, we demonstrate that
the proposed model is robust and does not require that
the magnitude of the ODD kernel be carefully titrated
to a precise value. Model performance on the 2AFC
amoeba/no-amoeba task, measured by the area under
the ROC curve (AUC) for increasing numbers of itera-
tions (k = 0, 1, 2, 3, 4), was plotted for different values of
the strength of the ODD kernel, given by the total in-
tegrated strengths of the equal and opposite target and
distractor contributions (Figure 8). The number of radial
frequencies was fixed at K = 4. Qualitatively similar per-
formance was obtained for ODD kernel strengths ranging
from 300 to 400. The ODD kernel used in the present
study, whose strength was set to 325, produced near opti-
mal performance and also exhibited monotonic improve-
ment with increasing numbers of iterations. That perfor-
mance was generally insensitive to the value of the main
free parameter in the model provides strong evidence for
the robustness of the proposed contour detection mecha-
nism based on multiplicative lateral interactions.
A second aspect of the model that merits scrutiny
is the detailed structure of the ODD kernels, which
SOA=20msSOA=40msSOA=80ms01Falsepositiverate01TruepositiverateK=2K=4K=6K=8SOA=120msHumanPerformanceSOA=160msSOA=200msk=0k=1k=201Falsepositiverate01TruepositiverateModelPerformanceK=2K=4K=6K=8k=3k=4 10
FIG. 7: A comparison of human and model performance on the 2AFC amoeba/no amoeba task. Left: Average
human performance for different SOA in milliseconds. Right: Performance of model cortical association fields for increasing
numbers of iterations. Both panels: Accuracy, which is equivalent to area under the ROC curve, (error bars) fitted to single
sigmoidal functions (solid lines). The four curves from top to bottom correspond to K = 2, 4, 6, 8 radial frequencies.
were trained using computer-generated images in which
the pairwise edge statistics uniquely identifying globally
salient contours could be calculated directly. Previous
models of contour perception typically employed much
simpler patterns of lateral connectivity, in which excita-
tory interactions were either collinear or cocircular, and
inhibitory interactions were approximately independent
of relative orientation [8, 24, 27, 47 -- 49]. To determine if
the detailed structure of the ODD kernel was critical to
the observed performance, we repeated the amoeba/no-
amoeba experiment using a much simpler kernel whose
basic form was consistent with a number of previously
published models (see Figure 8). Specifically, we used a
"Bowtie" kernel in which excitatory connections fanned
out with an opening angle of π/6 and the difference in
the preferred orientations of the pre- and post-synaptic
elements differed by no more than ±π/6. Both exci-
tatory and inhibitory connection strengths fell off in a
Gaussian manner, with inhibition strength being insen-
sitive to orientation. Although the overall accuracy of
the Bowtie kernels was lower than that achieved by the
ODD kernels, performance on the amoeba/no-amoeba
tasks was qualitatively similar, particularly regarding the
general monotonic improvement with the number of it-
erations and the absence of a sensitive dependence on
kernel strength. Thus, we conclude that multiplicative
lateral interactions are able to preserve smooth closed
contours while suppressing clutter in a manner that is
robust to broad changes in model details.
Discussion
We have shown that simple models of neural activity
in primary visual cortex, enriched with lateral associa-
tion kernels, reproduce some of the behavioral features
regarding the human perception of broken closed con-
tours. Our results agree not only with the measured de-
pendence on contour complexity but also with the tem-
poral dependence of human perception as a function of
SOA, suggesting that horizontal connections in V1 may
play a non-trivial and global computational role in the
perception of closed contours on very fast timescales.
A number of studies relate to the potential contribu-
tion of cortical association fields to human contour per-
ception; these encompass a range of anatomical, physi-
ological, psychophysical, and theoretical techniques [2 --
5, 7 -- 10, 10, 11, 13, 16 -- 19, 50]. In particular, a number
of theoretical models have sought to account for human
contour perception at the level of biologically-plausible
neural circuits [8, 27, 28, 49, 51 -- 54], with most studies
incorporating some form of cortical association field con-
figured to reinforce smoothness [24]. Although biologi-
cally plausible models of cortical association fields have
been used to account for the dependence of contour vis-
ibility on key parameters controlling task difficulty, such
as smoothness, closure, and density of background clut-
ter [8], model cortical association fields have not been
directly compared to the time course of human contour
perception as a function of contour complexity. Here,
we used cortical association fields based on ODD ker-
nels, which were computed from differences in the pair-
wise coactivation statistics of orientation-selective ele-
ments arising from target as opposed to distractor im-
ages. While we designed the kernels specifically for the
amoeba-clutter disambiguation, we emphasize that the
algorithm for the ODD kernel construction is completely
general and can be used to improve detection of salient
image features in any situation where generative mod-
els of targets and distractors are known, or there exists
data sets of sufficient size to characterize the contour co-
occurrence statistics empirically for both targets and dis-
tractors. In our experiments, ODD kernels were able to
account for the experimentally observed variations in the
saliency of closed contours as a function of parametric
complexity and for the time course with which smooth
contours are processed by cortical circuits. Crucial for
these results was our use of a synthetic target/distractor
data set with controllable complexity and the absence
of top-down contextual features or local cues that might
04080120160200SOA(ms)0.500.751.00AccuracyK=2K=4K=6K=8Human01234IterationsK=2K=4K=6K=8Model give away target presence.
Here, we used a semi-supervised training scheme to
learn lateral connectivity patterns optimized for perform-
ing the amoeba/no-amoeba task. Necessarily, we sought
to model only a subset of the lateral interactions be-
tween orientation-selective neurons, namely, those hori-
zontal connections configured to reinforce smooth, closed
contours. We did not attempt to capture the full range of
spatial relationships between features extracted at early
cortical processing stages [24, 55]. Presently, databases
containing sufficient numbers of fully annotated and seg-
mented natural images needed to reproduce the weeks
(or months) of visual experience required to train the
full complement of horizontal connections in the primary
visual cortex do not exist. Moreover, the computational
resources to exploit such databases, even if they did ex-
ist, are highly non-trivial to assemble. Thus, we focused
here on a subset of horizontal connections for which it
was possible to construct synthetic surrogate images. At
most, the proposed model represents a subset -- and only
a subset -- of the lateral connections between orientation-
selective cortical neurons. Moreover, even a complete set
of such horizontal connections would, at most, represent
but a subset of the cortical mechanisms that contribute
to the time course and shape-dependence of contour per-
ception.
FIG. 8: A comparison of ODD and simpler "Bowtie"
kernel performance on the on the 2AFC, K = 4
amoeba/no amoeba task plotted as a function of the
number of iterations for a range of different kernel
strengths. Line width and marker size denote values on
kernel strength, which was the main free parameter in the
model. Kernel strength is a dimensionless constant. Black
lines: ODD kernel performance. Blue lines: "Bowtie" ker-
nel performance. Qualitative behavior was similar for both
kernels, demonstrating that multiplicative lateral interactions
act robustly to reinforce smooth closed contours.
The supervised training scheme employed here might
be related to perceptual learning phenomena, which take
place over time scales much shorter than those typi-
cally associated with developmental processes [56 -- 58]. It
is possible that known physiological mechanisms, such
11
as spike-timing-dependent plasticity (STDP), especially
with accounts for realistic conduction delays [59], could
mediate a rapid refinement of lateral connections so as
to facilitate the perception of amoeba targets. Moreover,
physiological plasticity mechanisms might produce differ-
ent patterns of connectivity for orientation-selective ele-
ments representing points of low as opposed to high local
curvature, thereby optimizing lateral interactions for con-
tours of varying complexity. Here, we made no attempt
to customize distinct ODD kernels for detecting contours
of varying complexity. Instead, a single ODD kernel was
trained using a complete set of images in which differ-
ent numbers of radial frequency components were equally
represented. Although we did not investigate whether, or
to what extent, the performance of human subjects im-
proved over the course of the amoeba/no-amoeba experi-
ment, such investigations might shed insight into the role
of perceptual learning in the detection of closed contours.
The question of how lateral connectivity based on ODD
kernels might be acquired during development was not
addressed explicitly. In principle, coactivation statistics
between pairs of orientation-selective neurons could be
accumulated over time in an unsupervised manner by
a Hebbian-like learning rule [60]. Under natural view-
ing conditions, we expect that contour fragments consis-
tent with smooth, closed boundaries would tend to occur
simultaneously, whereas contour fragments inconsistent
with object boundaries would tend occur at random tem-
poral delays. Thus, a Hebbian-like learning rule sensitive
to temporal correlations, such as certain mathematical
forms of STDP-like learning rules [61], might under nor-
mal developmental conditions lead to connectivity pat-
terns that reinforce smooth contours.
Of course, human contour perception may have noth-
ing to do with cortical association fields, or lateral in-
teractions may play a subordinate role. Early mod-
els showed how spatial filtering could enhance texture-
defined contours in the absence of orientation-specific
interactions [4] and short-range lateral interactions can
accentuate texture-defined boundaries [31, 62]. How-
ever, psychophysical studies employing implicit con-
tours [2, 7, 8], in which foreground and background ele-
ments are present at equal density and which lack explicit
texture cues, appear to rule out explanations that omit
long-range, orientation-specific interactions. An influen-
tial class of biologically-inspired computer vision models
achieves a degree of viewpoint-invariant object recogni-
tion by constructing feed-forward hierarchies to extract
progressively more complex and viewpoint invariant fea-
tures [33, 63]. By analogy with such models, scale- and
position-independent representations for detecting long,
smooth contours could in principle be constructed hier-
archically, starting with simple edge detectors and build-
ing up progressively longer, more complex curves using
a "bag-of-features" approach. Presently, there appear
to be insufficient data to decide whether human con-
tour perception involves primarily lateral, feed-forward,
or even top-down connections [30, 32, 64]. Hypothet-
01234Iterations0.500.751.00AccuracyODDBowtie30032540075125175 ically, the cortical association fields used in the present
study could have been implemented as a feed-forward ar-
chitecture, using a hierarchy of orientation-selective neu-
rons to link progressively more widely separated contour
fragments. Functionally, there may not exist a clean
distinction between lateral, feed-forward and feed-back
topologies, with the possibility that all three types of
connectivity contribute to human contour perception.
To quantify the temporal dynamics underlying visual
processing, we performed speed-of-sight psychophysical
experiments that required subjects to detect closed con-
tours (amoebas) spanning a range of shapes, sizes and
positions, whose smoothness could be adjusted paramet-
rically by varying the number of radial frequencies (with
randomly chosen amplitudes). To better approximate
natural viewing conditions, in which target objects usu-
ally appear against noisy backgrounds and both fore-
ground and background objects consist of similar low-
level visual features, our amoeba/no-amoeba task re-
quired amoeba targets to be distinguished from locally
indistinguishable open contour fragments (clutter). For
amoeba targets consisting of only a few radial frequen-
cies (K ≤ 4), human subjects were able to perform at
close to 100% accuracy after seeing target/distractor im-
age pairs for less than 200 ms, consistent with a number
of studies showing that the recognition of unambiguous
targets typically requires 150 − 250 ms to reach asymp-
totic performance [22, 23, 35], here likely aided by the
high intrinsic saliency of closed shapes relative to open
shapes [7]. Because mean inter-saccade intervals are also
in the range of 250 ms [34], speed-of-sight studies indi-
cate that unambiguous targets in most natural images
can be recognized in a single glance. Similarly, we found
that closed contours of low to moderate complexity read-
ily "pop out" against background clutter, implying that
such radial frequency patterns are processed in parallel,
presumably by intrinsic cortical circuitry optimized for
automatically extracting smooth, closed contours. As
saccadic eye movements were unlikely to play a signif-
icant role for such brief presentations, it is unclear to
what extent attentional mechanisms are relevant to the
speed-of-sight amoeba/no-amoeba task.
Our results further indicate that subjects perform no
better than chance at SOAs shorter than approximately
20 ms. Other studies, however, report above chance
performance on unambiguous target detection tasks at
similarly short SOA values [1, 23, 26, 33]. The discrep-
ancy may be attributed to the different masks employed.
Whereas the above cited studies used masks consisting
of either spatially filtered (e.g. 1/f ) noise, distractor
images, or scrambled versions of the target image set,
we constructed rotation masks that were optimized for
each target/distractor image pair [20]. Our working hy-
pothesis was that an optimized mask should completely
obscure the target object in the target-mask composite
image; also referred to as pattern masking. The require-
ment that the mask completely hide the target follows
from the assumption that at very short SOA, the tar-
12
get and mask images are likely to be effectively fused
due to the finite response time of neurons and recep-
tors in the early visual system [65]. For the amoeba/no-
amoeba task, we created optimized masks by rotating the
amoeba and clutter fragments with the goal of produc-
ing the maximum amount of interference in the responses
of orientation-selective cells. Presumably, maximum in-
terference occurs when orientation-selective neurons are
presented with randomly rotated contour fragments in
rapid succession. Although backward masks can have
heterogeneous effects, with performance in some cases
showing a U -shaped dependence on SOA [66], for the
masks used here performance always increased monoton-
ically with SOA. Empirically, the fact that performance
was no better than chance at 20 ms SOA suggests that
our optimized masks were able to effectively interrupt
the processing of smooth, closed contours at early corti-
cal processing stages. Indeed, the ability to drive overall
performance down to chance at SOA values shorter than
20 ms could provide an operational criteria for assessing
the degree to which a given backward pattern mask is
able to effectively interrupt visual processing.
The amoeba/no-amoeba task required the integration
of information over length scales spanning viewing angles
of approximately 3 − 4◦, larger than the classical excita-
tory receptive field size of parafoveal V1 neurons. The
amoeba/no-amoeba image set (see Figure 1) was con-
figured so that purely local information, such as a few
adjoining contour fragments, would not be sufficient to
solve the target detection problem. Rather, distinguish-
ing amoebas from clutter required integrating global in-
formation across multiple contour fragments. Our re-
sults suggest that such global integration can be accom-
plished via lateral interactions between local, orientation-
selective filters. Although the density of target and
clutter segments was not precisely equilibrated in our
amoeba/no-amoeba image set, the wide range of target
sizes and shapes spanned by our image generation algo-
rithm makes it unlikely that the near perfect performance
of human subjects at long SOA could have been attained
using density cues alone [4]. Here, lateral inputs were
used to modulate the bottom-up responses in a multi-
plicative fashion, so that our cortical association fields
acted primarily as gates that suppressed contour frag-
ments that did not receive sufficiently strong contextual
support. By preventing lateral inputs from producing ac-
tivity unless there was already a strong bottom-up input,
a multiplicative non-linearity prevented the activation of
contour fragments not present in the original image.
The phenomenon of illusory contours suggests that in
some cases contextual effects can produce activity even in
the absence of a direct bottom-up response [30]. The pre-
cise form of the multiplicative interaction used here was
adopted for algorithmic simplicity rather than for biolog-
ical realism. We observed that including a small additive
contribution from the lateral interactions did not funda-
mentally affect our conclusions. This suggests that ODD
kernels, if implemented more generally, might account for
the perception of illusory contours as well. However, a
more realistic description of the underlying cellular and
synaptic dynamics would likely be necessary to model a
relaxation process that includes both additive and mul-
tiplicative elements.
Both the model and the psychophysical experiments
employed a 2AFC design (see Figure 5) in which the
goal was to correctly identify which of a pair of images
contained an amoeba target. Since each trial involved
a forced choice between two images, the model used a
simple classifier that labeled the image with greater to-
tal activity as the target. For both human subjects and
the model, the number of radial frequencies K proved to
be a good control parameter for adjusting task difficulty
(see Figure 7). For targets of low to moderate complex-
ity, both model performance (as a function of number
of iterations) and human performance (as a function of
increasing SOA) monotonically approached nearly per-
fect asymptotic performance as described by a single sig-
moidal function with a characteristic scale, representing
either time or number of iterations, that increased with K
(see Figure 7). Based on comparison with human perfor-
mance at different SOA values, each iteration of the ODD
kernels was estimated to require approximately 37.6 ms
of cortical processing time, consistent with measured con-
duction delays between laterally connected cortical neu-
rons [13].
Prior to any lateral interactions, the stimulus was pro-
jected onto a retinotopic array of orientation-selective fil-
ter elements, providing a convenient representation for
learning cortical association fields by computing differ-
ences in pairwise coactivation statistics between target
and distractor images. We found that each iteration of
the ODD kernel increased the activity of contour frag-
ments that were part of amoebas compared to the activ-
ity of clutter fragments, so that after several iterations
the mean overall activity, summed across all orientation-
selective filter elements, was higher on average for target
images than distractor images (see Figure 4). Even in tri-
als that were incorrectly classified, contour fragments be-
longing to amoebas were typically still favored relative to
background clutter. Because the total number of contour
fragments varied from trial to trial, with only the average
number of fragments being fixed across the entire image
set, our relatively crude criterion for discriminating be-
tween target and distractor images sometimes led to clas-
sification errors even when amoeba fragments had been
partially segmented from the background clutter, sim-
ply because the distractor image initially contained more
fragments. A more sophisticated classifier might have
led to a closer correspondence between model and hu-
man performance. Although performance of the present
multiplicative model appeared to saturate after only a
few iterations of the ODD kernel (e.g. k ≤ 4), it is possi-
ble that a different implementation might have continued
to show improvements after additional iterations. How-
ever, the longer processing time implied by additional
iterations suggests that other physiological mechanisms,
13
particularly visual search, would likely come into play.
Granted, there is an apparent mismatch between the
fading of clutter elements in the model and the persis-
tence of such elements perceptually in human subjects.
To reconcile this apparent mismatch, it has been sug-
gested that the initial perception of brightness might be
driven by the initial bottom-up response of the individual
orientation-selective feature detectors, whereas persistent
responses across these same feature detectors might drive
salience [28].
The amoeba/no-amoeba image set was designed to al-
low for parameterized complexity (in terms of the amount
of clutter, number of radial frequencies, etc.) while avoid-
ing reference to exogenous world knowledge. Since the
amoeba/no-amoeba image set was machine generated, it
was possible to produce a very large number of train-
ing images; 40, 000 target and 40, 000 distractor images
at 256 × 256 pixel resolution were used to train ODD
kernels in the present study. Many computer vision sys-
tems employ standard image classification datasets such
as the Caltech 101 [67], which allows for uniform bench-
marking and thus facilitates direct comparison between
models. Datasets based on natural images, however, suf-
fer from several shortcomings. First, the resolution and
number of images are fixed when the set is created. While
some man-made datasets, such as MNIST [68]), consist
of tens of thousands of handwritten characters, anno-
tated sets of natural photographs ideal for speed-of-sight
experiments are typically limited to a few hundred im-
ages.
In contrast, humans are exposed to millions of
natural scenes during visual development. Biologically
motivated models that attempt to replicate human per-
formance might require similar numbers of examples. A
second shortcoming of natural image datasets is preva-
lence of high-level contextual information that utilizes ex-
ogenous world knowledge, such as the increased a priori
likelihood of finding a car on a road, or an animal in a for-
est. Exploiting such exogenous world knowledge posses
a formidable challenge for existing computational models
and, on tasks that employ natural images, may obscure
the ability of such models to extract behaviorally mean-
ingful information from low-level visual cues. Third, nat-
ural image datasets typically provide limited capability
for adjusting intrinsic task difficultly. For example, one
widely used dataset [33] includes photographs of animals
at different distances, but only a few discrete distances
are annotated and the relationship of target distance to
task difficultly is not easily quantified. Here, we illus-
trated how a synthetic set of images could be used to
compare model and human performance in a task with
parametric difficulty, potentially validating the use of ar-
tificial as opposed to natural images.
The present study addressed the role of cortical associ-
ation fields in the perception of closed contours, which are
presumably important for detecting visual targets based
on shape or outline. Although studies show that human
subjects can rapidly distinguish between images contain-
ing target and non-target object categories using only
the line drawings obtained by filtering natural scenes [1],
normal experience involves a number of complementary
visual cues, such as texture, color, motion and stereop-
sis. Presumably, cortical association fields also act to re-
inforce features representing these complementary visual
cues as well. Human subjects, for example, can distin-
guish whether pairs of texture patches were drawn from
the same natural object or two different natural objects
in a manner that exhibits a similar dependence on pair-
wise co-occurrence statistics as was found for orientated
edges [55]. We may speculate that an analysis of coacti-
vation statistics for features selective to a combination of
cues such as local orientation, texture, color, motion, and
disparity may lead to a more general and more powerful
set of kernels capable of fast and effective determination
of global object properties, which in turn can play an
important role in complex object identification.
Methods
Synthetic amoeba/no-amoeba image set
An amoeba is a type of radial frequency pattern [36]
consisting of a deformed circle in which the radius varies
as a function of the polar angle. By choosing the number
and relative amplitudes of the different frequency com-
ponents, the radius can describe an arbitrarily complex
shape, exactly analogous to how a Fourier basis can be
used to construct an arbitrary waveform on a finite in-
terval. Each radial frequency component was represented
by a sinusoidal function defined at C = 1024 discrete po-
lar angles, spaced uniformly on the interval [0, 2π). The
cutoff radial frequency used in constructing the closed
contour provided a control parameter for regulating the
complexity of the resulting figure, which ranged from
nearly circular, when only the 2 lowest radial frequen-
cies had non-zero amplitudes, to highly sinusoidal and
irregular, when the first 8 radial frequencies had non-
zero amplitudes. All amoeba shapes generated here may
be considered smooth, in that local curvature was always
bounded.
In detail, the radius of an amoeba at each polar angle
was:
r(φc) = A0 +
K(cid:88)
n=1
(cid:18) 2πn
C
(cid:19)
An cos
+ αn
.
(2)
All amplitudes An were initially drawn from normal dis-
tributions with 0 mean and unit variance. All phases
αn were drawn from uniform distributions over the in-
terval 0 and π/2. The resulting radial frequency pattern
was then linearly rescaled so that the maximum radius,
rmax, was equal to a random number drawn from a uni-
form distribution such that L/4 ≤ rmax ≤ L/2, where L
is the linear size of the square image (L = 256 pixels),
and the minimum radius was given by a second randomly
chosen value so that rmax/4 ≤ rmin ≤ rmax/2. Uniform
14
pseudo-random numbers were generated by the intrinsic
MATLAB 7.0 function RAND, or its Octave 3.2.3 equiv-
alent.
To facilitate the construction of locally indistinguish-
able clutter and model contour occlusion in natural im-
ages, amoeba contours were divided into 16 periodically-
spaced fragments by removing short sections whose
lengths varied within a specified range. Specifically, the
gaps between amoeba fragments varied from 16 to 32 in
units of discrete polar angle (2π/C). Amoeba contours
were then broken into fragments by periodically inserting
16 gaps of variable width ranging from 16 to 32, spaced
C/16 segments apart. Gaps were deleted from the un-
derlying contour, so that the polar angle subtended by
each fragment varied in accordance with the changes in
preceding gap width. The starting point of the first gap
was chosen randomly on the interval 1, C/G, so that over
the entire image set the inserted gaps were distributed
uniformly around the circle.
To create clutter fragments, an amoeba was first gen-
erated using the above procedure. Consecutive amoeba
fragments were then grouped, with the number of frag-
ments in each group determined by a Poisson process
with a mean value of 2 and an upper cutoff of 3. Each
group of amoeba fragments was then rotated about its
center of mass through random angles on the interval
π/8 to 7π/8. The resulting clutter consisted of the same
fragments as the original amoeba but rotated so that
collectively the rotated fragments no longer supported
the perception of a closed object. Clutter fragments
constructed in this manner were thus locally indistin-
guishable from amoeba fragments. To create clutter in
both target and distractor images, several amoebas were
first superimposed at random positions and then groups
of fragments rotated following the procedure described
above. All amoebas contained the same total number
of contour fragments (and therefore the same number of
gaps) but varied in both maximum diameter and total
contour length.
The center of each amoeba was chosen randomly un-
der the restriction that no contour be allowed to cross
an image boundary. Specifically, the x-coordinate of the
amoeba center, x0, was chosen randomly on a restricted
interval, rmax ≤ x0 ≤ L − rmax, and likewise for the y-
coordinate, y0. When groups of amoeba fragments were
randomly rotated to make clutter, portions of a contour
belonging to a clutter fragment would occasionally cross
an image boundary.
In such cases, any out-of-bounds
portions of a contour were reflected back into the image
region using mirror boundary conditions.
Target images always consisted of 1 set of amoeba frag-
ments and 2 sets of clutter fragments. Distractor images
consisted of 3 sets of clutter fragments and thus, averaged
over the entire image set, had the same mean luminance
and the same variance as the target images. Mask images
were constructed following a procedure nearly identical
to that used for constructing distractor images, except
that mask images consisted of 6 sets of clutter fragments,
obtained by randomly rotating the 6 original amoeba ob-
jects used in constructing the corresponding target and
distractor images. All contour fragments were initially
represented as a set of points in polar coordinates, cor-
responding to the radius at each discrete polar angle.
Points along the contour were then transformed back to
Cartesian coordinates and rounded to the nearest dis-
crete pixel value. MATLAB scripts for generating the
image set used in this study are publicly available at:
http://petavision.sourceforge.net.
Ethics statement
The Los Alamos National Laboratory (LANL) Human
Subjects Research Review Board (HSRRB) has reviewed
the following experimental protocol and determined that
it provides adequate safeguards for protecting the rights
and welfare of human subjects involved in the protocol.
The protocol was reviewed and approved in compliance
with the U.S. Department of Health and Human Ser-
vices (DHHS) regulations for the Protection of Human
Subjects, 45 CFR 46, and in accordance with the LANL
Federal Wide Assurance (FWA#00000362) with the Na-
tional Institutes of Health/Office for Human Research
Protections (NIH/OHRP). The identification number is
LANL 08-03 X.
Human psychophysics
evaluated using
Human performance was
two-
alternative forced choice (2AFC) psychophysical exper-
iments. There were 5 subjects, all with normal or
corrected-to-normal vision. One subject only contributed
data for a portion of the tested SOAs. Each subject was
seated in a dark room, at an approximate distance of 65
cm from a 19-inch nominal (36.2× 27 cm actual size) Hi-
tachi 751 CRT monitor. Images spanned a viewing angle
of approximately 7◦ × 7◦. The monitor resolution was
1024 × 768 pixels and the refresh rate was 100 Hz. The
display was driven by a dual-core 3.0 GHz Mac Pro, with
MATLAB 7.6 running Psychtoolbox [69].
After a short training period to familiarize the subject
with the task, one target image and one distractor image
were shown side by side, followed by a mask intended to
interrupt cognitive processing of the target and distrac-
tor images. Two separate sets of experiments were con-
ducted for each subject. In one set, the SOA was chosen
randomly from the values 40, 80, 120 ms. For the second
set of experiments, the SOA was chosen randomly from
the values 20, 160, 200 ms. The duration of the stimulus
was always the same as the SOA, and thus both the target
and distractor images remained visible until mask onset.
The duration of the mask was always 100 ms. Each sub-
ject was shown 1200 images divided into 10 blocks of 120
images, with rest breaks in between blocks (rest break du-
ration was at the discretion of each subject). The pace of
15
the experiment was under the control of the subject, who
initiated each trial using the space bar. A small temporal
jitter, chosen uniformly between 0 to 250 ms, was added
to the interval preceding each trial, to prevent entrain-
ment. Task conditions, consisting of variations in both
the SOA and the number of radial frequencies K, were
randomly interleaved such that each condition occurred
the same number of times over the course of the entire
experiment.
On each trial, subjects indicated which side contained
the target, using a mouse-driven slider bar to report con-
fidence (see Figure 5). The reported confidence values
were used to construct receiver operating characteristic
(ROC) curves, which plot the percentage of true posi-
tives (or hits) against the percentage of false positives (or
false alarms), with each true/false positive pair obtained
by setting a confidence threshold at a different location
along the slider bar. A correct response was not neces-
sarily considered a true positive: to generate one point
on the ROC curve, the reported confidence on each trial
was measured relative to the current threshold position,
which could be to either the left or to the right of center.
Thus, a trial might be labeled as incorrect, even though
the subject moved the slider bar in the correct direction,
as long as the threshold level was not exceeded. Specifi-
cally, whenever the reported confidence fell to the left of
threshold, the corresponding trial was treated as though
the subject reported the target as being to the left, even
if the threshold location had been set to the right of cen-
ter and the confidence bar had actually been slid to the
right. Likewise, when the reported confidence fell to the
right of the current threshold position, the trial was al-
ways treated as if the subject had reported the target to
the right, again regardless of how the subject moved the
slider bar relative to the center position. By choosing
a range of threshold positions, spanning the full range
of reported confidence values, a complete ROC curve
was obtained. Note that as the threshold was moved
closer to the left edge of the slider bar, the percentage of
true and false positives both approached minimum val-
ues, since only trials with very high reported confidence
could contribute to either the true positive or false posi-
tive rate (most trials were rejected as either true or false
negatives). As the threshold position moved closer to
the center of the confidence slider bar, the percentage of
true positives increased. Finally, as the threshold was
moved closer to the right edge of the slider bar, both
the true positive rate and the percentage of false posi-
tives approached maximum values. The true positive rate
averaged over all false positive rates, or the area under
the ROC curve (AUC), was used as an overall measure
of subject performance. The AUC is equivalent to the
probability that a randomly chosen target image will be
correctly classified relative to a randomly chosen distrac-
tor image, and thus directly predicts performance on the
2AFC task. Results for each SOA and for each value of
K were averaged over 5 subjects. Error bars denote the
standard deviation over the 5 subjects.
Model
rotation matrix,
Model cortical association fields were based on differ-
ences in the coactivation statistics of orientation-selective
filter elements drawn from target and distractor images.
Geisler and Perry measured co-occurrence statistics for
oriented edges in human segmented natural images [25],
and found a close correspondence to human judgments
as to whether pairs of short line fragments were drawn
from the same or different contours. Thus, we refer to
the difference in coactivation statistics between target
object and distractor images as Object-Distractor Dif-
ference (ODD) kernels. ODD kernels were trained using
40, 000 target and 40, 000 distractor images, each divided
into 4 sets of 10, 000 images each, with each set associ-
ated with a different value of K = 2, 4, 6, 8. The order in
which the images were presented had no bearing on the
final form of the ODD kernel; that is, there was no tem-
poral component to the training. Training with more
images did not substantively improve performance, al-
though small differences were observed in the ODD ker-
nels trained using a smaller number of images (10, 000
target and 10, 000 distractor images).
Each 256×256 pixel training image activated a regular
array of 256×256 retinal elements whose outputs were ei-
ther 0 or 1, depending on whether the corresponding im-
age pixel was ON or OFF, respectively. Each retinal unit
activated a local neighborhood of orientation-selective fil-
ters, which spanned 8 angles spaced uniformly between
0 and π. To mitigate aliasing effects, the orientation-
selective filters were rotated by a small, fixed offset, equal
to π/16, relative to the axis of the training images. All
orientation-selective filters were 7×7 pixels in extent and
consisted of a central excitatory subunit, represented by
an elliptical Gaussian with a standard deviation of 7.0 in
the longest direction and an aspect ratio of 7.0, flanked
by two inhibitory subunits whose shapes were identical
to the central excitatory subunit but were offset by ±1.4
pixels in the direction orthogonal to the preferred axis.
The weight Wθ(x1−x2, y1−y2), from a retinal element
at (x2, y2) to a filter element at (x1, y1) with dominant
orientation θ, was given by a sum over excitatory and
inhibitory subunits:
W θ(x1 − x2, y1 − y2) = W θ(r1 − r2)
(cid:20) 1
(cid:0)r1 − r2
(cid:0)r1 + f − r2
(cid:0)r1 − f − r2
the matrix σ =(cid:2) 1 0
(cid:1)T(cid:21)
(cid:1) · R−1
θ σ−1Rθ ·(cid:0)r1 − r2
(cid:1)T(cid:21)
θ σ−1Rθ ·(cid:0)r1 + f − r2
(cid:1) · R−1
(cid:1)T(cid:21)(cid:41)
(cid:1) · R−1
θ σ−1Rθ ·(cid:0)r1 − f − r2
(cid:3) describes the shape of the elliptical
where the position vector is given by ri = [xi, yi] and
(cid:20) 1
(cid:20) 1
− exp
− exp
(cid:40)
(3)
= A
exp
2
2
2
Gaussian subunits for θ = 0. In Eq. 3, Rθ is a unitary
0 7
(cid:20) cos θ
sin θ
− sin θ cos θ
(cid:21)
,
Rθ =
16
(4)
and f = [0.0, 1.4] is a translation vector in the direc-
tion orthogonal to the dominant orientation when θ = 0.
The amplitude A was determined empirically so that the
total integrated strength of all excitatory connections
made by each retinal unit equaled 20.0 (and thus the
total strength of all inhibitory connections made by each
retinal unit equaled −40.0). Mirror boundary conditions
were used to mitigate edge effects. The retinal input to
each orientation-selective filter element s(x1, y1, θ) was
then given by
s(x1, y1, θ) =
W θ(x1 − x2, y1 − y2)I (x1,y1)(x2, y2),
(cid:88)
x2,y2
(5)
where I (x1,y1) is the 7 × 7 binary input image patch cen-
tered on (x1, y1). The sum is over all pixels (x2, y2) that
are part of this image patch. The initial output of each
orientation-selective filter element z0(x1, y1, θ) was ob-
tained by comparing the sum of its excitatory and in-
hibitory retinal input to a fixed threshold of 0.5. Values
below threshold were set to 0 whereas values above unity
were set to 1.0. Thus
is an element-wise implementation of these thresholds.
The responses of all suprathreshold orientation-selective
filters contributed to the coactivation statistics, with only
the relative distance, direction, and orientation of filter
pairs recorded. Because of the threshold condition, only
the most active orientation-selective filters contributed to
the coactivation statistics.
For every suprathreshold filter element extracted from
the i-th target image, coactivation statistics were accu-
mulated relative to all surrounding suprathreshold filter
elements extracted from the same image. Thus the ODD
kernel G is given by
(cid:0)ρ(x−x0, y−y0), φθ0(x−x0, y−y0)(cid:1) =
(cid:88)
z0(x, y, θ),
Gti
θ−θ0
x,y
,
(8)
where the radial distance ρ is a function of the (x, y)
coordinates of the two filter elements, the direction φ is
the angle measured relative to θ0, the sum is over all
suprathreshold elements within a cutoff radius of 32, the
superscript ti denotes the i-th target image, and the dif-
ference in the orientations of the two filter elements θ−θ0
is taken modulo π. Because the amoeba/no-amoeba im-
age set was translationally invariant and isotropic, the
where the function,
z(x1, y1, θ) = g(cid:0)s(x1, y1, θ)(cid:1),
0
s < 0.5
0.5 ≤ s ≤ 1.0
s > 1.0
g(s) =
s
1
,
(6)
(7)
central filter element may without loss of generality be
shifted and rotated to a canonical position and orienta-
tion, so that the dependence on ρ0, φ0, θ0 may be omitted.
The coactivation statistics for the i-th target image can
then be written simply as Gti
θ (ρ, φ), where (ρ, φ) gives
the distance and direction from the origin to the filter el-
ement with orientation θ, given that the filter element at
the origin has orientation π/16. An analogous expression
gives the coactivation statistics for the j-th distractor im-
age Gdj
θ (ρ, φ). The ODD kernel Gθ(ρ, φ) is given by the
difference
(cid:88)
Gθ(ρ, φ) = W+
(cid:88)
θ (ρ, φ) − W−
Gti
Gdj
θ (ρ, φ),
(9)
i
j
where the sums are taken over all target and distractor
images and the normalization factors W+ and W− are de-
termined empirically so as to yield a total ODD strength
of 325 (see Figure 8 and Results), defined as the sum over
all ODD kernel elements arising from either the target or
distractor components. By construction, the sum over
all ODD kernel elements equals zero, so that the average
lateral support for randomly distributed edge fragments
would be neutral. Our results did not depend critically
on the RMS magnitude of the ODD kernel (see Figure 8).
To minimize storage requirements individual connection
strengths were stored as unsigned 8-bit integers, so that
the results of the present study did not depend on com-
putation of high precision kernels.
As described above, the canonical ODD kernel is de-
fined relative to filter elements at the origin with orienta-
tion π/16. Filter elements located away from the origin
can be accounted for by a trivial translation. To account
for filter elements with different orientations, separate
ODD kernels were computed for all 8 orientations then
rotated to a common orientation and averaged to pro-
duce a canonical ODD kernel. The canonical kernel was
then rotated in steps between 0 and π (offset by π/16)
and then interpolated to Cartesian x−y axes by rounding
to the nearest integer coordinates. Although it has been
demonstrated that global contour saliency is enhanced
for orientations along the cardinal axes [58], this bias is
by construction absent from this model.
ODD kernels were used to compute lateral support for
17
each orientation-selective filter element, via linear convo-
lution. The output of each filter element was then modu-
lated in a multiplicative fashion by the computed lateral
support. The procedure was iterated by calculating new
values for the lateral support s, which were again used
to modulate filter outputs in a multiplicative fashion:
Gθ(ρ, φθ)zk−1(x(cid:48), y(cid:48), θ(cid:48)
sk(x, y, θ) = zk−1(x, y, θ)
),
(cid:88)
x(cid:48),y(cid:48),θ(cid:48)
(10)
where the subscript k denotes the k-th iteration. The
same kernel was used for all iterations. All source code
used to train and apply cortical association fields is pub-
licly available at
http://sourceforge.net/projects/petavision/.
elements, T = (cid:80)
To measure model performance, in each trial 1 target
image and 1 distractor image were tested as a pair, so
as to emulate the 2AFC format of the human experi-
ments. The orientation-selective filter responses to both
test images were evaluated after k = 0, 1, 2, 3, 4 iterations
of the ODD kernel. The total activation across all filter
x,y,θ(cid:48) zk(x, y, θ(cid:48)), was used to compare
the two test images. Since the model cortical association
fields tended to support contour fragments belonging to
amoebas while inhibiting clutter fragments, the image
with higher total activation T was assumed to be the
target image. Error bars for the model performance (as
shown in Figure 7) were estimated using the standard
deviation of a binomial distribution with probability p
equal to percent correct and N equal to the number of
trials.
Acknowledgments
The authors wish to thank Steven Zucker for stimu-
lating discussions that helped initiate this project. This
work was supported by Los Alamos National Laboratory
LDRD program under project 20090006DR; the National
Science Foundation, grant ID 0749348; and the DARPA
NeoVision2 project. This publication qualified for un-
classified release under DUSA BIOSCI with LA-UR 11-
00499.
[1] L. Velisavljevi´c and J. H. Elder, J Vision 9 (2009).
[2] D. J. Field, A. Hayes, and R. F. Hess, Vision Res 33, 173
Computation.
[7] I. Kov´acs and B. Julesz, P Natl Acad Sci USA 90, 7495
(1993), ISSN 0042-6989.
(1993).
[3] G. Loffler, Vision Res 48, 2106 (2008), ISSN 0042-6989,
[8] M. W. Pettet, S. P. McKee, and N. M. Grzywacz, Vision
vision Res Reviews.
Res 38, 865 (1998), ISSN 0042-6989.
[4] R. Hess and D. Field, Trends Cogn Sci 3, 480 (1999),
[9] U. Polat and D. Sagi, Vision Res 33, 993 (1993).
ISSN 1364-6613.
[10] M. K. Kapadia, M. Ito, C. D. Gilbert, and G. West-
[5] D. Fitzpatrick, Curr Opin in Neurobiol 10, 438 (2000),
heimer, Neuron 15, 843 (1995).
ISSN 0959-4388.
[11] U. Polat, A. Terkin, and O. Yehezkel, Adv Cogn Psych
[6] P. Seri´es, J. Lorenceau, and Y. Fr´egnac, J Physiology-
Paris 97, 453 (2003), ISSN 0928-4257, neuroscience and
3, 153 (2008).
[12] P.-C. Huang and R. F. Hess, Vision Res 47, 3108 (2007).
18
[13] V. Bringuier, F. Chavane, L. Glaeser, and Y. Fr´egnac,
J. Neurosci. 26, 8254 (2006).
Science 283, 695 (1999).
[40] J. P. Jones and L. A. Palmer, J. Neurophys. 58, 1233
[14] J. R. Cavanaugh, W. Bair, and J. A. Movshon, J. Neu-
(1987).
rophys. 88, 2530 (2002).
[41] T. W. Troyer, A. E. Krukowski, N. J. Priebe, and K. D.
[15] J. R. Cavanaugh, W. Bair, and J. A. Movshon, J. Neu-
Miller, J Neurosci 18, 5908 (1998).
rophys. 88, 2547 (2002).
[42] A. Angelucci, J. B. Levitt, E. J. S. Walton, J.-M. Hup´e,
[16] A. Pooresmaeili, J. L. Herrero, M. W. Self, P. R. Roelf-
J. Bullier, and J. S. Lund, J Neurosci 22, 8633 (2002).
sema, and A. Thiele, J. Neurosci. 30, 12745 (2010).
[43] P. Azzopardi and A. Cowey, P Natl Acad Sci USA 94,
[17] W. H. Bosking, Y. Zhang, B. Schofield, and D. Fitz-
14190 (1997).
patrick, J. Neurosci. 17, 2112 (1997).
[18] C. Gilbert and T. Wiesel, J. Neurosci. 9, 2432 (1989).
[19] R. Malach, Y. Amir, M. Harel, and A. Grinvald, P Natl
[44] N. A. Macmillan and C. D. Creelman, Detection theory:
a user's guide (CUP Archive, Cambridge, 1991).
[45] J. H. R. Maunsell and J. R. Gibson, J. Neurophys. 68,
Acad Sci USA 90, 10469 (1993).
1332 (1992).
[20] R. F. Hess, W. H. A. Beaudot, and K. T. Mullen, Vision
[46] J. Bell, D. R. Badcock, H. Wilson, and F. Wilkinson,
Res 41, 1023 (2001), ISSN 0042-6989.
[21] C. Keysers, D.-K. Xiao, P. Foldi`ak, and D. I. Perrett, J
Cognitive Neurosci 13, 90 (2001).
[22] C. Keysers and D. I. Perrett, Trends Cogn Sci 6, 120
(2002), ISSN 1364-6613.
[23] N. Bacon-Mac´e, M. J.-M. Mac´e, M. Fabre-Thorpe, and
S. J. Thorpe, Vision Res 45, 1459 (2005), ISSN 0042-
6989.
Vision Res 47, 1518 (2007), ISSN 0042-6989.
[47] Z. Li, Neural Comput 13, 1749 (2001).
[48] Z. Li, Neural Comput 10, 903 (1998).
[49] T. N. Mundhenk and L. Itti, Biol Cybern 93, 188 (2005).
[50] W. Li, V. Piech, and C. D. Gilbert, Neuron 50, 951
(2006).
[51] S. Grossberg and E. Mingolla, Percept Psychophys 38,
141 (1985).
[24] O. Ben-Shahar and S. Zucker, Neural Comput 16, 445
[52] S. Ullman, R. L. Gregory, and J. Atkinson, Philos T R
(2004).
Soc Lon B 337, 371 (1992).
[25] W. S. Geisler and J. S. Perry, Visual Neurosci 26, 109
[53] S.-C. Yen and L. H. Finkel, Vision Res 38, 719 (1998),
(2009).
ISSN 0042-6989.
[26] S. Mandon and A. K. Kreiter, Vision Res 45, 291 (2005),
[54] P. J. Garrigues and B. A. Olshausen, in Adv Neur In
ISSN 0042-6989.
(2007).
[27] M. Ursino and G. E. L. Cara, Neural Networks 17, 719
[55] A. D. Ing, A. J. Wilson, and W. S. Geisler, J Vision 10,
(2004).
1 (2010).
[28] A. Sterkin, A. Sterkin, and U. Polat, J. Vis. 8, 1 (2008).
[29] W. Bair, J. R. Cavanaugh, and J. A. Movshon, J Neurosci
[56] H. Yao, L. Shi, F. Han, H. Gao, and Y. Dan, Nat. Neu-
rosci. 10, 772 (2007).
23, 7690 (2003).
[57] T. Hua, P. Bao, C.-B. Huang, Z. Wang, J. Xu, Y. Zhou,
[30] N. R. Zhang and R. von der Heydt, J. Neurosci. 30, 6482
and Z.-L. Lu, Curr Biol 20, 887 (2010).
(2010).
[58] W. Li and C. D. Gilbert, J. Neurophysiol. 88, 28462856
[31] L. Schwabe, K. Obermayer, A. Angelucci, and P. C.
(2002).
Bressloff, J. Neurosci. 26, 9117 (2006).
[59] A. Knoblauch and F. T. Sommer, Neurocomputing 58-
[32] A. Angelucci, J. B. Levitt, E. J. S. Walton, J.-M. Hupe,
J. Bullier, and J. S. Lund, J. Neurosci. 22, 8633 (2002).
[33] T. Serre, A. Oliva, and T. Poggio, P Natl Acad Sci USA
60, 185 (2004).
[60] P. O. Hoyer and A. Hyvarinen, Vision Res 42, 1593
(2002).
104, 6424 (2007).
[61] S. Song, K. E. Miller, and L. F. Abbott, Nat Neurosci 3,
[34] S. Martinez-Conde, S. L. Macknik, X. G. Troncoso, and
D. H. Hubel, Trends Neurosci 32, 463 (2009), ISSN 0166-
2236.
[35] E. T. Rolls and M. J. Tovee, P Roy Soc Lond B Bio 257,
9 (1994).
[36] F. Wilkinson, H. R. Wilson, and C. Habak, Vision Res
919 (2000).
[62] Z. Li, Trends Cogn Sci 6, 9 (2002), ISSN 1364-6613.
[63] K. Fukushima, Biol Cybern 36, 193 (1980).
[64] C. D. Gilbert and M. Sigman, Neuron 54, 667 (2007).
[65] D. Schneeweis and J. Schnapf, Science 268, 1053 (1995).
[66] J. T. Enns and V. D. Lollo, Trends Cogn Sci 4, 345
38, 3555 (1998), ISSN 0042-6989.
(2000), ISSN 1364-6613.
[37] W. S. Geisler, J. S. Perry, B. J. Super, and D. P. Gallogly,
[67] L. Fei-Fei, R. Fergus, and P. Perona, in CVPR 2004,
Vision Res 41, 711 (2001), ISSN 0042-6989.
Workshop on Generative-Model Based Vision (2004).
[38] E. Schneidman, M. J. Berry II, R. Segev, and W. Bialek,
[68] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, in P
Nature 440, 1007 (2006).
[39] J. Shlens, G. D. Field, J. L. Gauthier, M. I. Grivich,
D. Petrusca, A. Sher, A. M. Litke, and E. J. Chichilnisky,
IEEE (1998), vol. 86, p. 2278.
[69] D. H. Brainard, Spatial Vision 10, 433 (1997).
|
1206.0637 | 1 | 1206 | 2012-06-04T14:50:03 | A Quantum-mechanical description of ion motion within the confining potentials of voltage gated ion channels | [
"q-bio.NC",
"physics.bio-ph"
] | Voltage gated channel proteins cooperate in the transmission of membrane potentials between nerve cells. With the recent progress in atomic-scaled biological chemistry it has now become established that these channel proteins provide highly correlated atomic environments that may maintain electronic coherences even at warm temperatures. Here we demonstrate solutions of the Schr\"{o}dinger equation that represent the interaction of a single potassium ion within the surrounding carbonyl dipoles in the Berneche-Roux model of the bacterial \textit{KcsA} model channel. We show that, depending on the surrounding carbonyl derived potentials, alkali ions can become highly delocalized in the filter region of proteins at warm temperatures. We provide estimations about the temporal evolution of the kinetic energy of ions depending on their interaction with other ions, their location within the oxygen cage of the proteins filter region and depending on different oscillation frequencies of the surrounding carbonyl groups. Our results provide the first evidence that quantum mechanical properties are needed to explain a fundamental biological property such as ion-selectivity in trans-membrane ion-currents and the effect on gating kinetics and shaping of classical conductances in electrically excitable cells. | q-bio.NC | q-bio |
To appear in Journal of Integrative Neuroscience 11 (2), June 2012.
A Quantum-mechanical description of ion motion within the
confining potentials of voltage gated ion channels
Johann Summhammera, Vahid Salarib, Gustav Bernroiderc,1
a Atom Institute, Vienna University of Technology,
Stadionallee 2, A-1020 Vienna, Austria
b Kerman Neuroscience Research Center (KNRC),
Kerman, Iran
c Department of Organismic Biology, Neurosignaling Unit, University of Salzburg
Hellbrunnerstr 34,A-5020 Salzburg, Austria
Email: [email protected]
Abstract
Voltage gated channel proteins cooperate in the transmission of membrane potentials between nerve
cells. With the recent progress in atomic-scaled biological chemistry it has now become established
that these channel proteins provide highly correlated atomic environments that may maintain electronic
coherences even at warm temperatures. Here we demonstrate solutions of the Schrodinger equation
that represent the interaction of a single potassium ion within the surrounding carbonyl dipoles in the
Berneche-Roux model of the bacterial KcsA model channel. We show that, depending on the surrounding
carbonyl derived potentials, alkali ions can become highly delocalized in the filter region of proteins at
warm temperatures. We provide estimations about the temporal evolution of the kinetic energy of ions
depending on their interaction with other ions, their location within the oxygen cage of the proteins
filter region and depending on different oscillation frequencies of the surrounding carbonyl groups. Our
results provide the first evidence that quantum mechanical properties are needed to explain a fundamental
biological property such as ion-selectivity in trans-membrane ion-currents and the effect on gating kinetics
and shaping of classical conductances in electrically excitable cells.
Keywords: Quantum mechanics; quantum biology; quantum oscillations; Schrodinger equation; neural
signaling; molecular dynamics; ion channels; coherence dynamics; selectivity filter; filter gating.
1
Introduction
Ion channels are the building blocks of electrical membrane signals in the nervous system. They
are responsible for the controlled charge transition across cell membrane, building up the trans-
membrane potentials that propagate along the membranes as either dendritic potentials or action
potentials. The ’channels’ are proteins inserted into the cell membrane and host different sub-
domains that operate at highly different physical action orders, ranging from the quantum-scale
at the single atom level up to the classical scale of the entire protein [4, 5]. The main parts of the
protein consist of the pore domain gate, controlling the access of ions into the protein, the cavity
region that hosts a hydrated ion prior to its access to the narrow ’selectivity filter’ that provides
the last steps of charge transfer into the cell or outside the cell (see Figure 1). The selectivity
filter (SF) of ion channel proteins is responsible for the selective and fast conduction of ions
across neuronal cell membranes. After the determination of the atomic resolution structure of
the bacterial model KcsA channel by MacKinnon and colleagues [15, 1], it became increasingly
apparent that delicate atomic interactions within the highly ordered conduction pathway of the
SF are critical for the ability of the protein to provide a very high rate of conduction without loss
of selectivity for a particular ion species. Following these initial findings a long list of molecular
dynamics studies (MD) added important functional aspects to the static details of the channel
1corresponding author
1
protein as reviewed by [2, 10, 14, 18]. Density functional calculations and hybrid QM/MM
methods have substantially contributed in capturing the interaction between carbonyls and their
effect on backbone atoms [8], together with a significant charge-polarization effect of ions on
their filter ligands [12]. Taking into account these results, additional theoretical treatments of
ion complexation in water and diverse binding site models have revealed the topology of forces
that are engaged in providing a highly selective coordination of ions in the filter region [6].
The question became dominant how selectivity can emerge without compromising conduc-
tion. Within this context at least two sets of problems are at the center of interest. First, it
can be expected that the Coulomb type forces among ions , the adjacent water dipoles and the
neighboring oxygen atoms, with average distances of 0,285 nm in the KcsA filter [13], require a
quantum description, because the involved distances are near the de Broglie wavelength associ-
ated with the ions at thermal energies. Are there typical quantum interaction effects or quantum
interferences that are indispensible to explain the biological functioning of these channel proteins
? The second point involves the question, can quantum effects propagate into the classical states
of proteins ? The last question gains considerable significance from findings about different ’gat-
ing mechanisms’ of the channel. Whereas ’pore gating’ is a relatively slow process that controls
the access of ions to the cavity region of the filter, the filter itself is found to be able to change
between ’permissive’ and ’non-permissive’ conformational states at a much shorter time-scales
[3, 18]. Because ’selectivity’ seems to be organized within the filter that is located opposite the
’entrance’ of the proteins pore region, it can be expected that the mechanisms providing selec-
tivity must somehow become coordinated with the mechanisms controlling access of ions to the
cavity region. If quantum effects do play a critical role for filter-ion coordination, it is feasible
that these delicate interactions could leave their quantum traces in the overall conformation and
the molecular gating state of the entire protein.
In the present paper we investigate this question. We study the question if and how quantum
interferences in the filter region of a typical K-channel can influence the atomic environment
within the frame set by time-dependent potentials derived from carbonyl charges, water molecules
and other ions present in the atomic environment of the filter-region. By solving the Schrodinger
equation for the quantum mechanical states of K + ions, we demonstrate that, depending on the
size of the confining potential and the thermal energy of the ion, the ion’s wave-function can
become highly delocalized, with its probability distribution extending over a significant fraction
of the filter region. This in turn exerts an influence on the motion and induced polarization of
the carbonyl groups lining the filter in a way that is different from the interaction that could
be expected by a strictly localized classical ion. Due to the interaction of the ions with a time-
dependent potential, the energy of ions is not conserved. Instead, we observe tera-hertz (THz)
oscillations of ionic wave-packets that become damped, giving off their energy to the environment
either via the vibrational modes of the surrounding carbonyl dipoles or by radiation, or both.
These effects will cool down the ions in the filter domain dramatically. We finally demonstrate
the result of a ’replacement study’, where K + ions in the same environment are replaced by
Na+ ions. The results show that the oscillation frequency of the coordinating carbonyl dipoles
discriminates the kinetic energy minima for both ion species. This is the first indication of a
quantum mechanical effect on the ion selectivity function of voltage gated channel proteins.
2 Methods
2.1 The molecule
We represent the selectivity filter as an array of C=O dipoles provided by the backbone carbonyl
groups of the highly conserved TVGYG amino acid sequence (see Figure 1,2). Each filter segment
carries five dipoles. We account for the intermittent, nearest neighbor water molecules in the filter
2
Figure 1: The main physical operation facilitated by an ion channel-protein is an ion-species
selective charge transfer (A), where one charge is taken up from the cavity (shown left), dehy-
drated and transferred into the filter region (marked by double bars). The filter region hosts at
least two ions. The uptake of one ion leads to a knock-on effect on the ions present in the filter
region. One ion is taken up from the cavity site, and another ion is released to the (intra or
extra) cellular site (shown right). (B) The basic organization of voltage gated ion channels. The
protein consists of three subdomains, the pore, the cavity and the filter region.
3
Figure 2: A schematic illustration of a dimeric KcsA view with filter location and associated
trans-membrane helices. The carbonyl atoms are indicated by short segments and numbered
S0-S4 starting from the extracellular site. In (B)a demonstration of a free energy distribution
given in kcal/mol along the axial z-coordinates (in Angstroems) of the filter according to the
Berneche-Roux model is shown. The ion-filter configuration for the energy distribution in (B)
is demonstrated by the scheme(C). In the present paper the ion at location S1 (insert C) is
considered quantum-mechanically, the surrounding atoms from carbonyl groups, other ions in
the filter region and the intermittent water molecule are treated as classical particles.
by a confining potential for the ions. This water-derived potential behaves similar to repelling
atoms, degrading the spread of the ion’s wavepacket. The distances of carbonyl C-atoms to
the filter axis were preset to 0,345 nm according to the Garafoli-Jordan permeation model [11].
Effective charge distances of C=O dipoles were set to 0.122 nm. Bond directions along C=O
groups are variable with zero degrees direction perpendicular to the filter axis. In the present
hybrid model we consider one potassium ion as a quantum particle and treat the surrounding
nuclear motion classically.
2.2 The quantum model
The basic interaction between classical ions and C=O dipoles as well as a quantum-mechanically
described ion and it’s atomic surrounding is the electrostatic repulsion and attraction. As usual
the potential between two electrical point charges is
V (r) =
1
4πǫ
·
q1q2
r
(1)
We have used ′r′ for the distance between two particles and ’q’ for their charges. The symbol
′ǫ′ denotes the dielectric constant of the medium containing the charges. Taking into account a
total of 20 C=O dipoles the resulting potential for the region containing one K + ion becomes:
VF (r, t) =
qion
4πǫ
20
Xj=1
(cid:12)(cid:12)(cid:12)
ROj (t) − r(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
RCj (t) − r(cid:12)(cid:12)(cid:12)
Here the vector ¯r represents the ions position at time t, qOj for the charge of the j-th oxygen
qOj
qCj
+
(2)
4
and qCj for the charge of the j-th carbon atom respectively. Further ¯ROj denotes the time-
dependent position of the j-th oxygen and ¯RCj the position of the j-th carbon atom.
In additon we have to consider the potential Vi that is contributed by the classical ions in
the channel:
VI (r, t) =
qion
4πǫ
(cid:12)(cid:12)(cid:12)
RK1(t) − r(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
RK2(t) − r(cid:12)(cid:12)(cid:12)
qK
qK
+
(3)
The potential due to the water molecules on both sides of the ion is given by a potential
VH2O, which strongly depends on the form of the wave-packet. The time-dependent locations of
classical ions have been denoted by RKj . The total potential to which the quantum particle is
exposed is thus:
V (¯r, t) = VF (¯r, t) + VI (¯r, t) + VH2O(¯r, t)
(4)
From a modelistic view the overall axial symmetry of the channel (Figure 3) allows for a
restriction of the calculations to one spatial dimension, along the z-axis of the filter. This also
reduces the computational effort considerably. Within the frame set we obtain a Schroedinger
equation for one K + in the filter:
−¯h2
2mk
d2
dz2 ψ(z, t) + V (z, t).ψ(z, t) = i¯h
∂
∂t
ψ(z, t)
(5)
The water molecules interacting with the ion move with the ion and thus somewhat impede
the spreading of its wave-packet by an effective potential:
VH2O(z, t) = −EH2O.nZ d
−d
ψ(z + s, t)2 ds
(6)
Here, EH2O denotes the depth of the confining potential, set to 10 meV, which corresponds
to about half the thermal energy kBT . The symbol ’d’ represents the average distance between
the oxygen of a water molecule and the K + ion as available from both, experimental [16, 26]
and theoretical studies [26]. For the present study we have set a d = 0.28 nm. Further ’n’ is a
normalization constant which ensures that the minimum of VH2O(z, t) equals −EH2O. We have
calculated the wave-function together with the resulting space and time dependent probability
p = ψ(z, t)2 for different oscillations of C=O groups and locations of classical ions in the
filter region. All second order differential equations were solved by applying the Crank-Nicolson
formalism [23]which is known for it’s stability over a large range of calculational steps. Additional
controls were applied to check the consistency of calculations by doubling the temporal resolution
within a constant time interval (e.g. 678.500 steps with 0.006 fs to 2.750.000 steps with 0,0015 fs
resolution, see Figure 4). No numerical artifacts on the stability of the results could be observed.
Also, under constant, time independent potentials the wave-packets behavior was as can be
expected for harmonic oscillating potential forms, for which analytic solutions are well known.
The initial wave-function was approximated by a Gaussian distribution centered at one of the
possible filter locations (S0-S4).
3 Results
3.1 Conductive States
In the following figures we provide some snapshots of varying potentials confining the quantum-
mechanical ion KQ at location S1 of the filter region for different interactions with ’classical
ions’ KB and KA moving into the filter region. The filter sets out with a double occupancy of
5
Figure 3: A three-dimensional distance model for the carbonyl-ion coordination within the filter
region of the KcsA model channel as used for the present study. The 20 C=O dipoles are
symbolized by circles and are numbered on the left to specify the distance matrix that we used
for the present calculations and the 3-dim topology provided by the right-hand figure (dark lines
join the C=O dipoles from two lateral strands and white lines join opposing C=O dipoles)
Figure 4: Calculational consistency of the Crank-Nicolson approximation: a comparison of wave-
function solutions with different temporal scales, ranging from 687.500 steps with a time width of
0.006 fs (top) to 1.375.000 steps (at 0.003 fs) (middle) and 2.750.000 steps (at 0.0015 fs) (bottom).
Changes in the probability distribution of the ions spatial location (red curve at the bottom of
the inserts) are not detectable, demonstrating the robustness and stabilitiy of the Crank-Nicolson
method for solving equations of the type ut = a.uxx (e.g. the Schrodinger equation) as used in
the present study.
6
Figure 5: QM-wave-packet distribution of one ion centered at location S1 interacting with two
classical ions KA and KB that enter the filter region. The orientation of the filter atoms is from
inside the cell (left) to outside the cell (right) according to the scheme shown in Figure 2(C).
The wave-packet distribution of the ion KQ at location S1 is shown within dark stripes in order
to visualize the emerging interference (coherence) pattern of its wavefunction. The confining
potentials within these stripes are demonstrated in the following Figure 6 for four different time
step situations, labeled by (a-d). The circles next to the ions, with two rows above and two rows
below lining the filter, symbolize the location of C atoms (outside) and O atoms (inside) that
belong to the carbonyl C=O groups of the filter. Note the distortion of C=O dipoles induced by
the presence of the KQ ion in S1
ions at location S1-S3 and develops into ion configuration S2-S4 with the outermost quantum-
mechanically calculated ion close to S0. The snapshot shown in the following Figure 5 and Figure
6 cover a time-window of just 4.124 ps.
The time sequence of frames (a-d) in Figure 5 and their matching potentials (a-d) in Figure 6
cover the ’conductive’ situation with one ion (KB) approaching the filter region from the cavity
site, another ion (KA) located at S3 and the wave-packet of the ’quantum-ion’ (KQ) centered at
S1. During this time (KQ) oscillates as a wave around it’s center due to the thermal energy of
the ion. In Figure 2 (b) the external ion exerts a repulsive ’knock-on’ effect on the (K A). This
external knock-on is also felt by the quantum-ion ( K Q) as it now extends it’s wave-function
with some probability spreading into location S0. This non-local effect covers a z-extension in
the order of 10−10m. Finally, as seen in the frame Figure 5(d) the classical filter ion has reached
position S2 inducing a strong non-local effect on the spread of the wave-function of the quantum
ion, originally located at S1. There is now a considerable probability extension of 1nm (about
the entire linear z-extension of the filter) for the amplitude of ion KQ, with a spread into S0 and
into the extra-cellular site of the filter.
7
Figure 6: Reconstructed dynamics of potentials along the filter locations S1 and S0 that results
from the movement of 2 (classical) ions KA, KB and one ’quantum-mechanical ’ ion KQ during
four consecutive snapshots (a-d) with locations as shown in Figure 2C. The potentials are shown
within the range of 0.1 to 0.5 eV. The x-axis is provided by the z -coordinates of the filter region
as shown in Figure 2B. It should be noted that the confining potentials for the ions are not
pre-set, but the potential distribution emerges from the Coulomb interaction of ions with the
charges provided by the lining oxygen atoms and the other ions present in the filter following
the calculations by equations (2.2 - 2.4)
8
Q i ion and a N a+
Figure 7: Mean kinetic energy of a K +
Q ion as a function of oscillation frequencies
of filter C=O groups during 9 ps of simulation. The replacement of the ’intrinsic’ K +
Q ion with a
N a+
Q ion with different coordination distances to the surrounding atoms becomes reflected by a
different kinetic energy minimum along the oscillation frequency of carbonyl dipoles. The loss of
kinetic energy (’cooling’) is less for the replaced atom species and requires a different oscillation
frequency of C=O groups.
3.2 Non-conductive states
For the non-conductive states of the filter we assume that the classical ions occupy the cavity
position and the filter position S3 initially. The quantum-ion remains in position S1 (the so-called
S1-S3 filter configuration). We now allow the interacting C=O groups to oscillate thermally. This
will in turn induce an oscillation of local potentials around the quantum-ion at S1.
Cooling effects:
It can be expected that the translocation of ions, followed by a transient binding into a local
potential, will generally lead to a loss of energy of this ion and a subsequent ’cooling effect’.
In an attempt at understanding the role of the filter structure for these possible effects, we
have made a series of measurements on the dependence of mean kinetic energies on different
oscillation frequencies of the coordinating C=O groups. The initial settings placed the (KQ) ion
into S1 with a kinetic energy of 20 meV (1 kT = 26 meV at room temperature), the wave-packet
extension was preset to 0.025 nm and the oscillatory amplitude of C=O to 0.02 nm. We followed
the evolution of the K Q wave-function during a period of 9 ps at intervals of 0.006 fs (over 1.5
million steps).
In Figure 7 we show the mean kinetic energy of this ion at the end of the 9
ps cooling period as a function of the oscillation frequency of the C=O groups. The temporal
development of the kinetic energies over the time window of 9 ps is demonstrated in Figure 8.
From Figure 8 it is apparent that at optimal oscillation frequencies of C=O groups (open
circles in Fig.7) the kinetic energy approaches a minimum after a quarter cycle of the oscillation
period. At this point the potential energy reaches a maximum. However, because the potential
sink follows the movement of the ion, this potential energy is only partly returned into the kinetic
9
Figure 8: The temporal evolution of kinetic energies of the K +
Q ion as shown in Figure 7 at site
S1 for three different oscillation frequencies: unbroken line = no oscillations, broken line = 588
GHz oscillations and circles = the ’optimal frequency’ at 446 GHz with maximal cooling effects.
The time evolution of these curves shows that only for the oscillation frequency that provides a
clear cooling effect (Figure 7) at 446 GHz, the potential energy is not returned to the kinetic part
during the next oscillation period. Thus the concerted quantum-mechanical behavior of the ion
together with the coordinating carbonyl dipoles can account for this unique selection procedure
at the atomic scale.
10
term during the next quarter of the oscillation. This process continues until the kinetic energy of
the ion approaches a minimum, depending on the frequency of the associated C=O oscillations.
Further, particular oscillation frequencies of the ion-carbonyl environment and the associated
energy dissipation can discriminate different ion species for a given atomic filter environment.
Whereas a primary ion selection effect seems to be associated by different ion-species dependent
dehydration energies during the ions transfer from the cavity to the selectivity filter [17], the
present findings are strongly suggestive for an additional role of quantum oscillatory effects
within the filters atomic environment on ion selectivity. This points to a clear quantum effect
that is indispensible in explaining ion selectivity, a basic and indispensible biological function of
ion channel proteins. As expected, our results show that if an N a+ ion is placed in the same
position as a K + , it also can transfer energy to the C=O lattice (see Figure 7). However, since
the N a+ ion has a lower mass than the K + ion, the most efficient cooling occurs at a higher
oscillation frequency of the C=O groups. The natural oscillation frequency of the C=O groups
caused by thermal motion of the surrounding atoms is essentially determined by the binding
strength of the C=O groups to the back- bone of the protein strands. We can therefore expect
this oscillation to exert its cooling effect in the most efficient way for ions with a specific charge
and a specific mass. When hopping from one site to a neighboring site in the selectivity filter
only a particular ion species will be captured optimally by the coordinating C=O cage. However,
because the exchange of energy works both ways, the con-specific ion will also be able to extract
energy from the coordinating C=O group within a shorter time than any other ion species.
This in turn will allow shorter times for bridging the energy barrier to the neighboring site.
Further simulations will be required to test whether this effect can lead to a possible resonance
phenomenon accompanying a fast passage through the selectivity filter.
4 Conclusions
We have shown that, depending on the surrounding carbonyl derived potentials, alkali ions in
the filter region of ion channels lined by the conserved TVGYG amino acid sequence can become
highly ’quantum-delocalized’ depending on ion location and oscillatory motion of the coordinating
C=O groups. Further, we find that due to the interaction of ions with a time-dependent potential,
the energy of ions is not conserved. Depending on the frequencies of oscillating C=O surroundings
the ion can lose up to 1/2 of its kinetic energy, exerting a substantial ’cooling effect’. The transfer
of this energy to the C=O environment is shown to depend on a specific C=O resonant oscillation
frequency and within the present study is observed under a ’non-conductive’ filter state for an ion
at location S1. In the frame of previous work on energy transfer systems in proteins, quantum
and classical [7, 20, 21]we suggest that the observed effect will cool down the ions inside the filter
giving off their energy to the environment, i.e. to the so-called P-loop domain of the lining filter
region. Energy transfer along the P-loop backbone atoms can in turn change the symmetry of the
filter entrance [3, 9, 13]. As a result it was anticipated that even the location of a single ion in the
filter can become associated with the gating state of the selectivity filter [3, 9]. Here we provide
evidence for this situation based on a quantum-mechanical calculation and offer predictions about
the involved energies and resonant oscillations and their effect on channel filter states. These
results indicate that temporal development of quantum states in channel proteins can propagate
into classical ion-channel conformations that determine the electrical signal properties of neuronal
membranes. Because the filter states must be correlated with the pore-domain gating state of the
channel in order to facilitate the co-occurrence of conduction and access of ions to the channel
protein [24], the pore gating states of the channels can be interpreted as ’classical witness states’
[25] of an underlying quantum process in the brain, as suggested previously [4, 5, 19].
11
References
[1] Berneche S, Roux B, Energetics of ion conduction through the K channel, Nature 414:
73-77, 2001.
[2] Berneche S, Roux B, A microscopic view of ion conduction through the K channel, PNAS
100: 8644-8648, 2003.
[3] Berneche S, Roux B, A gate in the selectivity filter of potassium channels, Structure 13:591-
600, 2005.
[4] Bernroider G, Roy S, Quantum-classical correspondence in the brain: scaling, action dis-
tances and predictability behind neural signals, Forma 19: 55-68, 2004.
[5] Bernroider G, Roy S, Quantum entanglement of K + ions, multiple channel states, and the
role of noise in the brain, SPIE 5841: 205-213, 2005.
[6] Bostick DL, Brooks III CL, Selectivity in K channels is due to topological control of the
permeant ion’s coordinated state, PNAS 104:9260-9265, 2007.
[7] Briegel HJ, Popescu S, Entanglement and intra-molecular cooling in biological systems? -
A quantum thermodynamic perspective, arXiv :0806.4552v2 [quant-ph], 2009.
[8] Bucher D, Raugei S, Guidoni L, Dal Peraro M, Rothlisberger U, Carloni P, Klein ML,
Polarization effects and charge transfer in the KcsA potassium channel, Biophys.Chemistry
124:292-301, 2006.
[9] Cordero-Morales JF, Cuello LG, Perozo E, Voltage-dependent gating at the KcsA selectivity
filter, Nature Structural & Molecular Biology 13:319-322, 2006.
[10] Corry B, Understanding ion channel selectivity and gating and their role in cellular sig-
nalling, MolBioSyst2:527-535, 2006.
[11] Garofoli S, Jordan PC, Modelling permeation energetics in the KcsA potassium channel,
Biophys J 84:2814-2830, 2003.
[12] Guidoni L, Carloni P, Potassium permeation through the KcsA channel: a density functional
study Biochim Biophys Acta 1563:1-6, 2002.
[13] Kona J, Minozzi M, Torre V, Carloni P, A gate mechanism indicated in the selectivity filter
of the potassium channel KscA, Theor Chem Acc 117:1121-1129, 2007.
[14] Kuyucak S, Andersen OS, Shin-Ho Chung, Models of permeation in ion channels, Rep Prog
Phys 64: 1427-1472, 2001.
[15] MacKinnon R, Potassium channels, FEBS Lett 555:62-65, 2003.
[16] Mahler J, Persson I, A study of the Hydration of the Alkali Metal Ions in Aqueous Solutions,
Inorg Chem (online pre publication),2011.
[17] Rempe SB, Pratt LR, Ion Hydration Studies Aimed at Ion Channel Selectivity, Los Alamos,
Special Feature April 2000, Theoretical Chemistry & Molecular Physics T-12, 2000.
[18] Roux B, Schulten K, Computational studies of membrane channels, Structure 12:1343-1351,
2004.
12
[19] Roy S, Llinas R, Relevance of quantum mechanics on some aspects of ion channel function,
C.R. Biologies 332:517-522, 2009.
[20] Salari V, Summhammer J, Bernroider G, Coherent excitation energy transfer (EET) along
the P-loop of KcsA ion channel models: Feasibility and possible effects on filter gating,
QuEBS2010, Abstracts, 2010.
[21] Salari V , Tuszynski J, Rahnama M, Bernroider G, Plausibility of quantum coherent states
in biological systems, J Phys: Conf Series 306:012075, 2011.
[22] Soper AK, Weckstrom K, Ion Solvation and Water Structure in Potassium Halide Aqueous
Solitions. Biophys Chem 124: 180-191,2006.
[23] Thomas JW, Numerical Partial Differential Equations: Finite Difference Methods, in Text-
books in Applied Mathematics, 22, Berlin, New York , 1995.
[24] VanDongen AM, K channel gating by an affinity-switching selectivity filter, PNAS 101:
3248-3252, 2004.
[25] Vedral V, Quantifying entanglement in macroscopic systems, Nature 453:1004-1007,2008.
[26] Zhou Y, Morais-Cabral JH, Kaufman A, MacKinnon R, Chemistry of ion coordination and
hydration revealed by a K + channel-Fab complex at 2.0 A resolution, Nature 414: 43-48,
2001.
13
|
1711.01423 | 1 | 1711 | 2017-11-04T10:45:14 | What can neuronal populations tell us about cognition? | [
"q-bio.NC"
] | Nowadays, it is possible to record the activity of hundreds of cells at the same time in behaving animals. However, these data are often treated and analyzed as if they consisted of many independently recorded neurons. How can neuronal populations be uniquely used to learn about cognition? We describe recent work that shows that populations of simultaneously recorded neurons are fundamental to understand the basis of decision-making, including processes such as ongoing deliberations and decision confidence, which generally fall outside the reach of single-cell analysis. Thus, neuronal population data allow addressing novel questions, but they also come with so far unsolved challenges. | q-bio.NC | q-bio | What can neuronal populations tell us about cognition?
Iñigo Arandia-Romero1, Ramon Nogueira1, Gabriela Mochol1, and Rubén Moreno-Bote1,2
1 Center for Brain and Cognition & Department of Information and Communications
Technologies, University Pompeu Fabra, 08018 Barcelona, Spain
2Serra Húnter Fellow Programme, 08018 Barcelona, Spain
Corresponding author:
Rubén Moreno-Bote
Email: [email protected]
Contributions: All authors discussed and wrote the paper
Acknowledgements: This work has been supported by the Spanish PSI2013-44811-P and
SlowDyn FLAGERA-PCIN-2015-162-C02-02 grants from MINECO to R. M. B., and IJCI-
2014-21937 grant from MINECO to G. M.
Highlights
Neuronal population
(NP)
level analysis
is slowly pervading systems
neuroscience
NPs provide both unprecedented temporal and spatial resolution to study
decision-making
However, large NPs come with challenging data analysis and interpretative
problems
Challenges can be alleviated by using models of NP activity
Abstract
Nowadays, it is possible to record the activity of hundreds of cells at the same time in
behaving animals. However, these data are often treated and analyzed as if they
consisted of many independently recorded neurons. How can neuronal populations be
uniquely used to learn about cognition? We describe recent work that shows that
populations of simultaneously recorded neurons are fundamental to understand the
basis of decision-making, including processes such as ongoing deliberations and decision
confidence, which generally fall outside the reach of single-cell analysis. Thus, neuronal
population data allow addressing novel questions, but they also come with so far
unsolved challenges.
Introduction
Single-neuron electrophysiology has provided golden ages in neuroscience passing from the
discoveries of Lord Adrian [1] to those by Hubel and Wiesel [2]. This approach relies on
finding mappings between the activity of a single neuron and external world, body or internal
variables, most often summarized with the so-called 'tuning curve', that is, the mean neuron
response as a function of the relevant variable. As such, single-cell electrophysiology has
proven to be a central pillar over which many theories of the brain rest [2-8].
In the last 20 years, there has been a surge of new technologies that allow recording
the activity of hundreds of cells within and across brain areas, either with dense electrodes or
with imaging techniques [9-16] and pioneering work has shown the benefits of these neuronal
population (NP) recordings [17-22]. New emerging technologies even promise several orders
of magnitude improvement in the number of recorded cells, leaping from hundreds to
thousands of cells [23, 24]. However, these promises come with new problems. The most
prominent challenges arise in experimental settings that involve animal behavior, where the
number of trials is limited to avoid animal satiation and exhaustion, and to minimize
overtraining to better study naturalistic behaviors. Considering the activity of each neuron as
one dimension in the neuronal state space, how should we analyze and model multi-
dimensional neuronal activity with just a few tens of trials per condition? How can we benefit
from the many dimensions that NP offers? And, most importantly, what are the new
questions that can be addressed with NPs about cognitive processes that cannot, or at least
hardly, be asked with single neurons?
We review recent literature that can help to answer the above questions with the focus
on decision-making. We provide examples of a better understanding of the transformation
from stimulus to choices by using the activity of simultaneously recorded cells. We argue
about the difficulty, and even impossibility, of drawing qualitatively similar conclusions
using single-cell electrophysiology. Then, we talk about the challenges that we face with
large NPs and limited number of trials, as those expected from experiments that combine
both animal behavior and NPs recordings. Finally, we point to past and more recent work
that could benefit from using NP recordings and discuss possible promising future directions.
Advantages of NPs
The success of NP approaches has been clearly exemplified in brain-machine interface
applications [25-28]. Recent studies have shown how it is possible to restore hand control in a
quadriplegic patient [29] or aid non-human primates in walking [30] by reading out the
population activity of motor cortex and converting these signals into motor commands
bypassing injured nerves. 'Reading out' or 'decoding' means computing an online estimate of
the value of an external world or internal variable from the activity of the recorded neurons
[31-33]. For instance, decoding the population activity of V1 neurons with linear classifiers
can predict with high accuracy the orientation of a stimulus drifting grating in a trial-by-trial
basis [20, 34, 35]. Recent literature tends to favor approaches based on decoding over
information theory [19, 20, 34, 36-43], mostly because decoding techniques can be applied
with relatively small number of trials while information-theoretic methods require computing
probability distributions over as many dimensions as neurons in the NP, a problem daunted
by the curse of dimensionality [33]. In addition, decoding techniques provide trial-by-trial
estimates of external and internal variables that can be directly interpreted and compared to
their actual values [44].
Despite the success of NP approaches to understand the basics of attention [36],
working memory [45] and decision-making [37, 38, 46], they are sometimes contemplated
with suspicion. Indeed, in some applications it could be argued that NPs do not offer radically
new knowledge compared to single-neuron approaches. For instance, while decoding analysis
can provide better estimates of information content in a NP by taking into account the trial-
by-trial correlated noise among cells, rough estimates can be obtained by extrapolating the
information of singles neurons to small NPs of a few tens of cells [47] (but not for
populations larger than 100-1000 cells due to the presence of 'differential correlations' [48]).
Moreover, research based on NP recordings sometimes ends up using traditional single-
neuron-based analysis [49, 50]. What are the unique opportunities that NPs can offer? Here
we give examples of how NPs can provide new knowledge that is very hard - and in some
cases impossible - to acquire from single-neuron-based analysis.
There are at least two broad families of problems that cannot be addressed with
single-neuron approaches: 1) internal neuronal dynamics that is not time-locked to observable
stimulus or body variables, and 2) variables that are encoded at the population level and not
at the single-neuron level. The first category relates to neuronal processing that is internally
or externally triggered without experimental control. These phenomena might include trial-
by-trial fluctuations of attention that are not cued experimentally, mind wandering, and any
other neuronal processing that is not time-locked to the stimulus or body variables [34, 36,
37, 39, 41, 45, 51-63]. Because of their lack of significant correlation to observable
experimental variables, these phenomena cannot be fully studied using single-neuron
approaches. The second category relates to neuronal processing that requires coordination of
neurons in a NP. For instance, stable representation of a world variable, such as world-
centered body location of mice [64], can occur despite unstable representation of that
variable at the single-cell level (Fig. 1a). Similarly, theoretical work shows that stable
representation of working memory signals can be achieved in NPs despite unstable single-
neuron representations [65]. Thus, observing single neurons can lead to the erroneous
conclusion that the encoding of the variable is unreliable, while the reality is different,
namely, the variable is reliably encoded at the NP level. Further, theoretical studies suggest
that the encoding of variables can involve temporal delays between neurons without
concomitant changes of activity in any single cell [66] (Fig. 1b), and conclusions on
information content drawn from combining single cells recordings can be misleading if there
is shared response variability among cells in the NP [32, 48] (Fig. 1c).
Opportunities from NPs to understand decision-making
NPs can provide new ways to study the processes underlying decision-making and cognition.
These novel opportunities with NP arise because crucial decision-making variables might not
be generally representable by single neurons. . An example of such a variable is 'decision
confidence' [67-71], that is, the confidence of a decision maker about her choice being
correct based on accumulated noisy evidence [68]. A large family of mathematical models for
two-alternative decision-making tasks called 'race models' requires at least two types of
neurons racing to generate the choice [67, 68]. Here each racing neuron represents at each
time point the evidence for each option accumulated so far (Fig. 2). For this family of
models, decision confidence is encoded in the difference between the activity of two racing
neurons (in addition to elapsed time), the so-called 'balance of evidence', a quantity that
cannot be represented by the activity of neither of the neurons alone [68, 72]. The activity of
any single racing neuron has little information about confidence. For example, it is possible
to have no change in firing rate with confidence (Fig. 2a,b, cell 1) or even increase in firing
rate (Fig.2, cell 2) for lower confidence (lower panels of Fig 2a vs Fig. 2b). More realistic
models based on probabilistic population codes [73], which incorporate the Poisson-like
nature of neuronal responses, also represent both choices and confidence simultaneously in
the difference in activity of different neuronal pools [74].
Recent experimental work has shed new light into how changes of mind might be
encoded in neuronal populations during decision-making, a process that can also guide
further studies in the encoding of decision confidence on NPs. Kiani and colleagues recorded
the activity of large NPs in area 8Ar of prearcuate gyrus while macaque monkeys performed
a motion discrimination task [37]. The authors used decoding techniques to find a hyperplane
in multidimensional activity space that separated well animals' choices on a trial-by-trial
basis (Fig. 3a). Importantly, the distance of the population activity vector from the hyperplane
could be used to infer in which trials the animal 'changed its mind' (Fig 3b). This method
allows studying the ongoing process of choice at unprecedented levels of temporal resolution.
Decision confidence theories predict that confidence should be related to the same NP
activity distance, but these predictions remain to be tested.
A recent paper by Rich and Wallis has studied the deliberation process in economic-
based choices at a surprising level of temporal resolution using up to 16 electrodes in areas 11
and 13 in the monkey orbitofrontal cortex [38]. Following a smart experimental design, a
linear decoder learned to discriminate between different offers presented on forced-choice
single-offer trials, showing how offers were encoded in the NP when they were presented
separately. This decoder was then used without retraining to predict the internal deliberation
process in free-choice trials where two offers were simultaneously available, uncovering
alternations over time between the states representing the available options (Fig. 3c). The
untested assumption in this work is that the representation of offers is the same in forced and
free choice trials. However, consistently with this assumption, longer total duration of the
state corresponding to the chosen offer was observed compared to the unchosen one. Thus,
the activity of NPs reveals a rich dynamic representation of choice options, which would be
very hard to obtain with single neurons. Similarly complex dynamics has been also reported
in other tasks and areas during decision-making [41, 45].
Challenges from NPs
The above examples illustrate the importance of NP analysis. But what is the NP size that we
ought to record from to address a given question? A priori, there is no limit to the desired
size: the larger, the better, although in some cases with just a bunch of simultaneously
recorded cells it is possible to address questions that are unfeasible with just a single cell, as
described above for relevant theories of decision confidence. As a rule of thumb, the larger
the NP, the more precisely the state of the neuronal network can be assessed [34, 53, 75], the
more accurately stimulus or body variables can be decoded [20, 34, 35], and the more
variables can be studied simultaneously [29, 30].
However, the promises of population-based analysis do not come free of challenges.
The most important challenge is model specification given large number of parameters and
limited number of trials. Consider, for instance, a population of 100 neurons with 200 trials
per distinct condition and a binary classification task (400 trials in total). Using a quadratic
discriminant for this task requires estimating a number of parameters that scales as the square
of the number of neurons, around 5 thousand parameters, which is outside the reach of this
limited amount of data. A linear classifier could be still used in this condition because 400
trials is in general sufficient to estimate its 101 free parameters. However, consider now a
population of 1000 neurons with again 200 trials per condition and the same binary
classification task. In this case, even for a linear classifier, we need to estimate 1001
parameters, which is above the number of data points collected. Roughly speaking and based
on our experience, we would need at least around 5 times more trials per class and per neuron
recorded to train a linear classifier (5neuronsclasses), which for large NPs will certainly
be difficult to achieve. The problem is even bigger for more complex decoders with many
more free parameters, such as deep feedforward networks [76]. Deep learning approaches
that have boosted machine learning in the last few years can be hardly suitable for neuronal
data based on a few trials per neuron and condition [76]. Novel approaches that would work
with standard 'few-trials' data need to be developed.
What approaches are then suitable for 'few-trials' data? This is a domain ripe for
research, and here we discuss two recently used methods. The first one is to use linear
decoders together with regularization techniques [37, 52, 77]. In this way, parameters that
provide none or little explanatory power are pushed to zero or very small values, therefore
reducing the complexity of the decoder and thus dampening overfitting. Regularization has
been shown to be useful for NPs approximately as large as 150 neurons with more than 500
trials per condition [37]. However, a possible danger is that regularization implicitly
introduces priors, which constraint the family of solutions, but might turn to be wrong. The
second method consists of building models with increasing complexity starting from basic
principles and performing model comparison using cross-validation techniques [34, 78, 79].
Basic principles refer here to the knowledge that has been acquired across many years of
research. For instance, knowledge about translational invariance of real-world objects
incorporated into convolutional networks allows to significantly reduce their number of free
parameters, thus improving dramatically their classification performance [76]. Another
example of prior knowledge that can be included in decoders is the Poisson-like nature of
neuronal responses [4, 73, 80-83] and the gain modulations induced by global fluctuations in
population activity [34, 51, 54, 75, 78, 84]. For instance, a recent model of NP activity with
Poisson-like firing and both heterogeneous multiplicative and additive gains has been shown
to explain better macaque V1 responses than alternative models ('multi-gain' model [34])
(Fig. 4). The first important advance is that this analysis directly exploits NP data, as
otherwise it is impossible to study the modulation of single-neuron activity with global gain.
Secondly, the resulting model has culminated research that has shown that global gain
modulations can be either purely additive [51, 75], purely multiplicative [84], homogeneous
multiplicative and additive [54], or a mixture of homogeneous multiplicative and
heterogeneous additive gains [78] (Fig. 4).
Finally, it is important to avoid some potential pitfalls when using NPs. First, as larger
NPs are available, computing 'noise correlation' matrices (correlations between all pairs of
neurons at fixed stimulus conditions, [32, 85, 86]) will become unfeasible, because the
covariance matrix becomes singular when the number of neurons exceeds the number of
trials [87], unless strong priors are used [88]. There is also a tendency to build large
surrogate NPs from independently recorded single-neurons [63, 89, 90]. Surrogate NPs can
be useful in some special circumstances [63, 89-91], but overusing them could lead to
erroneous conclusions. This can be the case when neuronal populations contain 'differential
correlations', which can be very small and yet have a huge impact on information and
computations, and are absent in surrogate populations [48].
Conclusions: what can NPs offer in the future?
We have shown that NP analysis is slowly pervading neuroscience, but fulminating strides
are expected to occur soon. First, ongoing debates on the role of certain brain areas in
decision-making and on its neuronal mechanisms will dramatically benefit from measuring
the activity of very large NPs [92-95]. Second, decision-making is characterized by complex
behavior that in real-life conditions involve a myriad of variables, such as complex stimuli
and body variables. The many-fold dimensions of neuronal activity will allow studying the
simultaneous encoding of those variables in the brain. Dimensionality reduction techniques
[96], including decoders, will leave room to new methods (e.g., [97]) more suitable when the
complexity of the experimental design does not permit for such reductionist, low
dimensional, view of neuronal activity, as computational neural network models predict [98-
100]. Finally, it is also expected that new-generation decoders and other approaches suitable
for 'few-trials' datasets, typical for behavioral and systems neuroscience, will allow precise
estimates of sensory information from increasingly large NPs [48, 86].
Figures and legends
Figure 1: Neuronal populations (NPs) are necessary to understand the neuronal code.
(a) NPs can represent stable information while the tuning of individual neurons changes
[64]. The activity of cell 1 (red) and cell 2 (blue) for different positions can change from the
first day (top-left) to the tenth day (top-right), while the information represented by the NP
remains constant (bottom). The ordering of the cells by preferred position differs from the
first (bottom-left) to the tenth (bottom-right) day due to changes in individual tuning, but
jointly they represent the same map. To recover the information represented by the NP a
readout neuron (black circle at the top) adapts the weights (𝑤𝑖-s) to the tuning of each neuron
(red and blue circles) at a given day. The information recovered across days is the same.
(b) Stimulus information in theory can be encoded in the relative spike timing between
neurons [66]. Vertical bars represent the spiking activity of cell 1 (red) and cell 2 (blue) in
three cases while two stimuli (orange and apple) are presented consecutively (top). The first
two cases show the activity recorded from single cells. Here, the readout neuron (black circle
on the right) is not able to recognize the stimulus that has been presented. In the third case, in
contrast, the readout neuron can measure the difference in the timing of spikes from cell 1
and cell 2, and recover the stimulus by using the relative spike latencies.
(c) Pair-wise correlations can help decoding [32]. Each plot shows the mean activity (crosses)
and the variability in the response (standard deviation represented by clouds) of two neurons
when two different stimuli are presented (orange and green). The responses in the left panel
have circular clouds, meaning that the variability is the same in all directions, and therefore
the correlation between the two neurons is zero for both stimuli. The panel in the right shows
two correlated neurons with the same amount of variability as in the left panel, but spread
along the same direction (ellipses). The red line marks the decoding boundary, dividing the
space between the area which corresponds to the orange stimulus and to the green one. The
clouds overlapping in the left panel mean that in some trials where the green stimulus has
been presented the decoder will predict the orange one, and vice-versa. However, for the
correlated activity case (right) most of the trials are correctly classified, as the clouds do not
overlap. Therefore, predicting the performance of a neuronal code depends on neuronal
correlations, a problem that lies beyond single-neuron electrophysiology.
Figure 2: Decision confidence in a 'race model'.
(a) A decision maker choses between two offers (green apple and orange), whose values are
represented by the activity of two neurons (red and green lines), growing over time as more
evidence is accumulated. The decision rule is based on what option depicts the highest
activity at decision time (dashed vertical line). In this case, the choice was easy because the
values of the offers were very distinct, and therefore the confidence on the decision
(represented by the distance between the activity of the neurons) was high [68].
(b) When the two offers have similar values (red vs green apple), the choice is more difficult,
and thus decision confidence (distance between the races) is lower. The estimation of
confidence would not be possible if the two neurons were not recorded simultaneously.
Comparing the high and low confidence cases (a-b) the activity of the neuron preferring the
green apple (cell 1) is identical, and thus this neuron alone does not correlate with decision
confidence at the time of the decision. Although the neuron that prefers the lower value offer
changes its activity in the two cases, the direction of this change is opposite to the change in
confidence (cell 2 has higher rate for the low confidence trial) and cannot give a good
approximation of the latter. nor provide information about the state of the other racing
neuron, making confidence estimation impossible from single neuron recordings.
Figure 3: Decision making in neuronal response space.
(a) The dynamics of decision making of an animal performing a two-alternative task can be
visualized in the neuronal response space, that is, the space where each direction corresponds
to the activity of a recorded neuron. At a given time window the population activity is a
single point in this space (blue and red dots) A linear decoder is trained to find a hyperplane
(in this case, a plane in three dimensions) which best separates the points belonging to the
two classes (red for T1 choice, and blue for T2 choice). At the beginning of each trial (faded
colors) the population activity provides weak evidence about stimulus conditions, and thus
the points are close to the hyperplane. As more evidence is accumulated, the population
activity diverges away from the hyperplane, making easer to predict the animal's choice. The
distance between population activity and the classification hyperplane defines the decision
variable (DV, black line).
(b) The temporal dynamics of DV reveals 'changes of mind'. In a given trial, the DV crossed
the ambiguity point equal to zero (indicated by an arrow), suggesting the presence of a
change of mind, that is, an instance in which the hypothetical intended choice flips (two trials
are shown). Tracking these changes of mind would not be possible with single-neuron
analysis. Reproduced from [37] with the permission of the publisher.
(c) Ongoing deliberation inferred NP activity in monkey orbitofrontal cortex. A linear
decoder, trained in forced-choice trials, was used to predict choices between simultaneously
presented offers in free-choice trials. The posterior probability derived from the decoder for
chosen (red), unchosen (blue) and unavailable options (gray, average of both unavailable
options), are shown in six different trials. This analysis allows tracking the ongoing
deliberation processes at unprecedented temporal resolution. Reproduced from [38] with the
permission of the publisher
Figure 4: Population activity models based on Poisson-like firing and global gain
modulation. A large variety of models have been recently used to characterize the activity of
NP including the effect of the population activity on single neuronal activity in primary visual
cortex (V1) [34, 75, 78, 84].
(a) Population activity fluctuates across trials for the same stimulus (left panels). The tuning
of a V1 neuron (right) is modulated with population activity (called here global gain), which
is estimated as the sum of spikes across all neurons excluding the activity from the neuron for
which tuning is being characterized; thus, this analysis cannot be done with single neurons, it
requires NPs. Trials were ranked from high (top left) to low (bottom left) global gain for each
stimulus condition. The activity of the selected neuron (green spike trains) was averaged
across either the top (red box) or bottom (pink box) 50th percentile of trials, and the averages
were plotted as a function of stimulus orientation (rightmost panel). The tuning was
modulated with global gain (red vs. pink lines). Points and error bars are mean responses and
s.e.m., respectively; lines are von Mises fits. Reproduced from [32] with the permission of
the publisher.
(b) Global gains can affect the tuning of V1 cells in a multiplicative (top-left) and/or an
additive (bottom-left) way. The modulatory effect of the population activity on the tuning
curves is heterogeneous, with different multiplicative (top-right) and additive factors
(bottom-right) for different neurons. ***: p < 0.001. Reproduced from [32] with the
permission of the publisher.(c) Models of NP activity including global gains. Neurons are
assumed to fire as independent Poisson processes (or variations of this, see [78, 84])
conditioned to gain factors and/or , which reflect modulatory factors in the NP. The
mean firing rate of each cell i in the NP is given by the function 𝑖 . The function 𝑖
determines the tuning of the cells when the gain factors are zero. The gains can produce
multiplicative effects on the tuning curves if 𝑖 is non-zero, and additive effects if 𝑖 is non-
zero (see panel b left)). A recent model for NP activity suggests heterogeneous multiplicative
and additive modulations (multi-gain model [34]). The multiplicative gain model assumes
that there is not additive gain and that the multiplicative gain is identical for all neurons
( 𝑖 𝑖 for all i). The affine model allows for heterogeneous additive gains, but the
multiplicative gain is still uniform for all neurons ( 𝑖 for all i). The multi-gain model
permits heterogeneous multiplicative and additive gains (it also uses ). The multi-
gain model provides better fits of V1 population activity than other models. Cross-validated
log-likelihood is shown.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
Adrian, E.D. and Y. Zotterman, The impulses produced by sensory nerve endings:
Part 3. Impulses set up by Touch and Pressure. J Physiol, 1926. 61(4): p. 465-83.
Hubel, D.H. and T.N. Wiesel, Receptive fields, binocular interaction and functional
architecture in the cat visual cortex. Journal of Physiology London, 1962. 160: p.
106-154.
Wilson, F.A., S.P. Scalaidhe, and P.S. Goldman-Rakic, Dissociation of object and
spatial processing domains in primate prefrontal cortex. Science, 1993. 260(5116): p.
1955-8.
Shadlen, M.N. and W.T. Newsome, The variable discharge of cortical neurons:
implications for connectivity, computation, and information coding. J Neurosci, 1998.
18(10): p. 3870-96.
Desimone, R., Face-selective cells in the temporal cortex of monkeys. J Cogn
Neurosci, 1991. 3(1): p. 1-8.
Rizzolatti, G. and L. Craighero, The mirror-neuron system. Annu Rev Neurosci, 2004.
27: p. 169-92.
O'Keefe, J. and J. Dostrovsky, The hippocampus as a spatial map. Preliminary
evidence from unit activity in the freely-moving rat. Brain Res, 1971. 34(1): p. 171-5.
Miyashita, Y., Inferior temporal cortex: where visual perception meets memory. Annu
Rev Neurosci, 1993. 16: p. 245-63.
Maynard, E.M., C.T. Nordhausen, and R.A. Normann, The Utah intracortical
Electrode Array: a recording structure for potential brain-computer interfaces.
Electroencephalogr Clin Neurophysiol, 1997. 102(3): p. 228-39.
Kelly, R.C., et al., Comparison of recordings from microelectrode arrays and single
electrodes in the visual cortex. J Neurosci, 2007. 27(2): p. 261-4.
Blanche, T.J., et al., Polytrodes: high-density silicon electrode arrays for large-scale
multiunit recording. J Neurophysiol, 2005. 93(5): p. 2987-3000.
Csicsvari, J., et al., Massively parallel recording of unit and local field potentials with
silicon-based electrodes. J Neurophysiol, 2003. 90(2): p. 1314-23.
Stosiek, C., et al., In vivo two-photon calcium imaging of neuronal networks. Proc
Natl Acad Sci U S A, 2003. 100(12): p. 7319-24.
Shoham, D., et al., Imaging cortical dynamics at high spatial and temporal resolution
with novel blue voltage-sensitive dyes. Neuron, 1999. 24(4): p. 791-802.
Buzsaki, G., Large-scale recording of neuronal ensembles. Nat Neurosci, 2004. 7(5):
p. 446-51.
Kerr, J.N. and W. Denk, Imaging in vivo: watching the brain in action. Nat Rev
Neurosci, 2008. 9(3): p. 195-205.
Schoenbaum, G. and H. Eichenbaum, Information coding in the rodent prefrontal
cortex. II. Ensemble activity in orbitofrontal cortex. J Neurophysiol, 1995. 74(2): p.
751-62.
18. Warland, D.K., P. Reinagel, and M. Meister, Decoding visual information from a
19.
20.
21.
population of retinal ganglion cells. J Neurophysiol, 1997. 78(5): p. 2336-50.
Nikolic, D., et al., Distributed fading memory for stimulus properties in the primary
visual cortex. PLoS Biol, 2009. 7(12): p. e1000260.
Graf, A.B., et al., Decoding the activity of neuronal populations in macaque primary
visual cortex. Nat Neurosci, 2011. 14(2): p. 239-45.
Aertsen, A.M., et al., Dynamics of neuronal firing correlation: modulation of
"effective connectivity". J Neurophysiol, 1989. 61(5): p. 900-17.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
Riehle, A., et al., Spike synchronization and rate modulation differentially involved in
motor cortical function. Science, 1997. 278(5345): p. 1950-3.
Ahrens, M.B., et al., Brain-wide neuronal dynamics during motor adaptation in
zebrafish. Nature, 2012. 485(7399): p. 471-7.
Kim, T.H., et al., Long-Term Optical Access to an Estimated One Million Neurons in
the Live Mouse Cortex. Cell Rep, 2016. 17(12): p. 3385-3394.
Lebedev, M.A. and M.A. Nicolelis, Brain-machine interfaces: past, present and
future. Trends Neurosci, 2006. 29(9): p. 536-46.
Andersen, R.A., et al., Cognitive neural prosthetics. Trends Cogn Sci, 2004. 8(11): p.
486-93.
Collinger, J.L., et al., High-performance neuroprosthetic control by an individual with
tetraplegia. Lancet, 2013. 381(9866): p. 557-64.
Sadtler, P.T., et al., Neural constraints on learning. Nature, 2014. 512(7515): p. 423-
6.
Bouton, C.E., et al., Restoring cortical control of functional movement in a human
with quadriplegia. Nature, 2016. 533(7602): p. 247-50.
Capogrosso, M., et al., A brain-spine interface alleviating gait deficits after spinal
cord injury in primates. Nature, 2016. 539(7628): p. 284-288.
Abbott, L.F., Decoding neuronal firing and modelling neural networks. Q Rev
Biophys, 1994. 27(3): p. 291-331.
Averbeck, B.B., P.E. Latham, and A. Pouget, Neural correlations, population coding
and computation. Nat Rev Neurosci, 2006. 7(5): p. 358-66.
Quian Quiroga, R. and S. Panzeri, Extracting information from neuronal populations:
information theory and decoding approaches. Nat Rev Neurosci, 2009. 10(3): p. 173-
85.
**34. Arandia-Romero, I., et al., Multiplicative and Additive Modulation of Neuronal
Tuning with Population Activity Affects Encoded Information. Neuron, 2016. 89(6): p.
1305-16.
This work demonstrates the influence of the population activity on the tuning of
individual monkey V1 neurons. Population activity enhances or reduces information
in small neuronal ensembles, depending on whether neurons are multiplicatively or
additively modulated by population activity.
Berens, P., et al., A fast and simple population code for orientation in primate V1. J
Neurosci, 2012. 32(31): p. 10618-26.
35.
*36. Cohen, M.R. and J.H. Maunsell, A neuronal population measure of attention predicts
behavioral performance on individual trials. J Neurosci, 2010. 30(45): p. 15241-53.
**37. Kiani, R., et al., Dynamics of neural population responses in prefrontal cortex
38.
indicate changes of mind on single trials. Curr Biol, 2014. 24(13): p. 1542-7.
Rich, E.L. and J.D. Wallis, Decoding subjective decisions from orbitofrontal cortex.
Nat Neurosci, 2016. 19(7): p. 973-80.
*39. Tremblay, S., et al., Attentional filtering of visual information by neuronal ensembles
40.
in the primate lateral prefrontal cortex. Neuron, 2014. 85(1): p. 202-15.
Adibi, M., et al., Population decoding in rat barrel cortex: optimizing the linear
readout of correlated population responses. PLoS Comput Biol, 2014. 10(1): p.
e1003415.
**41. Morcos, A.S. and C.D. Harvey, History-dependent variability in population dynamics
42.
during evidence accumulation in cortex. Nat Neurosci, 2016. 19(12): p. 1672-1681.
Sussillo, D., et al., Making brain-machine interfaces robust to future neural
variability. Nat Commun, 2016. 7: p. 13749.
43. Montijn, J.S., et al., Population-Level Neural Codes Are Robust to Single-Neuron
Variability from a Multidimensional Coding Perspective. Cell Rep, 2016. 16(9): p.
2486-98.
Churchland, A.K. and R. Kiani, Three challenges for connecting model to mechanism
in decision-making. Curr Opin Behav Sci, 2016. 11: p. 74-80.
44.
*45. Stokes, M.G., et al., Dynamic coding for cognitive control in prefrontal cortex.
46.
47.
Neuron, 2013. 78(2): p. 364-75.
Nogueira, R., et al., Lateral orbitofrontal cortex anticipates choices and integrates
prior with current information. Nat Commun, 2017. 8: p. 14823.
Ecker, A.S., et al., Decorrelated neuronal firing in cortical microcircuits. Science,
2010. 327(5965): p. 584-7.
**48. Moreno-Bote, R., et al., Information-limiting correlations. Nat Neurosci, 2014.
17(10): p. 1410-7.
In this study, it is characterized the type of neuronal pairwise correlations that limit
information. A particular pattern of trial-by-trial shared variability is generated when
the input layer is noisy itself, the so-called differential correlations, which make
information saturate for large neuronal ensembles.
Hosokawa, T., et al., Single-neuron mechanisms underlying cost-benefit analysis in
frontal cortex. J Neurosci, 2013. 33(44): p. 17385-97.
Sul, J.H., et al., Distinct roles of rodent orbitofrontal and medial prefrontal cortex in
decision making. Neuron, 2010. 66(3): p. 449-60.
Arieli, A., et al., Dynamics of ongoing activity: explanation of the large variability in
evoked cortical responses. Science, 1996. 273(5283): p. 1868-71.
Kiani, R., et al., Natural grouping of neural responses reveals spatially segregated
clusters in prearcuate cortex. Neuron, 2015. 85(6): p. 1359-73.
Okun, M., et al., Diverse coupling of neurons to populations in sensory cortex.
Nature, 2015. 521(7553): p. 511-5.
49.
50.
51.
52.
53.
*54. Mochol, G., et al., Stochastic transitions into silence cause noise correlations in
55.
56.
57.
cortical circuits. Proc Natl Acad Sci U S A, 2015. 112(11): p. 3529-34.
Scholvinck, M.L., et al., Cortical state determines global variability and correlations
in visual cortex. J Neurosci, 2015. 35(1): p. 170-8.
Jezzini, A., et al., Processing of hedonic and chemosensory features of taste in medial
prefrontal and insular networks. J Neurosci, 2013. 33(48): p. 18966-78.
Ponce-Alvarez, A., et al., Dynamics of cortical neuronal ensembles transit from
decision making to storage for later report. J Neurosci, 2012. 32(35): p. 11956-69.
**58. Engel, T.A., et al., Selective modulation of cortical state during spatial attention.
Science, 2016. 354(6316): p. 1140-1144.
This work demonstrates the influence of the population activity on the tuning of
individual monkey V1 neurons. Population activity enhances or reduces information
in small neuronal ensembles, depending on whether neurons are multiplicatively or
additively modulated by population activity.
Abeles, M., et al., Cortical activity flips among quasi-stationary states. Proc Natl
Acad Sci U S A, 1995. 92(19): p. 8616-20.
Seidemann, E., et al., 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, 1996. 16(2): p. 752-68.
Kaufman, M.T., et al., Cortical activity in the null space: permitting preparation
without movement. Nat Neurosci, 2014. 17(3): p. 440-8.
59.
60.
61.
*62. Kaufman, M.T., et al., Vacillation, indecision and hesitation in moment-by-moment
decoding of monkey motor cortex. Elife, 2015. 4: p. e04677.
*63. Mante, V., et al., Context-dependent computation by recurrent dynamics in prefrontal
cortex. Nature, 2013. 503(7474): p. 78-84.
*64. Ziv, Y., et al., Long-term dynamics of CA1 hippocampal place codes. Nat Neurosci,
65.
66.
67.
2013. 16(3): p. 264-6.
Druckmann, S. and D.B. Chklovskii, Neuronal circuits underlying persistent
representations despite time varying activity. Curr Biol, 2012. 22(22): p. 2095-103.
VanRullen, R., R. Guyonneau, and S.J. Thorpe, Spike times make sense. Trends
Neurosci, 2005. 28(1): p. 1-4.
Vickers, D., Decision processes in visual perception. 1979, New York: London
Academic Press. 406.
*68. Moreno-Bote, R., Decision Confidence and Uncertainty in Diffusion Models with
Partially Correlated Neuronal Integrators. Neural Computation, 2010. 22(7): p.
1786-1811.
*69. Kiani, R. and M.N. Shadlen, Representation of confidence associated with a decision
by neurons in the parietal cortex. Science, 2009. 324(5928): p. 759-64.
*70. Kepecs, A., et al., Neural correlates, computation and behavioural impact of decision
71.
72.
confidence. Nature, 2008. 455(7210): p. 227-31.
Drugowitsch, J., et al., The cost of accumulating evidence in perceptual decision
making. J Neurosci, 2012. 32(11): p. 3612-28.
Drugowitsch, J., R. Moreno-Bote, and A. Pouget, Relation between belief and
performance in perceptual decision making. PLoS One, 2014. 9(5): p. e96511.
73. Ma, W.J., et al., Bayesian inference with probabilistic population codes. Nat
74.
Neurosci, 2006. 9(11): p. 1432-8.
Beck, J.M., et al., Probabilistic population codes for Bayesian decision making.
Neuron, 2008. 60(6): p. 1142-52.
*75. Ecker, A.S., et al., State dependence of noise correlations in macaque primary visual
cortex. Neuron, 2014. 82(1): p. 235-48.
Goodfellow, I., Y. Bengio, and A. Courville, Deep Learning. 2016: MIT Press.
Bishop, C.M., Pattern recognition and machine learning. 2006, Singapore: Springer.
76.
77.
*78. Lin, I.C., et al., The Nature of Shared Cortical Variability. Neuron, 2015. 87(3): p.
80.
81.
79.
644-56.
Rabinowitz, N.C., et al., Attention stabilizes the shared gain of V4 populations. Elife,
2015. 4: p. e08998.
Softky, W.R. and C. Koch, The highly irregular firing of cortical cells is inconsistent
with temporal integration of random EPSPs. J Neurosci, 1993. 13(1): p. 334-50.
Tolhurst, D.J., J.A. Movshon, and A.F. Dean, The statistical reliability of signals in
single neurons in cat and monkey visual cortex. Vision Res, 1983. 23(8): p. 775-85.
Geisler, W.S. and D.G. Albrecht, Visual cortex neurons in monkeys and cats:
detection, discrimination, and identification. Vis Neurosci, 1997. 14(5): p. 897-919.
*83. Moreno-Bote, R., Poisson-like spiking in circuits with probabilistic synapses. PLoS
82.
Comput Biol, 2014. 10(7): p. e1003522.
*84. Goris, R.L., J.A. Movshon, and E.P. Simoncelli, Partitioning neuronal variability.
85.
86.
87.
Nat Neurosci, 2014. 17(6): p. 858-65.
Cohen, M.R. and A. Kohn, Measuring and interpreting neuronal correlations. Nat
Neurosci, 2011. 14(7): p. 811-9.
Kohn, A., et al., Correlations and Neuronal Population Information. Annu Rev
Neurosci, 2016. 39: p. 237-56.
Kanitscheider, I., et al., Measuring Fisher information accurately in correlated neural
populations. PLoS Comput Biol, 2015. 11(6): p. e1004218.
88.
Yatsenko, D., et al., Improved estimation and interpretation of correlations in neural
circuits. PLoS Comput Biol, 2015. 11(3): p. e1004083.
*89. Rigotti, M., et al., The importance of mixed selectivity in complex cognitive tasks.
*90.
91.
92.
93.
94.
95.
96.
97.
Nature, 2013. 497(7451): p. 585-90.
Ince, R.A., S. Panzeri, and C. Kayser, Neural codes formed by small and temporally
precise populations in auditory cortex. J Neurosci, 2013. 33(46): p. 18277-87.
Panzeri, S., et al., Neural population coding: combining insights from microscopic
and mass signals. Trends Cogn Sci, 2015. 19(3): p. 162-72.
Latimer, K.W., et al., Single-trial spike trains in parietal cortex reveal discrete steps
during decision-making. Science, 2015. 349(6244): p. 184-7.
Latimer, K.W., et al., Response to Comment on "Single-trial spike trains in parietal
cortex reveal discrete steps during decision-making". Science, 2016. 351(6280): p.
1406.
Shadlen, M.N., et al., Comment on "Single-trial spike trains in parietal cortex reveal
discrete steps during decision-making". Science, 2016. 351(6280): p. 1406.
Bollimunta, A., D. Totten, and J. Ditterich, Neural dynamics of choice: single-trial
analysis of decision-related activity in parietal cortex. J Neurosci, 2012. 32(37): p.
12684-701.
Cunningham, J.P. and B.M. Yu, Dimensionality reduction for large-scale neural
recordings. Nat Neurosci, 2014. 17(11): p. 1500-9.
Kobak, D., et al., Demixed principal component analysis of neural population data.
Elife, 2016. 5.
*98. Boerlin, M., C.K. Machens, and S. Deneve, Predictive coding of dynamical variables
in balanced spiking networks. PLoS Comput Biol, 2013. 9(11): p. e1003258.
99. Moreno-Bote, R. and J. Drugowitsch, Causal Inference and Explaining Away in a
Spiking Network. Sci Rep, 2015. 5: p. 17531.
100. Grabska-Barwinska, A., et al., A probabilistic approach to demixing odors. Nat
Neurosci, 2016. 20(1): p. 98-106.
|
1906.01767 | 1 | 1906 | 2019-06-05T00:41:06 | Codes, communication and cognition | [
"q-bio.NC"
] | Brette (2019) criticizes the notion of neural coding because it seems to entail that neural signals need to be decoded by or for some receiver in the head. If that were so, then neural coding would indeed be homuncular (Brette calls it dualistic), requiring an entity to decipher the code. But I think the plea of Brett to think instead in terms of complex, interactive causal throughput is preaching to the converted. Turing (not Shannon) has already shown the way. In any case, the metaphor of neural coding has little to do with the symbol grounding problem. | q-bio.NC | q-bio | Commentary on "Is coding a relevant metaphor for the brain?" (Brette 2019). To appear in Behavioral and Brain
Sciences.
Codes, communication and cognition
Stevan Harnad
Department of Psychology
Université du Québec à Montréal
and
Department of Electronics and Computer Science
University of Southamton
[email protected]
Abstract: Brette criticizes the notion of neural coding because it seems to entail that neural
signals need to be "decoded" by or for some receiver in the head. If that were so, then neural
coding would indeed be homuncular (Brette calls it "dualistic"), requiring an entity to decipher
the code. But I think Brette's plea to think instead in terms of complex, interactive causal
throughput is preaching to the converted. Turing (not Shannon) has already shown the way. In
any case, the metaphor of neural coding has little to do with the symbol grounding problem.
Both Shannon's (1948) information and Turing's (1936) computation are important in cognitive
science. Shannon is concerned with the faithfulness of signal transmission in communication
and Turing is concerned with what algorithms can do. Cognitive science is concerned with what
organisms (hence their brains) can do, and how.
Cells (including neurons) transmit signals. This is already true in plants (Baluska & Mancuso
2009) and of course also in machines. And organisms certainly do things. Which of the things
organisms do are "cognitive" and which are "vegetative" is mostly just a definitional matter, but
it is probably overstretching the notion to say that paramecia or hearts are "cognizing." The
examples are nevertheless instructive for cognitive science, because paramecia, hearts and
organisms with brains are all systems that can do things. So are computers and robots, for that
matter. Hence finding a causal explanation of how one of them does what it does may provide
useful lessons for explaining the others.
Let's start with the heart, an example used by Brette. What does the heart do? It pumps blood.
No metaphors. The heart literally pumps blood, and cardiac science has successfully reverse-
engineered the heart (to a close approximation). We know how the heart does it -- and part of
the proof that we know how is that we can apply and test our hypotheses about how the heart
pumps blood by building a synthetic model of a heart, plugging it into the heart's inputs and
outputs, and testing whether it can pump blood. If it can, the artificial heart passes the "Turing
Test" for cardiac function.
1
So what does the (human) brain (and body) pump? Human behavior. Or, rather, human
behavioral capacity. What people can do. Let's forget about what portion of that capacity
counts as cognitive and what proportion is just vegetative (like cardiac function): It all consists
of the capacity of a (living) system to do certain things. Now the challenge is to explain how.
Turing (1950) provided the ground rules: You have an explanation if you can design a system
that can do everything a human being can do, indistinguishably -- to a human -- from a human.
If your interest is just in "cognitive" capacities, then just generate those, ignoring the vegetative
capacities (or at least those that are not essential for generating the cognitive capacities).
Cognition, like Justice Potter Stewart's pornography, may be hard to define, but we know it
when we see it. And the capacity to interact with the dynamic world of objects and events and
their properties (including words describing those objects, events and properties)
indistinguishably from the way humans do, is surely cognitive, if anything is.
There is one more thing: Humans don't just do: They also feel. It feels like something, to a
human, to be seeing and doing what humans can see and do. But the capacity to feel eludes
Turing's program for cognitive science. It's something our brains pump invisibly. Turing (1950)
accordingly brackets it. But it keeps making disruptive peekaboo appearances in our attempts
to reverse-engineer cognition, as we shall see.
One of the main hypotheses about how the brain pumps cognitive capacity is via computation,
Turing computation. Computation is the manipulation of "symbols" (arbitrary formal objects)
on the basis of rules operating only on the symbols' shapes ("syntax"), not their meanings
("semantics"), in order to generate certain symbolic outputs from certain symbolic inputs.
That's what algorithms do. (An intuitive example is the rule we all learned in school for
extracting the roots of quadratic questions: "minus b plus or minus the square root of…".)
Algorithms are like recipes: apply them to the symbolic ingredients and you can explain how to
bake a symbolic cake. Computation is very powerful; just about everything in the universe can
be encoded symbolically and explained computationally, including cardiac function. The right
algorithm can pump symbolic blood. And you can show that the algorithm really works by
applying it to build a synthetic heart that really passes the cardiac Turing Test and pumps blood.
But to do that, you have to "interpret" the symbolic code and implement it in material form,
just as a formal recipe for a cake needs to be implemented in material form, using the real
ingredients referred to by the symbols, in order to generate a real cake.
So, despite its enormous power, computation cannot be all there is to cognition. Searle (1980)
showed, famously (in this journal), that a computer is not cognizing even if it can pass the
Turing Test (TT) because Searle too could pass the Chinese TT by executing the symbolic code
without understanding a word of Chinese. Why can't he understand? Because there is no
connection between the symbols in the code and the objects in the world that they are
interpretable as being about. Interpretable by whom? The user or the executor of the code. But
the meaning itself is not in the code.
2
That is the symbol grounding problem (Harnad 2006). Simple solution: The TT must not be
merely symbolic (verbal). It must test not only what the candidate can say, but also all the other
things a human cognizer can do in interacting with the objects in the world that the verbal TT is
merely chatting about. The candidate has to be a robot. And a Turing robot is not just a
computer, manipulating formal symbols; it is a dynamical system, able to interact with the
objects in the world. Its symbols are grounded in its capacity to identify and interact with their
referents indistinguishably from the way we do.
Now to neural "codes": Brette is right that it would be homuncular (although he calls it
"dualistic") to think of input to sensory receptors, activity along sensory pathways to sensory
and sensorimotor regions in the brain -- and then onward to motor regions and pathways to
motor effectors -- as encoded signals being transmitted in order to be decoded by a receiver, as
in telegraphic communication of morse code from a sender to a receiver. There is no
homunculus on the receiving end. It's all just a dynamic causal process constituting the
organism's capacity to do what it can do, some of it output in response to immediate sensory
input, some of it generated by endogenous processes.
But it is harmless to call the neural activity along sensory input pathways a "neural code."
Shannon's communication theory is about the end-to-end fidelity of signal transmission (of
analog or digital signals); it is not about cryptography, let alone about the interpretation of
computational algorithms or of natural language. To show that there is a substantive issue
involved here, Brette would have to show that there is a nontrivial chunk of performance
capacity (even the detection of interaural time difference: ITDs) that cannot be explained
causally if we insist on calling the activity occurring along the sensory input pathways a "neural
code." (Brette's preferred notion of "neural representations," by the way, sounds just as
homuncular to me as the idea of neural codes: "representation of what, to whom?" Ditto for
"internal model.")
Let me close with Brette's fleeting mention of "percepts." This is an instance of the "peekaboo"
influence of homuncular thinking. Psychophysics, too, can only study what the organism does
(input/output), not whether or how it feels like something to do it. Sensorimotor activity is only
perceptual if it is felt. I don't doubt that it feels like something to detect an IDT, just as it feels
like something to understand Chinese. But although symbol-grounding and Turing-testing may
be "easy" (in principle, if not in practice), explaining how and why organisms feel rather than
just do is and remains notoriously hard.
References
Baluška, F., & Mancuso, S. (2009). Plant neurobiology: from sensory biology, via plant
communication, to social plant behavior. Cognitive Processing, 10(1), 3-7.
Brette R. (2019) Is coding a relevant metaphor for the brain? Behavioral and Brain Sciences.
February 2019 1 -- 44.
Harnad, S. (2006). The symbol grounding problem. Encyclopedia of Cognitive Science. Wiley
3
Harnad, S. (2009) The Annotation Game: On Turing (1950) on Computing, Machinery, and
Intelligence. In: Epstein, R., Roberts, G., & Beber, G. (Eds.). Parsing the Turing test.
Springer Netherlands.
Shannon CE (1948) A mathematical theory of communication. Bell system technical
journal, 27(3): 379-423.
Turing, A. M. (1936). On computable numbers, with an application to the
Entscheidungsproblem. Proceedings of the London mathematical society, 2(1), 230-265.
Turing , A. M. (1950) Computing Machinery and Intelligence. Mind 49: 433-460.
4
|
1211.0249 | 3 | 1211 | 2013-01-21T15:27:58 | Finite element analysis of neuronal electric fields: the effect of heterogeneous resistivity | [
"q-bio.NC"
] | Simulation of extracellular fields is one of the substantial methods used in the area of computational neuroscience. Its most common usage is validation of experimental methods as EEG and extracellular spike recordings or modeling of physiological phenomena which can not be easily determined empirically. Continuous experimental work has been re-raising the importance of polarization effects between neuronal structures to neuronal communication. As this effects relies on very small potential changes, better modeling methods are necessary to quantify the weak electrical fields in the microscopic scale in a more realistic way. An important factor of influence on local field effects in the hippocampal formation is the heterogeneous resistivity of extracellular tissue. The vast majority of modeling studies consider the extracellular space to be homogeneous while experimentally, it has been shown that the stratum pyramidale has two times higher resistivity then other hippocampal layers. Common simulation methods for extracellular electrical fields based on the point source approximation are bound to describe the resistance of the space with a single, linear factor. We propose that models should be based on the space- and time-dependent Maxwell equations in order to account for heterogeneous properties of the extracellular space and specific arrangements of neurons in dense hippocampal layers. To demonstrate the influence of heterogeneous extracellular resistivity and neuronal spatial orientation on modeling results, we combine solutions of classical compartment models with spatiotemporal PDEs solved by the FEM. With the help of these methods, we show that the inclusion of heterogeneous resistivity has a substantial impact on voltages in close proximity to emitting neurons, increasing the extracellular potentials substantially compared to the homogeneous variant. | q-bio.NC | q-bio | Finite element analysis of neuronal electric fields:
the effect of heterogeneous resistivity
Pavol Bauer, Sanja Mikulovic, Stefan Engblom, Katarina E Le ao, Frank Rattay, Richardson N Le ao
1
3
1
0
2
n
a
J
1
2
]
C
N
.
o
i
b
-
q
[
3
v
9
4
2
0
.
1
1
2
1
:
v
i
X
r
a
Abstract —Simulation of extracellular fields is one of the sub-
stantial methods used in the area of computational neuroscience.
Its most common usage is validation of experimental methods
as EEG and extracellular spike recordings or modeling of
physiological phenomena which can not be easily determined
empirically. Continuous experimental work has been re-raising
the importance of polarization effects between neuronal struc-
tures to neuronal communication. As this effects rely on very
small potential changes, better modeling methods are necessary
to quantify the weak electrical fields in the microscopic sca le in
a more realistic way.
An important factor of in fluence on local field effects in the
hippocampal formation is the heterogeneous resistivity of extra-
cellular tissue. The vast majority of modeling studies consider the
extracellular space to be homogeneous while experimentally, it
has been shown that the stratum pyramidale has two times higher
resistivity than other hippocampal layers. Common simulation
methods for extracellular electrical
fields based on the poi nt
source approximation are bound to describe the resistance of
the space with a single scalar. We propose that models should
be based on the space- and time-dependent Maxwell equations
(Partial Differential Equations, PDEs) in order to account for
heterogeneous properties of the extracellular space and speci fic
arrangements of neurons in dense hippocampal layers.
To demonstrate the in fluence of heterogeneous extracellula r
resistivity and neuronal spatial orientation on modeling results,
we combine solutions of classical compartment models with
spatiotemporal PDEs solved by the Finite Element Method
(FEM). With the help of these methods, we show that the
inclusion of heterogeneous resistivity has a substantial impact
on voltages in close proximity to emitting hippocampal neurons,
substantially increasing the change in extracellular potentials
compared to the homogeneous variant.
Index Terms—extracellular fields finite element method neu-
ronal arrangement
I . INTRODUCT ION
Numerous computational studies have investigated time-
varying currents in homogeneous extracellular space [1], [2],
[3] as well as the role of neuronal morphology in uniform
electric field stimulation [4], [5], [6]. However, most stud ies
use models of the extracellular milieu that may not be ac-
curately applied when modeling brain structures with highly
diverse extracellular resistivity and neuronal arrangement.
As emphasized by Lopez-Aguado and Bokil [7], [8] it
has been traditionally neglected that currents propagate in all
directions in an extracellular medium and that inward and
outward currents originate from tissue regions having large re-
sistivity differences. A usual argument for this approach is that
resistivity in fluences extracellular electrical fields min
imally
and that extracellular space can be assumed as homogeneous
in current-source density (CSD). However, extracellular non-
homogeneous resistivity has been shown experimentally in
several regions of the brain, for example, in the hippocampus
and the cerebellum [9]. Additionally,
tissue swelling has
been observed after intense neural activity that in turn could
lead to an increase in extracellular resistivity [10]. Previous
studies [1], [11], [12], [13] have attempted to examine how
extracellular electrical fields affects neuronal activity although
with the help of quasi-static approximation and an assumed
homogeneous tissue resistivity.
To quantify the effect of extracellular heterogeneous resis-
tivity and neuronal spatial distribution on strength of neu-
ral fields, we are proposing a simple modeling pathway to
couple compartment-based neural models with the COMSOL
Multiphysics simulation environment. Here, we solve time-
dependent Maxwell’s equations using the FEM to analyze the
change of electrical fields as it occurs in the extracellular
space
surrounding neurons.
Our results indicate that inhomogeneous resistivity of the
extracellular milieu significantly in fluences the change of
ex-
tracellular potentials (EPs) in the hippocampus. By computing
the resulting voltage change due to outgoing transmembrane
currents of an exemplar CA1 pyramidal cell model in ho-
mogeneous and inhomogeneous extracellular resistivity, we
observed a maximal difference in EP change of 60% in the
hippocampal pyramidal layer. Furthermore, the here proposed
method offers the possibility to efficiently simulate the ef fects
of superimposed extracellular potentials created by neurons in
divergent positions relative to each other. We will also discuss
advantages and drawbacks of the used method and propose
alternatives and possible improvements towards more realistic
modeling of electrical fields of the brain.
Pavol Bauer and Stefan Engblom are with Division of Scientifi c Comput-
ing, Department of Information Technology, Uppsala University, SE-751 05
Uppsala, Sweden, e-mail: [email protected]
Sanja Mikulovic, Katarina E Le ao and Richardson N Le ao are with
Department of Neuroscience, Uppsala University Box 593, 751 24 Uppsala,
Sweden, e-mail: [email protected]
Sanja Mikulovic and Frank Rattay are with Institute for Analysis and
Scientific Computing, Vienna University of Technology Vien na, Austria
Katarina E Le ao and Richardson N Le ao are with Brain Instit ute, Federal
University of Rio Grande de Norte Natal-RN, Brazil
I I . MAT ER IAL S AND M E THOD S
A. Pyramidal neuron model
To demonstrate the effect of trans-membrane currents on
the effect of extracellular fields, a Hodgkin-Huxley (HH)-l ike
hippocampal CA1 pyramidal cells was adapted from [14] with
the addition of a current Ih from [15] and a current Im from
[16].
Ih current:
u∞ = [1 + exp((V + 76)/7)]−1 ,
τu = 104/[237 exp((V + 50)/12)+
17 exp(−(V + 50)/25)] + 0.6,
Ih (t, V ) = ghu(t, V ) (V − Eh ).
Im current:
s∞ = [1 + exp(−(V + 22.53)/10)]−1,
τs = 4135.7/[164.64 exp((V − 0.05) · 0.12)+
0.33 exp(−(V − 0.05)/10)] + 35.66,
Im (t, V ) = gm s(t, V ) (V − EK ).
(1)
(2)
(3)
(4)
(5)
(6)
Our model contained one somatic (r = 10 µm), 20 dendritic
(r = 1 –3 µm, l = 5 µm), two axon initial segments, AI S1
(r = 0.5 µm, l = 60 µm ), AI S2 (r = 0.5 µm, l = 60 µm),
and finally 32 axonal compartments (r = 0.5 µm, l = 5 µm),
see Figure 1. Note that here AI S1 refers to the part of AIS
from 0 to 60 µm, and AI S2 from 60 to 120 µm from the cell
body. Additionally, AI S2 compartments contain high N a+
channel density [17]. 20 dendritic compartments are used for
the dendritic tree, while additional compartments are included
in the branching analysis. The N a+ channel density was
varied between 986 and 2943 µA
cm2 (corresponding to AI S2 ).
The following conductance values were used: gN a = 8, gkdr
= 5, gm = 1 mS/cm2 for the soma; gC a = 10, gKC = 15,
gKAHP = 0.8, gO = 0.625, gN a = 0.07, gh = 0.2 mS/cm2
for the dendrite; gN a = 50, gkdr = 10, gm = 10 mS/cm2 for
the AI S2 and gN a = 9, gkdr = 10, gm = 10 mS/cm2 for the
rest of the axon compartments. The leakage conductance was
set to 0.1 for all of the compartments. Finally, the equilibrium
potentials were set to EN a = 60 mV, EK = -85 mV, Eh = -43
mV and Eleakage = -65 mV.
B. Creation of morphology and model coupling
The three-dimensional neuronal geometry was constructed
in COMSOL Multiphysics 4.3 with the help of the interface
to MATLAB ( “LiveLink ”) by morphological additions and
boolean unification of simple geometric volumes. As we aim
to represent the 3D-morphology as an exact counterpart of the
compartmental model, each section is recreated as a cylinder
with the same length and diameter as in the compartmental
de finition. If the cylinders were added on top of each other
with no change of the rotation vector, as shown in Figure 1, no
joining geometrical primitives were added in between them.
If the rotation differs, as for example in Figure 4, a sphere
is added in between the cylinders, followed by a removal of
the interior boundaries of both the cylinders and the sphere.
The reason for constructing the geometry in this way is that
the mesh engine otherwise respects the internal boundaries
such that the resulting mesh becomes unnecessary complex.
In some cases the occurrence of multiple internal boundaries
can even make the meshing procedure abort.
Ionic currents of single neuronal compartments model were
determined by solving Hodgkin-Huxley Ordinary Differential
Equations (HH ODEs) specified in equations 1 to 6 using a
2
2
0
µ
m
COMPARTMENT
MODEL
IONIC
CURRENT
FEM
MODEL
ICa, IK,
IK_AHP,
IK_O,
INa, Ih
ICa, IK,
IK_AHP,
IK_O,
INa, Ih
IKdr,
Im,
Ina
IKdr,Kdr
IKdr,
Im,
Im,
Ina
Ina
IKdr,
Im,
Ina
IKdr,
Im,
Ina
Dend
2,...,n
DendDend
1
Soma
AIS
AIS
1
1
AIS
2
Axon
1,...,n
Fig. 1. Mapping trans-membrane currents from compartmental model to
FEM boundary mesh. A compartmental model containing 20 dendritic, one
soma, two axon initial segments and 32 axonal compartments. Ionic currents
for each compartment are shown on the left. The compartmental model outputs
I (t) which is used as current boundary sources Qj (t) in the corresponding
three-dimensional volumes in the FEM model.
Runge-Kutta algorithm from the MATLAB ODE-suite (Math-
works) with a constant step-size T . The sum of all currents
per compartment for each time step of the simulation I0 is
afterwards normalized to an absolute value of units A/m2
and stored in a matrix I of size Ncompartments × Ntimesteps .
Next, by using the interface to COMSOL, each row of
the matrix I was mapped adequately to the corresponding
cylindrical domain as a boundary current source Qj (t). Note
that we hereby assume that transmembrane currents are the
only cause of change of extracellular potential, which is surely
not the case in a real neuron, as for example synaptic calcium-
mediated currents are suspected to contribute to a large fraction
of the extracellular signature [18].
The extracellular volume was modeled by cylindrical ob-
jects covering the neuronal morphology with either homo-
geneous (0.3 S/m) or heterogeneous resistivity as shown in
Figure 2A. In this case we constructed the extracellular volume
by taking the union of cylindrical objects with increasing
resistivity in the y -axis, according to [7]. It is assumed that the
conductivity of extracellular tissue is frequency-independent in
the used range of neural activity (10 –100 Hz) [19].
C. Electrostatic formulation and the Finite Element Method
To simulate non-homogeneous distributions of electrical
fields produced by single neurons and neuronal networks,
we used an electrostatic formulation of Maxwell’s equations
discretized with finite elements. Chie fly, in the finite eleme
nt
approach, Maxwell’s equations are solved by discretizing the
incorporated volumes (in this case the neuronal compartments)
into finite tetrahedral volume elements [20]. We seek the elec-
tric field intensity E in terms of the electric scalar potential
V ,
E = −∇V .
(7)
The relevant dynamic form of the continuity equation with
current sources Qj is given by
∂ ρ
∂ t
with J and ρ the current density and electric charge density,
respectively. Further constitutive relations include
∇ · J = −
+ Qj ,
(8)
and Ohm’s law
D = ε0 εrE,
J = σE,
(9)
(10)
in which D denotes the electric flux density . Finally, Gauss’
law states that
∇ · D = ρ.
(11)
Upon taking the divergence of (10) and using the continuity
equation (8) we get
+ Qj .
∇ · σE = −
∂ ρ
∂ t
Rewriting the electric charge density using Gauss’ law together
with the constitutive relation (9) and finally applying the g auge
condition (7) twice we arrive at the time-dependent potential
formulation
(12)
(13)
∇V (cid:19) = Qj .
− ∇ · (cid:18)σ∇V + ε0 εr
∂
∂ t
This is the formulation used in COMSOL Multiphysics [21].
The values for the electric conductance σ and the relative
permittivity εr were obtained from [2]. The source currents Qj
were computed from the compartmental model as described
above.
The formulation (13) 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 independent manner up to time t, it is then
straightforward to solve a single static PDE to arrive at the
potential V itself.
As for boundary conditions we took homogeneous Neu-
mann conditions (electric isolation) everywhere except for in
a single point which we choosed 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
underneath the neuronal geometry. This procedure ensures that
the formulation has a unique solution (it is otherwise only
specified up to a constant).
3
A tetrahedral mesh was applied to discretize space (using
the “
finest” mesh setting; resolution of curvature 0.2, reso
-
lution of narrow regions 0.8). The simulations were verified
against coarser mesh settings in order to ensure a practically
converged solution. As a final note, in the Time discrete solv er
the time step was set to the same step size T as used in
the ODE-based solution of the Hodgkin-Huxley equations,
thus ensuring a correct transition between both simulation
environments.
I I I . RE SULT S
A. The effect of heterogeneous extracellular fields
In order to examine the in fluence of heterogeneous ex-
tracellular space, we first constructed a three-dimensiona l
active neuron model in an homogeneous and heterogeneous
extracellular milieu (Figure 2). Neurons in the hippocampus
have an intricate spatial orientation that propitiates strong field
potentials: high density of pyramidal cell dendrites running
in parallel
in the stratum radiatum (SR), densely packed
pyramidal cell somas in the stratum pyramidale (SP) while
pyramidal cell axons run almost in parallel or crossing each
other in SP or stratum oriens (SO).
We measured the voltage on four de fined point probes
placed parallel to the dendrites, soma, AI S2 and axon terminal
compartments during the peak of an action potential (AP) of
80mV, by varying the distance between active neuron and point
probes from 1 µm to 80 µm (Figure 2A).
In the first set of simulations, we analyzed the effect of
the aforementioned four neuronal regions on the de fined poin t
probes assuming a widely accepted homogeneous resistivity of
350 Ω cm [2], [3]. Note that here applied boundary currents
of three-dimensional neuronal compartments correspond to the
peak transmembrane current during an AP. By doing so, the
peak voltage of 0.25 mV was obtained in the point probe
parallel to the AI S2 compartment, followed by soma, dendrite
and axon terminal. Consequently, the heterogeneous case was
examined by placing the neuron and point probes into a
heterogeneous extracellular space representing hippocampal
spatial order, where resistivity values for different strata were
obtained from [7] and shown in Figure 2A. The active neuron
was positioned in the center of SP and point probes were
moved along the x axis of the extracellular space. In the case
of non-homogeneous extracellular resistivity (Figure 2C) the
largest voltage change was measured in point probes parallel
to AI S2 , although the values in point probes placed in SP
were 60% larger in close distance than in the homogeneous
extracellular space scenario and 28% higher considering the
spatial mean over the total distance. The point probes placed in
SO and SR were as well affected by the higher restivity of the
pyramidal layer, showing an average increase of 4% in parallel
to the axon and 7% in parallel to the dendrite. The appreciable
difference between the voltage changes suggests that non-
homogeneous distribution of resistivity is an important aspect
of extracellular field effects.
A
B
D
S
AIS
2
A
0.2
0.15
0.1
0.05
)
V
m
(
e
g
a
t
l
o
V
SR
SP
SO
100
300
200
500
400
Resistivity (Ωcm)
600
D
S
0.4
0.3
0.2
0.1
0
0
0.2
0.15
0.1
0.05
0
0
10
20
30
40
A
10
30
20
Distance (µ m)
40
0
0
10
20
30
40
AIS 2
)
V
m
(
e
g
a
t
l
o
V
0.4
0.3
0.2
0.1
0
0
10
30
20
Distance (µm)
40
Fig. 2. Heterogeneous extracellular resistivity modulates the strength
of polarization effects in proximity of the firing neuron . A, Schematic
representation of a model neuron impact on defined point prob es. Point probes
act as receivers and are being shifted away from the active neuron along
the x-axis (left). Non-homogeneous resistivity distribution of hippocampal
regions (right). B, Relationship between membrane potential at different point
probes affected by dendritic (D), somatic (S), axon initial segment (AI S2 ) and
axon terminal (A) neuronal compartments (color coded as in A) for distance
varying between 0 and 80 µm. Dashed: the resulting voltage change assuming
homogenous extracellular resistivity (= 350 Ω cm), while solid represents the
result of assuming a heterogenous resistivity.
B. Analysis in respect to spatial orientation and neural mor-
phology
According to the superposition principle electrical fields
(EFs) are produced by summation of single neuronal activity.
Thus,
interaction between neuron and field is
the mutual
strongly modulated by the spatial orientation of neuronal
assemblies [22]. To analyze EFs considering a realistic spatial
order, we began simulating two neurons with parallel axons
(Figure 3A) where N eurona represents an active structure
with boundary current source Qj mapped to its surface while
N euronp is a passive measurement structure. We measured
Ve at four points along the axon of N euronp and detected
the maximal voltage amplitude of 0.28 mV. Note that here
Ve corresponds to extracellular voltage measured on the cell
membrane. Next we tested the in fluence of four neighboring
cells with parallel axons on the N euronp (Figure 3B). In this
case, having four active neuronal neighbors (distance between
the axons =10 µm) firing non-synchronously, a maximal
voltage of 0.49 mV in homogeneous and 0.68 mV in heteroge-
neous extracellular medium of N euronp was computed. Syn-
chronous firing of adjacent neurons produced a peak voltage
of 0.62 mV in homogeneous and 0.84 mV in heterogeneous
0.3
0
−0.1
0.3
0
−0.1
0
H
o
m
o
g
e
n
o
u
s
H
e
t
e
r
o
g
e
n
o
u
s
Time (ms)
100
Unsynchronized
Synchronized
0.7
0
-0.45
0.85
0
-0.6
40
0
Time (ms)
N
1
N
2
N
3
N
4
N
1
N
2
N
3
N
4
40
Unsynchronized
Synchronized
1.2
0
-0.8
1.4
0
-0.9
40
N
1
N
2
N
3
N
4
40
0
Time (ms)
N
1
N
2
N
3
N
4
40
4
H
o
m
o
g
e
n
o
u
s
H
e
t
e
r
o
g
e
n
o
u
s
H
o
m
o
g
e
n
o
u
s
H
e
t
e
r
o
g
e
n
o
u
s
Neuron
p
Neuron p
Neuron p
0.5
0
-0.3
0.7
0
-0.35
0
0.8
0
-0.5
0.9
0
-0.6
0
A
SO
SP
SR
B
SO
SP
SR
C
SO
SP
SR
Fig. 3. Analysis of extracellular fields in dependence of spike timi ng
and spatial orientation. A. Potential change calculated at different regions
(colored circles) of a passive neuron (N euronp ) in response to EPs triggered
from action potentials in the initial phase of an active neuron. Lines in the
right panel are color coded as the circles in the left panel. Somas were
placed at the stratum pyramidale and extracellular resistivity followed the
same distribution as shown in Figure 2C. Upper trace of the right panel
shows potential changes in homogeneous, while lower trace in heterogeneous
extracellular space. B, Same as in A but the N euronp surrounded by four
active neurons. Potential change at different regions of N euronp when four
active neurons fire asynchronously (middle panel) and synch ronously (right
panel). Insets on the middle and right panels show the firing of active neuron s).
C, Same as in B, but the active neurons where oblique (but not intersecting,
nearest interaxonal distance = 2 µm) at the position of AI S2 to N euronp .
medium of AI S2 . When axons crossed each other at the AI S2
region (Figure 3C), we observed a maximal voltage amplitude
of 0.89 mV in the AI S2 of Axonp during asynchronous
activity of the neighboring cells, whereas synchronous activity
produced 1.41 mV in heterogeneous medium. These results
suggest that for analyzed neuronal arrangement soma and
AI S2 have considerable in fluence on the strength of neural
fields for small distances ( ≤ 10 µm) and that values in
heterogeneous extracellular space are in mean ∼ 25% larger
than in homogeneous extracellular medium when averaged
over time.
Furthermore, we simulated the in fluence of pyramidal
neuron position within different hippocampal layers on the
strength of neural fields (Figure 4). We positioned somas in
SP with the smallest distance between the two axon initial
segments con fined to SO. This geometrical arrangement was
motivated by confocal images of two proximal pyramidal
neurons located parallel to each other with an axonal bend
A
27.43 µm
2 µm
13.53 µm
6.99 µm
B
A
SR
SP
SO
C
SR
SP
SO
0
0.1
0.2
0.3
0.4
0.5
0
0.1
0.2
0.3
0.4
0.5
Voltage (mV )
Voltage (mV )
Fig. 4. Modeling the effect of bended axons. A, Confocal image of
two neurons in the hippocampus filled with neurobiotin for post hoc spatial
analysis showing two parallel neurons with bent axons situated at a minimal
distance of 2 µm. B, A simplified representation of both neurons when the
nearest interaxonal region was confined in stratum oriens. I n this case, an
action potential in the active neuron generates a voltage transient of 0.25
mV at the passive neuron soma and 0.35 mV at the neuron AI S2 . The figure
shows the extracellular voltage distribution at the peak amplitude of the active
cell action potential. C, Same as B, but axon initial segments were confined at
the stratum pyramidale (higher resistivity). In this case, a 0.25 mV and 0.45
mV voltage transients were observed at the passive neuron soma and axon,
respectively.
starting about 50 µm behind the soma (Figure 4A). As
known from previous analytical studies [23] the bend of
axonal structures generally amplifies endogenous fields aro
und
neurons. We were able to con firm this effect in our simulation
as we simulated the field effect resulting from trans-membra ne
currents of an active neuron measured on the parallel neuron.
In this scenario, peak voltage of 0.35 mV was calculated at the
position of AI S2 while the smallest voltages were registered
at the passive neurons between the dendrites (0.12 mV) in SR
and axon terminals (0.16 mV). If AI S2 were con fined to SP,
the potential computed at the passive neuron was 25% larger
than in the case in which AI S2 were positioned in the SO
(Figure 4B).
IV. D I SCU S S ION
In this work, we use the Finite Element Method coupled
to the HH-equations to simulate how neuronal geometry,
arrangement and heterogeneous extracellular properties affect
the strengths of neural fields. We first show that there is a
difference in voltage transients produced by firing of neigh -
boring cells if the extracellular space is considered to be
homogeneous or heterogeneous. Additionally, we demonstrate
that the spatial orientation of specific cellular compartme nts is
an important determinant of the strength of neural fields.
5
In our computations, the highest change of extracellular
potential arises in the pyramidal layer, in proximity to the
AI S2 compartment. Action potentials are generally initiated in
the AIS due to a higher density of voltage-gated N a+ channels
[17], [24], [25], which is re flected by the parameter setting s
for this compartment in the occupied neuronal model. Ad-
ditionally, the spatial arrangement of hippocampal pyramidal
neurons also propitiates the proximity of AI S2 in both SP and
SO (see Figure 4, [26]).
Holt and Koch [27] showed that interactions near cell bodies
are more important than interactions between axons by using
standard one-dimensional cable theory and volume conductor
theory. Another study [28] reported a 4.5 mV change (in the
AIS) caused by extracellular interactions, a change more than
4× larger than what we have found in our simulations of
heterogeneous extracellular media. Additionally, by applying
analytical methods, Bokil et al [8] have shown that, in the
olfactory system, an action potential of 100 mV amplitude
in one axon could produce depolarization in other axons in
the bundle sufficient to initiate an action potential. Howev er,
using FEM simulations it was not possible to trigger spikes in
neurons solely by electrical fields mediated by in- and outflo w
of trans-membrane currents, in agreement with the work by
Traub and colleagues (in which maximal voltages in a sink
axon during synchronized activity of four neighboring neurons
is ≈ 1.2mV ).
Hence, simulations relying on the point source approxima-
tion to describe hippocampal neural fields may be distorted a s
the potential change is more than 28% greater in stratum pyra-
midale and in average 6% greater in other hippocampal layers
if a heterogeneous tissue is used instead of a homogeneous
one [7], [1].
For varying extracellular resistivity a numerical procedure
referred to as ‘the method of images’ [29] has been proposed
as an extension to the point source equation. Although this
method has been used in computations of extracellular action
potentials in a previous study [11], its practical use is limited
to a rather low number of resistivity layers and non-complex
geometry.
The requirement of FEM to model passive current flow
between neurons was suggested elsewhere [2], [3], but im-
plementation issues and the lack of adequate software tools
may have precluded its usage in the past.
Extracellular resistivity could also contribute to the ampli-
tude of local field potentials (LFP), and in fact, in vitro LFP
registered in hippocampal slices are greater in the SP than in
other layers [30]. Interestingly, we observed that hippocampal
slices in interface-type recording chambers (where slices are
not completely submerged in the buffer solution [31], [32]) are
more than 60% less conductive than in the chambers where
slices are submerged (unpublished results). This could help
explaining why LFP recorded in interface-type cambers are
far greater than in submerged-type counter parts [16], [32].
Here we show that the use of FEM software (COMSOL
Multyphysics), with an interface to Hodgkin and Huxley-
type ODE model in Matlab,
is a powerful tool to verify
macroscopic effects of neural fields. However, this approac h
may fail to simulate small nuances of extracellular interactions
(e.g. ion channels apposing active compartments may suffer
more from the effect of passive current flow than channels
in the opposite side). Nonetheless, simulations using PDEs
could increase the level of realistic models of membrane
dynamics. The idea of translating membrane dynamics to
PDE was proposed by Hodgkin and Huxley themselves [33],
and later pursued by Kashef and Bellman [34]. However,
implementing membrane dynamics in combination with the
Maxwell equation interface of the PDE simulator has been, to
our experience, quite cumbersome. For example, due to the
relatively high computational cost of the solution phase of the
finite element method, a representation of fully reconstruc ted
morphologies with volumetric cylindrical elements was not
possible. The high amount of (small) cylindrical elements
required to accurately model complex dendritic branches typ-
ically caused the meshing algorithm to break.
The software used in our simulations (COMSOL Multi-
physics) has helped to popularize the use of FEM in neurosci-
entific problems [35], [36]. However, the software is geared
towards industrial applications and the steep learning curve
associated to the adaptation of COMSOL to neuroscience
related problems may preclude its widespread usage by the
neuroscience community. Possible combination of Neuron
Simulation Environment (NSE) with a FEM simulator would
offer great possibilities by allowing both realistic neural mor-
phologies and extracellular space modeling. However, cur-
rently the existence of a direct interface between NSE and a
FEM simulator has not been reported; future studies should
focus on this task. Furthermore, hippocampal extracellular
stimulation has been proposed as therapeutic strategy to con-
trol several disorders, including temporal lobe epilepsy [37],
[38]. Developing models that consider as many as possible
physical tissue properties would help to improve extracellular
stimulation in epilepsy patients. In summary, our work adds to
the recently published studies attempting to reveal important
parameters determining the strength of extracellular electric
fields. Models that do not use space and time-dependent
differential equations when modeling neuronal interactions
may have failed to replicate the changes in measured voltages
caused by passive current flow in heterogeneous extracellul ar
tissue, especially when more than two neurons are modeled
simultaneously.
ACKNOW L EDGMENT
This work was supported by Kjell and M arta Beijers Foun-
dation and the BMWF Marietta Blau stipend. PB and SE
were supported by the Swedish Research Council within the
UPMARC Linnaeus center of Excellence.
RE F ERENCE S
[1] C. C. McIntyre and W. M. Grill, “Excitation of central ner vous system
neurons by nonuniform electric fields,”
Biophysical Journal, vol. 76,
no. 2, pp. 878–888, Feb. 1999, PMID: 9929489. [Online]. Avai lable:
http://www.ncbi.nlm.nih.gov/pubmed/9929489
[2] C. B ´edard, H. Kr oger, and A. Destexhe, “Modeling extra cellular field
frequency-filtering properties of extra cellular
potentials
and the
no.
space,”
Journal,
vol.
86,
3,
pp.
1829–
Biophysical
1842, Mar.
2004,
PMID:
14990509.
[Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/14990509
6
[3] —
in brain
“Model of low-pass filtering of local field potentials
—,
tissue,” Physical Review. E, Statistical, Nonlinear, and Soft Matter
Physics, vol. 73, no. 5 Pt 1, p. 051911, May 2006, PMID: 16802971.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/16802971
[4] C. Y. Chan, J. Hounsgaard, and C. Nicholson,
“Effects of e lectric
fields on transmembrane potential and excitability of turtl e cerebellar
The Journal of Physiology, vol. 402,
purkinje cells
in vitro,”
pp. 751–771, Aug. 1988, PMID: 3236254.
[Online]. Available :
http://www.ncbi.nlm.nih.gov/pubmed/3236254
[5] T. Radman, Y. Su, J. H. An, L. C. Parra, and M. Bikson,
“Spik e
timing amplifies the effect of electric fields on neurons: imp
lications
for endogenous field effects,”
The Journal of Neuroscience: The
the Society for Neuroscience , vol. 27, no. 11,
Official Journal of
pp. 3030–3036, Mar. 2007, PMID: 17360926. [Online]. Availa ble:
http://www.ncbi.nlm.nih.gov/pubmed/17360926
[6] T. Radman, R. L. Ramos, J. C. Brumberg, and M. Bikson, “Rol e of
cortical cell type and morphology in subthreshold and suprathreshold
uniform electric field stimulation in vitro,”
Brain Stimulation, vol. 2,
no. 4, pp. 215–228, 228.e1–3, Oct. 2009, PMID: 20161507. [On line].
Available: http://www.ncbi.nlm.nih.gov/pubmed/20161507
[7] L. L ´opez-Aguado, J. M. Ibarz, and O. Herreras,
“Activit y-dependent
changes of tissue resistivity in the CA1 region in vivo are layer-
Neuroscience, vol. 108,
specific: modulation of evoked potentials,”
no. 2, pp. 249–262, 2001, PMID: 11734358.
[Online]. Availab le:
http://www.ncbi.nlm.nih.gov/pubmed/11734358
[8] H. Bokil, N. Laaris, K. Blinder, M. Ennis, and A. Keller,
“ Ephaptic
interactions in the mammalian olfactory system,”
The Journal of
Neuroscience: The Official Journal of the Society for Neuros cience,
vol. 21, no. 20, p. RC173, Oct. 2001, PMID: 11588203. [Online].
Available: http://www.ncbi.nlm.nih.gov/pubmed/11588203
[9] Y. C. Okada, J. C. Huang, M. E. Rice, D. Tranchina, and C. Nicholson,
“Origin of the apparent
tissue conductivity in the molecula r and
granular layers of the in vitro turtle cerebellum and the interpretation of
current source-density analysis,”
Journal of Neurophysiology, vol. 72,
no. 2, pp. 742–753, Aug. 1994, PMID: 7983532. [Online]. Avai lable:
http://www.ncbi.nlm.nih.gov/pubmed/7983532
[10] A. M. Autere, K. Lamsa, K. Kaila, and T. Taira, “Synaptic
activation
of GABAA receptors induces neuronal uptake of ca2+ in adult rat
Journal of Neurophysiology, vol. 81, no. 2,
hippocampal slices,”
pp. 811–816, Feb. 1999, PMID: 10036281.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/10036281
“On
[11] C. Gold, D. A. Henze, C. Koch,
and G. Buzs ´aki,
the
the origin of
action potential waveform: A
extracellular
modeling study,”
Journal of Neurophysiology, vol. 95, no. 5,
pp. 3113–3128, May 2006, PMID: 16467426. [Online]. Availab le:
http://www.ncbi.nlm.nih.gov/pubmed/16467426
[12] C. Gold, D. A. Henze,
“Using extracellular
and C. Koch,
to constrain compartmental models,”
action potential
recordings
vol. 23, no. 1, pp.
Journal of Computational Neuroscience,
39–58, Aug.
2007,
PMID:
17273940.
[Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/17273940
[13] C. A. Anastassiou, S. M. Montgomery, M. Barahona, G. Buzs ´aki,
and C. Koch,
“The effect of spatially inhomogeneous extrace llular
electric fields on neurons,”
The Journal of Neuroscience: The
the Society for Neuroscience , vol. 30, no. 5,
Official Journal of
pp. 1925–1936, Feb. 2010, PMID: 20130201.
[Online]. Availa ble:
http://www.ncbi.nlm.nih.gov/pubmed/20130201
“Intrinsic and network rhyth mogenesis in
[14] P. F. Pinsky and J. Rinzel,
a reduced traub model for CA3 neurons,”
Journal of Computational
Neuroscience, vol. 1, no. 1-2, pp. 39–60, Jun. 1994, PMID: 8792224.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/8792224
[15] R. N. Leao, K.
Svahn, A. Berntson,
and B. Walmsley,
“Hyperpolarization-activated (I) currents in auditory br ainstem neurons
of normal and congenitally deaf mice,” The European Journal of
Neuroscience, vol. 22, no. 1, pp. 147–157, Jul. 2005, PMID: 16029204.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/16029204
[16] R. N. Le ao, H. M. Tan, and A. Fisahn, “Kv7/KCNQ channels
control
action potential phasing of pyramidal neurons during hippocampal
gamma oscillations
in vitro,”
The Journal of Neuroscience: The
the Society for Neuroscience , vol. 29, no. 42,
Official Journal of
pp. 13 353–13 364, Oct. 2009, PMID: 19846723. [Online]. Avai lable:
http://www.ncbi.nlm.nih.gov/pubmed/19846723
[17] L. M. Palmer and G. J. Stuart,
i nitiation
“Site of action potential
in layer 5 pyramidal neurons,” The Journal of Neuroscience: The
the Society for Neuroscience , vol. 26, no. 6,
Official Journal of
pp. 1854–1863, Feb. 2006, PMID: 16467534.
[Online]. Availa ble:
http://www.ncbi.nlm.nih.gov/pubmed/16467534
7
epilepsy,” Stereotactic and functional neurosurgery, vol. 89, no. 2, pp.
111–122, Apr. 2011, PMID: 21336007.
[38] H. Luna-Mungua, A. Meneses, F. Pea-Ortega, A. Gaona, and L. Rocha,
“Effects of hippocampal high-frequency electrical stimul ation in mem-
ory formation and their association with amino acid tissue content and
release in normal rats,” Hippocampus, vol. 22, no. 1, pp. 98–105, Jan.
2012, PMID: 20882549.
[18] K. C. Buszaki G, Anastassiou CA,
in the CN S
“Field effects
play functional roles,” Nature Neurosciecne, vol. 6, 2012. [Online].
Available: http://www.ncbi.nlm.nih.gov/pubmed/22595786
“Intrin sic
and G. T. Einevoll,
[19] H. Lind ´en, K. H. Pettersen,
filtering gives
dendritic
local
low-pass power
spectra of
fie
ld
Journal of Computational Neuroscience, vol. 29, no. 3,
potentials,”
pp. 423–444, Dec. 2010, PMID: 20502952.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/20502952
[20] J. Jin, The Finite Element Method in Electromagnetics, 2nd ed. New
York: John Wiley & Sons, 2002.
[21] AC/DC Module User’s Guide, Comsol, 2012, version 4.3.
[22] F. Fr ohlich and D. A. McCormick,
fields
“Endogenous electric
may guide neocortical network activity,” Neuron, vol. 67, no. 1,
pp. 129–143,
Jul. 2010, PMID: 20624597.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/20624597
“A model for the polariza tion of
[23] D. Tranchina and C. Nicholson,
neurons by extrinsically applied electric fields,”
Biophysical Journal,
vol. 50, no. 6, pp. 1139–1156, Dec. 1986, PMID: 3801574. [Onl ine].
Available: http://www.ncbi.nlm.nih.gov/pubmed/3801574
central
[24] B.
P. Bean,
“The
action
potential
in mammalian
vol.
Nature Reviews. Neuroscience,
neurons,”
6, pp.
8, no.
451–465,
[Online]. Available:
Jun.
2007,
PMID:
17514198.
http://www.ncbi.nlm.nih.gov/pubmed/17514198
[25] M. H. P. Kole and G. J. Stuart,
thres hold
“Is action potential
Nature Neuroscience, vol. 11, no. 11,
lowest
in the
axon?”
pp. 1253–1255, Nov. 2008, PMID: 18836442. [Online]. Availa ble:
http://www.ncbi.nlm.nih.gov/pubmed/18836442
[26] S. Ram ´on y Cajal, Histologie du Systeme Nerveux de l’Homme et des
Vertebretes, 1911.
[27] G. R. Holt and C. Koch, “Electrical interactions via the
extracellular
Journal of Computational Neuroscience,
potential near cell bodies,”
vol. 6, no. 2, pp. 169–184, Apr. 1999, PMID: 10333161. [Onlin e].
Available: http://www.ncbi.nlm.nih.gov/pubmed/10333161
[28] R. D. Traub,
F. E. Dudek, C.
P. Taylor,
and W. D.
“Simulation of hippocampal afterdischarges sync hronized
Knowles,
by electrical
interactions,”
Neuroscience, vol. 14, no. 4, pp.
1033–1038, Apr.
[Online]. Available:
2987752.
1985, PMID:
http://www.ncbi.nlm.nih.gov/pubmed/2987752
[29] E. Weber, Electro Magnetic Fields: Theory, Applications. New York:
John Wiley & Sons, 1950.
[30] A. Fisahn, F. G. Pike, E. H. Buhl, and O. Paulsen,
“Cholin ergic
induction of network oscillations at 40 hz in the hippocampus in vitro,”
Nature, vol. 394, no. 6689, pp. 186–189, Jul. 1998, PMID: 9671302.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/9671302
“Volume-conducted epileptiform events
Inaba and M. Avoli,
[31] Y.
adjacent
between
necortical
slices
in
an
interface
tissue
chamber,”
Journal of Neuroscience Methods, vol. 151, no. 2,
pp. 287–290, Mar. 2006, PMID: 16143402.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/16143402
[32] N. H ´ajos, T. J. Ellender, R. Zemankovics, E. O. Mann, R. Exley,
“Maintaining netwo rk
S. J. Cragg, T. F. Freund, and O. Paulsen,
activity in submerged hippocampal slices:
importance of oxygen
supply,”
The European Journal of Neuroscience, vol. 29, no. 2,
pp. 319–327,
[Online]. Availabl e:
Jan. 2009, PMID: 19200237.
http://www.ncbi.nlm.nih.gov/pubmed/19200237
“A quantitative descript ion
[33] A. L. HODGKIN and A. F. HUXLEY,
current
of membrane
and its
application
to conduction
and
excitation in nerve,”
The Journal of Physiology, vol. 117, no. 4,
pp. 500–544, Aug. 1952, PMID: 12991237.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/12991237
[34] B. Kashef and R. Bellman, “Solution of the partial diffe rential equation
of the Hodgkin-Huxley model using differential quadrature,” Mathemat-
ical Biosciences, vol. 19, no. 1-2, pp. 1–8, 1974.
[35] T. C. Zhang and W. M. Grill,
“Modeling deep brain stimula tion:
the
point source approximation versus realistic representation of
vol.
electrode,”
of Neural Engineering,
7,
no.
6,
Journal
p. 066009, Dec. 2010, PMID: 21084730.
[Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/21084730
[36] N. Yousif, N. Purswani, R. Bayford, D. Nandi, P. Bain, and
X. Liu,
the deep brain stimulation
“Evaluating the impact of
induced electric field on subthalamic neurons: a computatio nal
Journal of Neuroscience Methods, vol. 188, no. 1,
modelling study,”
pp. 105–112, Apr. 2010, PMID: 20116398.
[Online]. Availabl e:
http://www.ncbi.nlm.nih.gov/pubmed/20116398
[37] P. H. Stypulkowski, J. E. Giftakis, and T. M. Billstrom, “Development
of a large animal model for investigation of deep brain stimulation for
|
1002.4903 | 1 | 1002 | 2010-02-26T00:33:16 | Sparse incomplete representations: A novel role for olfactory granule cells | [
"q-bio.NC"
] | Mitral cells of the olfactory bulb form sparse representations of the odorants and transmit this information to the cortex. The olfactory code carried by the mitral cells is sparser than the inputs that they receive. In this study we analyze the mechanisms and functional significance of sparse olfactory codes. We consider a model of olfactory bulb containing populations of excitatory mitral and inhibitory granule cells. We argue that sparse codes may emerge as a result of self organization in the network leading to the precise balance between mitral cells' excitatory inputs and inhibition provided by the granule cells. We propose a novel role for the olfactory granule cells. We show that these cells can build representations of odorant stimuli that are not fully accurate. Due to the incompleteness in the granule cell representation, the exact excitation-inhibition balance is established only for some mitral cells leading to sparse responses of the mitral cell. Our model suggests a functional significance to the dendrodendritic synapses that mediate interactions between mitral and granule cells. The model accounts for the sparse olfactory code in the steady state and predicts that transient dynamics may be less sparse. | q-bio.NC | q-bio | Sparse incomplete representations: A novel role for olfactory granule cells
Alexei A. Koulakov1 and Dmitry Rinberg2
1Cold Spring Harbor Laboratory, Cold Spring Harbor, NY 11724
2HHMI Janelia Farm Research Campus, HHMI, Ashburn, VA 20147
Mitral cells of the olfactory bulb form sparse representations of the odorants and transm it this information to the cortex.
The olfactory code carried by the mitral cells is sparser than the inputs that they receive. In this study we analyze the
mechanisms and functional significance of sparse olfactory codes. We consider a model of olfactory bulb containing
populations of excitatory m itral and inhibitory granule cells. W e argue that sparse codes may emerge as a result of self
organization in the network leading to the precise balance between m itral cells' excitatory inputs and inhibition prov ided by
the granule cells. We propose a novel role for the olfactory granule cells. We show that these cells can build
representations of odorant stimuli that are not fully accurate. Due to the incompleteness in the granule cell representation,
the exact excitation-inhibition balance is established only for some mitral cells leading to sparse responses of the mitral
cell. Our model suggests a functional significance to the dendrodendritic synapses that mediate interactions between
mitral and granule cells. The model accounts for the sparse olfactory code in the steady state and predicts that transient
dynamics may be less sparse.
INTRODUCTION
Mammalian olfactory bulb
the
in
f irst stage
the
is
processing of information about odorants. The surface of
olfactory bulb
is covered by several
thousands of
glomeruli. Each glomerulus receives inputs from a set of
sensory neurons expressing the same type of olfactory
receptor protein. The inputs into each glomerulus are
therefore substantially correlated (Shepherd et al., 2004;
Lledo et al., 2005). The glomerulus-based modularity is
preserved further by the mitral cells. Most of them receive
direct excitatory inputs from a single glomerulus only.
Mitral cells are also a major output class of the olfactory
bulb that carries information about odorants to olfactory
cortex (Figure 1).
In 1950, Adrian reported that the spontaneous activ ity of
the mitral cells in awake rabbit is much higher than that in
an anesthetized animal, and in awake animal odor
responses vanishe on the background of this spontaneous
activ ity (Adrian, 1950). These observations were confirmed
and more thoroughly analyzed in mice in recent study by
Rinberg et.al. (Rinberg et al., 2006). The odor responses
of mitral cells are sparse and state-dependent (Kay and
Laurent, 1999; Rinberg et al., 2006; Fuentes et al., 2008).
Only a small fraction of mitral cells responds to the
indiv idual odorants despite receiv ing substantial odorant
related input.
How can sparseness emerge mechanistically? One of the
mechanisms for the generation of sparse responses
involves cells' f iring near threshold (Rozell et al., 2008). If
the inputs into a cell fall below threshold, the cell does not
fire, which leads to the sparse responses. However, this
explanation does not apply in the case of mitral cells,
because they exhibit high spontaneous firing rates in the
awake behav ing animals, i.e. specifically in the state when
the responses are most sparse.
It is common to explain sparse codes as satisfying the
constraint of metabolic energy efficiency (Lennie, 2003;
Olshausen and Field, 2004). Although the information
content of sparse representations is poor compared to the
full codes (Baraniuk, 2007), sparse codes can use less
metabolic energy. However, the high spontaneous rates
coexisting with sparse representations makes the olfactory
code both informationally poor and metabolically costly.
The functional significance of sparse olfactory codes is
therefore unclear.
These arguments point to neurons other than m itral cells
as likely sources of sparseness. Amongst the bulbar
neurons, the granule cells (GC) are thought to implement
lateral inhibition between m itral cells belonging to different
glomeruli through a mechanism based on dendro-dendritic
reciprocal synapses (Figure 1) (Shepherd et al., 2004).
Such interactions are thought to facilitate discriminations
between similar stimuli and mediate competitive
interactions between coactive neurons (Arev ian et al.,
2008). Because GCs are the most abundant cell type of
the olfactory bulb it is likely that they are substant ially
involved in these tasks.
In this study we propose a novel role for the GCs of the
olfactory bulb. W e show that granule cells can form
complimentary to mitral cells representations of olfactory
stimuli. Exact balance between
the excitation
from
receptor neurons and the inhibition from the granule cells
eliminates odor responses in some mitral cells. However,
some mitral cells retain the ability to respond to odors, due
1
to incompleteness of the granule cells representation. The
olfactory code on the level of mitral cells becomes sparse.
This function is facilitated by the network architecture
based on dendrodendritic reciprocal synapses between
the mitral and granule cells.
A
Glomeruli
Mitral
cells
Granule
cells
miW
imW
ig
mx
my
B
Granule to mitral cell synaptic weights:
W
G ranule cell index
x
e
d
n
i
l
l
e
c
l
a
r
t
i
M
Figure 1. Mitral – granule cell network model. A. Mitral cells
(gray triangles) receive excitatory inputs, mx , from glomeruli,
(large circles). The mitral cell output, my , is sent to the brain.
Mitral cells and granule cells (small solid circles) are connected
by reciprocal dendrodendritic synapses. Granule cells receive the
excitatory input from the mitral cells, which defined by synaptic
imW
, and m itral cells are inhibited by granule cells,
weight matrix
miW . B. Granule to mitral cell synaptic weight
with weight matrix
miW . Here we argue that GCs represent mitral cell inputs as a
superposition of columns in the weight matrix, such as the one
marked by the dotted rectangle.
RESULTS
Our goal is to describe the behav ior of two ensembles of
neurons: mitral and granule cells (Figure 1A). Mitral cells
are projection neurons that receive inputs from olfactory
glomeruli and send information about odorants to olfactory
cortices for further processing. Mitral cell inputs from the
mx (here and below
glomeruli are contained in vector
2
index m represents the m itral cell number running
between 1 and M , the total number of mitral cells). These
inputs
represent synaptic
inputs
from
the
receptor
neurons. The output firing rate of the mitral cells is denoted
by my . The purpose of the present study is to understand
the relationship between m itral cell inputs mx and their
outputs my in the presence of granule cell inhibition.
interneurons. GC responses are
GCs are
inhibitory
ia with the index i running
described by the vector
between 1 and N , the number of granule cells. We
assume that the granule cells are much more abundant
than mitral cells, i.e. N M (Tolley and Bedi, 1994). The
mitral cells receive inhibitory inputs from the granule cells
miW (Figure 1B). Each
determined by the weight matrix
miW describes the strength of synapses
element of matrix
between a granule cell number i and a mitral cell number
m . Because of the relative abundance of granule cells,
this matrix is rectangular as shown schematically in Figure
1B. Similarly, the reverse synaptic weights from the mitral
imW
to granule cells are described by the matrix
. Two
imW
miW and
reciprocal weight matrices,
, although
attributed to the same dendrodendritic synapse, are not
necessarily the same.
The system of mitral cells and granule cells connected
by reciprocal dendro-dendritic synapses can be
described by a Lyapunov function. The behav ior of the
network can be simplified if the synaptic weights from
imW
mitral cells to granule cells
are proportional to the
miW :
reverse synaptic weights from granule cells
This assumption is not unrealistic because both weights
are determined by active zones within
the same
dendrodendritic synapse (Shepherd et al., 2007), and can,
in principle, be regulated to be proportional. In the
Methods section we show that under this condition the
dynamics of the network is described by minimization of
the cost function that is called Lyapunov function (Hertz et
al., 1991). Lyapunov function is a standard construct in the
neural network theory that has been extensively used to
study the properties of complex networks. The steady
states of the neural activ ities can be obtained as m inima of
this function. Thus, by f inding minima of the Lyapunov
function one can understand the steady-state responses of
W
im
(1)
W
mi
.
.
mitral and GCs to odorants. The Lyapunov function for
mitral - GC network has the following form:
2
1
M
N
2
i
i
m
1
1
W a
mi
i
C a
i
(2)
x
m
The first term in the function contains the sum of squared
differences between the excitatory inputs to the mitral cells
from receptor neurons mx and the inhibitory inputs from
the granule cells. The inhibitory inputs are proportional to
ia weighted by synaptic
the activ ity of granules cells
miW . The f irst term therefore reflects the balance
matrix
between excitation and inhibition on the inputs to mitral
cells. Minimization of this term due to the system 's
dynamics leads to the establishment of an exact balance
between excitation and inhibition. The second term of the
Lyapunov function represents a cost imposed on the
ia , which
0
granule cell firing. It reaches m inima when
correspond to minimization of number of active GCs.
Granule cells represent mitral cell
inputs as a
superposition of dendrodendritic weights. W e will first
assume that the second term in the Lyapunov function is
negligible. This condition corresponds to the case when
the granule cells are easy to activate and, therefore, the
cost associated with their firing is small. Minimizing the first
term of the Lyapunov function without the cost leads to the
mx
precise
balance
between
excitation
and
W a
[see equation (2)]:
inhibition
i
mi
i
i
According to this equation the dendrodendritic weights can
iW
be v iewed as vectors
defined in the space of mitral cell
inputs (gray arrows in Figure 2A). The firing rates of
ia can be v iewed as unknown coefficients.
granule cells
The problem posed by equation (3) can be restated as
follows: Represent the vector of mitral cell excitatory inputs
x
iW
as a superposition of vectors
by f inding the
ia (Figure 2). We conclude that the network
coefficients
dynamics in olfactory bulb leads to the representation of
the bulbar
inputs by
the granule cells using
the
dendrodendritic synaptic weights as basis vectors. In this
model the granule cells play the role of coding cells that
carry representations of mitral cell inputs.
W a
mi
i
x
m
(3)
.
2x
x
iW
1x
Figure 2. The olfactory system with two mitral and five GCs. The
inputs into mitral cells from the receptor neurons are represented
by the 2D vector x
. The weights from the GCs are shown by five
iW
2D vectors
. Equation (3) leads to the representation of the
input vector in terms of the weight vectors, as shown by the
dashed lines, with the coefficients given by the activities of the
GCs. In this example, the second and the third GCs are active,
although other solutions are also possible.
iW
is equal to the
The number of the basis vectors
number of granule cells, N . The dimensionality of the
input vector x
is equal to a number of the mitral cells, M ,
which is much smaller than N . Therefore, the problem of
representation of mitral cell inputs as a superposition of
basis vectors may have several solutions. Thus, in Figure
2 f ive granule cells (gray arrows) have to represent inputs
1x and
2x ). Clearly there are many
into two mitral cells (
ways how a 2D vector can be decomposed into five
components.
This picture suggests the following interpretation of the
role of the bulbar GC. The strength of dendrodendritic
synapses represents the receptive fields of indiv idual GCs,
i.e. specific features in the glomerular inputs that these
cells
respond
to. The bulbar network attempts
to
decompose
its
inputs
into a superposition of
these
indiv idual features. Because of the abundance of the GCs,
no unique representation of the inputs in terms of GC
receptive fields can be found. The olfactory code as
defined by the GCs is therefore underdetermined. As
suggested
underdeterm ined
of
studies
the
by
(overcomplete) systems the unique solution can be found
if one includes the cost imposed on the f iring rate vector
into consideration. This is the second term in equation (2)
that we so far neglected.
Mitral cell firing rates result from the balance between
excitation and inhibition. In our model the response of a
mitral cell is defined as the dev iation of the firing rate of a
ym :
ym ,
rate,
spontaneous
from
cell,
mitral
3
ym ym ym . This response results from the excitatory
inputs from receptor neurons mx and inhibitory inputs from
the granule cells
.
(4)
y
m
N
x
W a
m
mi
i
i
1
my of the mitral cell is a result of
Therefore the response
the balance between excitatory and inhibitory inputs. If
such a balance is established precisely, i.e. if the granule
cells represent exactly the bulbar inputs according to
equation (3), no mitral cell is expected to respond to
odorants. On one hand, this observation implies that the
responses of mitral cells are ultimately sparse - they
completely disregard their excitatory inputs. This model
therefore contains the feature of the olfactory code
observed experimentally, i.e., sparseness. Sparseness in
this model emerges as a result of self-organizing balance
between excitation and inhibition. On the other hand, it is
hard to imagine a utility of this type of code, because the
mitral cells do not respond to odorants at all. Next we
identify the features of the model that could allow mitral
cells to respond to odorants.
Sparse mitral cell responses result from incomplete
granule cell representations. How can responses of
mitral cells emerge in our model? If GCs represent the
excitatory inputs of the mitral cells exactly, mitral cells are
unresponsive. Therefore GCs have to fail to represent
glomerular inputs exactly for the mitral cells to respond to
odorants. The GC code therefore has to be incomplete
leading to an inaccurate odorant representation by the
GCs.
Here we present the simplest example of such an
incomplete GC code. In Figure 3 we illustrate the synaptic
currents emerging in the simplistic olfactory system with
six mitral cells and only one GC. Vector x
represents the
excitatory inputs received by all six mitral cells. In Figure
3A the odorant supplies inputs into three m itral cells out of
six. We then assume that the granule-mitral cell weights
W match exactly this input pattern. As a result, when the
into to
GC is activated, it will return an inhibitory input x
the mitral cells that exactly matches the pattern of
excitation. The responses on the m itral cells to the odorant
will be very weak as a result of such a compensation
(some small residual responses are needed to drive the
GC).
Figures 3B and C show examples of two odorants that
activate similar subsets of mitral cells. The pattern of
however is exactly the same as in Figure 3A,
inhibition x
because with the single GC present, the inhibitory currents
are constrained to a single pattern. As a result of
subtraction between receptor neuron excitatory and GC
inhibitory inputs into the mitral cells, only a single mitral
cell responds to each of these odorants. Thus, the
responses of mitral cells become sparse. Some mitral cells
relay their inputs, while others are almost completely
inhibited by the GCs. That some mitral cells respond is a
consequence of the incompleteness of the GC code.
Mitral
cell
input
x
Granule
cell
synapses
W
Granule cel l
to mitral cel l
inhibition
x
Mitral ce ll
response
excit.-inhib.
x
x
y
A
r
e
b
m
u
n
l
l
e
c
l
a
r
t
i
M
B
C
D
Figure 3. Simple olfactory system that contains six mitral and
one granule cell. Gray areas indicate strong activity. Light gray
areas on the right indicate weak responses due to possible small
correction induced by the cost function. (A-D) responses to four
different odorants represented by the combinatorial inputs shown
on the left.
The result of sparsening of the m itral cell responses is the
reduction of redundancy of the odorant representation.
Thus, although m itral cells may receive highly redundant
inputs (Figure 3B and C, lef t column), only few may
respond (right column). The sparseness in the mitral cell
responses may be
important
for reducing overlaps
between odorant representations.
in
the example
In
Figures 3B and C, responses of mitral cells to two
odorants share no overlap. Figure 3D shows
the
responses to an odorant that are inhibitory (below the
baseline activ ity). Therefore, the incompleteness in the GC
odorant representation may emerge from the restriction on
the diversity in the GC weights.
4
Granule cell nonlinear input-output relationship may
be responsible for sparse granule cell coding. The
example presented above contained only one GC.
Therefore GC code was incomplete leading to the mitral
cell firing. If the number of GCs was equal or larger than
the number of mitral cells, there are enough vectors in the
basis to represent any glomerular input. Because the
olfactory bulb contains many more GC than mitral cells
one
should
seek
some other
source
for
the
incompleteness of GC representation. We propose that a
source of incompleteness is in the constraint that GC firing
rates cannot be negative.
To satisfy non-negativ ity condition for the GCs firing rate,
iC a in equation (2) should be infinite
the cost function
iC a is defined by the
for negative ai . In general, cost
non-linearity in the cells’ input-output relationship. The
simplest example is shown at Figure 4.
For the positive firing rates the cost function increases
proportionally to the GC threshold for firing , making it
more costly to fire for the cells with a large threshold. This
feature selects which GC will be active in the stationary
state with the higher threshold cells less likely to fire. The
exact relationship between the input-output relationship
and the cost function is given by equation (11) of the
Methods.
A
a
( )g u
B
( )C a
a
a
0
u
Figure 4. Granule cell input-output relationships (A) and the
( )C a (B). u is the granule cell's
corresponding cost-functions
input current. a is the resulting firing rate. The cost function
prohibits negative firing rates, which reflects the feature of the
input-output relationship. For positive firing rates the cost-function
is approximately linear.
Because GC firing rates cannot be negative, some
odorants cannot be exactly represented by the GCs.
Indeed, consider the gray regions in Figure 2A. The input
vector x
in this area cannot be represented as a
superposition of GC weights (gray arrows) with posi tive
coefficients. To exactly represent a vector in this area, at
least one GC has to have a negative firing rate, which is
not allowed. Therefore, the inputs within gray area will lead
to an incomplete representation by the granule cells. The
exact balance between inhibition and excitation on the
input of some mitral cells will be destroyed and, according
to equation (4), some mitral cells will respond to these
inputs.
A more high-dimensional example is shown in Figure 5.
From this example it is clear that substantial responses of
mitral cells emerge when the input from receptor neurons
falls outside the conical domain spanned by the GC weight
vectors (gray vectors in Figure 5B). In this case the
inhibitory currents from the granule cells to the m itral cells
on conical domain that
are given by the nearest vector x
can be represented as a superposition of GC weights with
positive coefficients. The response of the mitral cells to the
odorant it the result of the difference between excitatory
receptor neuron inputs x
and inhibitory inputs from the
. Thus, substantial mitral cell response is
granule cells x
, i.e. there is no precise balance of
possible if x
x
inhibition and excitation on the input to mitral cells. In this
case the representation of the receptor neuron inputs by
the GC is incomplete.
The responses of GC are sparse for the incomplete
representations. Thus, in Figure 5B, for the 3D space
characterizing inputs into three mitral cells only two GCs
out of eight (red arrows) are active. This is because GCs
approx imate the vector lying on the surface of the region
spanned by GC weights that is 2D. In multidimensional
space, when M mitral cells are present, the active GCs
weight vectors lie on the surface of the domain spanned by
all GCs, so the number of active GCs is smaller or equal
1M .
Because the number of GCs is much larger that the
number of mitral cells N
M
that
, one expects
responses of GCs are very sparse. This reasoning is
confirmed more rigorously in the Theorem 1 of Methods.
Complete GC representation
A
x
x
Incomplete GC representation
B
x
y
x
iW
iW
Figure 5. The olfactory bulb with three mitral cells (3D input
space) and eight GCs. The inputs into mitral cells from receptor
neurons are represented by the black arrows. The synaptic
weights from eight GCs to the mitral cells are shown by the gray
arrows. (A) If inputs lie within the cone restricted by the weight
vectors (dashed) the input vector can be represented as a
superposition of weight vectors exactly. The GC representation of
5
To demonstrate our point we will use simplified model of
bulbar network containing only one granule cell (Figure 3)
This network has an advantage that an exact solution can
be found and understood even when a f inite threshold for
firing is present in the activation of the GCs. Indeed,
consider input configuration shown in Figure 3B. Assume
for simplicity that all of the non-zero weights and mitral cell
inputs have unit strengths. Then, the Lyapunov function for
the activ ity of the single GC a is
K
a
a
( )
(1
.
(5)
2
a . K is the
0
Here, of course, we have to assume that
3K in Figure
number of non-zero weights for the GC (
3). By minimizing the Lyapunov function we obtain
a
K
K
1
/
, for
, and zero otherwise. The activ ity
of the mitral cell number two (Figure 3B and 6) cannot be
affected by the GC, because the latter makes no synapses
onto this cell. The activ ities of mitral cells 1, 4, and 6 are
given by
a
)
2
(6)
y
1, 4,6
a
1
K
K
for
. This formula means that the mitral cells will
increase their firing rate slightly to activate the GC. The
amount of increase is equal to the threshold for activation
of the GC div ided by the number of mitral cells contributing
to the input current, i.e. K . The activ ity of the GC is also
assumed to raise fast above the threshold so that it
suppresses all firing increases on its input above this value
[equation (6)]. The responses of m itral cells as functions of
the threshold are shown in Figure 6.
Figure 6 shows that for large thresholds all mitral cells
that receive receptor inputs will respond to the odorant. In
this regime the GC is silent due to the high threshold and,
therefore, the mitral cells relay the receptor neuron inputs
directly. The m itral cell code is not sparse in this case. W e
argue that the case of large thresholds corresponds to the
network in the anesthetized animal. When the threshold for
GC firing is small, the m itral cell firing becomes sparse
(Figure 6). This is the regime considered throughout this
paper, which, we argue, corresponds to the awake animal.
According to this model therefore the transition from
awake to anesthetized state is accomplished by an
increase in the thresholds of GC firing, which could be
mediated by the centrifugal projections.
mitral cell inputs is exact and the mitral cells are expected to
respond weakly due to exact balance between inhibition and
excitation. (B) If the input vector is outside the cone of GC
weights (dashed) the GCs cannot represent the inputs precisely.
This is because the GC firing rates (coefficients of expansion)
cannot be negative. The best approximation of the inputs is
is the nearest vector on the cone
. Vector x
shown by vector x
to the input vector x
. This vector is formed by two GCs (red).
Because x
, the GC code is
differs from GC representation x
representation
Incompleteness of
incomp lete.
to
leads
.
substantial mitral cell firing y
x
x
Incomplete granule cell representations are typical for
random network weights. It may appear that incomplete
representations are lim ited to the edges of the parameter
space, such as the gray area in Figure 2A. The situation
shown in Figure 5A may seem to be more typical than that
in Figure 5B. Here we argue that this intuition applies only
to input spaces of small dimensionality such as shown in
Figures 2 and 5. For a multidimensional input space, which
corresponds to large number of m itral cells, random inputs
and network weights create incomplete representations.
This implies that it becomes almost impossible to expand a
random input vector with positive components into the
basis containing vectors with positive components using
only non-negative coefficients without the loss of precision.
In the Methods section we show that the number of
coactive GC for random binary inputs with M mitral cells
is ~ M . Because a precise representation of the M -
dimensional random input requires M vectors, this result
implies that the representation of odorants by the GCs is
typically imprecise. The GC code is therefore incomplete.
for sparse GC-to-mitral cell
We also show
that
connectiv ity, when only K M
weights are non-zero,
the number of coactive GC
larger
is somewhat
/M K . This number is still substantially smaller that
~
what is required for complete GC code. We conclude that
GC cannot represent m itral cell inputs precisely for the
case of random connectiv ity, which implies ubiquity of
incomplete representations.
The state dependence of the granule cell code. Above
we have discussed two reasons for the incompleteness of
GC representation. First, the incompleteness may result
from the restricted diversity of the GC weights (Figure 3).
Second, we suggested that the non-negativ ity of the GC
firing rates may lead to an inaccurate representation of
odorants implying incompleteness. Here we discuss the
third reason why GC may yield an inaccurate code. W e
suggest that the presence of f inite (non-zero) threshold for
GC
firing
(Figure 4A) can
lead
to an
inaccurate
representation. This phenomenon allows us to explain the
transition between the awake and anesthetized responses.
6
1L norm is called sparse and
respond to a stimulus. Thus,
the corresponding neural representation is called sparse
overcomplete.
It was found that a network of inhibitory neurons can
implement sparse overcomplete codes (Rozell et al.,
2008). To show this, the dynam ics of the system was
represented as m inimization of a cost function called a
Lyapunov function. The basic idea of this method goes
back to John Hopf ield, who was the f irst to evaluate the
Lyapunov function for a recurrent network (Hopf ield, 1982,
1984). Hopfield network has attractor states that allow to
implement associative memory (Hopf ield, 1982; Hertz et
al., 1991). In contrast to Hopfield networks, in purely
inhibitory networks, the recurrent weights entered the
Lyapunov function with a m inus sign, which abolished the
attractor memory states and made the network purely
sensory (Rozell et al., 2008). Minimization of Lyapunov
function in realistic recurrent networks with inhibition was
suggested
to
implement
the parsimony constraint
mentioned above.
To implement sparse overcomplete representations with
realistic networks of neurons at least two requirements on
the network connectiv ity have to be met (Rozell et al.,
2008). First, the feedforward weights between the input
layer of the network and the inhibitory neurons have to
contain the dictionary elements (Figure 7A). Thus, the
inhibitory neurons that are responsible for a particular
dictionary element will be driven strongly when it is present
in the input, due to high overlap between the stimulus and
the feedforward weight. Second, the recurrent inhibitory
weight between any pair of neurons has to be proportional
to the overlap between their dictionary elements (Figure
7A). Because inhibition between neurons implements
competition,
tuned
that similarly
implies
feature
this
neurons compete most strongly. Therefore, the two types
of network weights, feedforward and recurrent, have to be
in close match, one of them constructed as the overlap of
the other. It is not clear how this match is established in a
biological system, because these are different sets of
synapses. Here we argue that the olfactory bulb network
architecture that is based on dendrodendritic synapses
can ensure that the feedforward and recurrent connect iv ity
are in close match.
Mi tral cell responses
y
2y
1, 4 , 6y
0
K
Figure 6. Sparsification of responses as a function of parameter
(threshold for GC firing). Example shown in Figure 3B is used.
We argue that for large thresholds the response of mitral cells is
less sparse than for small thresholds. Because the threshold for
GC firing can be regulated by the centrifugal inputs from cortex,
increase in firing threshold may mediate the transition between
awake and anesthetized states.
DISCUSSION
Sparse overcomplete codes. At the core of these
representations is an assumption that a sensory input can
be decomposed into a superposition of primitives that are
called dictionary elements (Field, 1994; Olshausen and
Field, 2004; Rozell et al., 2008). The decomposition is
sought in the form of a set of coefficients with which
different dictionary elements contribute to the input. These
coefficients are thought of as responses of neurons in a
high level sensory area that indicate whether given feature
is present in the stimulus or not. Because the number of
primitives
several
large,
usually
is
available
decompositions are consistent with the inputs. If the set of
primitives is overcomplete, the decomposition problem
generally has no unique solution. To make sensory
representation unambiguous some form of parsimony
principle is added to the model in the form of cost function
on the coefficients/responses. The solution that yields a
minimum of the cost function is assumed to be chosen by
the nervous system. The decomposition is found to be
dependent on the cost function. The general form of a cost
in power :
is a sum of
function
firing
rates
. For L2 , a simple sum of squared
L
ai
i
coefficients, it can be shown that generally all neurons will
respond to any stimulus, and, therefoore, the code is not
1L norm when the cost is equal to the sum of
sparse. For
absolute values of the responses, the solution is found to
be sparse. This means that only a few neurons will
7
Sparse overcomplete representation
networks
Recurrent
layer
A
Input
layer
imd
G
ik
?
m
d d
im km
Olfactory bulb
B
Mitral
cells
Granule
cells
imd
G
ik
m
d d
im km
Figure 7. A possible functional significance of dendro-dendritic
synapses. (A) Networks implementing sparse overcomplete
representations. These networks decompose the input pattern
into a parsimonious superposition of dictionary elements.
Response of each cell in the recurrent layer indicates whether
given dictionary element is present in the input. The feedforward
excitatory weights contain the dictionary elements. The recurrent
inhibitory weights are proportional to the overlap between
dictionary elements. More similar
input patterns compete
stronger. It is not clear however how this condition on the
recurrent weights is implemented biologically. (B) In the networks
with dendrodendritic synapses the connectivity between granule
cells is inhibitory. Because every two granule cells are two
synapses away from each other, the strength of pair wise
inhibition contains overlap between dendrodendritic weights thus
automatically satisfying the constraint on the recurrent inhibitory
weights shown in (A).
Dendrodendritic synapses facilitate the representation
of odorants by the granule cells. GCs of the olfactory
bulb receive excitatory inputs from the mitral cells through
the dendrodendritic synapses (Shepherd et al., 2004).
These synapses encode patterns that can drive strongly
indiv idual GCs. At the same time the effective connectiv ity
between GCs is inhibitory (granule to mitral and m itral to
granule
synapses
are
inhibitory
and
excitatory
respectively). Here we argue that the strength of mutual
inhibition between GCs may be proportional to the overlap
between their input weights. The reason for this is that
indiv idual GCs are two synapses away from each other.
To calculate their mutual inhibition one has to sum over all
intermediate synapses, which leads to evaluation of a
convolution or overlap (Figure 7B). Therefore, GCs inhibit
each other stronger if their inputs from mitral cells are
8
to
more sim ilar. Thus, both requirements necessary
implement sparse overcomplete representations stated
prev iously are met. This implies that the function of GCs
may be to detect specific patterns of activ ity in the inputs
that m itral cells receive. The GCs
then build
the
parsimonious representation of mitral cell inputs. The
parsimony is ensured by the mutual inhibition between
GCs, with more sim ilar GCs inhibiting each other stronger.
The
latter condition
is
facilitated by
the network
architecture based on dendrodendritic synapses. This
observation prov ides a potential explanation
the
for
existence of these synapses that so far have been
discovered in the olfactory bulb only (Shepherd et al.,
2004).
Mitral cells respond to odorants because granule cell
code is incomplete. Two problems emerge if we assume
that granule cells implement sparse overcomplete codes.
First, granule cells are interneurons and, as such, cannot
directly transmit their representation to the upstream
network. The significance of
these
representations
becomes unclear. Second, we find that if granule cells
establish an absolutely accurate representation of their
inputs, the mitral cells will respond to odorants very
weakly. This is because granule cells can eliminate these
responses
from
the mitral
cells'
firing. These
considerations suggested to us that the representations by
the granule cells are incomplete. This implies that granule
cells cannot find accurate representations of their inputs,
for example, because this would require for the firing rates
to become negative. If granule cells' codes are incomplete
the m itral cells transmit only the unfinished portion of the
representation to the upstream olfactory networks. As a
consequence the representations of odorant by the m itral
cells becomes sparse. The redundancies in the m itral cell
codes are reduced and the overlaps in representations of
similar odorants are erased. As a result the patterns of
mitral responses
to similar odorants become more
distinguishable from each other.
The Lyapunov function. Lyapunov functions are standard
tools in the neural network theory (Hertz et al., 1991).
(Seung et al., 1998) have shown that the network that
contains
two populations of neurons,
inhibitory and
excitatory, can be described by the Lyapunov function.
This model can be mapped onto the system of mitral-GCs.
Here we have eliminated the dynamics of mitral cells from
consideration and reduced the description to GCs only. By
doing so we have shown that GC could be used to extract
features from mitral cell inputs. (Lee and Seung, 1997)
have proposed a network mechanism that implements
conic encoders, i.e. the ones that represent inputs under
non-negativ ity constraint. Although the non-linear network
mechanism is somewhat different from the one used here,
the set of encoding/error neurons in Lee and Seung can
be v iewed as analogs of granule/mitral cells, respectively,
in our model. In our study we propose a mapping of the
conic networks onto the olfactory bulb network and study
the implications of this mapping for the olfactory code.
Feedback
from higher brain areas and state
dependence of the olfactory code. The sparseness of
the mitral cell responses depends on the nonlinearity of
the GCs, and specifically on the GC activation threshold
. In this study we assumed that GCs have similar
activation threshold that is small enough for them to be
easily activated by small levels of activ ity in the mitral cells.
If the thresholds for activation of indiv idual GCs are
different, one can env ision a mechanism how olfactory
code carried by both m itral and granule cells can be
controlled to adapt to a particular task. Thus, if the
threshold for activation is raised for a subset of GCs, these
cells will no longer be active and, therefore, their activ ity
will not be extracted from the firing of mitral cells. If, for
example, the threshold for all of the GCs is increased thus
making them unresponsive, the olfactory code carried by
the mitral cells replicates their inputs from receptor
neurons. Conversely, if the activation threshold is lowered
for a subset of GCs, these cells will efficiently extract their
activ ity from the m itral cells' responses. Thus, a particular
redundancy amongst similar odorants can be excluded in
a task-dependent manner. Therefore, the thresholds for
GC activation may regulate both an overall sparseness of
mitral cell responses and the f ine structure of the bulbar
olfactory code.
GC threshold values depend on cellular properties, but
also can be effectively modulated by additional input into
these cells. GCs in the mammalian olfactory bulb are
recipients of the signals from the efferent projections from
the cortex and other brain areas (Dav is and Macrides,
1981; Luskin and Price, 1983). Additional signal to GCs
can change the effective threshold values. If a GC
receives excitatory inputs from the cortex, the m itral cell
signal is closer to the threshold value and the granule cell
is easier to excite by the odorant-related inputs. These
features prov ide an opportunity to control sparseness and
information content of the mitral cell signal by regulating
the input to the GCs from the cortex.
Implicit experimental ev idence for the presence of such
regulation already exists. In the anesthetized state, when
the efferent signal coming to the bulb from the cortex is
minimal, the mitral cells respond strongly to the odor
stimulation. In the awake state, when the cortex is active,
the mitral cell code becomes much sparser (Adrian, 1950;
Rinberg et al., 2006). Cutting lateral olfactory tract and
eliminating feedback from the brain in awake rabbit made
mitral cells respond in the similar manner as in an
anesthetized rabbit (Moulton, 1963). Therefore centrifugal
projection may indeed regulate sparseness of the olfactory
code.
Some ev idence also points towards the possibility of finer
network tuning by specific activation or deactivation of
to enhance extraction of relevant
subsets of GCs
information. First,
(Doucette and Restrepo, 2008)
demonstrated that the mitral cell responses to odorants
change as animals learn the task. Second, Fuentes et.al.
show that the number of mitral cell responses depends on
the task (Fuentes et al., 2008). When a rat is involved in
odor discrimination task the average number of mitral cell
excitatory responses is less then in the rat passively
smelling an odor. The assumption is that when an animal
discriminates odorants,
to
it may be advantageous
suppress redundant m itral cell responses and enhance
those, which carry
the most behav iorally
relevant
information. Our model proposes the neuronal mechanism
for this phenomenon.
Transient regime. In this study we explicitly considered
the steady state regime and used Lyapunov function
approach to find the response of the mitral and GCs to the
stimulus in the stationary state. Obv iously the system 's
response to odorants is dynamic. What happens when the
stimulus appears, when an animal first sniffs an odorant?
The thorough analysis of this problem is beyond the scope
of this paper, however, we would like to present some
obv ious points, which
lead
to specif ic experimental
predictions.
If the signal first appears on the input to mitral cells, it
causes elevation of the mitral cell firing rates, which in turn
causes activation of GCs and suppression of m itral cells
by feedback inhibition. The initial responses of the mitral
cells are therefore represented by quick bursts of activ ity
followed by decay to the steady state described in this
study. The olfactory bulb may transmit information about
the stimulus by these quick bursts of activ ity. This
observation
leads
to several questions. First, what
information about the stimulus is sent to the cortex by
transient bursts and of activ ity and the consequent steady
state? Second, what are the experimental conditions for
observations of such bursts?
While it is hard to say conv incingly anything about the role
of different modes of information transmission, we may
speculate about the second question. The mitral cell bursts
of activ ity presumably are short and stand on top of the
high spontaneous activ ity. In order to observe them
to synchronize spike trains with
reliably one needs
stimulus delivery.
In mammals, stimulus delivery
is
controlled by sniffing. In the prev iously studies (Kay and
Laurent, 1999; Rinberg et al., 2006; Doucette and
Restrepo, 2008; Fuentes et al., 2008),
the authors
9
synchronized their recordings with stimulus onset, but not
with sniffing/breathing. This approach may
to
lead
smearing of the short bursts, and emphasize the steady
state responses of the network. In such regime the odor
responses should be sparse as predicted by the model.
New ev idence by (Shusterman et al., 2009) and (Cury and
Uchida, 2009) suggests the presence of fast bursts of
mitral cell activ ity synchronized with sniffing.
Conclusions. In this study we addressed theoretically the
experimental ev idence of sparse odor codes carried by the
mitral cells in awake behav ing rodents. W e proposed a
novel role for the granule cells of the olfactory bulb in
which they collectively build an incomplete representation
of the odorants in the inhibitory currents that they return to
the m itral cells. Because the representation formed by
granule cells
is
incomplete, mitral cells can carry
information to the cortex in the form of sparse codes. This
function is facilitated by the network architecture that
includes bidirectional dendrodendritic mitral-to-granule cell
synapses. W e suggest that the synaptic strengths in the
indiv idual synapses are correlated between
the
two
directions.
Acknowledgments. The authors are grateful to Dmitry
Chklovskii, Barak Pearlmutter, and Sebastian Seung for
helpful discussions.
METHODS
Derivation of the Lyapunov function.
Our model is based on the following equations describing
the responses of mitral and granule cells respectively:
N
y
x
W a
m
m
mi
i
i
1
M
W y
im
m
1
g u
a
(
)
(9)
,
i
i
iu is the membrane voltage of the granule cell and
where
y
y
y
. We use the following abbrev iation for the
m
m
m
u . We assume therefore
/du dt
temporal derivatives
that the dynamics of mitral cell responses is fast enough to
reflect the instantaneous values of inputs. If one def ines a
function of granule cell firing rates
1
M
a
( )
(cid:0)
2
i
m
1
where the cost function is defined as
a
a da
g
C a
')
0
W a
mi
i
C a
i
N
i
1
u
, (10)
(11)
(7)
(8)
x
m
u
,
1
(
m
m
m
2
'
2
0
(12)
one can show that the behav ior or the system can be
v iewed as
i.e.
form of gradient descent,
a
ui / ai . The time derivative of the Lyapunov
function is therefore always negative
ai
( ui )
ai
ai
i
i
ai
g1 ( ai )
i
( )g u is a monotonically
The last inequality holds if
increasing function. Therefore the Lyapunov function
always decreases if the system behaves according to
equations (7) through (9). Since the function is limited from
below, the stable states of the system are described by the
minima of the Lyapunov function.
threshold-linear neuronal
the
function
for
The cost
u
g u
]
[
( )
activation function
shown in Figure 4A
can be approximated by a linear function when the firing
rate of the granule cell are not too large
a a
C a
0
( )
,
(13)
For negative values of firing rates a the cost function is
infinitely big, reflecting the fact that negative firing rates are
not available.
The importance of condition (1)
For our model to have a Lyapunov function a more general
condition
than (1) may hold.
Indeed,
the sufficient
condition for the Lyapunov function to ex ist is that the
WimWmk
Gik
network weight matrix
is symmetric
i
(Hertz et al., 1991). This will be true if , for example,
W E
nmE is an arbitrary symmetric M
W
where
mi
nm
in
by M matrix. Thus condition (1) is sufficient but not
necessary for the network to have a Lyapunov function.
We argue however that this condition is necessary for the
system to be described by the Lyapunov function in the
form of equation (2).
The number of coactive granule cells
Here we will distinguish two types of Lyapunov functions.
First, we will consider the homogeneous case, when the
Lyapunov function has the form
1
M
a
( )
(cid:0)
0
2
i
m
1
In the homogeneous case we disregard by the contribution
of the cost function other than constraining the firing rates
to be non-negative. This condition corresponds
to
vanishingly small threshold for the activation of the granule
cells. The non-homogeneous Lyapunov function is
W a
mi
i
(14)
x
m
.
2
10
(15)
a
a
a
( )
( )
(cid:0) (cid:0)
i
i
0
i
ia . We will prove here two
0
with the same constraint
theorems that limit from above the number of coactive
ia ). A quantity
0
granule cells (i.e. the ones for which
that will be important to us is the error in approx imation
W a
x
x
that is also, according to equation
i
mi
m
m
i
(4), the response of mitral cells.
Theorem 1: Sparseness of the homogeneous solution
Assume that the M -dimensional receptive fields of the
iW
are the vectors of general position, i.e.
granule cells
every subset of M vectors
iW
are linearly independent.
Then, in the m inimum of the homogeneous Lyapunov
mx for all m (i.e. mitral cells do
0
function (14), either
not respond, granule cell representation is complete) or
fewer that M granule cells are active. In the former case
mx ) all of the granule cells may be active.
0
(
Proof: assume that M or more GCs are simultaneously
active. Let us vary slightly the activ ity of only one GC:
ka
. The corresponding variation in the Lyapunov
x W
O
(
)
2
(cid:0)
. Because we are
function is
k
0
considering the minimum of the Lyapunov function, all of
the scalar products
k
x W
have to be zero, which is
kW
x
0
or the number of vectors
possible only if
less than M .
Theorem 2: Sparseness of the inhomogeneous solution
the set of N M -dimensional vectors
Consider
iW
,
. Assume that these vectors are of general
i
i
1M of these vectors is
position, i.e. any subset of
linearly
independent. Then
in
the minimum of
the
inhomogeneous Lyapunov function (15) no more than M
granule cells can be simultaneously active.
Proof: For any active granule cell the following equation is
valid:
a
W x
/
(cid:0)
i
mi
m
i
m
Assume that more than M granule cells are active. Then
1M such equations for M unknowns
we have at least
mx . Such a system
1M
in general case (if
(16)
is
0
.
11
i
corresponding vectors
are independent) is inconsistent
and has no solution. Thus the number of coactive granule
cells cannot exceed M .
Note: Consider the case of small but non-zero thresholds
of firing of granule cells . In this case two regimes can
be distinguished. If vector x
can be expanded in terms of
with positive coefficients, the f iring rates of M
iW
vectors
granule cells are generally ~ 1 but the responses of m itral
cells are small ( ~ ). This is the regime of sparse
overcomplete representation. If the glomerular input vector
x
cannot be represented as a superposition of granule
iW
cell weights
with positive coefficients (incomplete
representation), the responses of cells are essent ially
(ignoring contributions ~ ) given by the solution of
homogeneous problem (14), which explains our attention
to this problem. In this case, according to Theorem 1,
fewer than M GCs have large firing rates and only one
has a small firing rate ( ~ ).
Coactive granule cells
the case of sparse
in
connectivity. Assume that the weight matrix between
granule and mitral cells is sparse, which implies that out of
miW for the i -th GC only K M
M matrix elements
are non-zero. This corresponds to the situation when the
GC can contact a limited subset of only K m itral cells.
Here we will calculate the number of GCs that are coactive
in the case of sparse connectiv ity. Clearly, fewer than M
GCs have to be coactive according to Theorem 2. W e
argue however that the number of coactive GC can be
substantially smaller than M .
0
i (16) leads to the following approx imate
W ith
ia -s
equation that is only valid for non-zero
Ka
K M
a
p
(
/
)
2
,
j
i
j
p W x
where
is the projection of the input vector on
i
i
synaptic weight. W e can find therefore the activ ities of the
GCs approx imately as functions of the threshold T :
. W e
S
a
T
SK M
p T K
a
2 /
(
) /
,
, where
i
i
i
i
can replace summation in the last equation with integration
of
distribution
the
over
projections,
K
N
p
p K
K
(
exp[ (
)
) /
] /
2
(we will assume
that the sparseness of input vectors is 50% for simplicity).
For T K
K
we obtain
(17)
i
Lennie P (2003) The cost of cortical computation. Curr Biol
13:493-497.
Lledo PM, Gheusi G, Vincent JD (2005) Information
processing in the mammalian olfactory system.
Physiol Rev 85:281-317.
Luskin MB, Price JL (1983) The topographic organization
of associational fibers of the olfactory system in
the rat, including centrifugal fibers to the olfactory
bulb. J Comp Neurol 216:264-291.
Moulton DG (1963) Electrical activ ity in the olfactory
system of rabbits with indwelling electrodes. In:
Olfaction and Taste I (Zotterman Y, ed), pp 71-84.
Oxford: Pergamon Press.
Olshausen BA, Field DJ (2004) Sparse coding of sensory
inputs. Current Opinion in Neurobiology 14:481-
487.
Rinberg D, Koulakov A, Gelperin A (2006) Sparse odor
coding
in awake behav ing mice. J Neurosci
26:8857-8865.
Rozell CJ, Johnson DH, Baraniuk RG, Olshausen BA
(2008) Sparse coding v ia thresholding and local
competition in neural circuits. Neural Comput
20:2526-2563.
Seung HS, Richardson TJ, Lagarias JC, Hopf ield JJ
(1998) Minimax and Hamiltonian dynam ics of
excitatory-inhibitory networks.
In: Advances
in
neural information processing systems 10, pp 329
- 335.
Shepherd GM, Chen WR, Greer CA (2004) Olfactory Bulb.
In: The synaptic organization of
the brain
(Shepherd GM, ed), pp 165-216. New York:
Oxford University Press.
Shepherd GM, Chen WR, W illhite D, Migliore M, Greer CA
(2007) The olfactory granule cell: from classical
enigma to central role in olfactory processing.
Brain Res Rev 55:373-382.
Shusterman R, Koulakov A, Rinberg D (2009) The
temporal structure of mitral/tufted cell odor
responses in the olfactory bulb of the awake
mouse.
In: Society
for Neuroscience annual
meeting. Chicago.
Tolley LK, Bedi KS (1994) Undernutrition during early life
does not affect the number of granule cells in the
rat olfactory bulb. J Comp Neurol 348:343-350.
(18)
.
(19)
exp
K
ln
N K
/
S T
( )
N K
T K
) /
(
2
T K
(
)
2
This equation has to be solved self-consistently with
T
SK M
2 /
to obtain
the
threshold
for
the
input
projection T for granule cells to be active. Because the
number of coactive granule cells can also be found from
)p we have
(
integrating
S T K M
GN T
(
( )
)
REFERENCES
Adrian ED (1950) The electrical activ ity of the mammalian
olfactory bulb. EEG Clin Neurophysiol 2:377-388.
Arev ian AC, Kapoor V, Urban NN (2008) Activ ity-
dependent gating of lateral inhibition in the mouse
olfactory bulb. Nat Neurosci 11:80-87.
Baraniuk RG (2007) Compressive sensing. Ieee Signal
Proc Mag 24:118-+.
Cury K, Uchida (2009) Odor coding based on spike timing
with respect to respiration-coupled oscillations
during
for
In: Society
sampling.
active
Neuroscience annual meeting. Chicago.
Dav is BJ, Macrides F (1981) The organization of
centrifugal projections from the anterior olfactory
nucleus, ventral hippocampal
rudiment, and
piriform cortex to the main olfactory bulb in the
hamster: an autoradiographic study. J Comp
Neurol 203:475-493.
Doucette W , Restrepo D (2008) Profound context-
dependent plasticity of mitral cell responses in
olfactory bulb. PLoS Biol 6:e258.
Field D (1994) What is the goal of sensory processing?
Neural Computation 6.
Fuentes RA, Aguilar MI, Aylwin ML, Maldonado PE (2008)
Neuronal activ ity of mitral-tuf ted cells in awake
rats during passive and active odorant stimulation.
J Neurophysiol 100:422-430.
Hertz J, Krogh A, Palmer RG (1991) Introduction to the
theory of neural computation. Cambridge, Mass.:
Perseus.
Hopfield JJ (1982) Neural networks and physical systems
with emergent collective computational abilities.
Proc Natl Acad Sci U S A 79:2554-2558.
Hopfield JJ (1984) Neurons with graded response have
collective computational properties like those of
two-state neurons. Proc Natl Acad Sci U S A
81:3088-3092.
Kay LM, Laurent G (1999) Odor- and context-dependent
modulation of mitral cell activ ity in behav ing rats.
Nat Neurosci 2:1003-1009.
Lee DD, Seung HS (1997) Unsuperv ised learning by
convex and conic coding. In: Advances in Neural
Information Processing Systems 9 pp 515-521.
12
|
1311.3211 | 1 | 1311 | 2013-11-13T17:04:41 | Stochastic inference with deterministic spiking neurons | [
"q-bio.NC",
"cond-mat.dis-nn",
"cs.NE",
"physics.bio-ph",
"stat.ML"
] | The seemingly stochastic transient dynamics of neocortical circuits observed in vivo have been hypothesized to represent a signature of ongoing stochastic inference. In vitro neurons, on the other hand, exhibit a highly deterministic response to various types of stimulation. We show that an ensemble of deterministic leaky integrate-and-fire neurons embedded in a spiking noisy environment can attain the correct firing statistics in order to sample from a well-defined target distribution. We provide an analytical derivation of the activation function on the single cell level; for recurrent networks, we examine convergence towards stationarity in computer simulations and demonstrate sample-based Bayesian inference in a mixed graphical model. This establishes a rigorous link between deterministic neuron models and functional stochastic dynamics on the network level. | q-bio.NC | q-bio |
Stochastic inference with deterministic spiking neurons∗
Mihai A. Petrovici1*, Johannes Bill2*, Ilja Bytschok1, Johannes Schemmel1, Karlheinz Meier1
1Kirchhoff Institute for Physics, University of Heidelberg
2Institute for Theoretical Computer Science, Graz University of Technology
(Dated: June 12, 2018)
The seemingly stochastic transient dynamics of neocortical circuits observed in vivo have been
hypothesized to represent a signature of ongoing stochastic inference. In vitro neurons, on the other
hand, exhibit a highly deterministic response to various types of stimulation. We show that an
ensemble of deterministic leaky integrate-and-fire neurons embedded in a spiking noisy environment
can attain the correct firing statistics in order to sample from a well-defined target distribution.
We provide an analytical derivation of the activation function on the single cell level; for recurrent
networks, we examine convergence towards stationarity in computer simulations and demonstrate
sample-based Bayesian inference in a mixed graphical model. This establishes a rigorous link between
deterministic neuron models and functional stochastic dynamics on the network level.
PACS numbers: xxx-xxx
Introduction
In responding to environmental sensory stimuli, brains
have to deal with what is typically limited, noisy and
ambiguous data. Based on such imperfect information,
animals need to predict and react to changes in their
environment. The recent hypothesis that the brain han-
dles these challenges by performing Bayesian, rather than
logical inference [1–3], has been strengthened by electro-
physiological data which identified neural correlates of
the involved computations [4, 5] and theoretical work on
spiking network implementations [6–8].
In probabilistic inference, the potential values of a
quantity are described by a random variable (RV) zk
and all knowledge about dependencies between ran-
dom variables is stored in a joint probability distri-
bution p(z1, . . . , zK).
The Bayesian belief about a
set of unobserved RVs {z1, . . . , zM} given an observed
set of RVs is represented by the posterior distribution
p(z1, . . . , zM zM +1, . . . , zK). In particular, the posterior
contains information not only on the most likely conclu-
sion and potential alternatives, but also on the level of
uncertainty of the outcome.
Theoretical work [2] has argued in favor of sample-
based representations of probability distributions in the
brain.
In this representation, instead of providing the
entire distribution at any point in time, only samples
z(t) ∼ p(z1, . . . , zK) are used as a proxy. When model-
ing large systems, this offers three important advantages.
First, approximate solutions can be provided at any time,
with increasingly reliable results as the calculation pro-
gresses ("anytime computing"). Secondly, in a sample-
based representation marginalization comes at no cost, as
p(zk) can be determined by simply neglecting the values
of all other RVs. Thirdly, some sampling algorithms such
as Gibbs sampling support a high degree of paralleliza-
tion. In particular, the resulting algorithmic structure is
reminiscent of neural networks [9].
FIG. 1. (A) Interpretation of spike patterns as samples of a
binary random vector z. The variable zk is active for dura-
tion τon (grey bar) after a spike of neuron k. (B) In stochastic
neuron models, internal state variables modulate the instan-
taneous firing probability (red).
In contrast, deterministic
integrate-and-fire neurons elicit a spike when the membrane
potential crosses a threshold voltage (blue). The probability
of firing as a function of the respective internal variable is
represented by the grayscale in the background.
Recently, a theory has been suggested which combines
these advantages by implementing Markov chain Monte
Carlo sampling in networks of abstract model neurons
[7]. In this framework, spike patterns are interpreted as
samples of binary RVs as follows (see Fig. 1A):
k = 1 ⇔ Neuron k fired in (t − τon, t]
z(t)
(1)
.
The duration τon of the active state following a spike is
a free parameter; in cortex, τon ≈ 10 ms is a good esti-
mate for the timescale on which a spiking neuron affects
the membrane potential of downstream cells. The neu-
ron model underlying [7] is inherently stochastic (Fig. 1B
top), with an instantaneous firing rate defined by
rk(t) = lim
∆t→0
(cid:26) 1
p(spike in [t, t + ∆t))
=
τ exp(vk)
0
∆t
if zk = 0
if zk = 1 ,
(2)
where vk represents an abstract membrane potential.
In contrast to this approach, in vitro experiments have
demonstrated the largely deterministic nature of single
neurons [10]. Similarly, microscopic models of neural cir-
cuits typically rely on deterministic dynamics of their
constituents. An often-used mechanistic model is the
leaky integrate-and-fire (LIF) neuron
rewrite (3) as
τeff
du
dt
= ueff − u ,
2
(5)
wki (Erev
i − uk) δ(t− ts) , (4)
τeff
du
dt
=
I ext + glEl
(cid:104) gtot (cid:105) +
Erev
i
i
i gnoise
(cid:104) gtot (cid:105)
− u ,
(6)
Cm
duk
dt
= gl(El − uk) + I
,
(3)
with capacitance Cm, membrane potential uk, leak po-
tential El,
leak conductance gl and input current I.
The spiking condition is deterministic as well: when uk
crosses a threshold ϑ from below, a spike is emitted and
uk is reset to for a refractory period τref (see Fig. 1B
bottom). For conductance-based synapses, the synaptic
input current to a neuron is typically modelled as
dI syn
k
dt
= − I syn
k
τsyn
+
(cid:88)
(cid:88)
syn i
spk s
with the synaptic time constant τsyn, the synaptic weight
wki and the reversal potential of the ith synapse Erev
.
i
The aim of this letter is to demonstrate how a network
of deterministic neurons in a biologically plausible spik-
ing noisy environment can quantitatively reproduce the
stochastic dynamics required for sampling from a well-
defined distribution p(z1, . . . , zK) and perform inference
given observations. We start by calculating the dynam-
ics of a single LIF neuron in a spiking noisy environ-
ment and derive its activation function by describing the
spike response as a first passage time (FPT) problem.
This establishes an equivalence to the abstract, inher-
ently stochastic units (2). On the network level, we show
how biologically realistic conductance-based synapses (4)
approximate the interaction for sampling from a well-
defined target distribution. We complement our study
with a demonstration of probabilistic inference by im-
plementing the posterior of a small graphical model for
handwritten digit recognition in a recurrent network of
LIF neurons.
Deterministic neurons in a noisy environment
k
+ I ext
obey eqn. (4), the current I ext
k . While the synaptic currents I rec
The total input current Ik to a neuron can be for-
mally partitioned into recurrent synaptic input, diffuse
synaptic noise and additional external currents: Ik =
k + I noise
I rec
and
I noise
captures additional
k
current stimuli. We start by considering a single neuron
that receives diffuse synaptic noise I noise
in the form of
random spikes from its surrounding. The capacity of re-
current networks to produce such noise has been shown
in [11]. Throughout the following analysis of individual
neurons we omit the index k and set I rec = 0.
k
k
k
When a conductance-based LIF neuron receives
strong synaptic stimulation, it enters a so-called high-
conductance state (HCS, [12]), characterized by acceler-
ated membrane dynamics. It is therefore convenient to
time
effective
replaced by a smaller
where the membrane time constant τm = Cm/gl
is
constant
τeff = Cm/gtot, with the total conductance gtot subsum-
ing both leakage and synaptic conductances. In a HCS,
τeff governs the decay towards an effective leak poten-
i + I ext)/gtot, where gnoise
represents the total conductance at the ith synapse. In
tial ueff = (glEl +(cid:80)
a high input rate regime, (cid:112)Var(gtot)/(cid:104) gtot (cid:105) → 0 and
i gnoise
Erev
i
i
the equation governing the membrane potential can be
written as
(cid:80)
with (cid:104)·(cid:105) denoting the mean. In a first approximation,
τeff can be considered very small in the HCS, resulting in
u ≈ ueff , with the effective potential ueff simply being a
linear transformation of the synaptic noise input.
Using methods similar to [13], it can be shown that,
if stimulated by a large number of uncorrelated spike
sources, the synaptic current I noise – and therefore, also
ueff – can be described as an Ornstein-Uhlenbeck (OU)
process:
du(t) = θ · (µ − u(t)) + Σ · dW (t) ,
(7)
(8)
(9)
with parameters
θ = τsyn
I ext + glEl +(cid:80)
(cid:88)
(cid:2)wnoise
νi
i
µ =
Σ 2 =
Erev
i
τsyn
i
i νiwnoise
(cid:104) gtot (cid:105)
(Erev
i − µ)(cid:3)2
τsyn /(cid:104) gtot (cid:105)
.
(10)
i
where νi represents the input rate at the ith noise synapse
and wi its weight.
The activation function as an FPT problem
The inherently stochastic neuron model (2) leads to a
logistic activation function for constant potential v:
p(z = 1) = σ(v) := [1 + exp(−v)]
−1
.
(11)
In the following, we derive the activation function of the
deterministic LIF neuron in a spiking noisy environment.
Similarly to the abstract model [7], we define the refrac-
tory state of a neuron as z = 1.
An example of membrane potential dynamics with re-
set is shown in Fig. 2A. Two modes of firing can be ob-
served: the "bursting" mode, where the effective mem-
brane potential after the refractory period is still above
threshold, and the freely evolving mode, where the neu-
ron does not spike again immediately after the refractory
period. Denoting the relative occurrence of burst lengths
n by Pn and the average duration of the freely evolving
mode that follows an n-spike-burst by Tn, we can identify
the following relation:
(cid:80)
(cid:80)
n Pn · n · τon
n Pn · (n · τon + Tn)
p(z = 1) =
.
(12)
Given the parameters of the associated OU process, we
can derive a recursive expression for Pn and Tn, thereby
ultimately allowing the calculation of p(z = 1):
Pn = p(un < ϑ, un−1 ≥ ϑ, . . . , u1 ≥ ϑu0 = ϑ)
(13)
dun−1p(un−1un−1 ≥ ϑ)
(cid:33)(cid:90) ∞
Pi
ϑ
(cid:35)
dunp(unun−1)
,
dun−1p(un−1un−1 ≥ ϑ)
=
Tn =
i=1
(cid:32)
1 − n−1(cid:88)
(cid:34)(cid:90) ϑ
(cid:90) ∞
(cid:34)(cid:90) ϑ
−∞
ϑ
−∞
dunp(unun < ϑ, un−1)(cid:104) T (ϑ, un)(cid:105)
(cid:35)
(14)
.
Fig. 2B displays an intuitive picture of the integrals in
(13) and (14). The transfer function p(unun−1) is the
Green's function of the OU process for t = τref :
(cid:20) (un−(un−1−µ) exp(−θτref )−µ)2
(cid:21)
− θ
Σ 2
p(unun−1) = C e
with the normalization C = (cid:112)θ/πΣ 2(1 − e−2θτref ).
1−exp(−2θτref )
, (15)
T (ub, ua) denotes the time the membrane needs to reach
ub starting from ua. This FPT problem has been ex-
tensively discussed in literature and its moments can be
given in closed form [14].
To improve the prediction of the activation function,
we further take into account small, but finite τeff ,
in
which case the membrane potential no longer directly fol-
lows the input current, but is a low-pass-filtered version
thereof. By using an expansion in (cid:112)τeff /τsyn, a first-
order correction to the FPT can be calculated [15]:
(cid:104) T (ϑ, u)(cid:105) = τsyn
√
π
dx exp(x2)[erf(x) + 1]
with the effective threshold ϑeff ≈ ϑ−ζ(cid:0) 1
u−µ
σ
2
(cid:1)(cid:113) τeff
2τsyn
, (16)
, where
ζ denotes the Riemann Zeta function. A comparison of
the predicted p(z = 1) with results from a numerical
simulation is shown in Fig. 2C. Here, the average effective
potential ¯u = (cid:104) ueff (cid:105) was established through an external
current I ext. For the translation from the LIF domain to
σ(cid:90)
ϑeff −µ
3
FIG. 2.
(A) Membrane potential u(t) (blue) and resulting
spike activity (black) of a LIF neuron in a spiking noisy en-
vironment. (B) Overlay of u (blue) and ueff (red). Due to
the small τeff in the HCS, the two curves are nearly identical
when the neuron is not refractory. At the end of each refrac-
tory state (corresponding to z = 1, grey), the predicted prob-
ability distribution for ueff is plotted in pink. The normalized
subthreshold area (dark pink) is used for the propagation in
(13). (C) Theoretical prediction (red) vs. simulation results
(blue); errors are smaller than the symbol size. A logistic
function σ(¯u) (green) has been fitted to the prediction.
the abstract stochastic model (2) we identify
¯u − ¯u0
α
,
v =
(17)
¯u0
the
value
¯u for which
where
denotes
p(z = 1) = σ(0) = 1
2 and α represents a scaling fac-
tor between the two domains. We conclude that a single
LIF neuron in a spiking noisy environment can closely
reproduce the activation function (11).
of
Sampling via recurrent networks of LIF neurons
with conductance-based synapses
k
We next connect the neurons to a recurrent network.
In addition to noise stimuli, a LIF neuron receives synap-
tic currents I rec
from other neurons in the ensemble.
Synaptic interaction introduces correlations among the
spike response of different neurons, i.e. the random vari-
ables z1, . . . , zK are not independent. For certain con-
nectivity structures, it is possible to specify a target dis-
tribution [7, 16] for the network states z(t) that occur
under the network dynamics .
In the following, we use the emulation of Boltzmann
machines as an example case. The joint distribution
reads:
pB(z) =
1
Z
exp
zTW z + zTb
,
(18)
(cid:18) 1
2
(cid:19)
K(cid:88)
(cid:17)
z(cid:48) exp
(cid:80)
(cid:16) 1
2 z(cid:48)TW z(cid:48) + z(cid:48)Tb
where W is a symmetric zero-diagonal weight ma-
trix, b is a bias vector, zTW z and zTb denote bi-
linear forms over W and 1 respectively, and Z =
is the partition function
that ensures correct normalization. This probabilistic
model underlies state-of-the-art machine learning algo-
rithms for image [17] and speech recognition [18]. For
the abstract neuron model (2), it had been proven [7]
that a membrane potential of the form
vk = bk +
Wkj zj
(19)
j=1
leads to the desired target distribution (18) of network
for t → ∞. This finding uses the fact
states z(t)
that individual neurons can sample from the condition-
als p(zk = 1 z\k) = σ(vk), with z\k = {zj j (cid:54)= k}, in a
Gibbs sampling inspired updating scheme.
As shown above, LIF neurons in a spiking noisy envi-
ronment closely approximate this logistic activation func-
tion if the synaptic currents I rec shift the mean mem-
brane potential ¯uk according to the linear interaction
(19). Using the linear transformation (17) between vk
and ¯uk, and estimating the impact of a pre-synaptic spike
on the post-synaptic neuron through conductance-based
synapses of weight wkj, we arrive at the following pa-
rameter translation between the abstract and the LIF
domain:
bk =(¯ub
Wkj =
k − ¯u0
wkj
1
k)/α
(cid:16)
(cid:17)
kj − µ
Erev
(cid:0)e−1 − 1(cid:1) − τeff
1 − τsyn
τeff
τsyn
αCm
(cid:104)
(20)
(cid:16)
e
(cid:17)(cid:105)
− τsyn
τeff − 1
,
(21)
4
FIG. 3.
(A) Spike pattern of a recurrent network of LIF
neurons during sampling from a randomly generated Boltz-
mann machine. (B) Sampled distribution pN(z) of network
states (blue bars) and analytically calculated target distribu-
tion pB(z) (in red). (C) Kullback-Leibler divergence between
the sampled distribution pN and the target distribution pB as
a function of integration time T for 10 trials (thin lines). The
red dotted line shows convergence for a network of intrinsi-
cally stochastic (theoretically optimal) neurons that is guar-
anteed to converge to pB for T → ∞. (D) Kullback-Leibler
divergence DKL (pN pB) when sampling for T = 103 s from
100 different randomly generated target distributions.
where ¯ub
k is the mean free potential ¯uk in Fig. 2C that
establishes p(zk = 1 z\k = 0) = σ(bk), and Erev
kj denotes
the reversal potential for synapse wkj. The idea behind
(21) is to match the integrals of individual postsynaptic
potentials (PSPs) on vk and ¯uk. We furthermore employ
short-term synaptic depression to approximate the the-
oretically optimal rectangular PSP shape also in case of
consecutive spikes (bursts).
The sampling process with networks of LIF neurons
was examined in computer simulations. Fig. 3A shows
the spike pattern of a recurrent network of K = 5 LIF
neurons that sample from a randomly generated Boltz-
mann machine. The parameters bk and Wkj were drawn
from the interval [−0.6, 0.6]. The thus defined target
distribution pB(z) is approximated by the distribution
pN (z) of network states when the spike pattern is inter-
preted as samples z(t) by convolution with a τon = 10 ms
box kernel. Fig. 3B shows the average network distri-
bution pN(z) (blue bars) after T = 10 s simulation time
alongside the target values pB(z) (red lines) calculated
from (18). Sampled probabilities pN(z) depict the mean
over 10 independent simulation runs, errorbars reflect
stochastic variations between individual runs. The cho-
sen integration time T = 10 s displays a conservative es-
timate of the maximum duration a neuronal ensemble
experiences stable stimulus conditions in a behaving or-
ganism and can thus be expected to sample from a sta-
tionary distribution. We find that the recurrent network
of LIF neurons accurately encodes the target distribution
over several orders of magnitude, within the precision im-
posed by the sample-based representation.
Fig. 3C shows how the network distribution becomes
increasingly more reliable as more samples are consid-
ered. After few samples, the network has generated a
coarse approximation of pB(z) that could serve as an "ed-
ucated guess" in online computation tasks. For simula-
tion times T well beyond biologically relevant timescales,
systematic errors in pN (z) become apparent: The KL di-
vergence saturates on a non-zero value, while the (the-
oretically ideal) abstract model [7] further converges to-
wards the target distribution (red dotted line). Fig. 3D
shows that the sampling quality holds for a variety of
similarly generated target distributions.
Demonstration of probabilistic inference
We conclude our investigation of sampling in recurrent
networks of LIF neurons with an example of Bayesian
inference based on incomplete observations. A fully con-
nected Boltzmann machine of K = 144 neurons, aligned
on a 12x12 grid, was trained as an associative network
[19] to store three prototypic patterns, namely hand-
written digits 0, 3 and 4.
In the network, each cell is
assigned to one pixel of the image. Statistical correla-
tions between the pixels as well as their mean intensities
were encoded in the weights Wkj and biases bk of the
corresponding joint distribution p(z). This distribution
reflects "prior knowledge" imprinted in the network.
The probabilistic model is augmented by adding real-
valued input channels for each pixel. Like the network
variables zk, input channels are associated with random
variables yk ∈ R, 1 ≤ k ≤ K. The resulting graphical
model is sketched in Fig. 4A and entails the following
structure for a full probabilistic model:
p(y, z) = p(z) · K(cid:89)
p(yk zk)
.
(22)
k=1
The full model p(y, z) connects the network variables
zk of the prior to inputs yk by means of the likelihood
p(yk zk). We have chosen a Gaussian likelihood with
unit variance (see Fig. 4B):
p(yk zk) = N (yk; µ = zk − 1
2
, σ2 = 1) .
(23)
The task for the network is to implement the posterior
distribution that follows from Bayes' rule: p(z y) ∝
p(z) · p(y z). The posterior combines two sources of
information: The likelihood p(yk zk) tends to align the
network state with the observation, i.e. zk = 1 for yk > 0,
while the prior p(z) reconciles the observations with
knowledge on consistent activation patterns z.
In this
way, the posterior p(z y) evaluates all possible outcomes
z y simultaneously by assigning a belief to each of them,
and thus captures the model's (un-)certainty about dif-
ferent solutions. A short derivation shows that the poste-
rior p(z y) is a Boltzmann machine for any input y, thus
being compatible with the sampling dynamics of spiking
networks. More specifically, we obtain the following ab-
stract membrane potential:
vk = bk + yk +
Wkj zj .
(24)
j
In the LIF domain, the sum bk + yk is equivalent to an
effective bias (20) and corresponds to an external cur-
rent I ext
k that shifts ¯uk appropriately. Thus,
k = I b
k + I y
(cid:88)
5
k
+ I b
k + I y
k .
In case of I y
a network neuron receives synaptic input from recurrent
connections and noise sources, as well as an external cur-
k + I noise
rent, i.e., Ik = I rec
k = 0 ∀k, the network samples from
the prior distribution p(z) = p(z y = 0). A two-
dimensional projection of network states z(t) ∼ p(z)
is shown in Fig. 4C. The sampled distribution exhibits
three distinct peaks that correspond to the three hand-
written digits stored in the recurrent weight matrix. A
closer look at the network trajectory reveals how the
system stays in one mode for some duration, traverses
the state space and then samples from a different mode
of the distribution. These dynamics also reflect in the
marginals of the network variables under a 20 ms box
k dt(cid:48), shown in the color
filter, ¯zk(t) = 1
maps.
(cid:82) t
t−20 ms z(t(cid:48))
20 ms
A typical scenario of stochastic inference on incomplete
observations is shown in Fig. 4D. Four input channels,
located at the center of the grid, were picked to inject
positive currents I y
k > 0 to the network while all other
inputs remained uninformative. Positive currents I y
k en-
code positive values of the respective input pixels yk and
were chosen such that the observation appeared incom-
patible with the digit 0, and remained ambiguous with re-
spect to digits 3 and 4. Accordingly, in the posterior dis-
tribution p(z y) the 0-mode is significantly suppressed
while uncertainty about the provided cue is expressed by
two distinct modes in the 3 and 4 directions.
Discussion
We have shown how recurrent networks of determin-
istic neurons in a spiking noisy environment can per-
form probabilistic inference through sampling from a
well-defined posterior distribution. Our approach builds
on theoretical work by Buesing et al. [7] and extends
Bayesian spiking network implementations to determin-
istic neuron models widely used in computational neu-
roscience. For the analytical derivation and the com-
puter simulations we have employed leaky integrate-and-
fire neurons with conductance-based synapses. However,
the analysis can be readily transferred to other neuron
models [20]. The essential diffusion approximation relies
on high-frequency spiking inputs that could be provided
by the surrounding network and lead to strong synaptic
conductances and fast membrane dynamics. Thereby,
our derivation identifies a potential functional role of bi-
ologically observed high-conductance states within a nor-
mative framework of brain computation.
For mathematical tractability, simplifying modeling
assumptions had to be made. The neuron model only
uses an absolute refractory time τref , which matches
the activation time constant τon, and neglects any ad-
ditional gradual recovery effects after a spike. On the
network level, we have assumed statistically indepen-
6
FIG. 4. (A) Graphical model used for the probabilistic inference task. The network implements a Markov random field over
latent variables z. Observables yk are conditionally independent given the network state. (B) A mixture-of-Gaussians likelihood
model provides input to the sampling neurons. (C) Two dimensional projection of sampled states z(t) when sampling from
the prior p(z). The network preferentially spends time in modes close to the stored hand-written digits 0, 3, 4. Solid line:
network trajectory over 200 ms. Color maps: Marginals ¯zk(t) averaged over 20 ms. The time arrow covers the duration of
the red trajectory and consecutive snapshots are 20 ms apart. (D) As in (C) when sampling from the posterior p(z y) with
incomplete observations y. The provided input y is incompatible with digit 0 and ambiguous with respect to digits 3 and 4.
dent noise sources and instantaneous axonal transmis-
sion. Furthermore, post-synaptic potentials mediated
through conductance-based synapses differ from the the-
oretically optimal rectangular shape and can lead to de-
viations from the target distribution outside of the high
noise regime [21]. However, computer simulations indi-
cate that in most scenarios the above approximations are
not critical.
Beyond neuroscience, the ability to perform proba-
bilistic inference with deterministic neurons displays a
promising computing paradigm for neuromorphic hard-
ware systems. Originally designed as neuroscientific
modeling tools, these systems typically implement a
physical model of integrate-and-fire neurons [22, 23],
which renders the application of the proposed networks
straightforward.
In particular, the distributed nature
of the sampling algorithm allows to exploit the inher-
ent parallelism of neuromorphic architectures, fostering
an application of neuromorphic hardware to online data
evaluation and robotics.
∗ The first two authors contributed equally to this work.
We thank W. Maass for his essential support and Ste-
fan Habenschuss for helpful comments. This research
was supported by EU grants #269921 (BrainScaleS),
#237955 (FACETS-ITN), the Austrian Science Fund
FWF #I753-N23 (PNEUMA) and the Manfred Stark
Foundation.
[5] P. Berkes, G. Orb´an, M. Lengyel, and J. Fiser, Science
331, 83 (2011).
[6] S. Deneve, Neural computation 20, 91 (2008).
[7] L. Buesing, J. Bill, B. Nessler, and W. Maass, PLoS
computational biology 7, e1002211 (2011).
[8] R. P. Rao, in Advances in neural information processing
systems (2004) pp. 1113–1120.
[9] P. O. Hoyer and A. Hyvarinen, Advances in neural infor-
mation processing systems , 293 (2003).
[10] Z. F. Mainen and T. J. Sejnowski, Science 268, 1503
(1995).
[11] N. Brunel, Journal of computational neuroscience 8, 183
(2000).
[12] A. Destexhe, M. Rudolph, and D. Pare, Nature Reviews
Neuroscience 4, 739 (2003).
[13] L. M. Ricciardi and L. Sacerdote, Biological Cybernetics
35, 1 (1979).
[14] L. M. Ricciardi and S. Sato, Journal of Applied Proba-
bility 25, 43 (1988).
[15] N. Brunel and S. Sergi, Journal of Theoretical Biology
195, 87 (1998).
[16] D. Pecevski, L. Buesing, and W. Maass, PLoS compu-
tational biology 7, e1002294 (2011).
[17] R. Salakhutdinov and G. E. Hinton,
in International
Conference on Artificial Intelligence and Statistics (2009)
pp. 448–455.
[18] A.-R. Mohamed, G. E. Dahl, and G. Hinton, Audio,
Speech, and Language Processing, IEEE Transactions on
20, 14 (2012).
[19] G. Hinton, P. Dayan, B. Frey, and R. Neal, Science 268,
1158 (1995).
[20] R. Brette and W. Gerstner, Journal of neurophysiology
94, 3637 (2005).
[21] W. Gerstner and J. L. van Hemmen, Network: Compu-
[1] K. P. Kording and D. M. Wolpert, Nature 427, 244
tation in Neural Systems 3, 139 (1992).
(2004).
[22] S. Mitra, S. Fusi, and G. Indiveri, Biomedical Circuits
[2] J. Fiser, P. Berkes, G. Orb´an, and M. Lengyel, Trends
and Systems, IEEE Transactions on 3, 32 (2009).
in cognitive sciences 14, 119 (2010).
[3] K. Friston, J. Mattout, and J. Kilner, Biological cyber-
netics 104, 137 (2011).
[4] T. Yang and M. N. Shadlen, Nature 447, 1075 (2007).
[23] D. Bruderle, M. A. Petrovici, B. Vogginger, M. Ehrlich,
T. Pfeil, S. Millner, A. Grubl, K. Wendt, E. Muller, M.-
O. Schwartz, and et al., Biological Cybernetics 104, 263
(2011).
|
1710.07613 | 1 | 1710 | 2017-10-15T07:24:12 | Quantum field inspired model of decision making: Asymptotic stabilization of belief state via interaction with surrounding mental environment | [
"q-bio.NC"
] | This paper is devoted to justification of quantum-like models of the process of decision making based on the theory of open quantum systems, i.e. decision making is considered as decoherence. This process is modeled as interaction of a decision maker, Alice, with a mental (information) environment ${\cal R}$ surrounding her. Such an interaction generates "dissipation of uncertainty" from Alice's belief-state $\rho(t)$ into ${\cal R}$ and asymptotic stabilization of $\rho(t)$ to a steady belief-state. The latter is treated as the decision state. Mathematically the problem under study is about finding constraints on ${\cal R}$ guaranteeing such stabilization. We found a partial solution of this problem (in the form of sufficient conditions). We present the corresponding decision making analysis for one class of mental environments, the so-called "almost homogeneous environments", with the illustrative examples: a) behavior of electorate interacting with the mass-media "reservoir", b) consumers' persuasion. We also comment on other classes of mental environments. | q-bio.NC | q-bio |
Quantum field inspired model of decision making:
Asymptotic stabilization of belief state via interaction
with surrounding mental environment
Fabio Bagarello 1
Irina Basieva 2
Andrei Khrennikov 3
Abstract
This paper is devoted to justification of quantum-like models of the process of decision
making based on the theory of open quantum systems, i.e. decision making is consid-
ered as decoherence. This process is modeled as interaction of a decision maker, Alice,
with a mental (information) environment R surrounding her. Such an interaction gen-
erates "dissipation of uncertainty" from Alice's belief-state ρ(t) into R and asymptotic
stabilization of ρ(t) to a steady belief-state. The latter is treated as the decision state.
Mathematically the problem under study is about finding constraints on R guaranteeing
1 Dipartimento di Energia, Ingegneria dell'Informazione e Modelli Matematici, Scuola Politecnica, Universit`a
di Palermo, I-90128 Palermo, INFN, Sezione di Napoli, ITALY and
Department of Mathematics and Applied Mathematics, Cape Town, South Africa
e-mail: [email protected] Home page: www1.unipa.it/fabio.bagarello
2 Department of Psychology, City University, London, UK
e-mail: [email protected]
3 International Center for Mathematical Modeling, in Physics and Cognitive Science
Linnaeus University, Vaxjo, Sweden
National Research University of Information Technologies, Mechanics and Optics (ITMO)
St. Petersburg, Russia
e-mail: [email protected]
such stabilization. We found a partial solution of this problem (in the form of suffi-
cient conditions). We present the corresponding decision making analysis for one class of
mental environments, the so-called "almost homogeneous environments", with the illus-
trative examples: a) behavior of electorate interacting with the mass-media "reservoir";
b) consumers' persuasion. We also comment on other classes of mental environments.
keywords: decision making; quantum-like model; mental (information) environment; open
quantum systems; dissipation of uncertainty; voters' behavior; consumers' persuasion
1
Introduction
The recent years were characterized by explosion of interest in applications of the mathematical
formalism of quantum theory to studies in cognition, decision making, psychology, economics,
finance, and biology, see, e.g., the monographs [1]-[6] and a few representative papers [7]-[33]
(the first steps in this direction were done long time ago, see, e.g., [34]).
The approach explored in such mathematical modeling is known as quantum-like. In this
approach an agent (human, animal, or even cell) is considered as a black box processing infor-
mation in accordance with the laws of quantum information and probability theories. Thus the
quantum-like modeling is basically quantum informational modeling (although this character-
istic feature is typically not emphasized, cf., however, with [35]).1
Quantum-like models have to be sharply distinguished from genuinely quantum physical
models of cognition which are based on consideration of quantum physical processes in the
brain, cf. with R. Penrose [38] and S. Hameroff [39]. Although the quantum physical models
have been criticized for mismatching between the temperature and space-times scales of the
quantum physical processes and neuronal processing in the brain, see especially Tegmark [40],
they cannot be rejected completely and one may expect that quantum-like models of cognition
will be (soon or later) coupled with real physical processes in the brain, see [41]-[47] for some
steps in this direction.
The quantum-like approach generated a variety of models of cognition and decision mak-
ing. In the simplest model [34], [1], the mental state (the belief state) of an agent, Alice, is
represented as a quantum state ψ and questions or tasks as quantum observables (Hermitian
operators). Answers to the questions are given with probabilities as determined by Born's rule.
1See, e.g. D' Ariano [36] and Plotnitsky [37] for the information approach to quantum mechanics.
2
For Hermitian operator A with the purely discrete spectrum, Born's rule can be written as
p(A = αk) = kPαkψk2 = hPαk ψ, ψi,
(1.1)
where αk is an eigenvalue of A and Pαk is the projector on the eigenspace corresponding to this
eigenvalue.
This model does not describe dynamics of the belief state in the process of decision making.
Consideration of dynamics was introduced in the works of Khrennikov [7, 8], and Pothos and
Busemeyer [10]. In their dynamical model as in the previous models, an observable A corre-
sponding to a question (task) faced by Alice is represented as a Hermitian operator. Then
Hamiltonian H generating unitary dynamics
ψ(t) = U(t)ψ0, U(t) = e−itH
(1.2)
of the initial belief state ψ0 is introduced, and Alice's decision is represented as measurement
of the observable A at some instant of time. The authors of [10] presented cognitive arguments
supported by experimental studies to determine the instant tm of measurement. Here the
probability of a particular answer is also determined by the Born rule, but applied to belief
state ψ(tm). Of course, this is an important issue, since different values ot tm can give rise to
completely different results. In spite of the partial progress in determination of tm in the article
of Pothos and Busemeyer [10], this complex problem cannot be considered completely solved.
Moreover, in solving this problem Pothos and Busemeyer had to go beyond the quantum theory
and to appeal to psychological theoretical and experimental studies. It would be attractive to
solve this problem entirely in the quantum framework.
We remark that the Schrodinger equation describes the dynamics of an isolated system. In
the presence of an environment, the dynamics of the system is non-unitary. Approximately (un-
der some sufficiently natural conditions) it is described by the Gorini-Kossakowski-Sudarshan-
Lindblad (GKSL) equation (often called simply the Lindblad equation), the simplest version
of the quantum master equation. One of the main distinguishing features of such dynamics is
that it does not preserve the pure state structure:
it (immediately) transforms a pure initial
state ψ0 into a mixed state given by the density operator:
ρ(t) = U(t)ρ0, U(t) = e−itL,
(1.3)
where L is the generator of the GKSL-evolution and ρ0 = ψ0ihψ0. This dynamics is in general
non-unitary. This equation describes the process of the system adaptation to the surrounding
environment. This is the complex dynamical process combining the internal state dynamics
3
of the system with adaptation to signals received from the environment. If the dynamics is
discrete with respect to time, then it can be represented as a chain of unitary evolutions and
(generalized) quantum Bayesian updates.2
In this paper we cannot discuss this interesting issue
in more detail, see [16] for detailed consideration of a two dimensional example with application
to the evolution theory.
In fact, GKSL-dynamics does not contradict the Schrodinger equation structure of the quan-
tum evolution. Let us denote the system under study by S and the surrounding environment
by R ("reservoir" for S.) Suppose that initial state of the compound system S + R is pure and
separable. Dynamics in the state space of S + R is still unitary and given by a Hamiltonian
for S + R. The main distinguishing feature of this unitary dynamics is that (in the presence of
interaction between S and R) it induces entanglement and the state of the compound system
becomes not more separable. Hamiltonians for the composite system S + R are very complex,
since they includes, in general, an infinite number of degrees of freedom of R. Typically it is im-
possible to solve the Schrodinger equation for the state of composite system S + R. (Although
in some special cases, as those considered in this paper and others discussed in [4] analytic
solutions can be found.) Therefore, most studies are restricted to the dynamics of the state
ρ(t) of S alone, which is described (approximately) by the GKSL-equation. But even if one
were able to solve the Schrodinger equation for S + R, the solution would be a very complex
infinite-dimensional state vector. Since we are interested in behavior of S, we would then take
the trace with respect to all degrees of freedom of R and obtain the state of S. A simple math-
ematical theorem implies that in presence of entanglement this trace-state cannot be pure, i.e.,
the state is descirbed by density operator. Its dynamics under the GKSL-equation is known as
decoherence: decreasing of state's purity (or coherence) in the process of interaction with an
environment.
Since consideration of an isolated cognitive system is even a higher degree idealization than
consideration of an isolated physical system, it is natural to modify the dynamical scheme of de-
cision making based on unitary Schrodinger dynamics [7, 8, 10] and consider general dynamics
of the belief-state, either by using the approximative GKSL-dynamics or by tracing the state of
2As was quickly understood in quantum physics, the Luders projection postulate describes only one very
special class of the quantum state updates resulting from measurements. We remark that already von Neumann
accepted applicability of this straightforward form of the state update only for observables with non-degenerate
spectra. Generally, in the case of an observable with degenerate spectrum, a pure pre-measurement state can
be transferred into a mixed state [48]. Later these considerations of von Neumann were elaborated in the form
of the theory of quantum instruments, see [49] for non-physicist friendly presentation. The most consistent
justification of this theory is obtained in the framework of the theory of open quantum systems.
4
the compound system. Roughly speaking there is no choice: either one has to ignore the pres-
ence of environment or consider non-unitary dynamics, e.g., (1.3). Of course, such non-unitary
dynamical model of decision making is much more mathematically complicated. However, it
has one very important advantage. Here the aforementioned problem of determination of tm is
solved automatically, although not straightforwardly.
For a natural class of quantum master equations, in the limit t → ∞ the system's state
ρ(t) approaches a steady state ρout. Diagonal elements of this operator (with respect to the
"pointer basis" corresponding to the observable A) under measurement give the probabilities
of measurement outputs. The model of decision making as decoherence was represented in a
series of papers of Asano et al., see, e.g., [51], [52] [13], [6]. The model is purely informational.
Both a "quantum-like system", Alice, and her mental (or information) environment ("bath",
reservoir") are represented by quantum states, ψ and φ. (It can be assumed that initially
these states are pure.) We are not interested in their physical or neurophysiologic realizations.
Asano et al. [51] applied the theory of open quantum systems and the GKSL-equation to model
experimental data collected in decision making experiments. The initial belief state is typically
represented as a pure state of complete uncertainty,
ψ = (0i + 1i)/√2,
(1.4)
where ii, i = 0, 1, are eigenstates of the dichotomous observable A representing a "yes"/"no"
question. A proper decision making dynamics should asymptotically (t → ∞) drive the belief
state of Alice ρ(t) to the output (mixed) state represented by the density operator ρout. Its
diagonal elements in the basis of eigenvectors of the operator A give probabilities of possible
decisions.
We remark that the open quantum systems approach to decision making can be considered
as a possible realization of the contextual treatment of cognition, cf. Khrennikov [1], [2] and
Dzhafarov et. al.
[12], [17]. The surrounding mental environment represents a measurement
context for decision making.
Of course, Alice cannot wait for t → ∞ to make a decision. And, as we know in physics,
a quantum system relaxes to a steady state very quickly. Here we have to point to difference
between a mathematical model and its applications to real phenomena. The notion of limit
is a mathematical abstraction. Of course, it is not applicable to real physical or cognitive
systems. In practice, the limit procedure encodes the process of approaching some quantity.
The existence of a limit for the density operator ρ(t) guarantees that its fluctuations decrease. If
the magnitude of fluctuations becomes smaller than some ǫ > 0, then such ρ(tǫ) can be selected
5
as a good approximation of the steady state (the latter is also a mathematical abstraction, in
real physical processes it is never approached exactly). This tǫ plays the role of tm from the
unitary dynamical model of decision making. Moreover, more complicated considerations based
on the theory of decoherence demonstrate that this instant of time tǫ can be identified with the
relaxation time T. The latter is determined by the structure of interactions between a system
and its environment. Typically in quantum physics the time interval [0, T ] is very short and a
system relaxes very quickly to its steady state. Since such considerations would make the paper
even more complex mathematically, we shall not proceed in this direction. (See, e.g., [53] for
derivation of an analytical expression for the relaxation time as a function of the heat-bath and
interaction parameters.) The main message of these considerations to a reader is that limt→∞
is just an abstraction (at t = ∞ all fluctuations disappear completely). In reality a system
relaxes very quickly to its steady state.
The main problem of the quantum(-like) decision theory is to construct an operator repre-
sentation of Hamiltonians and "Lindbladians" (the latter represents interactions of Alice with
her mental reservoir). In contrast to quantum physics, there is no analog of the classical phase
formalism for cognition and decision making, i.e., we cannot use the quantization procedure to
transform functions on the classical phase space into operators - the procedure of Schrodinger's
quantization is not applicable.
In quantum theory one can also use the quantization procedure based on the algebra of
operators of creation and annihilation. It is especially useful for second quantization (quantum
field theory), see appendix 1 for brief presentation of the basics. This sort of quantization
can be successfully applied to quantum-like modeling. Consider again a dichotomous question
A, with the values "no"/ "yes". Creation operator a⋆ creates "yes" from "no", annihilation
operator a transforms "yes" to "no". We can compound Hamiltonians and observables with the
aid of creation and annihilation operators (similarly, e.g., to quantum optics) . This approach
to construction of operators for problems of decision making was pioneered by Bagarello [4] and
coauthors who applied it, see, e.g., [28], [26], [27], to a variety of problems. Bagarello et al. were
interested in time dynamics of averages; in particular, probabilities were treated statistically.
In series of works [28], [26], [27], a system was interpreted as an agent and the reservoir had a
purely information interpretation, see especially [54] for a discussion.
The same approach based on creation-annihilation operators can be explored to model
decision making at individual level (with the subjective interpretation of probabilities). Here
one can explore the scheme which was approved in the works of Asano et al. [6]: consideration
of the asymptotic dynamics of the belief state and using the asymptotically output state (for
6
t → ∞) as the basis of decision making. The first step in this direction was done in paper [51].
The aforementioned decision making scheme (based on asymptotic stabilization of the mental
(belief) state of an agent) generates interest to study the asymptotic dynamics of the system
interacting with a reservoir. This is a nontrivial mathematical problem and its partial solution
(for specially designed information reservoirs) is presented in this paper.
In section 2 we discuss several aspects of the process of decision making for a dichotomous
question A through belief-state stabilization. Then, in section 3, the mathematical problem is
discussed in more details. More explicitly, in section 3.1 we prove stabilization of the belief state
in the case of the "almost homogeneous mental environment" of Alice. We start with the case
of a question' A having an infinite number of possible outcomes labeled as na = 0, 1, ..., n, ... .
Then in section 3.3 we consider the case of a "dichotomous question" A having two possible
outcomes, na = 0, 1. In section 3.2 we generalize our theory to the case of environments having
a more complex structure. Section 4 contains our conclusions, while appendices 1 and 2 are
devoted to few introductory remarks on canonical anti-commutation relations (CAR) and some
extra mathematical details.
2 Decision making as decoherence
Consider the case of two dichotomous questions posed to Alice, A = 0, 1 ("no", "yes") and
the mental environment R similarly composed of dichotomous degrees of freedom. In section
3.3 this situation will be modeled by using operators a, a† (Alice's operators) and b, b† (R's
operators) satisfying the CAR:
{a, a†} = I,{b, b†} = I,{a, b} = 0,
(2.1)
where the anti-commutator of two operators x, y is defined as {x, y} = xy + yx. Moreover,
a2 = b2 = 0. In quantum field theory, see appendix 1, these operators are known as the operators
of creation and annihilation.3 The operators related to R can depend on some parameter k
(discrete or continuous) representing degrees of freedom of R, b = b(k), b† = b†(k). Typically
environment R has a huge number of degrees of freedom and belongs to infinite-dimensional
Hilbert state space K. In the case of dichotomous questions asked to Alice, it can be assumed
that her state space H has the dimension of two, the qubit space of quantum information theory.
In quantum field theory the terminology "creation-annihilation operators" has coupling to
real physical systems. The operators represent the processes of creation-annihilation of quantum
3In physics these operators represent the processes of creation and annihilation of fermions, e.g., electrons.
7
particles (e.g., photons or electrons), see appendix 1. As was emphasized in introduction, our
model is of the purely informational nature. Therefore the operators a, a† and b, b† have to
be treated as the formal representation of the processes of "creation" and "annihilation" of
information states.4
Consider, e.g., Alice's operators a, a†. Let φ0 = 0i, φ1 = 1i be the orthonormal basis
in Alice's state space H corresponding to the answers "no"/"yes" to the question A. As a
consequence of the CAR, the operators a, a† act on these basis vectors in the following way:
a†0i = 1i, a†1i = 0; a1i = 0i, a0i = 0.
The operators a and a† modify Alice's attitude to selection of alternatives. In this framework
these operators are considered as reflection operators [54]. We represent the question A as the
number operator of Alice, na = a†a. The eigenvalues of this operator, na = 0, 1, correspond to
the choices of Alice at t = 0. Therefore it is natural to name na the decision operator.5
Consider the simplest situation which can be modeled in this framework: Alice is asked a
question A and she is surrounded by a population R whose members are asked the same question
A. Alice interacts with this population R and she gets to know behavior of its members with
respect to this question. Of course, she cannot "scan" R completely to know the concrete
answers of people to A. She just gets to know a sort of average nb of answers to A. In the
simplest case of a homogeneous population this average is a constant.
In the general case
average nb can non-trivially depend on the parameter k encoding various population clusters
R(k) : nb = nb(k). Here Alice (through interaction with this population) obtains information
about averages for the answers to A corresponding to different clusters R(k) of R.
In principle, R needs not be combined of physical agents. As was emphasized in introduc-
tion, our model is of the purely informational nature. The environment R can, for example,
represent mass-media's image of the question under consideration: the image created in TV-
debates and shows, web-blogs, newspapers. For example, A can be a referendum question:
"vote for Brexit or against?" (or the recent USA-election question: "vote for Trump or not?").
By analysing (generally unconsciously) mass-media's opinions Alice estimates the average nb
(or in general the averages nb(k)) and she makes her decision based on these averagess.
In a more general situation the members of R express their attitude with respect to a variety
of dichotomous questions and the A, the question asked to Alice, can be among them, but not
necessarily. In the latter case Alice makes her decision based on behavior of the surrounding
4In principle, we can simply call these operators ladder operators as is done in formal mathematical theory.
5In quantum field theory the number operator has the meaning of the number of physical particles, see
appendix 1. In our model it represents states indexed by numbers.
8
environment R (physical or informational) without even understanding that her decision is
made through influence of R.
Now we consider dynamics of decision making. In section 3 we shall use the Heisenberg
picture, i.e., we are concerned with the dynamics of operators. We are interested in the dynamics
of the reflection operators, a†(t), a(t), with initial conditions a†(0) = a†, a(0) = a. Based on the
reflection operators dynamics we can find the the decision operator dynamics na(t) = a†(t)a(t)
and, hence, its average
na(t) :=< na(t) ⊗ I >=< a†(t)a(t) ⊗ I >
(2.2)
with respect to some initial state of the compound system, see section 3 for details.
In the case of dichotomous questions the average na(t) has the straightforward probabilistic
meaning. To see this, let us expand expression (2.2). Denote the state space of the compound
system, Alice and her mental environment, by the symbol H⊗K and the operator of Heisenberg
evolution of this system by Ut. For an arbitrary initial state h., .iR, formula (2.2) reads as
na(t) =< Ut(a† ⊗ I)U †
t Ut(a ⊗ I)U †
t >R= TrUt(a†a ⊗ I)U†
tR.
By using the cyclic property of the trace we get:
tRUt = TrR(t)(a†a ⊗ I) =< a†a ⊗ I >R(t),
na(t) = Tr(a†a ⊗ I)U†
where R(t) = U †
t RUt represents the state dynamics (so we transferred the operator dynamics
into the the state dynamics, from the Heisenberg picture to the Schodinger picture).6 Consider
now the partial trace of R(t) with respect to the environmental degrees of freedom, ρ(t) =
TrKR(t). It represents the dynamics of Alice's belief-state in the process of her interaction with
the environment. Consider the average of the decision operator na = a†a = a†(0)a(0) with
respect to this belief-state:
< na >ρ(t)= Trρ(t)na.
This average is directly related to probability. Since the decision operator na has two eigenval-
ues, na = 0, 1, the average coincides with the probability:
Pt(A = 1) =< na >ρ(t)=< ρ(t)φ1, φ1 > .
(2.3)
6The Schrodinger picture providing the belief-state dynamics is basic for the probabilistic interpretation.
However, equations in the Heisenberg picture are easier for analytic treatment. Therefore in section 3 we shall
work in the Heisenberg picture.
9
We recall that here the question A is represented by the operator na. The eigenvector φ1
corresponds to the answer "yes" to this question. Pt(A = 1) is the subjective probability
in favor of this answer which Alice assigns at the moment of time t. We remark that this
assignment can be done unconsciously. Alice will consciously report only her final decision, see
considerations below.
Now we make a very general remark about the partial trace. For any operator M acting in
H, the following equality holds:
TrρM = TrR(M ⊗ I), where ρ = TrKR.
Therefore the average na(t) equals to the average of Alice's decision operator:
na(t) =< na >ρ(t) .
(2.4)
In section 3 we obtain the averages of the decision operator na(t) for its eigenstates φna (for
operators satisfying CAR, we have: na = 0, 1). Thus initially Alice has a definite belief about
possible answers to the question A.
In section 3.1 we consider the case of the "almost homogeneous reservoir" - a mental (in-
formation) environment R which is behaviorally almost homogeneous in the following sense
(see section 3.1 for the mathematical formalization). The basic parameter characterizing R is
deduced by < b†(k)b(q) >= nb(k)δ(q − k), the average for k-type agents. For an almost homo-
geneous R, nb(k) is constant. In general, however, it can vary with k 7. It must be stressed
that the reservoir considered in section 3.1 is not completely homogeneous. Its members are
not all identical. They have different internal dynamical scales represented by the function
ω(k) = ωbk which really depends on k, even if this dependence is relatively mild (ω(k) is linear
in k, so that ω(k1) ≃ ω(k2) if k1 ≃ k2), see again section 3.1.
For such environment R, the average na(t) of the decision operator na was found, see formula
(3.7) section 3.1 (with the corresponding generalization to the CAR-case in section 3.3). This
formula in combination with equalities (2.3), (2.4) gives the probabilistic dynamics:
Pt(A = 1) = na(t) = na e− 2λ
ω
b
2
π
t + nb(cid:18)1 − e
2
π
2λ
ω
b
t(cid:19) = nb + (na − nb) e− 2λ
ω
2
π
b
t,
(2.5)
where λ is the constant of interaction between Alice and R, see (3.1). As we will see in the
7In the simplest model all agents in R are asked the same question A as Alice and nb(k) is the average of
their answers to A. In a more complex environment R each cluster R(k) of agents labeled by the parameter k
is asked its own question A(k) and nb(k) is the average of their answers to A(k).
10
next section, this is the result of a reasonable choice of the Hamiltonian operator, where the
interactions between Alice and her reservoir are included8.
Consider, e.g., the case na = 1, e.g., initially Alice was for Brexit. Suppose that her
surrounding environment is almost homogeneous9 and nb = 1/3, i.e., its members express
sufficiently strong anti-Brexit attitude. Then Alice following the dynamics expressed by (2.5)
would lose her inclination to vote for Brexit and finally she may accept the surrounding attitude
to vote against Brexit. It is interesting to notice that the rapidity of the decision process is
directly related to the strength of the interaction between Alice and her environment, and
inversely connected to ωb. This is in agreement with the fact, see [4], that ωb describes a sort
of inertia of the system.
The "decision probability" is given by the expression:
P (A = 1) = lim
t→∞
Pt(A = 1) = nb.
(2.6)
We remark that dependence on the initial belief-state of Alice disappeared. It does not matter
whether she was in the belief-state φ0 (she was firmly determined to reply "no") or in the belief-
state φ1 (she was firmly determined to reply "yes"). Finally, she set the subjective probability
P (A = 1) = nb to reply "yes". Moreover, as we will discuss later, even if initially Alice was in
a superposition belief-state (i.e., she started with a belief-state of uncertainty), it would lead
to setting of the same probability in favor to reply "yes". This behavior of Alice is natural: she
interacts with a homogeneous reservoir, an ensemble of agents, where all agents have the same
probability to favor the "yes" answer; Alice simply follows them.
In spite of its simplicity, the belief state dynamics of a decision maker interacting with an
almost homogeneous information environment can have interesting applications. It describes
well not only modern mass-media campaigns on the political arena, but can also be used in
economics and finance.
8This approach, which is used everywhere in [4], is the standard way in which the dynamics of any system,
micro- or macroscopic, is deduced in classical and in quantum mechanics. What we are doing, in fact, is to
adapt the same basic idea to decision making processes.
9The latter assumption is not so unusual. Although the modern information (mental) environment represents
huge variety of information flows, an individual is typically coupled to the concrete flow, e.g., she has the custom
to see only particular TV-channels and follow only specially selected Internet resources. Even her physical
human environment is typically homogeneous. A professor of a university is surrounded by rather homogeneous
population with liberal views, including Brexit, Trump or Putin. One of the authors of this paper was visiting
USA just before the second Bush-vote. The university environment expressed generally the belief that Bush
would not be elected. For such R, the average nb could be estimated as approximately equal to one.
11
For example, in this framework we can model the process of persuasion of consumers. A
company wants to persuade consumers to buy some commodity C. So the question A is "to
buy or not to buy?" To approach the goal, the company creates an information environment
R (almost homogeneous) with the parameter nb ≈ 1. 10 Then our model shows that, for
consumers, it is possible to approach the same level of confidence to buy C as in the surrounding
information environment R. As was already remarked, the information structure of R need not
include directly the question A, i.e., the information campaign need not be straightforwardly
oriented to advertising of the commodity C. The R can be concerned with other questions,
probably quite different from A. Then the main task of the company driving the persuasion
campaign is to establish interaction of consumers with R which can be formally represented by
the interaction Hamiltonian considered in this paper.
Persuasion of consumers need not be reduced to the concrete commodity C. A group of
corporations can perform persuasion regarding a class of commodities, e.g., a new generation
of mobile phones, or wind energy, or electric cars. The strategy is the same: creation of
an information environment and coupling it with people's opinion in such a way that would
generate stabilization of subjective probabilities Pt(A = 1) (in a long run campaign) to the
desired value of nb.
3 The stabilization procedure
This section is devoted to justification of the model of decision making as decoherence, stabiliza-
tion of the belief-state ρ(t) of Alice. In the Heisenberg picture this problem can be transformed
into the problem of stabilization of reflection (creation-annihilation) operators a(t), a†(t) and
consequently the decision operator na(t) = a†(t)a(t) (the number operator). Of course, such
stabilization is possible only for special classes of information (mental) environments and in-
teractions between decision maker and surrounding environment. Mathematically the problem
is very complicated and we were able to solve it only partially. This section is of the purely
mathematical (and quantum physical) nature.
If S is a closed quantum system driven by a time-independent, self-adjoint, hamiltonian H,
it is natural to suspect that only periodic or quasi-periodic effects can take place, during the
time evolution of S. This is because the energy of S is preserved, and this seems to prevent
any damping effect in S. For instance, if we work in the Schrodinger representation (SR), the
10Typically this is done through a massive campaign in the mass-media. However, the information content
of R is not reduced to direct advertising of C, see considerations below.
12
time evolution Ψ(t) of the wave-function of the system is simply Ψ(t) = e−iHtΨ(0) and, since
the operator e−iHt is unitary, we do not expect Ψ(t) to decrease (in some suitable sense) to
zero when t diverges. Nevertheless, we will show that a similar decay feature is possible if S is
coupled to a reservoir R, but only if R is rather large as compared to S, in particular, if R has
an infinite number of degrees of freedom.
To see this in details, we start by considering a first system, S, interacting with a second
system, S, and we assume for the time being that both S and S are of the same size: to be
concrete, this means here that S describes a single particle and, analogously, S describes a
second particle. To model dynamics, we introduce the operators a, a† and na = a†a (for the
first particle) and the operators b,b† and nb = b†b (for the second particle).11 These operators
obey the following canonical commutation relations (CCR):
[a, a†] = [b, b†] = I,
while all the other commutators are assumed to be zero.
A natural choice for a Hamiltonian of the compound system S ∪ S is the following:
h = ωana + ωbnb + µ(cid:0)a†b + b†a(cid:1) ,
where ωa, ωb and µ must be real quantities for h to be self-adjoint. Recall that losing self-
adjointness of h would produce a non unitary time evolution, and this is out of the scheme
usually considered in ordinary quantum mechanics. The hamiltonian h contains a free part
plus an interaction part. The latter is such that, if the eigenvalue of na increases by one unit
(during time evolution), then the eigenvalue of nb decreases by one unit, and viceversa. This is
because [h, na + nb] = 0, so that na + nb is an integral of motion.
Let us now consider dynamics of operators in the Heisenberg representation. The equations
of motion for a(t) and b(t), can be easily deduced and turn out to be
a(t) = i[h, a(t)] = −iωaa(t) − iµb(t),
b(t) = i[h, b(t)] = −iωbb(t) − iµa(t),
a(0) = a,
b(0) = b.
The solution can be written as a(t) = aαa(t) + bαb(t) and b(t) = aβa(t) + bβb(t), where the
functions αj(t) and βj(t), j = a, b, are linear combinations of eλ±t, with λ± = −i
2 (ωa + ωb −
p(ωa − ωb)2 + 4µ2). Moreover αa(0) = βb(0) = 1 and αb(0) = βa(0) = 0, in order to have
11Cognitive meaning of these operators in the quantum-like model of decision making will be discussed in
section 2.
13
a(0) = a and b(0) = b. Hence we see that both a(t) and b(t), and,as a consequence, na(t) =
a†(t)a(t) and nb(t) = b†(t)b(t), are linear combinations of oscillating functions, so that no
damping is possible within this simple model.
Suppose now that the system S is replaced by an (infinitely extended) reservoir R, whose
particles are described by an infinite set of CCR-operators b(k), b†(k) and n(k) = b†(k)b(k),
k ∈ R. Each k labels one of the elements of the reservoir surrounding S. Hence, for k1 6= k2,
we are considering two different elements which may have different characteristics. This is why,
in (3.1), we are introducing two k-depending functions, ω(k) and f (k). The hamiltonian of
S ∪ R extends h above and is now taken to be
H = H0 + λHI, H0 = ωna +ZR
where [a, a†] = I, [b(k), b†(q)] = Iδ(k − q), while all the other commutators are zero. It could
be useful to notice that here, rather than the CAR used in section 2, we are assuming CCR.
We will see later, in section 3.3, that this does not affect our main conclusions. The reason
for the different choices is to show that the output of the model does not really depend on the
commutation rules adopted. All the constants appearing in (3.1), as well as the regularizing
function f (k), are real, so that H = H †. Notice that an integral of motion exists also for S ∪R,
na +RR n(k) dk, which extends the one for S ∪ S, na + nb. With this choice of H, the Heisenberg
ω(k)n(k) dk, HI =ZR(cid:0)ab†(k) + a†b(k)(cid:1) f (k)dk,
equations of motion are
(3.1)
( a(t) = i[H, a(t)] = −iωa(t) − iλRR f (k) b(k, t) dk,
b(k, t) = i[H, b(k, t)] = −iω(k)b(k, t) − iλf (k) a(t),
(3.2)
They are supplemented by the initial conditions a(0) = a and b(k, 0) = b(k). In particular, the
last equation can be rewritten in an integral form as
b(k, t) = b(k)e−iω(k)t − iλf (k)Z t
0
a(t1)e−iω(k)(t−t1) dt1.
(3.3)
3.1 An almost homogeneous reservoir
In this section we fix f (k) = 1 and ω(k) = ωbk, where ωb ∈ R+. This is a standard choice in
quantum optics, [55]. We now insert b(k, t) in (3.3) in the first equation in (3.2), change the order
of integration, and use the integral expression for the Dirac deltaRR e−iωbk(t−t1) dk = 2π
δ(t− t1),
as well as the equalityR t
2g(t) for any test function g(t). Then, we conclude
that
ωb
0 g(t1)δ(t− t1) dt1 = 1
a(t) = −(cid:18)iω +
πλ2
ωb (cid:19) a(t) − iλZR
b(k) e−iωbkt dk.
(3.4)
14
This equation can be solved, and the solution can be written as
a(t) =(cid:18)a − iλZR
dk η(k, t)b(k)(cid:19) e−(cid:16)iω+ πλ
b (cid:17)t,
ω
2
(3.5)
ρ(k)(cid:0)eρ(k)t − 1(cid:1) and ρ(k) = i(ω− ωbk) + πλ2
where η(k, t) = 1
. Using complex contour integration
it is possible to check, see appendix 2, that [a(t), a†(t)] = I for all t: this means that natural
decay of a(t), described in (3.5), is balanced by the reservoir contribution. This feature is
crucial since it is a measure of the fact that the time evolution is unitarily implemented in our
approach, even if a(t) apparently decays for increasing t.
ωb
Let us now consider a state over S ∪R, h XS ⊗ XRi = hϕna, XSϕnai h XRiR, in which ϕna is
an eigenstate of the number operator na and < >R is a state of the reservoir, which is assumed
to satisfy, among other properties (see [55, 4]),
(cid:10) b†(k)b(q)(cid:11)R = nb(k)δ(k − q).
(3.6)
This is a standard choice, see for instance [55], which extends the choice we made for S. Here
XS ⊗ XR is the tensor product of an operator of the system, XS, and an operator of the
reservoir, XR. Then, if for simplicity we take the function nb(k) to be constant in k we get,
calling na(t) :=< na(t) >=< a†(t)a(t) >,
na(t) = na e− 2λ
ωb
2
π
t + nb(cid:18)1 − e
2λ
2
π
ωb
t(cid:19) ,
(3.7)
which goes to nb as t → ∞. Hence, if 0 ≤ nb < na, the value of na(t) decreases with time.
If, on the other hand, nb > na, then the value of na(t) increases for large t. This is the
exponential rule which, as discussed before, cannot be deduced if R has not an infinite number
of degrees of freedom. Notice that, in particular, if the reservoir is originally empty, nb = 0,
then na(t) = na e− 2λ
t decreases exponentially to zero: the system becomes empty. On the
ω
π
2
b
other hand, since na +RR n(k) dk is a constant of motion, the reservoir starts to be filled up.
Remark 1. The continuous reservoir considered here (k ∈ R) could be replaced by a
discrete one, describing again an infinite number of particles, but labeled by a discrete index.
In this case, to obtain a Dirac delta distribution, which is the crucial ingredient in the derivation
above, we have to replace the integralRR e−ik(t−t1) dk = 2πδ(t− t1) with the Poisson summation
formula, which we write here as Pn∈Z einxc = 2π
In the above model we see that bath is essentially characterized by three functions: f (k),
ω(k) and nb(k). In particular, in what we have done so far, we have taken f (k) = 1, ω(k) = ωb k
c Pn∈Z δ(cid:0)x − n 2π
c (cid:1), for all non zero c ∈ R.
15
and nb(k) = nb. This choice can be interpreted as follows: the bath is almost homogeneous,
meaning that the functions f (k) and nb(k), which in principle could depend on k (remember
that each k labels an element of the bath), are constant in k: different members of the bath
all share the same values of f (k) and of nb(k). However, these members are not all completely
identical, since ω(k) depends on k, as we have already observed in section 2.
3.2 Changing reservoir
To keep the situation under control as much as possible, we now discuss what happens if we
keep f (k) and ω(k) fixed as before, but now allow for a different dependence of nb(k) on k. In
other words, we do not assume that nb(k) is constant in k, while we still take f (k) = 1 and
ω(k) = ωb k. With this choice, a(t) is given again as in (3.5), and we conclude, in the same
way, that [a(t), a†(t)] = I. The function na(t) is the one in (4.5) but, this time, we cannot use
(4.4) since nb(k) is no longer a constant function. The computation of RR nb(k)η(k, t)2dk is
quite similar to that of RR η(k, t)2dk, with the obvious difference due to the presence of nb(k)
inside the integral. What is essential for us is that the integral is real-valued for all real choices
of nb(k). To use again the complex integration techniques, it is useful to assume that nb(k) is
analytic in k, and that nb(k) does not diverge when k diverges. For concreteness, we take
(3.8)
nb(k) =
nb
k2 + α2 ,
for some positive nb and for some α > 0. Notice that we could safely take α < 0, too; our choice
is fixed only for demonstration. On the other hand, the condition nb > 0 is essential because
it follows from the origin of nb(k), which came from a positive operator. The computation of
RR nb(k)η(k, t)2dk is performed again with the techniques discussed in details in appendix 2,
with the only difference that we have two singularities of the integrating functions both in the
upper and in the lower complex semi-planes, so the result of the integral is the sum of two
residues.
The computation can be performed thoroughly, and an analytic form of na(t) could be given
for all t. However, what is more relevant for us is large time limit of na(t), which turns out to
be
na(∞) := lim
t,∞
na(t) =
.
(3.9)
Notice that this result makes sense only if α and λ are not zero. This suggests that the complex
pole in nb(k), and the interaction between Alice and her bath, are really essential to get some
nb(αω2
b + πλ2)
+(cid:16)α + πλ2
ω2
b (cid:17)2(cid:19)
αλ2ω2
b(cid:18) ω2
ω2
b
16
stabilization. We observe also that the asymptotic value of the decision function na(t) strongly
depends on the various parameters of the model, which could be adjusted to fit experimental
data, if needed.
Notice that the nb(k) in (3.8) describes a bath which is far from being uniform. Here, in
particular, nb(k) → 0 for k → ∞. Hence not all the parts of the bath interact with Alice in
the same way: serious differences arise.
Remark 2. In this section we have considered a bath which the amplitude nb(k) goes to
zero as k−2, see (3.8). In fact, we could think of a function nb(k) decreasing like k−1, but this
is not compatible with the reality of the function nb(k), at least if we still want to use complex
integration techniques to compute na(t). The reason is that, to avoid singularities in the real
axis, we are forced to choose nb(k) = µ
k−iα, for some complex µ and for some α > 0. This
ωb (cid:17)(cid:17), for some
produces na(∞) = γωb, at least if µ is chosen as follows: µ = γ(cid:16)ω − i(cid:16)αωb + πλ2
real γ. So we see that, with this peculiar choice of nb(k), we still get a real na(∞), but what
we cannot exclude that for finite time na(t) can be complex.
Remark 3. Both (3.7) and (3.9) suggest that the asymptotic limit of na(t) does not really
depend on whether, at t = 0, the system is in a pure state or in a combination of states. This
was already anticipated in section 2. We are using the Heisenberg representation. This means
that the state does not change, while the observables depend on time.
3.2.1 A remark on f (k) and ω(k)
So far we have fixed f (k) = 1 and ω(k) = ωb k. This is not the only natural choice, in particular
if we want to stress the difference between different parts of the reservoir. What is technically
useful for us, is that Dirac's delta function appears when deducing the differential equation for
a(t), see appendix 2. In fact, this is also possible under other conditions. Suppose that the two
functions f (k) and ω(k) are such that:
ω(k) → ±∞ for k → ±∞,
and
dω(k)
dk
=
1
β
f 2(k),
for some real β 6= 0. These assumptions are satisfied if f (k) = 1 and ω(k) = ωbk, but also in
many other situations. Hence
ZR
f 2(k)e−iω(k)(t−t1) dk = βZR
e−iω(k)(t−t1) dω(k) = 2πβδ(t − t1),
17
and we can significantly simplify the equation for a(t), which now becomes
a(t) = −(cid:0)iω + βπλ2(cid:1) a(t) − iλZR
f (k)b(k) e−iω(k)t dk.
(3.10)
This is the equation which replaces (3.4) in this new situation. The solution can be found as
before, and we get
a(t) =(cid:18)a − iλZR
dk η(k, t)b(k)f (k)(cid:19) e−(iω+βπλ2)t,
(3.11)
ρ(k)(cid:0)eρ(k)t − 1(cid:1) and ρ(k) = i(ω − ω(k)) + βπλ2. At a first sight, the situation
where η(k, t) = 1
is not particularly different from those discussed earlier and in appendix 2. However, a serious
technical problem now arises: to find [a(t), a†(t)] and na(t), we need to compute integrals and
the poles of the integrating functions are strongly dependent on the analytic expression of ω(k):
in particular, if ω(k) is not linear, the computations become rather hard!
3.3 What if we use CAR?
In the Hamiltonian (3.1) the operators a and b(k) are assumed to satisfy CCR. However, from
the point of view of a DM procedure, it might be more interesting to consider the case in
which Alice's decisions are described by CAR-operators. This means that, first of all {a, a†} =
a a† + a†a = I, with {a, a} = 0. It is then natural to consider a CAR-bath as well, i.e. to
assume that the operator b(k) satisfies the following rules:
{b(k), b†(q)} = δ(k − q) I,
{b(k), b(q)} = 0,
{a♯, b♯(k)} = 0,
x♯ being either x or x†. Assuming the same Hamiltonian (3.1), differential equations of motion
for the annihilation operators a(t) and b(k, t) can be deduced. They turn out to be the same
as in (3.2), except that λ must be replaced by −λ. In other words, we get
( a(t) = i[H, a(t)] = −iωa(t) + iλRR f (k) b(k, t) dk,
b(k, t) = i[H, b(k, t)] = −iω(k)b(k, t) + iλf (k) a(t).
(3.12)
As in the CCR-case, we assume that in (3.6) the function nb(k) is constant, nb(k) ≡ nb ≥ 0.
The CAR-setting implies that this constant cannot exceed one, 0 ≤ nb ≤ 1 (this is the average
of the CAR-number operator).
As we can see from (3.7), for the almost homogeneous reservoir, that na(t) depends on λ2
and not on λ. The same is true also for the choice (3.8) of the reservoir. This is evident from
18
(3.9) for large values of t, but can also be checked for any finite t. In other words, even with this
different choice of nb(k), we observe that na(t) depends on λ2 and not on λ. The conclusion,
therefore, is clear: the CCR-CAR nature of our model does not affect the main results we have
deduced. One choice or the other should be related to the nature of the decision we want to
model:
in case of a binary question (yes or not), the natural settings is probably the CAR-
one. But if we assume that several possible answers (normally, infinite!) are possible, then we
should use a CCR-version of the model. In this case, the existence of a quadratic integral of
motion can limit the number of possible answers to the original questions, giving rise to a sort
of finite-dimensional Hilbert space, similarly to what is discussed in details in [24].
4 Concluding discussion
In modeling of the process of decision making, the main output of this paper is the description
of properties of an information reservoir (environment) R leading to asymptotic stabilization.
In other words, this is a description of a context surrounding an agent, Alice, which guaranties
that she would make some decision, i.e., her mind would not fluctuate for ever between a variety
of alternative answers to a question (solutions of a problem) A.
The presented results on asymptotic stabilization can be used to support mathematically
the model of decision making as decoherence - in its quantum field version. Here the theory
of open quantum systems is applied straightforwardly in the Hamiltonian framework. Decision
making dynamics is given by the Heisenberg equations for operators describing reflections of
an agent, Alice, and the operators representing the information reservoir. The most interesting
for decision making are constraints imposed on information reservoirs, environments, guaran-
teeing asymptotic stabilization. Our present study is only a first step in this direction; further
studies of sufficient and necessary conditions of asymptotic stabilization are needed.12 We hope
that this paper will attract attention of experts in mathematical physics (especially working
on mathematical problems of quantum field theory) to the problem of asymptotic stabiliza-
tion. Novel applications to social science, microeconomics, and finance, generate new laws for
functions f (k), nb(k), ω(k) which have not been considered in physical applications.
12 We remark that applications of quantum field formalism to modeling of decision making generate new
dynamical models based on algebras of qubit creation-annihilation operators [57]. These operators satisfy the so
called qubit commutation relations which combine both CAR and CCR-features. To the best of our knowledge,
there are no results concerning asymptotic stabilization for such dynamical processes and this is an interesting
topic for research.
19
To extend the readership of this paper, the calculations in section 3 were done at the "phys-
ical level of rigor". A complete mathematical treatment of the problem should involve consider-
ation of operator-valued distributions (generalized functions). Such treatment definitely must
be performed at some stage of future development of quantum field inspired models of decision
making.
In general, the great success of quantum field theory of physics rises the expectations that
this formalism will fruitfully contribute to modeling of decision making and cognitive processes.
Appendix 1: Second quantization: CAR
We briefly review some basic facts on the so–called canonical anti-commutation relations (CAR),
which were originally introduced in connection with what in quantum physics is called sec-
ond quantization for identical particles with half-integer spin. We say that a set of operators
{aℓ, a†
ℓ, ℓ = 1, 2, . . . , L} satisfy the CAR if the conditions
{aℓ, a†
n} = δℓnI,
{aℓ, an} = {a†
ℓ, a†
n} = 0
(4.1)
hold true for all ℓ, n = 1, 2, . . . , L. Here, I is the identity operator and {x, y} := xy + yx is the
anticommutator of x and y. These operators are those which are used to describe L different
modes of fermions. From these operators we can construct nℓ = a†
ℓ=1 nℓ, which
are both self–adjoint. In particular, nℓ is the number operator for the ℓ–th mode, while N is the
number operator of S. Compared with bosonic operators, the operators introduced here satisfy
a very important feature: if we try to square them (or to rise to higher powers), we simply get
zero: for instance, from (4.1), we have a2
ℓ = 0. This is related to the fact that fermions satisfy
the Fermi exclusion principle.
ℓaℓ and N =PL
The Hilbert space of our system is constructed as follows: we introduce the vacuum of
the theory, that is a vector ϕ0 which is annihilated by all the operators aℓ: aℓϕ0 = 0 for all
ℓ = 1, 2, . . . , L. Then we act on ϕ0 with the operators a†
ℓ (but not higher powers, since these
powers are simply zero!):
ϕn1,n2,...,nL := (a†
1)n1(a†
2)n2 · · · (a†
L)nLϕ0,
(4.2)
nℓ = 0, 1 for all ℓ. These vectors give an orthonormal set and are eigenstates of both nℓ and N:
ℓ=1 nℓ. Moreover,
nℓϕn1,n2,...,nL = nℓϕn1,n2,...,nL and Nϕn1,n2,...,nL = Nϕn1,n2,...,nL, where N = PL
using the CAR, we deduce that
nℓ (aℓϕn1,n2,...,nL) = (nℓ − 1)(aℓϕn1,n2,...,nL)
20
and
nℓ(cid:16)a†
ℓϕn1,n2,...,nL(cid:17) = (nℓ + 1)(a†
l ϕn1,n2,...,nL),
for all ℓ. Then aℓ and a†
ℓ are called the annihilation and the creation operators. In fact, they
transform the vector ϕn1,n2,...,nℓ,...,nL into a different vector, proportional to ϕn1,n2,...,nℓ±1,...,nL,
annihilating or creating one particle in the ℓ-th mode. However, in some sense, a†
ℓ is also an
annihilation operator since, acting on a state with nℓ = 1, we destroy that state.
The Hilbert space H is obtained by taking the linear span of all these vectors. Of course,
H has a finite dimension. In particular, for just one mode of fermions, dim(H) = 2. This also
implies that, contrarily to what happens for bosons, the fermionic operators are bounded.
The vector ϕn1,n2,...,nL in (4.2) defines a vector (or number) state over the algebra A as
ωn1,n2,...,nL(X) = hϕn1,n2,...,nL, Xϕn1,n2,...,nLi,
(4.3)
where h , i is the scalar product in H.
Appendix 2: Explicit computations
In this appendix we give some details of our computations just sketched in Section 3.1, useful
for those who are not familiar with this kind of interacting open systems.
Our starting point is b(k, t) in (3.3), with f (k) = 1 and ω(k) = ωb k. Hence
b(k, t) = b(k)e−iωbkt − iλZ t
0
a(t1)e−iωbk(t−t1) dt1.
When we replace this expression in the equation for a(t), a(t) = −iωa(t) − iλRR b(k, t) dk, we
have to compute, in particular, the double integral
ZR(cid:18)Z t
0
a(t1)e−iωbk(t−t1) dt1(cid:19) dk =Z t
0
a(t1)(cid:18)ZR
a(t1)δ(t − t1)dt1 =
π
ωb
e−iωbk(t−t1) dk(cid:19) dt1 =
a(t),
=
2π
ωb Z t
0
and (3.4) now follows. An easy way to solve (3.4) consists in using the change of variable
a(t) = A(t) expn−(cid:16)iω + πλ2
ωb (cid:17) to since, in terms of A(t), (3.4) can be rewritten as
A(t) = −iλZR
b(k) e−iωbkt dk,
⇒
A(t) = A(0) − iλZ t
0 (cid:18)ZR
b(k) e−iωbkt1 dk(cid:19) dt1,
21
which returns, after few simple computation, the solution in (3.5).
To check now that the CCR are preserved during time evolution, i.e. that [a(t), a†(t)] = I
for all t, we observe that the operators at the right hand side of (3.5) are the initial (t = 0)
operators, so that they satisfy the commutation rules [a, a†] = I, [b(k), b†(q)] = δ(k − q), while
[a, b(k)] = [a†, b(k)] = ... = 0. Therefore
[a(t), a†(t)] = e− 2πλ
ωb
2
t(cid:18)1 + λ2ZR η(k, t)2dk(cid:19) I.
In our case, after some simple algebra, we deduce that
ZR η(k, t)2dk =ZR
ω2
ωb ± i πλ2
where k± = ω
we conclude that
ω2
b
dk
b (k − k+)(k − k−)(cid:20)(cid:18)e
2
2πλ
ω
b
2
t + 1(cid:19) − e(−iω+ πλ
ω
b
)teiωbkt − e(iω+ πλ
ω
b
2
)te−iωbkt(cid:21) ,
. Using complex integration we can compute each contribution here and
ZR η(k, t)2dk =
2
2πλ
ωb
1
λ2 (cid:18)e
t − 1(cid:19) ,
(4.4)
so that our claim easily follows.
To prove now (3.7), we first observe that the mean values of contributions like a†b(k) and
ab†(k) on our states are zero. Therefore, using (3.5), we get
na(t) = e− 2πλ
ω
b
2
t(cid:18)(cid:10)ϕna, a†aϕna(cid:11) h IRiR + λ2 hϕna, IϕnaiZRZR
2
η(k, t)η(q, t) (cid:10) b†(k)b(q)(cid:11)R dk dq(cid:19) =
Then, if nb(k) = nb, we can use the result in (4.4) and recover the result in (3.7).
= e− 2πλ
ω
b
t(cid:18)na + λ2ZR
nb(k)η(k, t)2dk(cid:19) .
(4.5)
Acknowledgments. This paper is based on the discussions between the coauthors during
the visit of two of them (IB and KHR) to the university of Palermo in April 2017. IB was
supported by Marie Curie Fellowship at City University of London, H2020-MSA-IF-2015, grant
N 696331; AKH was supported by the EU-project Quantum Information Access and Retrieval
Theory (QUARTZ), Grant No. 721321. FB thanks the Gruppo Nazionale di Fisica Matematica
of I.N.d.A.M., and by the University of Palermo, for support. F.B. also thanks the Distinguished
Visitor Program of the Faculty of Science of the University of Cape Town, 2017.
References
[1] A. Khrennikov, Information dynamics in cognitive, psychological, social, and anomalous
phenomena. Ser.: Fundamental Theories of Physics, Kluwer, Dordreht, 2004.
22
[2] A. Khrennikov, Ubiquitous quantum structure:
from psychology to finance. Springer,
Berlin-Heidelberg-New York, 2010.
[3] Busemeyer J R, Bruza P D. 2012 Quantum models of cognition and decision. Cambridge:
Cambridge University Press.
[4] Bagarello F. 2012 Quantum dynamics for classical systems: with applications of the number
operator. New York: J. Wiley.
[5] Haven E, Khrennikov A. 2013 Quantum social science. Cambridge: Cambridge University
Press.
[6] Asano M, Khrennikov A, Ohya M, Tanaka Y, Yamato I. 2015 Quantum adaptivity in
biology: from genetics to cognition. Heidelberg-Berlin-New York: Springer.
[7] Khrennikov, A., 2004. On quantum-like probabilistic structure of mental information. Open
Systems and Information Dynamics 11 (3), 267-275.
[8] Khrennikov, A., 2006. Quantum-like brain: Interference of minds. BioSystems 84, 225-241.
[9] Busemeyer J R, Wang Z, Townsend J T. 2006 Quantum dynamics of human decision
making. J. Math. Psychol. 50,220-241
[10] Pothos E M, Busemeyer J R. 2009 A quantum probability explanation for violation of
rational decision theory. P. Roy. Soc. Lond. B Bio. 276, 2171-2178
[11] Busemeyer J R , Pothos E M, Franco R, Trueblood J. 2011 A quantum theoretical expla-
nation for probability judgment errors. Psychol. Rev. 118, 193-218
[12] Dzhafarov, E.N. & Kujala, J.V. 2012 Selectivity in probabilistic causality: Where psychol-
ogy runs into quantum physics. Journal of Mathematical Psychology 56, 54-63.
[13] Asano M, Ohya M, Tanaka Y, Basieva I, Khrennikov A. 2012 Quantum-like dynamics of
decision-making. Physica A 391, 2083-2099
[14] Hawkins R J and Frieden B R. 2012 Asymmetric Information and Quantization in Financial
Economics. Int. J. Math. and Math. Sc. 32012, Article ID 470293.
[15] Pothos E M , Busemeyer J R.2013 Can quantum probability provide a new direction for
cognitive modeling? Behav. Brain Sci. 36, 255–327
23
[16] Asano M, Basieva I, Khrennikov A, Ohya M, Tanaka Y, Yamato I 2013. A model of
epigenetic evolution based on theory of open quantum systems. Systems and Synthetic
Biology 7, 161-173
[17] Dzhafarov E N and Kujala J V 2014 On selective influences, marginal selectivity, and
Bell/CHSH inequalities. Topics in Cognitive Science 6, 121-128
[18] Kvam P D, Busemeyer J R, Lambert-Mogiliansky A. 2014 An empirical test of type-
indeterminacy in the Prisoners Dilemma. Lecture Notes in Computer Science 8369,213-
224
[19] Khrennikov A, Basieva I. 2014 Possibility to agree on disagree from quantum information
and decision making. J. Math. Psychol. 62–63, 1–15
[20] Plotnitsky A. 2014 Are quantum-mechanical-like models possible, or necessary, outside
quantum physics? Phys. Scripta T163, 014011
[21] A. Khrennikov, Basieva, I. (2014). Quantum Model for Psychological Measurements : From
the Projection Postulate to Interference of Mental Observables Represented As Positive
Operator Valued Measures. NeuroQuantology. 12. 324-336
[22] A. Khrennikov, Basieva, I., Dzhafarov, E.N., Busemeyer, J.R. 2014. Quantum Models for
Psychological Measurements : An Unsolved Problem. PLoS ONE. 9. Article ID: e110909.
[23] J. R. Busemeyer, Z. Wang, A. Khrennikov, I. Basieva, I. 2014. Applying quantum principles
to psychology. Physica Scripta. T163. 014007.
[24] Bagarello F., Oliveri F., An operator-like description of love affairs, SIAM Jour. Appl.
Math., 70, No. 8, pp. 32353251, (2010)
[25] Bagarello F., Gargano F., Modeling interactions between political parties and electors, Phys.
A, dx.doi.org/10.1016/j.physa.2017.04.035
[26] Bagarello F. 2015 An operator view on alliances in politics. SIAM J. Appl.. Math. 75(2),
564-584
[27] Bagarello F, Haven E. 2016 First results on applying a non-linear effect formalism to
alliances between political parties and buy and sell dynamics. Physica A 444, 403-414
24
[28] Bagarello F. 2015 A quantum-like view to a generalized two players game. Int. J. Theor.
Phys. 54 (10), 3612-3627
[29] Denolf J and Lambert-Mogiliansky A. 2016. Bohr complementarity in memory retrieval.
J. Math. Psych. 73, 28-36
[30] E. Haven and A. Khrennikov, Statistical and subjective interpretations of probability in
quantum-like models of cognition and decision making. J. Math. Psych. 74, 8291 (2016).
[31] R. J. Hawkins and B. R. Frieden, 2017 Quantization in financial economics: An
information-theoretic approach. The Palgrave handbook of quantum models in social sci-
ence. Applications and Grand Challenges. Haven E and Khrennikov A (eds). Palgrave
Macmillan UK, pp. 19-38.
[32] Lambert-Mogiliansky A. 2017. Quantum-like type indeterminacy: A constructive approach
to preferences a la Kahneman and Tversky. Haven E and Khrennikov A (eds). Palgrave
Maxmilan UK, pp. 229-250.
[33] Danilov V I, Lambert-Mogiliansky A and Vassili Vergopoulos V. 2016 Dynamic consistency
of expected utility under non-classical(quantum) uncertainty. https://halshs.archives-
ouvertes.fr/halshs-01324046/
[34] A. Khrennikov, Classical and quantum mechanics on information spaces with applications
to cognitive, psychological, social and anomalous phenomena. Found. of Physics, 29, N.
7, 1065-1098 (1999).
[35] M. Asano, I. Basieva, A. Khrennikov, M. Ohya, Y. Tanaka, I. Yamato, Quantum infor-
mation biology: from information interpretation of quantum mechanics to applications in
molecular biology and cognitive psychology. Found. Phys. 45, N 10, 1362-1378 (2015).
[36] D' Ariano, G. M.: Physics as information processing. In: Jaeger, G. et al. (eds.), Advances
in Quantum Theory, AIP, Melville, NY, 1327, pp. 7–16 (2011)
[37] Plotnitsky, A.: Niels Bohr and Complementarity: An Introduction. Springer, Berlin and
New York (2012)
[38] Penrose, R.: The Emperor's new mind. Oxford Univ. Press, New-York (1989)
[39] Hameroff, S.: Quantum coherence in microtubules. A neural basis for emergent conscious-
ness? J. of Cons. Stud. 1, 91-118 (1994)
25
[40] Tegmark, M. (2000). Importance of quantum decoherence in brain processes. Phys. Rev.
E 61 (4): 41944206
[41] de Barros A J , Suppes P. 2009 Quantum mechanics, interference, and the brain. J. Math.
Psychol. 53, 306-313.
[42] A. Khrennikov, Quantum-like model of processing of information in the brain based on
classical electromagnetic field. Biosystems 105(3), 250-262 (2011).
[43] de Barros, A. J. Quantum-like model of behavioral response computation using neural
oscillators. Biosystems 110, 171-182 (2012)
[44] Takahashi T. Toward a physical theory of quantum cognition. Topics in Cognitive Science
6, 104107 (2014).
[45] Melkikh, A.V., 2013. Biological complexity, quantum coherent states and the problem of
efficient transmission of information inside a cell. BioSystems. 111, 190-198.
[46] Melkikh, A.V., 2014. Congenital programs of the behavior and nontrivial quantum effects
in the neurons work. BioSystems. 119, 10-19.
[47] Busemeyer, J. R., Fakhari, P., Kvam, P. 2017. Neural implementation of operations used
in quantum cognition. Progress in Biophysics and Molecular Biology, in press.
[48] von Neumann J., Mathematical Foundations of Quantum Mechanics, Princeton Univ.
Press, Princenton, 1955.
[49] Basieva I and Khrennikov A. 2017 Decision-making and cognition modeling from the theory
of mental instruments. The Palgrave Handbook of Quantum Models in Social Science.
Applications and Grand Challenges. Haven E and Khrennikov A (eds). Palgrave Maxmilan
UK, pp. 75-93.
[50] Khrennikov A. 2017 Why quantum? The Palgrave Handbook of Quantum Models in Social
Science. Applications and Grand Challenges. Haven E and Khrennikov A (eds). Palgrave
Maxmilan UK, pp. 321-334.
[51] Asano M, Ohya M, Tanaka Y, Basieva I, Khrennikov A. 2011 Quantum-like model of
brain's functioning: decision making from decoherence. J. Theor. Biol. 281, 56-64
26
[52] M. Asano, M. Ohya, A. Khrennikov, Quantum-Like Model for Decision Making Process
in Two Players Game - A Non-Kolmogorovian Model. Found. Phys., 41, 538-548 (2011).
[53] Skomski R, Kashyap A and Sellmyer D J 2012 A quantum-mechanical relaxation model.
J. Appl. Phys.111, 07D507
[54] Bagarello F, Haven E., Khrennikov A., A model of adaptive decision making from repre-
sentation of information environment by quantum fields. Phil. Trans. A, to be published
in 2017.
[55] S.M. Barnett, P.M. Radmore, Methods in theoretical quantum optics, Clarendon Press,
Oxford, 1997
[56] A. Khrennikov, "Social laser:" action amplification by stimulated emission of social energy.
Phil. Trans. Royal Soc. 374, N 2054, 20150094 (2016).
[57] P. Khrennikova, Modeling behavior of decision makers with the aid of algebra of qubit
creation-annihilation operators J. Math. Psych. published on line (2016)
27
|
0902.2020 | 3 | 0902 | 2011-11-01T01:49:42 | Feature selection in simple neurons: how coding depends on spiking dynamics | [
"q-bio.NC",
"q-bio.QM"
] | The relationship between a neuron's complex inputs and its spiking output defines the neuron's coding strategy. This is frequently and effectively modeled phenomenologically by one or more linear filters that extract the components of the stimulus that are relevant for triggering spikes, and a nonlinear function that relates stimulus to firing probability. In many sensory systems, these two components of the coding strategy are found to adapt to changes in the statistics of the inputs, in such a way as to improve information transmission. Here, we show for two simple neuron models how feature selectivity as captured by the spike-triggered average depends both on the parameters of the model and on the statistical characteristics of the input. | q-bio.NC | q-bio |
Feature selection in simple neurons: how coding
depends on spiking dynamics
Michael Famulare
University of Washington
Department of Physics
Box 351560
Seattle, WA 98195-1560
[email protected]
Adrienne Fairhall
University of Washington
Department of Physiology and Biophysics
HSB G424, Box 357290
Seattle, WA 98195-7290
October 29, 2018
Abstract
The relationship between a neuron's complex inputs and its spiking
output defines the neuron's coding strategy. This is frequently and ef-
fectively modeled phenomenologically by one or more linear filters that
extract the components of the stimulus that are relevant for triggering
spikes, and a nonlinear function that relates stimulus to firing proba-
bility. In many sensory systems, these two components of the coding
strategy are found to adapt to changes in the statistics of the inputs, in
such a way as to improve information transmission. Here, we show for
two simple neuron models how feature selectivity as captured by the
spike-triggered average depends both on the parameters of the model
and on the statistical characteristics of the input.
Neuronal dynamics are characterized by nonlinearities that lead to large,
approximately stereotyped voltage excursions, or spikes, that are the basis
for interneuronal signaling. Capturing the relationship between inputs and
the resulting pattern of spike outputs from a given neuron in the form of
a reduced functional model is a focus of sensory neuroscience. In the sense
1
that such a model provides a general mapping from input to output, it can
be thought of as the neuron's "coding strategy".
Reverse correlation methods (Bryant and Segundo, 1976; de Boer and Kuyper,
1968; Sakai, 1992; Hunter and Korenberg, 1986) provide a means to sam-
ple the statistical characteristics of stimuli that tend to trigger spikes; in
the simplest case, the mean, or spike-triggered average stimulus (STA), is
the optimal linear kernel for predicting the firing rate from the stimulus
(Rieke et al., 1996). Using reverse correlation, one may obtain an approx-
imate functional model for the neuronal input/output transformation in
terms of the input features that drive the system (Meister and Berry II,
1999; de Ruyter van Steveninck and Bialek, 1988; Simoncelli et al., 2004).
These methods may be applied not only to determine how neural systems
are driven by external stimuli, but to extract a model for how specific pat-
terns of synaptic current inputs drive single neurons. This allows one to de-
termine the role that a single neuron with a characteristic complement of ion
channels plays in a circuit: the integration of inputs over a certain timescale
(Slee et al., 2005; Svirskis et al., 2003; Prescott et al., 2006), the detection
of sudden change or highly synchronous events (Abeles, 1982; Slee et al.,
2005; Svirskis et al., 2004), or the selection of certain frequency components
in the input (Izhikevich, 2001; Prescott et al., 2006).
Here, we will derive explicit expressions for the outcome of such a statis-
tical analysis applied to two simple neuron models. We have two goals. The
first is to develop a general framework for understanding how the details
of neuronal dynamics establish or influence the features in the input that
trigger spikes. Second, neuronal systems show adaptation to statistics, in
the sense that the neuron's coding strategy often changes when driven by
stimuli with different statistical properties. In the case of single neurons,
such effects can modulate or gate the effective computation of the neuron
according to the statistical properties of the signal or the background inputs
(Hasenstaub et al., 2007; Destexhe and Pare, 1999; Fellous et al., 2003). To
identify the rules governing this process, one would like to know to what ex-
tent the observed changes may result from time-independent neuronal non-
linearities and to what extent they must be due to changes in underlying
neuronal parameters. To study this, we will compute how the experimentally
obtained features of two fixed models depend on the statistical properties of
the stimulus, focusing on the variance of a white noise input.
The key points of this paper are:
• The relevant linear filter corresponding to a nonlinear spiking neuron
model is determined by the nonlinear dynamics linearized in a manner
2
consistent with the typical operating regime of the system, which is
determined both by its dynamics and by the stimulus conditions. To
characterize this regime, we compute the voltage probability distribu-
tions from the Fokker-Planck interpretation of the models.
• We then use a novel application of the technique of stochastic lin-
earization to map the nonlinear models onto a set of linear models.
By studying both the mapping, determined by an optimization func-
tion relating the linear and nonlinear models, and the related STA
predictions for the equivalent linear system, we can delineate the roles
of different nonlinearities on spike encoding.
• The form of the STA is influenced both by the subthreshold (non)linear
dynamics and the spike afterhyperpolarization.
• Models with similar phase space topology can have STAs whose form is
controlled by different mechanisms. A rapid-onset exponential integrate-
and-fire model (EIF) has no significant subthreshold nonlinearity, and
so its STA is almost completely determined by the probability current
due to spiking.
In contrast, the quadratic integrate-and-fire model
(QIF), while superficially similar to the EIF, has an STA whose form
is dominated by the sampling of the subthreshold nonlinearity, with
spiking effects playing a secondary role.
1 Models and numerical methods
Change in the effective feature selectivity with driving variance has been
studied for the case of the leaky integrate-and-fire (LIF) model (Paninski et al.,
2003; Paninski, 2006b; Yu and Lee, 2003). In the LIF model, the dynamics
are linear until the voltage reaches an imposed threshold after which the
voltage is immediately reset below threshold. Thus, the LIF contains no
intrinsic excitability, and further, does not allow for the possibility that the
system can cross threshold multiple times before spiking due to noisy inputs.
This discontinous behavior with respect to spike initiation is not found in
biological neurons.
Two simple models with more realistic spike initiation are the quadratic
and exponential integrate-and-fire models (QIF and EIF, respectively) (Ermentrout and Kopell,
1986; Fourcaud-Trocme et al., 2003). Both models are similar in spirit to
the LIF insofar as they replace the afterhyperpolarization mechanism with
a discontinous jump, or after-spike reset, but the point of reset in these
3
models occurs at the peak of the spike instead of at the threshold voltage.
This mitigates the effects of the pathological behavior in response to noise
that the discontinuity creates (Paninski, 2006b) by moving it away from
the interesting region of spike initiation. The models are described by an
equation of the form:
τm v = −v + f (v) + (Vr − Vs) δ(v − Vs) + s(t),
(1)
where v denotes the membrane voltage, τm is the membrane time constant,
Vs is the voltage that defines the spike height and Vr is the post-spike reset
voltage. The input current, s(t), is a zero-mean gaussian white noise (GWN)
process with correlation function hs(t)s(0)i = σ2τmδ(t). The delta-function
is shorthand for the act of resetting the voltage to Vr after it reaches Vs. All
of the spike-generating and nonlinear subthreshold dynamics are encoded in
f (v). For the two models studied here, we have:
f (v) =(αv2
g exph v−V∗
g i
for the QIF,
for the EIF.
(2)
We study parameters such that the resting potential is zero and the un-
stable fixed point is at α−1 for both models. This requires us to choose
V∗ = α−1 (1 + gα ln(gα)) and g ≪ V∗. Somewhat paradoxically, despite the
higher order nonlinearity, the choice of small g causes the exponential non-
linearity to turn on over a much tighter range in voltage than the quadratic
nonlinearity of the QIF. We will see that this leads to more linear behavior
of the EIF model below threshold. Thus the two models behave noticeably
differently below threshold while still having the same after-spike dynamics.
1.1 Reverse correlation analysis
Reverse correlation is used to determine characteristics of the stimulus that
are correlated with neuronal response. From a long, random stimulus pre-
sentation s(t) and the resulting spike response times ti, one collects the set
of N current traces that led to a spike, s(τ − ti), over an interval of time
τ = [0, −T ] prior to the spike where T is chosen appropriately to capture
all the stimulus history that is relevant to triggering the spike. The spike-
triggered average or STA, ¯s(τ ), is found by averaging these samples over
i:
N
¯s(τ ) =
1
N
s(τ − ti).
(3)
Xi=1
4
1.2 Defining spike times
The results of reverse correlation analysis can depend on how the spike
time is defined. Here, we will look at the STAs with a temporal resolution
that is short compared to the average spike width. Different choices of the
voltage threshold used to define spikes will accordingly lead to STAs that
differ from each other due to temporal jittering of the ensemble of spike-
triggered trajectories. Because we want to understand how the spiking of
the model determines the feature selected from the stimulus ensemble, we
are interested in choosing a threshold that yields an STA that best captures
the role of the stimulus on the approach to the spike but is not sensitive
to stimulus-driven variations in the spike itself. For both models considered
here, this is achieved by selecting the unstable fixed point that separates the
subthreshold region from the spiking region in the absence of noise. Since the
location of the unstable fixed point is a function of the mean input current
and the quadratic form of the nonlinearity, we will call it the dynamical
threshold in accordance with previous work (Izhikevich, 2000; Hong et al.,
2007). For the zero-mean inputs considered here, the dynamical threshold
is Vth = α−1. Thus, we define spike times as the time of the last upward
crossing of the dynamical threshold preceding an after-spike reset.
1.3 Model simulation
In discrete-time with time step h, the nonlinear models in equation 1 were
realized as:
vn+1 = vn +
h
τm
(−vn + f (vn)) + σr h
τm
if vn+1 ≥ Vs, then vn+1 → Vreset .
ξn ,
(4)
(5)
where the ξn are drawn from a gaussian distribution with zero mean and
unit variance. For all figures in this paper, simulations were run with a time
step of h = τm/200 until 2 × 105 spikes were accumulated. The noise was
generated with randn in Matlab R2007b. Parameters used in simulation:
α = 1, τm = 1, Vs = 25, Vr = −0.2, g = 1
10 , and V∗ = 0.77.
2 Numerical results
We computed the STA numerically for a range of values of the stimulus
standard deviation σ for both models. Results are shown in figure 1. The
STA at all values of σ has two components: an extended feature and a sharp
5
upward step at the time of the spike. For the feature, two different types of
behavior appear. For large σ, the STAs are approximately decaying expo-
nentials for which, as the standard deviation increases, the decay timescale
decreases and the amplitude increases. Thus, at larger σ, the QIF and EIF
models perform approximately linear leaky integration, where the effective
leakiness depends on the standard deviation. For very small σ, the STAs
are non-monotonic, with the peak amplitude occurring well before the spike
time.
3 Approximate STA for finite standard deviations
To best understand how details of the models influence the STAs, we would
like to be able to calculate the STAs analytically.
In the zero standard
deviation limit, the STA can be analytically calculated for the QIF via
a large deviations principle and path integral methods (Paninski, 2006a;
Badel et al., 2008a; Wilson and Steyn-Ross, 2008), but that type of analysis
does not extend to finite σ. However, path integral methods can be applied
for arbitrary σ to perfectly linear models with no reset. As noted previously
(Hong et al., 2007), the observation that the STA is an exponential implies
that the subthreshold dynamics of the model are effectively linear. Since the
STAs of the nonlinear models are roughly exponential for larger standard
deviations, we should be able to introduce a linear approximation to the
nonlinear models that captures the qualitative behavior of the STA and
helps explain in detail how the STA arises from the form of the nonlinearity.
The main idea is as follows. While it is impossible to derive complete,
time-dependent statistical distributions for these models, we can get the
steady-state distribution from the Fokker-Planck equation. This distribution
gives us information about how the properties of the stimulus and spiking
dynamics determine how the system samples its subthreshold nonlinearities.
We will then use the steady-state distribution to map the nonlinear models
onto linear models and thus compute an approximation to the STA. Since
the linear model follows from the steady-state distribution, we can think of
the linear model as describing the "time-averaged dynamics" of the nonlinear
models.
3.1 The steady-state distribution
A key ingredient for understanding the behavior of the spiking models and
for determining an analytically tractable mapping of a nonlinear model to
a linear model is the steady-state probability distribution for the voltage
6
in response to an input with given statistical characteristics. This proba-
bility distribution, pN (v), can be computed from the Fokker-Planck equa-
tion (Paninski et al., 2003; Brunel and Latham, 2003; Lindner et al., 2003;
Fourcaud-Trocme et al., 2003), which for models of the form given in equa-
tion 1 is:
∂pN (v, t)
∂t
=
∂
∂v (cid:20)(cid:18) v − f (v)
τm (cid:19) pN (v, t)(cid:21) +
σ2
2τm
∂2pN (v, t)
∂v2
+R(t) [δ(v − Vr) − δ(v − Vs)] ,
(6)
where R(t) is the time-dependent mean firing rate that needs to be deter-
mined self-consistently in solving the equation. This is a continuity equation
for pN (v, t) which expresses that the evolution of the distribution is driven
by the deterministic nonlinear driving force, diffusion, and spiking. We are
interested in the steady-state distribution, for which ∂pN
∂t = 0 and R(t) goes
to the mean rate R. Using standard methods (Risken, 1996), one can show
that the steady state distribution is:
pN (v) =
2Rτm
σ2 e
−1
σ2 (v2
−2F (v))Z Vs
max(v,Vr)
dv′ e
1
σ2 (v′2
−2F (v′)),
(7)
stant.
where F (v) = R f (v)dv, and the mean firing rate is the normalization con-
This distribution is the product of a Boltzmann factor, controlled entirely
by the nonlinear dynamics, F (v), preceding a spike, and a spiking flux term
which carries the dependence on the spike parameters Vr and Vs. Since
we are mainly interested in behavior below the unstable fixed point, or
dynamical threshold, and the models considered here have reset voltages,
Vr, near the resting potential, the contribution of the spiking flux term does
not depend strongly on the form of F (v), but does depend strongly on the
location of Vr.
3.2 Stochastic linearization
We turn to a set of techniques known as stochastic linearization (SL) (see
(Socha, 2005) for an extensive review) to model and understand the behavior
of the STA of the nonlinear models.
In the SL approach, one seeks the
parameters of a linear model that optimally capture the properties of the
nonlinear model in a regime of interest. In our case, we are interested in
the linear model that best captures the approach of a nonlinear model to
threshold for a given input standard deviation, but we are unconcerned with
7
the dynamics of the spike itself. Thus, we search for an equivalent linear
model of the form:
τm v = −kσv + cσ + s(t),
(8)
where we make no attempt to model the spike or the reset (Gerstner and Kistler,
2002). This linear model is simply an Ornstein-Uhlenbeck process (Risken,
1996), and it has the associated steady-state probability distribution:
pL(v) =r kσ
πσ2 exp"−
kσ
σ2 (cid:18)v −
cσ
kσ(cid:19)2# .
(9)
To determine the parameters of the optimal linear model, we must se-
lect an optimization function that maps the nonlinear model onto the linear
model. There are no unique methods for choosing optimization functions
that will yield good results (Socha, 2005), and different functions will gen-
erally yield different results. We focus on two optimization functions which
give weight to different properties of the nonlinear model.
3.3 Minimizing the Kullback-Leibler divergence
The Kullback-Leibler divergence (DKL) measures the similarity of two prob-
ability distributions (Cover and Thomas, 2006). To map the nonlinear mod-
els to sets of linear models, we can use the DKL to minimize the difference
between the subthreshold part of the nonlinear steady-state distribution
and the matched linear model's steady-state distribution. The DKL for this
problem is:
DKL (pN pL) = Z Vth
where ZN = Z Vth
−∞
−∞
dv
pN (v)
ZN
ln
pN (v)
ZN pL(v)
,
(10)
pN (v).
To find the optimal linear model with this criterion, we minimize the DKL
with respect to kσ and cσ. Doing so yields
kσ =
cσ = kσE [v] ,
,
(11)
where E [. . .] = Z −1
pN (v) [. . .] .
σ2
2(cid:16)E [v2] − E [v]2(cid:17)
N Z Vth
−∞
8
In the σ → 0 limit, k0 = 1 and c0 = 0, corresponding to the classical
linearization around the fixed point of a nonlinear model. These expressions
show that minimizing the DKL amounts to simply estimating the mean and
variance below threshold. This criterion is only sensitive to the probability
distribution itself and has no knowledge of the underlying dynamics.
3.4 Minimizing the energy below threshold
An alternative optimization criterion is to optimize the mean square error
in the energy, or first integral of the nonlinear models, below threshold. The
energy of the nonlinear model is
E =
v2
2
− F (v),
and so the optimization criterion for kσ and cσ is
I = E"(cid:18) v2
2
(1 − kσ) + cv − F (v)(cid:19)2# ,
(12)
where F (v) and E [. . .] are defined as before. This criterion amounts to
trying to match the Boltzmann part of the distributions, taking spiking into
account only through the bias it provides to the expectation value. This
piece is primarily sensitive to the specifics of the dynamics below threshold
and is less sensitive to the overall shape of the distribution than the DKL
is. Minimizing I yields:
kσ = 1 − 2
E(cid:2)v2F (v)(cid:3) E(cid:2)v2(cid:3) − E [vF (v)] E(cid:2)v3(cid:3)
E [v4] E [v2] − E [v3]2
cσ =
E [vF (v)] E(cid:2)v4(cid:3) − E(cid:2)v2F (v)(cid:3) E(cid:2)v3(cid:3)
E [v4] E [v2] − E [v3]2
(13)
Again, in the σ → 0 limit, k0 = 1 and c0 = 0. In this case, we see that
the optimal parameters are directly sensitive to the form of the nonlinearity
below threshold and that the shape of the probability distribution only enters
through the expectation values.
3.5 The meanings of the optimization criteria
The two optimization criteria give different weights to different roles of the
nonlinearity. The DKL criterion is sensitive to the net statistical distribu-
tion below threshold, regardless of whether it comes about due to the spike
9
or the subthreshold nonlinearity, whereas the energy criterion is primarily
sensitive to the form of the subthreshold nonlinearity. Accordingly, we can
expect that the linear model, for a given nonlinear model and input stan-
dard deviation, found by the different criteria will be different. Specifically,
for nonlinear models whose optimal linear equivalents are best described by
the energy criterion, the parameters kσ and cσ will be closely related to the
form of the subthreshold nonlinearity but may not be very sensitive to the
overall details of the voltage distribution below threshold. In contrast, for
models with linear equivalents that are best described by the DKL criterion,
the parameters may have essentially no relation with the subthreshold non-
linearity, but rather describe global statistical properties set by the mean
and variance of the voltage distribution.
3.6 STA of the linear model
To find the STA for the linear model and compare it to numerical simula-
tions, we move to discrete time by defining t = nh, where n is an integer and
h is the time step. For clarity of notation, we identify v(t) = v(nh) ≡ vn.
The linear model in equation 8 is equivalently described by the forward
transition probability distribution:
p (vn+1vn) =r τm
2πσ2h
exp"−
τm
2σ2h(cid:18)vn+1 −(cid:18)1 −
hkσ
τm (cid:19) vn −
hcσ
τm (cid:19)2# .
(14)
Also of use are the steady-state probability distribution, pL(v), given in
equation 9, and the backward transition probability distribution, p(vnvn+1),
which can be derived with Bayes' rule:
p (vnvn+1) =
p(vn+1vn)pL(vn)
pL(vn+1)
,
=r τm
2πσ2h
exp"−
τm
2σ2h(cid:18)vn −(cid:18)1 −
hkσ
τm (cid:19) vn+1 −
hcσ
τm (cid:19)2# ,
(15)
for small h
. Notice that the linear model is statistically reversible (Weiss,
τm
1975): the backward transition distribution is the time reversal of the for-
ward, vn ⇄ vn+1. Since the model is linear, the STA follows from the spike-
triggered voltage, ¯v, via:
¯s(t) = τm ¯v(t) + kσ ¯v(t) − cσ,
¯sn = (¯vn − ¯vn−1)
10
τm
h
+ kσ ¯vn−1 − cσ.
(16)
(17)
The spike-triggered voltages of the linear model can be found exactly with
the following recipe. We start at the spike time, t = 0 (n = 0) -- the first
time for which v > Vth and v > 0.
The mean voltage at the spike time, ¯v0, is given by:
¯v0 =Z ∞
−∞
dv0 v0 p(v0spike)
(18)
The spike-triggered voltage distribution, p(v0spike), follows from the threshold-
crossing condition. The probability of finding a voltage v0 at the spike time
is given by the probability that v0 is above Vth, multiplied by the probability
that v0 was arrived at from voltages v−1 that were below threshold, summed
over all possible subthreshold values of v−1:
p(v0spike) = Z −1
0 H(v0 − Vth)Z Vth
−∞
dv−1p(v0v−1)p(v−1),
(19)
where p(v−1) is the unconditioned distribution of voltages prior to the spike
and is given by the steady state distribution in equation 9, p(v0v−1) is
the forward transition distribution, H(v0 − Vth) is the Heaviside function
representing the probability for v0 to be above threshold, and Z0 is the
normalization constant.
The mean voltage at the time immediately preceding the spike, ¯v−1, is
determined by averaging over all voltages below threshold that can transition
to voltages above threshold in the next time step, and can be found from
¯v−1 = Z ∞
−∞
dv−1 v−1 p(v−1spike),
(20)
p(v−1spike) = Z −1H(Vth − v−1)Z ∞
Vth
dv0 p(v−1v0)p(v0spike), (21)
where Z is the normalization constant for this distribution. Similarly, the
remaining ¯vn for n ≤ −2 are given by:
¯vn =Z ∞
−∞
dvn . . . dv−1 vn p(vnvn+1) . . . p(v−1spike).
(22)
Equation 22 is exact for a linear model but is impractical to use. With-
out noticeable loss of accuracy for the simulations considered in this paper,
numerous simplifications can be made. For a discussion of approximations
to ¯v0 and ¯v−1, see the appendix.
11
Via the central limit theorem, the ¯vn for n ≤ −2 can be simplified as:
¯vn ≈Z ∞
−∞
dvn vn p(vn¯vn+1).
(23)
Since these are all gaussian integrals over an infinite domain, the mean
value is the most likely value and so the remaining averages are arrived at
recursively to give:
¯vn =(cid:18)1 −
hkσ
τ (cid:19) ¯vn+1 +
hcσ
τ
for n ≤ −2.
(24)
Using the fact that h
of the exponential function and gives:
τm ≪ 1, this recursion relation can be solved in terms
¯vn =
cσ
kσ
+(cid:18)¯v−1 −
cσ
kσ(cid:19) exp(cid:20) kσ(n + 1)h
τm
for n ≤ −1.
(25)
(cid:21)
The STA immediately follows from equation 16, and is:
¯sn =(2kσ(cid:16)¯v−1 − cσ
kσ(cid:17) exph kσ(n+1)h
kσ ¯v0 + (¯v0 − ¯v−1) τm
h
τm
i n ≤ −1,
n = 0.
(26)
For the linear model, the STA is simply given by the exponential filter up to
a singular piece at the spike time that arises from requiring that threshold
be crossed from below.
3.7 Comparison to numerics
For the QIF, the energy criterion qualitatively captures the time constant
at all σ and correctly matches the amplitude of the STA at high σ (see
figures 2A and 3C). Conversely, the DKL criterion is much less accurate.
While it too predicts qualitatively correct time constants, the amplitude
of the predicted STAs is much too large (not shown). This is because the
DKL criterion strongly weights the effects of the after-spike reset and total
probability mass, and thus biases the resting potential of the linear model
too far below threshold. Thus, the STA for the QIF is primarily determined
by the form of the subthreshold nonlinearity and is less sensitive to the
escape from the subthreshold domain due to spiking. Figure 3A shows how
the optimal linear model relates to the subthreshold nonlinearity in this
case.
For the EIF, however, the energy criterion gives kσ ≈ 1 for all σ. While
this is not surprising given the effectively linear subthreshold dynamics, it
12
does not agree with the numerics. The DKL criterion, on the other hand,
applied to the EIF leads to qualitative agreement between the linear mod-
els and numerics (see figures 2B and 3D). This confirms the idea that the
changes in the STA in the EIF can only be due to the reduction in the time
spent below threshold because of spiking, and that the small amount of non-
linearity below threshold for the parameters used is not relevant except at
small σ (see figure 3B for further discussion). These changes in the STA are
analogous to those studied previously by Paninski (Paninski et al., 2005).
The singular upward step at the spike time arises from the condition that
threshold must be crossed from below -- that v must be positive -- to elicit a
spike (Aguera y Arcas et al., 2003; Hong et al., 2007). This "delta-function"
component, shown in figures 1C and 1D, appears here so prominently be-
cause we have chosen the spike-defining threshold to be at a voltage for
which the stimulus is still relevant to spiking. This step does not vanish
in the continuous-time limit. The value of the step can be calculated ap-
proximately (see appendix) with good agreement with simulation data (see
figure 4). This mode occasionally appears in experimental STAs when the
spike waveform is slow (R. Mease, personal communication). It is usually
not seen because the threshold is generally drawn well into the intrinsically
excitable domain of the voltage and so a condition on v does not significantly
constrain the stimulus in that situation.
4 Discussion
Due to nonlinearity, LN characterizations of neural systems show depen-
dence on stimulus statistics, even without changes in the underlying dy-
namical parameters (Theunissen et al., 2000; Yu and Lee, 2003; Borst et al.,
2005; Gaudry and Reinagel, 2007; Gill et al., 2008; Westwick and Kearney,
2003).
In particular, by changing only the input standard deviation, the
effective computation changes its functional form and timescale. We have
explored the consequences of this for two reduced naturally-spiking neu-
ron models, the quadratic and exponential integrate-and-fire models.
In
determining the linear filter or filters characterizing the model, our work
differs significantly from previous approaches (Gerstner and Kistler, 2002;
Hong et al., 2007; Wilson and Steyn-Ross, 2008; Badel et al., 2008a) in that
the point of linearization is not taken, as is classically done, to be the
equilibrium point; rather, we allow the subthreshold voltage distribution
to determine the optimal point of linearization. This distribution carries
information about the form of the nonlinearities, the mean firing rate, and
13
the stimulus itself. These properties account for changes in the effective
linear model with stimulus variance. We find that despite these models' su-
perficial similarities, different mechanisms are primarily responsible for this
form of adaptation. The key difference between the models is that the QIF
is nonlinear below threshold, qualitatively corresponding to a neuron with
hyperpolarizing currents that are activated below threshold, while the EIF
is mostly linear below threshold. Both models have been successfully fit to
neuronal data from a variety of neuron types (Izhikevich, 2004; Rauch et al.,
2003; Badel et al., 2008b).
Thus, both the neuron's intrinsic properties and the statistics of the
background or of the driving stimulus ensemble determine the effective fil-
tering properties of the system. This shows one means by which modulating
the statistics of the input can effectively gate the transmission of different
types of input or stimulus features through the system (Hasenstaub et al.,
2007; Destexhe and Pare, 1999; Fellous et al., 2003). While this analysis fo-
cused on very simple model neurons, the methods we describe generalize to
more complex, higher dimensional neuronal models, although analytical so-
lutions are unlikely. These simple examples give a clear insight into intrinsic
modulation of feature selectivity.
Our previous treatments of this problem (Aguera y Arcas et al., 2003;
Aguera y Arcas and Fairhall, 2003; Hong et al., 2007) concentrated on the
case where spikes are well-separated, so that the effects of spike history are
explicitly separated from the role of the stimulus in determining the proba-
bility of generating a spike. Another approach to this problem is to include
an explicit spike-history term in the generative model (Gerstner and Kistler,
2002; Paninski et al., 2004; Powers et al., 2005). Here, the spike history
is incorporated into the computation of features due to the effects of the
mean firing rate on the steady-state distribution of threshold escape and
reset. These results underscore the difficulty in inferring information about
underlying biophysical parameters from the output of reverse correlation,
independent of a consideration of the stimulus properties.
5 Acknowledgments
We thank Brian Lundstrom, Sungho Hong, Liam Paninski, and our review-
ers for helpful comments and discussions. This research is funded by the
McKnight Endowment Fund for Neuroscience.
14
Appendix: Approximating the singular piece of the
STA
In numerical investigation, we find that there is a simple approximation for
the average voltage at the spike time, ¯v0, for the range of σ considered in this
paper and the use of the dynamical threshold for spike triggering. In our
hands, this relationship does not seem to follow from an obvious perturbative
calculation. For the QIF and EIF, we find to first order in σq h
τm :
(27)
¯v0 ≈ Vth −
cσ
kσ
+ f σr h
τm
where f = 0.85 is the result of a fit to the exact integral in equation 18 for
different σ, h, and τm.
The exact integral for the average voltage at the time immediately pre-
ceding the spike, ¯v−1, can also be approximated with an error of a few parts
in a thousand. The distribution, p(v−1spike), can be approximated as:
p(v−1spike) ≈
H(Vth − v−1)p(v−1¯v0)
Z Vth
−∞
dv−1p(v−1¯v0)
(28)
To first order in σq h
show that:
τm , where ¯v0 is given by equation 27, it is possible to
¯v−1 ≈ Vth −
cσ
kσ
− σr h
τm r 2
π
+ f(cid:18) 2
π
− 1(cid:19)!
(29)
Taken together, this shows that the singularity in the STA at the spike
h . See figure 4 for numerical results.
time, which is given by v, goes as σp τm
References
Abeles, M. (1982). Role of the cortical neuron:
integrator or coincidence
detector? Israel J. Med. Sci., 18:83 -- 92.
Aguera y Arcas, B. and Fairhall, A. L. (2003). What causes a neuron to
spike? Neural Computation, 15:1789 -- 1807.
Aguera y Arcas, B., Fairhall, A. L., and Bialek, W. (2003). Computation
in a single neuron: Hodgkin and Huxley revisited. Neural Computation,
15:1715 -- 1749.
15
Badel, L., Gerstner, W., and Richardson, M. J. (2008a). Spike-triggered
averages for passive and resonant neurons receiving filtered excitatory
and inhibitory synaptic drive. Phys. Rev. E, 78(011914).
Badel, L., Lefort, S., Berger, T., Petersen, C., Gerstner, W., and Richard-
son, M. (2008b). Extracting non-linear integrate-and-fire models from
experimental data using dynamic i-v curves. Biological Cybernetics, 99(4 --
5):361 -- 370.
Borst, A., Flannigan, V. L., and Sompolinksy, H. (2005). Adaptation with-
out parameter change: Dynamic gain control in motion detection. PNAS,
102(17):6172 -- 6176.
Brunel, N. and Latham, P. E. (2003). Firing rate of the noisy quadratic
integrate-and-fire neuron. Neural Computation, 15:2281 -- 2306.
Bryant, H. and Segundo, J. (1976). Spike initiation by transmembrane
current: a white-noise analysis. J. Physiol., 260:279 -- 314.
Cover, T. M. and Thomas, J. A. (2006). Elements of Information Theory.
Wiley.
de Boer, E. and Kuyper, P. (1968). Triggered correlation.
IEEE Trans.
Biomed. Engr., 15:169 -- 179.
de Ruyter van Steveninck, R. and Bialek, W. (1988). Real-time performance
of a movement sensitive neuron in the blowfly visual system: Coding and
information transfer in short spike sequences. Proc. R. Soc. London Ser.
B, 234:379 -- 414.
Destexhe, A. and Pare, D. (1999). Impact of network activity on the integra-
tive properties of neocortical pyramidal neurons in vivo. J. Neurophysiol.,
81:1531 -- 1547.
Ermentrout, G. B. and Kopell, N. (1986). Parabolic bursting in an excitable
system coupled with slow oscillation. SIAM J. Appl. Math., 46:233 -- 253.
Fellous, J., Rudolph, M., Destexhe, A., and Sejnowski, T. (2003). Synaptic
background noise controls the input/output characteristics of single cells
in an in vitro model of in vivo activity. Neuroscience, 122:811 -- 829.
Fourcaud-Trocme, N., Hansel, D., van Vreeswijk, C., and Brunel, N. (2003).
How spike generation mechanisms determine the neuronal response to
fluctuating inputs. J. Neurosci., 23(37):11628 -- 11640.
16
Gaudry, K. S. and Reinagel, P. (2007). Contrast adaptation in a nonadapting
LGN model. J. Neurophysiol., 98:1287 -- 1296.
Gerstner, W. and Kistler, W. M. (2002). Spiking Neuron Models: Sin-
gle Neurons, Populations, Plasticity. Cambridge University Press, Cam-
bridge.
Gill, P., Zhang, J., Woolley, S., Fremouw, T., and Theunissen, F. (2008).
Sound representation methods for spectro-temporal receptive field esti-
mation. J Comp. Neurosci., 21(1):5 -- 20.
Hasenstaub, A., Sachdev, R., and McCormick, D. (2007). State changes
rapidly modulate cortical neuronal responsiveness. J. Neurosci., 27:9607 --
9622.
Hong, S., Aguera y Arcas, B., and Fairhall, A. L. (2007). Single neuron
computation: from dynamical system to feature detector. Neural Com-
putation, 19:3133 -- 3172.
Hunter, I. and Korenberg, M. (1986). The identification of nonlinear bi-
ological systems: Wiener and hammerstein cascade models. Biological
Cybernetics, 55:135 -- 144.
Izhikevich, E. (2000). Neural excitability, spiking, and bursting. Interna-
tional Journal of Bifurcation and Chaos, 10:1171 -- 1266.
Izhikevich, E. (2001). Resonate-and-fire neurons. Neural Networks, 14:883 --
894.
Izhikevich, E. (2004). Which model to use for cortical spiking neurons?
Neural Networks, IEEE Transactions on, 15(5):1063 -- 1070.
Lindner, B., Longtin, A., and Bulsara, A. (2003). Analytic expressions for
rate and CV of a type I neuron driven by gaussian white noise. Neural
Computation, 15:1761 -- 1788.
Meister, M. and Berry II, M. (1999). The neural code of the retina. Neuron,
22:435 -- 450.
Paninski, L. (2006a). The most likely voltage path and large deviations
approximations for integrate-and-fire neurons. J. Comp. Neurosci., 21:71 --
87.
Paninski, L. (2006b). The spike-triggered average of the integrate-and-fire
cell driven by gaussian white noise. Neural Computation, 18:2592 -- 2616.
17
Paninski, L., Lau, B., and Reyes, A. (2003). Noise-driven adaptation:
in
vitro and mathematical analysis. Neurocomputing, 52-54:877 -- 883.
Paninski, L., Pillow, J., and Simoncelli, E. (2004). Maximum likelihood es-
timation of a stochastic integrate-and-fire neural encoding model. Neural
Computation, 16:2533 -- 2561.
Paninski, L., Pillow, J., and Simoncelli, E. (2005). Comparing integrate-
and-fire models estimated using intracellular and extracellular data. Neu-
rocomputing, 65-66:379 -- 385. Computational Neuroscience: Trends in
Research 2005.
Powers, R., Dai, Y., Bell, B., Percival, D., and Binder, M. (2005). Contri-
butions of the input signal and prior activation history to the discharge
behaviour of rat motoneurones. J. Physiol., 528:131 -- 150.
Prescott, S. A., Ratte, S., De Koninck, Y., and Sejnowski, T. J. (2006).
Nonlinear interaction between shunting and adaptation controls a switch
between integration and coincidence detection in pyramidal neurons. J.
Neurosci., 26(36):9084 -- 9097.
Rauch, A., Camera, G. L., Luscher, H., Senn, W., and Fusi, S. (2003).
Neocortical pyramidal cells respond as integrate-and-fire neurons to in
vivo-like input currents. J. Neurophysiol., 90(3):1598 -- 1612.
Rieke, F., Warland, D., de Ruyter van Steveninck, R., and Bialek, W. (1996).
Spikes: exploring the neural code. The MIT Press, Cambridge, MA.
Risken, H. (1996). The Fokker-Planck Equation: Methods of Solution and
Applications. Springer, Berlin, second edition.
Sakai, H. (1992). White-noise analysis in neurophysiology. Physiol. Rev.,
72:491 -- 505.
Simoncelli, E., Paninski, L., Pillow, J., and Schwartz, O. (2004). Character-
ization of neural responses with stochastic stimuli., pages 327 -- 338. The
MIT Press, Cambridge, MA, third edition.
Slee, S., Higgs, M., Fairhall, A., and Spain, W. (2005). Two-dimensional
time coding in the auditory brainstem. J. Neurosci., 26:9978 -- 9988.
Socha, L. (2005). Linearization in analysis of nonlinear stochastic systems:
Recent results -- part I: Theory. Applied Mechanics Reviews, 58(3):178 --
205.
18
Svirskis, G., Dodla, R., and Rinzel, J. (2003). Subthreshold outward currents
enhance temporal integration in auditory neurons. Biol. Cybern., 89:330 --
340.
Svirskis, G., Kotak, V., Sanes, D., and Rinzel, J. (2004). Sodium along
with low threshold potassium currents enhance coincidence detection of
subthreshold noisy signals in mso neurons. J. Neurophysiol., 91:2465 --
2473.
Theunissen, F. E., Sen, K., and Doupe, A. J. (2000). Spectral-Temporal
Receptive Fields of Nonlinear Auditory Neurons Obtained Using Natural
Sounds. J. Neurosci., 20(6):2315 -- 2331.
Weiss, G. (1975). Time-reversibility of linear stochastic processes. J. Appl.
Probab., 12(4):pp. 831 -- 836.
Westwick, D. T. and Kearney, R. E. (2003). Identification of Nonlinear Phys-
IEEE Press Series in Biomedical Engineering. IEEE
iological Systems.
Press.
Wilson, M. T. and Steyn-Ross, D. A. (2008). Subthreshold dynamics of a
single neuron from a Hamilonian perspective. Phys. Rev. E, 78(061908).
Yu, Y. and Lee, T. S. (2003). Dynamical mechanisms underlying contrast
gain control in single neurons. Phys. Rev. E, 68(1):011901.
19
Figure 1: The STAs for (A) the QIF and (B) the EIF (using the L2-norm
for easy visual comparison), triggered on Vth = α−1, for various σ with the
upward step at t = 0 removed. Note that at small σ, the STA is non-
monotonic while, at large σ, it is approximately a decaying exponential. As
representative examples, the STA in real units is shown (C) for the QIF for
σ = 0.3 and (D) for the EIF for σ = 3 with the last time-step included.
20
Figure 2: Comparison between the numerical STA (solid) and the STA
predicted by equation 26 (dashed).
(A) For the QIF, numerical results
(solid) are compared to the prediction from stochastic linearization via the
Energy criterion (dashed).
In this case, the DKL criterion predicts STA
amplitudes that are too large (not shown) (B) For the EIF, numerical results
(solid) are compared to the prediction from SL via the DKL criterion. In
this case, the Energy criterion predicts kσ ≈ 1 for all σ (not shown).
21
Figure 3: The left column refers to the QIF and its optimization via the
energy criterion, and the right, the EIF via the DKL. (A) This figure shows
how the linear model corresponding to the energy criterion for σ = 2 for the
QIF corresponds to the full nonlinear model. In gray, we see the steady-
state voltage distribution. The quadratic nonlinearity is shown dashed and
the optimal linear model as determined by the energy criterion is the solid
line. We see that the linear model in this case is closely related to the
average slope of the quadratic nonlinearity below threshold. (B) In contrast,
for the EIF, using the DKL criterion, the optimal linear model does not
correspond closely with the exponential nonlinearity. This is indicative of
the fact that the adaptation of the STA in the EIF is due primarily to
the spiking reset, as evinced by the STA results (see figure 2). (C,D) The
steady state distributions of the (QIF,EIF) models (solid lines) are compared
to their linear model approximations (dashed).
22
Figure 4: The value of the STA in the singular component at the time of the
spike, ¯s(0), for all cases studied in this paper. As explained in the appendix,
the value is approximately linear in σ and model-independent.
23
|
1804.01958 | 1 | 1804 | 2018-04-05T17:12:17 | The multiphysics of prion-like diseases: progression and atrophy | [
"q-bio.NC",
"q-bio.TO"
] | Many neurodegenerative diseases are related to the propagation and accumulation of toxic proteins throughout the brain. The lesions created by aggregates of these toxic proteins further lead to cell death and accelerated tissue atrophy. A striking feature of some of these diseases is their characteristic pattern and evolution, leading to well-codified disease stages visible to neuropathology and associated with various cognitive deficits and pathologies. Here, we simulate the anisotropic propagation and accumulation of toxic proteins in full brain geometry. We show that the same model with different initial seeding zones reproduces the characteristic evolution of different prion-like diseases. We also recover the expected evolution of the total toxic protein load. Finally, we couple our transport model to a mechanical atrophy model to obtain the typical degeneration patterns found in neurodegenerative diseases. | q-bio.NC | q-bio |
The multiphysics of prion-like diseases: progression and atrophy
†Stevens Institute of technology, New Jersey, USA
Johannes Weickenmeier†, Ellen Kuhl‡, and Alain Goriely∗
‡Living Matter Laboratory, Stanford University, Stanford USA,
∗Mathematical Institute, University of Oxford, Oxford UK .
Many neurodegenerative diseases are related to the propagation and accumulation of toxic proteins
throughout the brain. The lesions created by aggregates of these toxic proteins further lead to
cell death and accelerated tissue atrophy. A striking feature of some of these diseases is their
characteristic pattern and evolution, leading to well-codified disease stages visible to neuropathology
and associated with various cognitive deficits and pathologies. Here, we simulate the anisotropic
propagation and accumulation of toxic proteins in full brain geometry. We show that the same
model with different initial seeding zones reproduces the characteristic evolution of different prion-
like diseases. We also recover the expected evolution of the total toxic protein load. Finally,
we couple our transport model to a mechanical atrophy model to obtain the typical degeneration
patterns found in neurodegenerative diseases.
PACS numbers:
Introduction: Age-related neurodegenerative disorders
are extremely complex and multifaceted pathologies.
Yet, their evolution is known to be closely associated
with the progression of particular protein aggregates.
When these proteins misfold and/or aggregate they can
be transported into the brain tissue and become the
seed for further misfolding and aggregation [1]. Unless
these toxic proteins are removed or stabilized, this chain-
reaction proceeds. As toxic proteins invade the brain and
form larger aggregates, they prevent the proper function
of other proteins and create tissue lesions. over time, it
leads to a disruption of the proper function of the nervous
system, a loss of tissue structure, necrosis, brain atrophy,
and ultimately death [2].
In prion diseases, such as Creutzfeldt-Jakob disease,
the infectious agents are typically small soluble misfolded
isoforms of proteins that can aggregate and trigger the
misfolding of the same protein found in its benign form.
When these aggregates are large enough, they form char-
acteristic tissue lesions. A similar mechanism of pro-
tein corruption and propagation is found in both sec-
ondary injury following traumatic brain injury [3–5] and
common age-related neurodegenerative diseases such as
Alzheimer's disease, the most common form of demen-
tia [6]. For instance, it is known from in vitro exper-
iments that the amyloid-β protein can form aggregates
that propagate through the tissues. These aggregates
have a distribution of sizes from small soluble groups to
large insoluble fibrils. Their accumulation as plaques is
believed to play a key role in Alzheimer's disease. Tau
proteins also play an important role in these diseases
[7]. These small proteins are known to stabilize mi-
crotubules within the axon. Hyperphosphorylated tau
proteins can act as seeds for further misfolding and ag-
gregation. These aggregates are found within the axon
where they are rapidly transported through the usual
interneuronal transport mechanisms [17]. They are also
transported from the cell to the extracellular space where
they diffuse through secretion and damage of the host
cell, and from the extracellular space into cells by endo-
cytosis [8, 9]. These aggregates are subject to biological
clearance slowing down or arresting the overall spreading
process. The particular role of neuronal pathways in this
transport mechanism may provide an explanation for the
observation that the spreading of neurofibrillary tangles
follows a characteristic topographic pattern.
FIG. 1: Typical spatial progression of protein aggregates in
various neurodegenerative diseases. From Jucker [1].
Independently of their molecular origin or specific ac-
tions, these pathologies, known as prion-like diseases,
share the same macroscopic features when viewed as a
a. amyloid-β in Alzheimer's diseaseb. tau inclusions in Alzheimer's diseasec. α−synuclein inclusions in Parkinson's diseased. TPD-43 inclusion in amyotrophic lateral sclerosis spatio-temporal evolution process [10]. The progression
is believed to include the following steps: (i) seeding of
misfolded proteins, (ii) templated misfolding and aggre-
gation of native (homologous) proteins, (iii) growth of
aggregates, (iv) fragmentation and spatial propagation
of aggregates of different sizes, (v) assembly of misfolded
proteins into secondary structures (e.g. protofibrils and
fibrils), and (vi) formation of larger tertiary structures
(lesions) at the tissue level. A striking feature of each par-
ticular pathology is its systematic evolution pattern. We
posit that the main reason for such reproducibility is that
the overall progression of the disease is governed by basic
transport processes and that the key difference between
individual diseases, aside from the particular set of pro-
teins involved, is the original location of the seeds, lead-
ing to a characteristic spatio-temporal evolution (Fig. 1)
with respective pathological symptoms.
An essential feature of the dynamics is that small
proteins are preferentially transported transynaptically
but can also diffuse from the extracellular space to the
cells through endocytosis [11]. These effects lead to
fast transport along neuronal pathways and a slower
diffusion in directions perpendicular to the axon bundles.
Here, we couple an anisotropic transport process with
a mechanical model to take into account the relative
effect of seed location, anisotropy, and convoluted brain
geometry on the damage patterns and the resulting
atrophy of brain tissue [12, 13]. We show that a
minimal model for transport and atrophy is sufficient to
recover the main qualitative spatial features of different
neurodegenerative diseases as well as the overall increase
of key biomarkers.
The model: We couple a transport model to a mechani-
cal model of the brain. The geometry of the brain is taken
to be a connected domain Ω in either R2 (slice simulation)
or R3 (full brain simulation) composed of gray (Ωgray)
and white (Ωwhite) matters with different material prop-
erties and total volume (or area) V . We first consider
the transport process. Different models for the aggre-
gation/propagation process have been proposed ranging
from graph Laplacian diffusion on simplified structural
networks [14] to Smoluchowski-type equations [15, 16]
tracking the aggregate-size distribution. Here, we use
a minimal model for the propagation of a toxic protein
that captures the important physical characteristics of
the problem: transformation of existing native proteins
and anisotropic diffusion along axonal pathways. We as-
sume that the concentration c(x, t) ∈ Ωwhite follows the
Fisher-KPP equation
∂c
∂t
= ∇ · (D∇c) + αc(1 − c),
(1)
where D = d⊥1 + (d(cid:107) − d⊥)γ ⊗ γ is a transversely
anisotropic diffusion tensor chosen to have a preferential
direction along the axon bundle characterized by the unit
2
vector γ = γ(x, t). Here, α > 0, d⊥ is the regular tissue
diffusion and d(cid:107) (cid:29) d⊥ is the diffusion along the axons.
This model describes the overall increase of toxic protein
concentration, assuming that the pool of protein in the
native form is sufficiently large as not to be affected by
the increase in toxic protein. The nonlinearity provides
a saturation term expressing the maximal concentration
of toxic proteins achievable (taken to be 1 without loss
of generality). In one dimension (along the axonal bun-
dles), the system supports traveling fronts with a velocity
v = 2(cid:112)αd(cid:107) leading to a complete invasion in a charac-
teristic time T = L/v, where L is the typical length scale
for the entire brain, which is on the order of 16cm for hu-
man brains. In the gray matter, for x ∈gray, we assume
simple isotropic diffusion from cells to cells different from
the white matter, (Eq. (1) with d(cid:107) = d⊥ and α = 0).
We are interested in tracking two quantities related to
the concentration field c(x, t). First, the time τ (x, Ccrit)
at which the concentration first reaches a critical level
close to the saturation level Ccrit = (1 − ε) defined im-
plicitly by c(x, τ ) = Ccrit. Second, we investigate the
temporal evolution of the average amount of toxic pro-
tein as a possible biomarker, computed as
(cid:90)
Ω
C(t) =
1
V
c(x, t) dx.
(2)
We couple the transport process to the mechanics of
atrophy by assuming that gray and white matter tissues
are morphoelastic materials [18, 19] and assigning an
isotropic shrinking factor 0 < ϑ < 1 at each point
depending on the concentration of toxic proteins. Given
an initial reference configuration, atrophy is character-
ized by a deformation ϕ : B0 → Bt, x(X, t) = ϕ(X, t)
from the reference brain B0 to the aged brain after
shrinking Bt. Here, x(X, t) is the position at time t
of the material point originally located at X at time
t = 0. This deformation is obtained by first defining
the deformation gradient F = ∇Xϕ and assuming that
it can be decomposed as F = ϑ1/3A, where ϑ denotes
the volumetric change (ϑ < 1 represents shrinking and
ϑ > 1 would correspond to growth). The elastic defor-
mation tensor A, is computed by solving the Cauchy
equation div(T) = 0 for the Cauchy stress tensor
T = J−1A · ∂AW associated with the strain-energy
density W = W (A, x) and J = det(A). The amount
of shrinking is proportional to the total exposure of the
tissue to the toxic proteins, so that ∂tϑ = δc, where
δ > 0 is an overall parameter that describes the rate of
removal of material.
Computational investigation: We create three finite
element models from T2-weighted magnetic resonance
images of a 32-year old male using Simpleware: a two-
dimensional sagittal model, a two-dimensional coronal
model, and a fully three-dimensional, anatomically ac-
curate whole brain model. The sagittal model consists
of 13442 linear triangular elements and 7216 nodes; the
coronal model consists of 38223 linear triangular elements
and 19813 nodes; and the three-dimensional whole brain
model consists of 401940 linear tetrahedral elements and
80233 nodes. To model anisotropic toxic protein propa-
gation, we register the axonal fiber orientation from the
diffusion tensor magnetic resonance image of the subject
and create an element based fiber-orientation map.
For the initial seeding, we assume a concentration
c(x) = 1 for all x ∈ Ωseed ⊂ Ω and c(x) = 0 other-
wise. Then, using Eq. (1) we propagate the toxic protein
in space and time across the brain with a diffusivity con-
trast of d(cid:107) : d⊥ = 100 to ensure a faster propagation
along the axonal fiber direction. In the white matter, we
use d(cid:107) = 100 mm2/year and α = 0.5/year. In the gray
matter, we use d(cid:107) = d⊥ = 10 mm2/year.
We post-process the concentration field c to extract the
activation time τ (x, Ccrit). From those activation times,
we create activation maps of all four cases of prion-like
disease similar to Fig. 1. Using Eq. (2), from the spatio-
temporal concentration field c(x, t), we extract the inte-
gration of the toxic protein concentration in time C(t) to
create a temporal map of the toxic protein concentration
averaged over the entire brain [20].
FIG. 2: Activation maps of tau inclusions in brains of patients
with Alzheimer's disease in a 2D sagittal slice.
Finally, using the tissue atrophy model with different
atrophy rates δ in the gray and white matter tissue,
we utilize our two-dimensional coronal model to create
atrophy maps.
It is important to note that atrophy
takes place primarily in locations with elevated con-
centrations of toxic proteins. Therefore, the change of
volume does not significantly affect the time evolution
of the concentration since it is close to maximal at
the locations where atrophy takes place. We take
advantage of this property to compute tissue shrinking
3
as a post-processing step based on the values of the
concentration at different time points. Since we assume
that after each step of shrinking, there is total relaxation
of the stresses, the mechanical stresses generated are
small during shrinking and we can use a standard
compressible neo-Hookean model valid in small defor-
mations with parameters fitted to experimental data:
W = c1(tr(AT · A)− 3) + c2(J − 1)2 with c1 = 666.56 Pa
and c2 = 1777.5 Pa [21]. For the relative shrinking
rate,
it is known that the gray matter tissue shrinks
looses more volume than the white matter tissues [22].
For our computation, we use a ratio ϑgray : ϑwhite = 4 : 1.
2D simulations: We start with a two-dimensional
simulation on a sagittal slice (Fig. 2). The region in-
cludes the ventricle and cerebellum but, through no-flux
boundary conditions, it is assumed that no progression
takes place into these regions. Based on the description
of the evolution of tau inclusions in Alzheimer's disease,
we simulate the propagation of toxic tau proteins by
considering an initial seeding close to the entorhinal cor-
tex (Fig. 2a). We observe the quick progression into the
brain stem and the hippocampus (Fig. 2a), to paralimbic
and adjacent medial-basal temporal cortex (Fig. 2b),
to cortical association areas (Fig. 2c), and eventually
reaching primary sensory-motor and visual areas. The
relatively rapid progression along the hippocampus is
associated with a strong anisotropy along the axonal
pathways tangent to the ventricles. This progression
is consistent with the well-established Braak stages
of Alzheimer's disease [23]: Stage I-II transentorhinal
(Fig. 2a); Stage III-IV limbic (Fig. 2b); Stage V-VI
isocortical (Fig. 2c-d).
3D simulations: The previous analysis uses a high-
resolution mesh but lacks the three-dimensional geom-
etry of the brain. Next, we use the accepted knowledge
about different neurodegenerative diseases to compare
the progression of various toxic proteins. To test the
specific role of geometry in the progression, we assign
the same parameter values for all cases and only modify
the initial seeding region of the four cases discussed in
[1] and depicted in Fig. 1. For every individual seeding
region we show, in Fig. 3, the time τ to reach a critical
level of damage.
We observe generic progression trends in all cases: (i)
as soon as a sufficient level of toxic proteins reaches the
area close to the ventricles, there is a fast progression
in the limbic system (around the ventricles) leading to
a rapid invasion of the temporal and occipital lobes; (ii)
once a sufficient level of toxic protein is reached within
the cortex, further invasion through the cortex takes
place; (iii) if the parietal lobe is not directly involved
initially, it only becomes invaded in the last stages of the
disease.
Tau inclusions in brains of patients with Alzheimer's diseasettinitial seedingcerebelluma.b.c.d.ventricle 4
[20]. The variation of the typical parameters entering
the sigmoids are used to represent physiological quan-
tities such as time of outbreak, and rate of spreading.
These curves serve as a general guideline to understand
the variations between individuals, pathologies, and
the effect of various stimuli or therapeutics [25]. We
can extract these curves from the simulation using
(2) as shown in Fig. 4. Our simulations recover the
predicted behavior for this type of disease integrated
across the entire brain. This profile is not unexpected.
Indeed, we known that Equation (1) supports fronts in
one-dimension that have profile and similar evolution
equations for Alzheimer's diseases. A simple estimate
using the typical wave speed of front propagation in
one-dimension leads to a lower and upper bound for
the total invasion of 11.3 and 35.8 years, respectively,
obtained by using the front velocity with either the fast
axonal diffusion constant d(cid:107) ≈ 100 mm2/year or the
slow gray matter diffusion constant d ≈ 10 mm2/year.
These estimates are consistent with the characteristic
time scales observed in Fig. 4. Similar curves have also
been obrained on 2D slices in different geometries [16].
Here, we observe that the fully 3D evolution on an
actual brain model also follows the same general trend
of a progressive invasion.
Shrinking: We now turn to the effect of elevated toxic
proteins on the tissue. It is known that the formation of
large aggregates or high concentration of toxic proteins
prevents the proper function of neuronal cells, leading to
ischemia, and eventually tissue removal. For instance, in
the top row of Fig. 5, there is a marked and rapid atro-
phy associated with the disease. We use our propagation
model on a similar coronal slice to compute the activa-
tion time at three different time points (middle row).
For each of these time points we compute the shrinking
of the slice and show the change in geometry from the
previous one (bottom row). The degeneration-induced
atrophy patterns (bottom row) agree well with the at-
rophy pattern observed in Alzheimer's disease (top row).
Conclusions: Neurodegenerative diseases are known for
their extreme complexity and medical science has wres-
tled with this major challenge by amassing a consider-
able amount of information despite few successes. One
of the remarkable features that appear in these diseases
is their remarkable reproducible topographic propaga-
tion pattern. Each disease has a well-defined spatio-
temporal evolution that has been documented over the
years and catalogued into stages, each associated with
typical symptoms. This evolution appears much more
controlled than other diseases such as cancer and has
led to three main hypotheses in recent years. First, it
has been proposed that diseases share the same charac-
teristics as prion diseases and are based on the seeding,
propagation, and accumulation of toxic proteins involved
FIG. 3: Simulated activation maps for the spatial progres-
sion of toxic protein in full brain geometry for various initial
seeding regions matching Fig. 1 [1].
FIG. 4: Evolution of the total concentration for the simulation
of Fig 2 (2D sagittal slice) and of Fig 3b (3D tau concentration
in Alzheimer's disease).
An indirect way to follow the evolution of neurodegen-
erative diseases is to look at the averaged concentration
of toxic proteins or associated biomarkers. The paradigm
for this evolution is based on the so-called Jack's curves.
These curves are the expected, mostly hypothetical,
evolution of a typical biomarker concentration as a func-
tion of time [24]. They take the shape of sigmoid-like
functions: a slow evolution at first during incubation,
followed by a sharp increase during the outbreak and
leading to saturation when the disease is fully established
a. amyloid-β in Alzheimer's diseaseb. tau inclusions in Alzheimer's diseasec. α−synuclein inclusions in Parkinson's diseasettttd. TPD-43 inclusion in amyotrophic lateral sclerosis0102030400.00.20.40.60.81.0onset2Dslice3DbrainC(t)C(t)t (years) 5
[3] A. C. McKee, T. D. Stein, C. J. Nowinski, R. A. Stern,
D. H. Daneshvar, V. E. Alvarez, H.-S. Lee, G. Hall, S. M.
Wojtowicz, C. M. Baugh, et al., Brain 136, 43 (2013).
[4] D. H. Daneshvar, L. E. Goldstein, P. T. Kiernan, T. D.
Stein, and A. C. McKee, Molecular and Cellular Neuro-
science 66, 81 (2015).
[5] M. Cruz-Haces, J. Tang, G. Acosta, J. Fernandez, and
R. Shi, Translational neurodegeneration 6, 20 (2017).
[6] J. C. Watts, C. Condello, J. Stohr, A. Oehler, J. Lee,
S. J. DeArmond, L. Lannfelt, M. Ingelsson, K. Giles, and
S. B. Prusiner, Proceedings of the National Academy of
Sciences 111, 10323 (2014).
[7] M. Goedert, Science 349, 1255555 (2015).
[8] C. Soto, Neuron 73, 621 (2012).
[9] J. W. Wu, M. Herman, L. Liu, S. Simoes, C. M. Acker,
H. Figueroa, J. I. Steinberg, M. Margittai, R. Kayed,
C. Zurzolo, et al., Journal of Biological Chemistry 288,
1856 (2013).
[10] L. C. Walker and M. Jucker, Annual review of neuro-
science 38, 87 (2015).
[11] B. Frost, R. L. Jacks, and M. I. Diamond, Journal of
Biological Chemistry 284, 12845 (2009).
[12] A. Goriely, M. G. D. Geers, G. A. Holzapfel, J. Jayamo-
han, A. J´erusalem, S. Sivaloganathan, W. Squier,
J. A. W. van Dommelen, S. L. Waters, and E. Kuhl,
Biomech. Model. Mechanobiol. 14, 931 (2015).
[13] A. Goriely, S. Budday, and E. Kuhl, Advances in Applied
Mechanics 48, 79 (2015).
[14] F. Abdelnour, H. U. Voss, and A. Raj, Neuroimage 90,
335 (2014).
[15] P. C. Bressloff, Lecture Notes on Mathematical Mod-
elling in the Life Sciences (2014).
[16] M. Bertsch, B. Franchi, N. Marcello, M. C. Tesi, and
A. Tosin, Mathematical Medicine and Biology p. dqw003
(2016).
[17] P. C. Bressloff and J. M. Newby, Reviews of Modern
Physics 85, 135 (2013).
[18] E. Kuhl, J. Mech. Behavior Biomed. Mat. 29, 529 (2014).
[19] A. Goriely, The Mathematics and Mechanics of Biological
Growth (Springer Verlag, New York, 2017).
[20] C. R. Jack and D. M. Holtzman, Neuron 80, 1347 (2013).
[21] L. A. Mihai, L. Chin, P. A. Janmey, and A. Goriely, Jour-
nal of The Royal Society Interface 12, 20150486 (2015).
[22] P. M. Thompson, K. M. Hayashi, G. De Zubicaray, A. L.
Janke, S. E. Rose, J. Semple, D. Herman, M. S. Hong,
S. S. Dittmer, D. M. Doddrell, et al., Journal of neuro-
science 23, 994 (2003).
[23] H. Braak and E. Braak, Acta neuropathologica 82, 239
(1991).
[24] C. R. Jack Jr, D. S. Knopman, W. J. Jagust, R. C. Pe-
tersen, M. W. Weiner, P. S. Aisen, L. M. Shaw, P. Ve-
muri, H. J. Wiste, S. D. Weigand, et al., The Lancet
Neurology 12, 207 (2013).
[25] H. van den Bedem and E. Kuhl, Current Opinion in
Biomedical Engineering 1, 23 (2017).
[26] S. Leh´ericy, M. Marjanska, L. Mesrob, M. Sarazin, and
S. Kinkingnehun, European radiology 17, 347 (2007).
[27] M. Zamparo, A. Trovato, and A. Maritan, Physical re-
view letters 105, 108102 (2010).
[28] T. C. Michaels, S. I. Cohen, M. Vendruscolo, C. M. Dob-
son, and T. P. Knowles, Physical review letters 116,
038101 (2016).
[29] A. Lloret-Villas, T. Varusai, N. Juty, C. Laibe,
FIG. 5: Atrophy. Top row: magnetic resonance images show-
ing hippocampal atrophy in Alzheimers disease [26]: yearly
examination of the same subject is shown.
Increasing hip-
pocampal and atrophy is observed together with ventricular
enlargement and widening of cortical sulci. Middle row: acti-
vation time based on initial seeding in the brain stem. Bottom
row: shrinking based on the activation time and toxic protein
concentration with a ratio ϑgray : ϑwhite = 4 : 1.
in regular sub-cellular functions such as alpha-synucleins
or tau proteins. Second, it has been proposed that the
overall evolution of a typical biomarker follows regular
sigmoid-like curves. Third, a given disease is associated
with typical atrophy patterns.
There have been multiple mathematical models pro-
posed for neurodegenerative diseases, but most of them
focus on biochemical pathways, cellular interactions,
and the formation of amyloids [27–29]. Our approach
here is radically different as we study the problem at the
largest possible scale of the brain. Our minimal model
takes into account the main microscopic features of these
diseases, but incorporated the full brain geometry and
axonal directions from diffusion tensor imaging. Despite
the complexity of these diseases our model recovers the
three aforementioned key features of neurodegenerative
disease progression and suggests that brain geometry,
transport anisotropy, and mechanics play a central role
in neurodegeneration and atrophy pattern formation.
This work was supported by the NSF Grant CMMI
1727268 to Ellen Kuhl. The support for Alain Goriely by
the Engineering and Physical Sciences Research Council
of Great Britain under research grant EP/R020205/1 is
gratefully acknowledged.
[1] M. Jucker and L. C. Walker, Nature 501, 45 (2013).
[2] J. Brettschneider, K. Del Tredici, V. M.-Y. Lee, and
J. Q. Trojanowski, Nature Reviews Neuroscience 16, 109
(2015).
year 0year 1year 2year 3 N. Le Nov`Ere, H. Hermjakob, and V. Chelliah, CPT:
pharmacometrics & systems pharmacology 6, 73 (2017).
6
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.